OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation

OpenFlyV2 introduces a unified onboard vision-language-action (VLA) architecture for aerial robots that integrates semantic understanding, language-grounded perception, spatial reasoning, and continuous flight control wi…

→ View original source

AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers

The paper introduces AB-Attention, a training-free method that accurately approximates dense 3D full attention in video diffusion transformers at high sparsity levels, addressing fidelity degradation in dynamic block-spa…

→ View original source

IntReal: A Real-World Dataset and Benchmark for Multimodal Interleaved Generation

Researchers introduce IntReal, a benchmark addressing the underexplored task of multimodal interleaved generation where text and images are naturally interwoven. The framework comprises two components: IntReal-100K, a re…

→ View original source

TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving

TAO introduces a tool-aware joint scheduling framework for LLM-based agentic serving that addresses inefficiencies in existing systems, which ignore tool-return readiness and rely on heuristic GPU admission and placement…

→ View original source

Instruction Capacity: Behavior and Degradation of Language Models under Controlled Load

The paper introduces "instruction capacity" as a quantitative metric to characterize the degradation of language models' instruction-following ability under increasing load, addressing the lack of a standardized limit ex…

→ View original source

DTCA: Denoising Token Credit Assignment for Reinforcement Learning in Diffusion Language Models

Researchers introduce Denoising Token Credit Assignment (DTCA), a framework addressing token-level credit assignment challenges in reinforcement learning (RL) for diffusion language models (DLMs). Because DLMs generate t…

→ View original source

Can Discrete Tokens Serve as Semantically Lossless Representations of Visual Signals?

The authors investigate whether discrete visual tokens can serve as semantically lossless representations, challenging the prevailing view that quantization error inevitably degrades semantic fidelity compared to continu…

→ View original source

Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data

Researchers propose eliminating calibration data for channel selection in Mixture-of-Experts (MoE) decoding. MoE models face bottlenecks during interactive decoding due to loading routed-expert weights from memory. Tradi…

→ View original source

Beyond the Estimate: Wiener-Informed Decoding for Easier Training of Pixel-Space Flow Models

Wang and You introduce Wiener-informed Decoding (WiD), a method that leverages frozen second-order image statistics to accelerate and stabilize the training of pixel-space flow models. The approach integrates a trainable…

→ View original source

Did the Model Use That? Counterfactual Training Boosts Self-Attribution in Language Models

The study investigates self‑attribution in language models—whether models explicitly signal when their responses depend on supplied information. Evaluating eleven frontier models, the authors observed that most rarely at…

→ View original source

Constrained Decoding: Harmful or Beneficial? Revisiting Its Impact on LLM Performance

This research investigates the impact of constrained decoding on Large Language Model (LLM) performance, challenging the prevailing assumption that forcing specific output formats inherently degrades reasoning capabiliti…

→ View original source

Work Agents: Systems, Capabilities, and Optimization

The authors propose a system‑level view of work agents as compositions of foundation models and runtime harnesses that operate within stateful workspaces to produce verifiable outcomes. They define two orthogonal conditi…

→ View original source

Speculative Diffusion is a Tale of Two Tradeoffs

The paper introduces speculative decoding as a method to accelerate autoregressive LLM inference by using a faster draft model, traditionally requiring lossless acceptance relative to a verifier distribution, whereas rec…

→ View original source

The LLM-Attacker: A Unified Red-Team Framework for Privacy Evaluation of Split and Distributed LLM Systems

The LLM-Attack Suite provides a reproducible red‑team pipeline for privacy assessment of split and distributed large language models, where intermediate activations are transmitted to an external server. The framework lo…

→ View original source

Different Tasks, Different Timesteps: Temporal Task-Aware Quantization for Diffusion Language Models

Diffusion Language Models (DLMs) face distinct quantization challenges due to their iterative denoising process. This work demonstrates that quantization sensitivity varies across both downstream tasks and denoising time…

→ View original source

Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems

The authors of “Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems” evaluated the risk of “mosaic” disclosures where LLM agents combine information from multiple sources to infer and reveal sensitive …

→ View original source

Bridging Exploration and Persona: Information-Directed Dueling for Test-Time Agent Personalization

The paper introduces ID‑TAP, an information‑directed dueling‑bandit framework for test‑time personalization of large language model agents across large discrete persona spaces. ID‑TAP couples a query‑conditioned neural u…

→ View original source

From RL Papers to Framework Code: Inclusion and Review in Seven Open-Source Training Frameworks

A study linking commit histories from seven open-source reinforcement-learning post-training frameworks to 6,939 screened arXiv papers from January 2025 through June 2026 assessed how readily published RL methods become …

→ View original source

Building AI Research Teams for Autonomous Discovery of Linear Attention Architectures

In this work, the authors construct a persistent multi‑disciplinary team of AI researchers, each maintaining a private research log and sharing experimental results, while an adaptive allocation mechanism assigns researc…

→ View original source

Self-Credit: Credit Assignment for Long-Horizon Agentic RL via Intra-group Rescaling

The paper introduces Self-Credit , a novel agentic reinforcement learning (RL) framework designed to address sparse reward challenges in long-horizon tasks. Unlike traditional RLVR methods, Self-Credit generates dense su…

→ View original source

OverSeer: Multi-Agent World Models Beyond What Agents See

OverSeer, a multi-agent world model introduced by researchers from CUHK, Tsinghua University, and others, addresses the challenge of generating temporally consistent, agent-specific video streams by integrating auxiliary…

→ View original source

EvaOmni: An Evidence-Grounded Benchmark for Audio-Visual Reasoning

Researchers introduced EvaOmni, a new benchmark designed to evaluate audio-visual reasoning in multimodal large language models (Omni LLMs) by emphasizing evidence-grounded assessment. Unlike traditional benchmarks that …

→ View original source

Any 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale

The paper introduces a novel approach for 3D scene generation by leveraging high-dimensional implicit 3D latent spaces derived from 2D foundation models such as DINOv2, SigLIP2, and DA3, addressing limitations of traditi…

→ View original source

RoboValue: A Fine-Grained Sim-and-Real Benchmark for Unified Evaluation of Robotic Value Models

RoboValue introduces a unified benchmark for evaluating robotic value models across four capability dimensions: task-state understanding, temporal progress monitoring, failure and recovery reasoning, and value consistenc…

→ View original source

HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs

HarnessPro introduces a label‑free recursive self‑improvement framework that evolves agent harnesses to boost task performance while lowering execution cost through Pareto‑based optimization. Candidate harnesses are asse…

→ View original source

DLite: History-free Speculative Decoding for Agentic AI

DLite introduces a history‑free speculative decoding framework designed for agentic AI workloads, where long, evolving contexts and frequent multi‑turn interactions cause conventional draft models to lose up to 80% of th…

→ View original source

Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries

Librarian Agents introduce a two‑stage framework in which a builder LLM agent proactively reorganizes an unstructured corpus into a hierarchical filesystem of synthetic files and reference links, creating a curated works…

→ View original source

LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents

LEASH introduces a training‑free framework for weak‑to‑strong escalation and oversight of computer‑use agents. It operates in three stages: first, an initial task‑level uncertainty score routes the request to either a we…

→ View original source

Learning to Read Out: Unembedding Dynamics in Language Model Pretraining

The authors investigate when a language model actually acquires task‑relevant information versus when it can express that information through its output layer. They argue that tracking correct token production conflates …

→ View original source

Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models

The authors propose Mine‑Sample‑Render (MSR), a test‑time procedure that decouples diversity generation from the internal sampling of a generative model by turning the model into a conditional renderer. First, they mine …

→ View original source

Explaining the Roles of Experts in Mixture-of-Experts LLMs for Personalization

The authors personalize three open‑weight Mixture‑of‑Experts (MoE) large language models by further training them on a large‑scale personalization dataset. To make the resulting modifications interpretable, they train cr…

→ View original source

When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill

The study investigates remote KV‑cache tiers in a disaggregated GLM‑5.3 (FP8 MoE with multi‑head latent attention) deployment on 8×B300 nodes using prefill context parallelism (PCP8). Enabling the remote tier (LMCache wi…

→ View original source

From Failures to Learnable Worlds: Policy-Conditioned Environment Generation for Navigation Agents

The authors propose WEDGE, a weakness‑guided environment generation framework that tailors training data to the specific failure modes of a pretrained navigation policy. Given a policy, WEDGE leverages an LLM to extract …

→ View original source

More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning

The authors benchmark ten self‑distillation techniques for agentic reinforcement learning against the GRPO baseline on four text‑based environments—ALFWorld, ScienceWorld, Search‑QA, and WebShop—using three different bac…

→ View original source

Enabling Token-wise Importance Sampling in Off-Policy RLVR with Proximal KL Regularization

The paper studies off‑policy reinforcement learning with verifiable rewards (RLVR) where data arise from asynchronous generation, rollout‑training mismatch, and response reuse. It shows that token‑wise importance weighti…

→ View original source

Can an Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration in LLM Agents

The paper proposes a confidence‑calibration technique for LLM‑based agents that aligns the model’s self‑reported verbalized confidence with the eventual success or failure of a multi‑step episode rather than with the cor…

→ View original source

FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching

FlashVSR+ introduces a hybrid attention mechanism that combines strong local correlation from pretrained VSR models with sparse global interaction, achieving up to twice the speed of DiT and 45% lower peak memory, which …

→ View original source

Spectral-Driven Adaptive-Rank Optimizer for Large Language Model Reinforcement Learning with Verifiable Rewards

Zhipeng Chen, Wayne Xin Zhao, and Ji‑Rong Wen introduced Ares (Adaptive Rank‑Effective Spectral Optimization), an optimizer that dynamically adjusts the rank of gradient updates based on the spectral energy distribution …

→ View original source

Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training

The authors introduce an uncertainty‑guided progressive contraction framework to compress reasoning supervision during post‑training of large language models. First, they treat each teacher‑generated chain‑of‑thought not…

→ View original source

S-Space: Exploring Spatial Workspace in Multimodal Language Models

The authors investigated how multimodal language models (MLLMs) internally represent spatial information by identifying a linear subspace, termed S‑Space, in which the activations of object‑token embeddings encode contin…

→ View original source

MixTok: Adaptive Visual Tokenization with Global 1D Tokens and Sparse 2D Refinement

MixTok introduces a hybrid visual tokenizer that merges a compact 1D global token stream with a set of sparse 2D tokens to refine locally complex image regions. A content‑adaptive router evaluates candidate spatial locat…

→ View original source

Opera: A Verbal Critic Framework for Long-horizon Coding Agents

Opera is a verbal critic framework designed for long‑horizon coding agents that treats each correction as a persistent note and monitors it until the underlying problem is solved. It determines when to intervene using bo…

→ View original source

A GPU-Parallel Framework for Heterogeneous Multi-Task Reinforcement Learning

The authors present Hebero, a GPU‑parallel benchmark built on Isaac Lab that enables joint training and evaluation of a single reinforcement‑learning policy across 40 heterogeneous manipulation tasks. By scaling the numb…

→ View original source

Incidental information contaminates patient notes and disrupts clinical reasoning in large language models

The study investigated how incidental, non‑clinical information affects large language models (LLMs) when they generate ambient clinical notes or perform reasoning tasks. Researchers first analyzed 576 real‑world patient…

→ View original source

The Lattice of Transition Laws

The authors treat diffusion models, autoregressive (AR) models, and intermediate hybrids as trajectories on a shared corruption lattice, where each decoding schedule incurs a cost defined by the dependence discarded amon…

→ View original source

Predicting Cable Dynamics with Physical Attention Bias

The authors investigate learned simulators for deformable linear objects such as cables, focusing on improving generalization to unseen cables and stability over long rollouts. They note that most prediction error arises…

→ View original source

Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI

The paper introduces a model‑agnostic framework that enables frozen large language models (LLMs) and vision‑language models (VLMs) to acquire new expertise from deployment experience without modifying their internal weig…

→ View original source

Skill Constellations: Tracing the Supply Chain of Agent Skills on GitHub

The study constructs the first temporally resolved copy network of agent skills by mining the Git history of every *.skill* file in the GitSkills dataset, capturing 2,193,119 individual skill adoptions across GitHub. Thi…

→ View original source

A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization

The authors address the fragmented evaluation of LLM‑based agents for black‑box optimization (BBO) by creating AgenticBBO‑Bench, a unified finite‑budget benchmark that spans five heterogeneous domains: synthetic benchmar…

→ View original source

Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction

The paper introduces Multi-agent Egocentric World Model (ME-World) to extend egocentric world modeling from single-agent to multi-agent settings with fine‑grained embodied interactions. Existing approaches predict first‑…

→ View original source

Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning

The authors introduce Iris-3B, a 3‑billion‑parameter text‑to‑image transformer trained in pixel space rather than through a variational autoencoder bottleneck. Pretraining proceeds via a curriculum that upsamples from 25…

→ View original source

From Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery

The paper introduces Personalized Test-Time Scaling (PTTS), which treats test‑time computation allocation as a multidimensional optimization problem where users jointly specify accuracy, latency, and inference‑cost targe…

→ View original source

RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

RobotWorld introduces a simulation testbed comprising 84 distinct tasks that cover manipulation, mobile manipulation, locomotion, driving, and aerial control, each equipped with explicit interaction budgets and executabl…

→ View original source

UNREAL: Unifying Retrieval and Long-Context with a Single Model

UNREAL introduces a model‑internal evidence selection mechanism that works for both corpus‑scale retrieval and long‑context inference. It extracts retrieval queries from the frozen LLM’s internal representations by encod…

→ View original source

Closing the Context Gap: Activation Alignment for Tabular In-Context Learning

The paper introduces activation alignment, a technique that trains a lightweight linear map on synthetic unlabeled data to align the intermediate activations of a context‑constrained student model with those of a full‑co…

→ View original source

Noise Out, Bias In: Targeted Bias Injection in Diffusion Language Models via Closed-Loop Activation Steering

The study exploits the iterative denoising process of masked diffusion language models (dLLMs), which expose token‑level answer distributions at every step before a token is committed, unlike autoregressive models that r…

→ View original source

UndoBench: Separating Task Competence from Recovery Capability in Tool-Using AI Agents

UndoBench introduces a benchmark that isolates task competence from recovery capability in tool‑using AI agents by constructing 36 base workflows paired with 36 fault scenarios spanning eight enterprise domains. The desi…

→ View original source

Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability

The authors investigated whether training large language models to produce shorter chain‑of‑thought (CoT) rationales harms the transparency of those rationales. They fine‑tuned several pretrained models using three disti…

→ View original source

Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It

The study investigated source bias in LLM‑driven decision making by evaluating 12 language‑model agents across three distinct domains (product purchasing, hotel booking, and paper citation). In a controlled end‑to‑end se…

→ View original source

RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations

The authors introduce RealCompanion, a benchmark that evaluates how well an AI companion can understand a human partner by reasoning over extended, real‑world dialogues. They collected ten longitudinal relationships betw…

→ View original source

Video Generation Models: A Survey of Post-Training and Alignment

This survey examines post‑training and alignment strategies for video generation models, framing post‑training as a unified approach that adapts pretrained generators without full retraining. The authors distinguish impl…

→ View original source

MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization

MemFold introduces a compact soft memory mechanism for long-context personalization that optimizes a fixed-budget latent interface through on-policy reinforcement learning rather than reconstruction or imitation objectiv…

→ View original source

Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes

Where-OPD introduces a spatially guided on‑policy self‑distillation method for multimodal large language models (MLLMs) that leverages procedurally generated synthetic scenes providing automatic object identities and spa…

→ View original source

Pay for the Fault, Not the Flow: Label-Free In-Flow Multi-Agent Workflow Optimization

The article introduces InFlowOp, a label-free framework for optimizing multi-agent workflows in large language models (LLMs). Traditional workflow construction faces challenges in determining task granularity, agent assi…

→ View original source

Training LLM Judges from Language Feedback via Position-Selective Self-Distillation

The authors propose a method to train large language model (LLM) judges using natural language feedback through position‑selective self‑distillation. Unlike outcome‑supervised reinforcement learning approaches such as GR…

→ View original source

Safety of Latent Communication in Multi-Agent Systems

The paper investigates safety risks introduced by latent communication in multi‑agent systems, where agents exchange information directly in internal representation space via lightweight trainable links that map a sender…

→ View original source

Agent Error Dataset: Scaling 50,000 Error--Diagnosis Pairs for Failure Analysis and Error-Aware Post-Training

The Agent Error Dataset (AED) introduces 50,228 error-diagnosis pairs derived from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. Developed to enable fai…

→ View original source

PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents

PreviewDiff introduces a training-free test-time search method that reconfigures diffusion sampling from scalar selection into multimodal critic-guided search over intermediate latents, addressing persistent failures in …

→ View original source

LIFT: Layout-In-Future Video Generation under Large Viewpoint Change via On-Policy Self-Distillation

LIFT presents a unified image‑to‑video generation framework that augments conventional camera‑trajectory conditioning with Layout‑In‑Future (LIF) control, allowing users to define the semantic content and spatial arrange…

→ View original source

TabFM: A Zero-Shot Foundation Model for Tabular Data

TabFM introduces a 400‑million‑parameter foundation model that treats supervised tabular prediction as an in‑context learning problem, enabling calibrated zero‑shot forecasts in a single forward pass without any task‑spe…

→ View original source

SCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language Models

The authors study why aggressive token pruning hurts vision‑language model (VLM) reasoning despite retaining some visual tokens. Using a fixed‑context Pass@K analysis they show that repeatedly sampling from the same prun…

→ View original source

An RL View of OPD: Least Square Policy Distillation for Sample-Efficient LLM Reasoning

The authors reinterpret on‑policy distillation (OPD) as a reinforcement‑learning problem, showing that the reverse‑KL objective used in OPD is equivalent to a KL‑regularized policy‑optimization objective. Building on thi…

→ View original source

Learning to Learn from Context: Synthetic Training from Perturbed Public Documents

The article introduces a synthetic training pipeline to enhance large language models' (LLMs) context-learning capabilities, addressing the challenge of memorization when using public documents already present in pretrai…

→ View original source

IterSynth: Rethinking Deep Search Agents via Role-Decoupled Iterative Synthesis

IterSynth introduces a role‑decoupled iterative synthesis framework for LLM‑based deep‑search agents that separates planning from answer generation. A Planner module proposes sub‑queries or information needs, while a Syn…

→ View original source

Coding Agents for Generalized Task and Motion Planning Problems

The study investigates whether large‑language‑model coding agents can automate generalized task‑and‑motion planning (TAMP) by synthesizing programs that transfer across problem instances. Using a fixed synthesis budget, …

→ View original source

ViRDM: Taming Representation Distribution Matching for Few-Step Causal Video Generation

ViRDM proposes a teacher‑ and critic‑free post‑training method for few‑step causal video diffusion that adapts representation distribution matching (RDM) from one‑step image generation to multi‑step video rollouts. The a…

→ View original source

Parts-of-Speech as Emergent Categories in SAE Latent Space

The authors investigated whether part‑of‑speech (PoS) information is encoded in the latent space of Sparse AutoEncoders (SAEs) applied to language model representations. They treated PoS tags as a controlled linguistic p…

→ View original source

RGBD20K: A Large-Scale Benchmark for RGB-D Semantic Segmentation

The authors introduce RGBD20K, a large‑scale benchmark designed to advance RGB‑D semantic segmentation by providing richer semantic coverage, more training data, and cleaner annotations. The dataset contains 20,000 synch…

→ View original source

Rufus-Air: An Open LLM Post-Training Recipe

Rufus‑Air presents an open, reproducible post‑training recipe applied to the GLM‑4.5‑Air‑Base model (106B‑A12B parameters). The recipe consists of eight serial stages: supervised fine‑tuning (SFT), reasoning reinforcemen…

→ View original source

Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs

The authors investigate whether transformer‑based LLMs behave linearly when presented with a mixture of two independent text streams. They form a combined input by taking a weighted sum of the token embeddings (or hidden…

→ View original source

AV-GRPO: Modality-Anchored Decoupling Diffusion Reinforcement Learning for Joint Audio-Video Generation

AV‑GRPO introduces a modality‑anchored online diffusion reinforcement‑learning framework for joint audio‑video generation that tackles heterogeneous rewards, divergent modality dynamics, and synchronization evaluation bi…

→ View original source

Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures

The authors introduce Jev, a language model fine‑tuned via reinforcement learning for calibrated decisions (RLCD) to output calibrated probabilities for multiple typed questions about a single input in one forward pass. …

→ View original source

Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents

Qwen-Planner-Agent introduces a closed‑loop AI‑for‑AI framework that treats the planner agent itself as both the subject and the engine of development. The framework consists of three coupled modules. First, an AI‑driven…

→ View original source

Uranus: Building the Next-Generation Simulation Infrastructure for Embodied AI

Uranus introduces a data‑driven robot simulator that leverages a joint‑trajectory‑conditioned autoregressive diffusion model to generate visual observations on‑the‑fly. At each simulation step the model consumes a future…

→ View original source

All modalities are equal, but video is more equal: Closing the Cross-Attention Gap in Joint Video Generation

The paper investigates an asymmetry in multimodal diffusion transformers where companion modalities such as 3D body motion or audio form strong attentional links to video tokens, but the reverse links that allow those mo…

→ View original source

InternW0: A Foundational Physical World Model for Efficient Real-World Interactions

Researchers from Shanghai AI Laboratory have introduced InternW0, a foundational physical world model designed for efficient, actionable real-world interactions. The architecture utilizes an asymmetric video-action frame…

→ View original source

Six Layers Less: Encoder Pruning for Whisper with Label-Free Recovery

Researchers have introduced a novel method for optimizing OpenAI's Whisper Automatic Speech Recognition (ASR) models by specifically targeting the encoder architecture. While previous optimizations like whisper-large-v3-…

→ View original source

Calibration as a First-Class Criterion in LLM Evaluation

The paper argues that calibration— the agreement between a language model’s expressed confidence and its empirical correctness—should be treated as a core evaluation criterion alongside traditional performance metrics. I…

→ View original source

GeoPair: Geometry-Preserving Cross-Layer Factorization for Training-Free Transformer Compression

GeoPair introduces a training‑free compression framework that targets cross‑layer redundancies in transformer models by jointly optimizing weight pairings and learning a shared dictionary factorization. Rather than forci…

→ View original source

The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Researchers have introduced Taste-Bench, a novel benchmark designed to quantify "taste"—the ability of LLM agents to make optimal decisions during long-horizon engineering and research tasks. Unlike traditional end-to-en…

→ View original source

Blaming Across the Aisle: Political Contrasting and Blame Attribution in the Danish Parliament

This study investigates shifts in blame attribution within the Danish Parliament over a nearly three-decade period, spanning from 1997 to 2026. To analyze this longitudinal data, the researchers developed BlameBERT, a sp…

→ View original source

StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training

StableVQ introduces a novel framework designed to improve the training stability and performance of vector-quantized (VQ) tokenizers, which are essential for autoregressive and masked image generation models. The authors…

→ View original source

Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs

To address the inference bottlenecks inherent in Diffusion Large Language Models (dLLMs), the researchers introduced Flash-dLLM, a training-free acceleration framework designed to mitigate GPU memory I/O constraints. Rec…

→ View original source

ACLArena: Agent Continue Learning in Multi-stage Post-training

The ACLArena framework was introduced to investigate and optimize Agent Continual Learning (ACL) during multi-stage post-training, addressing the challenge of integrating diverse capabilities without catastrophic forgett…

→ View original source

Document Retrieval-Aware Chunking (D-RAC): Universal Retrieval-Aware Ingestion of Enterprise Documents via PDF Normalization and Multimodal Markdown Conversion

Researchers have introduced Document Retrieval-Aware Chunking (D-RAC), a framework designed to improve Retrieval-Augmented Generation (RAG) by ingesting heterogeneous enterprise documents such as PDFs, Word files, and sc…

→ View original source

onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

The researchers introduce onPanda, an interactive annotation framework designed to streamline the creation of alignment data for Large Language Models (LLMs) and autonomous agents. Unlike traditional manual post-editing,…

→ View original source

Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms

This research investigates why state-of-the-art video diffusion models frequently generate content that violates physical laws, identifying root causes within internal attention mechanisms rather than data deficiencies. …

→ View original source

OmniVBench: A Benchmark and Large-Scale Dataset for Omni Reference-to-Video Generation

To address the limitations of existing evaluation protocols in reference-to-video (R2V) generation, researchers have introduced OmniVBench and the Omni-R2V Dataset. Current benchmarks often fail to assess whether models …

→ View original source

APort Vault: Benchmarking AI Agent Payment Authorization with the Open Agent Passport

The APort Vault benchmark evaluates payment authorization security in tool-using AI agents by replaying 4,371 human-authored attacks against a live payment agent. The study utilized 14 different models from eight labs, c…

→ View original source

Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network

Researchers have introduced RefineEdit, a training-free, prompt-to-prompt image editing framework that leverages the inherent editability of Generative Refinement Networks (GRN). Unlike traditional diffusion-based editor…

→ View original source

WeVisDoc: From Coverage to Capability for Robust End-to-End Document Parsing

Researchers have introduced WeVisDoc, a novel two-stage data-centric framework designed to enhance the robustness of end-to-end document parsing across diverse layouts and acquisition conditions. To address the limitatio…

→ View original source

FAMOS: Feed-Forward 3D Articulation Modeling from Sparse Observations

Researchers have introduced FAMOS, a feed-forward framework designed to model 3D articulated objects from sparse, unordered sets of partial point clouds. Unlike traditional methods that depend on category-level shape pri…

→ View original source

Verifiable Social Reasoning for LLM Assistants

Researchers have introduced Fuse, a novel multi-agent simulation framework designed to evaluate the social reasoning capabilities of Large Language Models (LLMs) in user-mediated consultation settings. Addressing the dif…

→ View original source

Self-Evolving Search Index

Researchers have introduced SELF-INDEX, an autonomous framework designed to enable search indices to self-evolve without human intervention. Recognizing that retrieval quality depends on how effectively index keys repres…

→ View original source

Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL

Researchers have introduced ActObs, a novel supervised fine-tuning (SFT) approach designed to improve reinforcement learning (RL) performance by supervising both action and observation tokens within agent trajectories. W…

→ View original source

PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?

Researchers have introduced PACT (Pressure-Applied Compliance Testing), a new benchmarking framework designed to evaluate the reliability of enterprise-grade LLM agents in following regulatory rules under psychological o…

→ View original source

Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling

This study investigates how candidate-generation strategies influence the efficiency and performance of test-time scaling in large language models (LLMs). While increasing the number of candidates (N) improves reasoning …

→ View original source

UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation

To address the limitations of current evaluation methods in multi-modal image generation—which often assess textual and visual conditions in isolation—the researchers introduce UFO, a unified framework designed for simul…

→ View original source

VākQA: A Benchmark and Evaluation Study for Telugu Spoken Factoid Question Answering

Researchers have introduced VākQA, a novel benchmark designed to evaluate Spoken Factoid Question Answering (SQA) in Telugu, a language previously lacking dedicated spoken benchmarks. The dataset comprises 2,001 question…

→ View original source

When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation

This research investigates length inflation during on-policy distillation (OPD), a phenomenon where student models generate excessively long sequences that may exhaust generation budgets. The authors identify "terminatio…

→ View original source

Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model

This research evaluates MiniMax-H3, an omni-modal generative model utilizing a shared latent framework to unify text, image, video, and audio modeling. To assess whether multimodal alignment enhances physical world reaso…

→ View original source

The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction

Researchers have introduced Edge0, a streaming inference engine designed to mitigate the memory wall encountered when serving large Mixture-of-Experts (MoE) models on consumer-grade hardware. While 35B-class models requi…

→ View original source

CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents

CERA-MoA (Co-Evolving Router with continually learning Agents for Mixture-of-Agents) introduces an iterative reinforcement learning framework designed to synchronize query routing with the evolving capabilities of LLM ag…

→ View original source

ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments

To address the "scientific experience bottleneck" caused by fragmented toolchains and implicit domain conventions in scientific software, researchers introduced ScienceIDE, an infrastructure designed to transform complex…

→ View original source

A Zeroth-Order Paradigm for LLM Preference Alignment

Researchers have introduced Comparison-based Preference Optimization (ComPO), a novel zeroth-order paradigm designed to align large language models (LLMs) with human preferences. Unlike traditional direct alignment metho…

→ View original source

FLAT: Resampling Image and Text into 1D Flexible-Length Aligned Transmodal Tokens for Retrieval and Generation

The FLAT (Flexible-Length Aligned Transmodal representations) framework introduces a unified pre-training approach designed to bridge the gap between multimodal representation learning and generative modeling. Unlike tra…

→ View original source

AI for Games in the Foundation Model Era

This survey explores the transformative impact of foundation models and learned game-world models across the entire game development lifecycle. The authors categorize current AI applications into six distinct functional …

→ View original source

Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems

Researchers introduced Emergence World, a continuous multi-agent environment designed to stress-test the long-horizon resilience of autonomous systems. The study utilized eight parallel simulations containing ten agents …

→ View original source

HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness

Researchers have introduced HarnessVLN, a zero-shot, training-free framework designed to unify instruction-following and object-goal navigation by leveraging Multimodal Large Language Models (MLLMs). Rather than relying …

→ View original source

Not All Prompts Are Equal: Exploration-Guided Prompt Scaffolding for Multimodal Reinforcement Post-Training

Researchers have introduced an exploration-guided prompt scaffolding framework designed to optimize the reinforcement learning (RL) post-training of multimodal large language models (MLLMs). The authors identify a fundam…

→ View original source

RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments

Researchers have introduced RSIAgent, a training-free multi-agent framework designed to facilitate recursive self-improvement in novel environments through autonomous memory construction. The framework utilizes a coordin…

→ View original source

How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus

This study investigates the "lossless" claim of Orthrus, a hybrid autoregressive-diffusion architecture designed to accelerate language model inference through parallel token generation and an intra-model consensus mecha…

→ View original source

ModaLens: Measuring Image Sensitivity in Report-Conditioned Medical VLMs

Researchers introduced ModaLens, an image-swap audit framework designed to evaluate image sensitivity in report-conditioned medical vision-language models (VLMs). The study investigated whether models actually utilize vi…

→ View original source

SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking

Researchers have introduced Simple Attention Sparsification (SAS), a novel post-training method designed to mitigate the quadratic computational complexity of Transformer models by selecting a subset of context units. Un…

→ View original source

Ambient @ EgoProactive 2026 : Proactive Egocentric Assistance with Visually Grounded Supervision

In the ECCV 2026 Wearable AI Challenge, the "Ambient @ EgoProactive" submission achieved first place in the large-model division and second place in the <=2B parameter division by developing a proactive egocentric ass…

→ View original source

Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model

For the EgoLongQA track of the ECCV 2026 Wearable-AI Challenge, researchers developed "Ambient," a compact vision-language model designed for long-video perception in ten-minute egocentric clips. The system achieved firs…

→ View original source

Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation

Researchers have introduced Benchmark Radar, a comprehensive living database and search engine designed to streamline the discovery and evaluation of AI benchmarks. Addressing the challenges of locating relevant datasets…

→ View original source

Mi-Ripple: Restoring Images Degraded by Iterative AI Editing

The paper introduces Mi-Ripple, a specialized restoration workflow designed to mitigate "digital ripple" artifacts—specifically grid-like and granular textures—caused by iterative reference-conditioned AI image editing. …

→ View original source

Memory as Plans: World-Action Modeling with Memory-Grounded Planning

To address the limitations of Markovian policies in non-Markovian, long-horizon robotic manipulation, the authors propose MaP-WAM, a "Memory-as-Plans" framework. Unlike traditional methods that rely on growing visual win…

→ View original source

IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications

Researchers have introduced IdeaAMBIG, a new benchmark designed to evaluate the "codification readiness" of research-method specifications—specifically, whether a written method provides enough detail for a human or AI a…

→ View original source

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

Researchers investigated the role of image tokenizers as "visual languages" within unified multimodal models by developing a controlled, pure-autoregressive testbed. Rather than relying on isolated metrics, the study tra…

→ View original source

Beyond Solver Verdicts: Generative Reward Models for Autoformalization

Researchers have identified a critical vulnerability in neurosymbolic autoformalization termed Verdict-Preserving-Unfaithfulness (VPU), where incorrect formal encodings bypass verification by executing successfully and p…

→ View original source

FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation

The FreeFlow architecture introduces a novel approach to optical flow estimation by eliminating task-specific inductive biases such as correlation volumes, feature warping, or iterative refinement. Instead of relying on …

→ View original source

Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning

Researchers have addressed the English-centric bias in reasoning language models by focusing on L2 reasoning, which enables models to reason consistently within the language of a user's prompt. To investigate how data co…

→ View original source

Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking

Researchers have addressed performance degradation in multimodal entity linking, particularly regarding rare entities, by moving beyond popularity-based metrics like pageviews toward structural knowledge-graph metrics. B…

→ View original source

ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation

The researchers introduce ActReview, a novel post-training framework designed to transform LLM-generated peer reviews from simple diagnostics into actionable revision guidance. To address the limitations of current model…

→ View original source

T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks

Researchers have introduced T1, a 122B parameter Mixture-of-Experts (MoE) model specifically designed for long-horizon terminal tasks such as coding. The model operates within a cloud-based sandbox shell, capable of exec…

→ View original source

DRG-MAPPO: Hierarchical Dynamic Role-Graph Multi-Agent Reinforcement Learning for Cooperative Air Combat

To address the challenges of tactical coordination and ambiguous task allocation in complex air combat simulations, researchers have introduced DRG-MAPPO, a Hierarchical Dynamic Role-Graph Multi-Agent Proximal Policy Opt…

→ View original source

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation

Researchers have introduced X-AuT, a progressive compression framework designed to reduce the depth of audio encoders in speech large language models (LLMs) without compromising decoder-side embedding stability. To mitig…

→ View original source

Negative Self-Distillation: Learning to Reason by Avoiding Flaws

To mitigate the performance degradation caused by On-Policy Self-Distillation (OPSD) in complex reasoning tasks, researchers have introduced Negative Self-Distillation (NSD), a novel optimization framework. While OPSD of…

→ View original source

Reference-Based Bias Detection in LLMs via Relative Representations of Hidden States

Researchers have introduced a novel reference-based method for auditing bias in Large Language Models (LLMs) by analyzing internal hidden-state representations rather than relying solely on expensive output-level benchma…

→ View original source

Why Is Video Still So Expensive? A Survey of Inference-Efficiency Mechanisms in Video and Audiovisual LLMs

This survey explores the computational and memory challenges inherent in Video Large Language Models (VideoLLMs), which integrate video representations with pretrained LLMs to perform tasks like temporal grounding and ca…

→ View original source

Show-Harness: Just a VLM Agent Can Play Robots

Researchers have introduced Show-Harness, an embodied harness designed to bridge the gap between foundation vision-language models (VLMs) and robotic control via a compact semantic interface. Instead of requiring massive…

→ View original source

AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems

The researchers introduce AgentGrad, a novel prompt optimization framework designed to enhance the performance of Large Language Model (LLM)-based multi-agent systems (MAS). Addressing critical flaws in existing textual …

→ View original source

A*-Thought-V2: Efficient Latent Reasoning via Geometric Dynamics of LLM

Researchers have introduced A*-Thought-V2, a framework designed to optimize Chain-of-Thought (CoT) reasoning by reducing the computational and context costs associated with long reasoning chains. Rather than utilizing ha…

→ View original source

Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation

The Marigold V2 framework introduces an advanced approach to monocular depth estimation by repurposing Diffusion Transformer (DiT) architectures, originally designed for image generation, for dense regression tasks. To a…

→ View original source

NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness

The NeoHorse-1 project introduces a framework for recursive self-improvement (RSI) via agentic post-training, utilizing a heterogeneous model pool managed by an intelligent routing harness. The system captures user inter…

→ View original source

NOAH: Learning the Full Patient Journey. A Longitudinal Multimodal Time-Aware Model for Representation and Forecasting

Researchers have introduced NOAH, a task-agnostic, generative transformer model designed to represent and forecast longitudinal, multimodal patient trajectories. To address the limitations of existing discriminative mode…

→ View original source

UniMate: One Unified Model to Animate Diverse Skeletons

UniMate introduces a unified foundation model designed to synthesize articulated motion for arbitrary 3D skeletons using only a rigged asset and a text prompt, eliminating the need for per-skeleton fine-tuning or test-ti…

→ View original source

Unlocking Lossless Speedups in LLMs via Discrete Diffusion

Researchers have introduced Uno, a new class of diffusion-augmented Large Language Models designed to overcome the sequential bottlenecks of traditional autoregressive (AR) next-token prediction. Unlike standard models t…

→ View original source

Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal

Researchers have introduced a novel framework for "narrow-boundary safety," addressing the challenge of fine-tuning Large Language Models (LLMs) to exhibit context-specific refusal behaviors. Unlike traditional topic-lev…

→ View original source

EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents

The researchers introduced EmbodiedSkills, a unified framework designed to transform Vision-Language-Action (VLA) models into robust embodied agents capable of managing long-horizon tasks. Unlike traditional models that …

→ View original source

Refuse without Refusal: A Structural Analysis of Safety-Tuning Responses for Reducing False Refusals in Language Models

In the paper "Refuse without Refusal," researchers Minji Kim and Hyounghun Kim investigate the phenomenon of false refusals in large language models, where models incorrectly reject benign queries that contain superficia…

→ View original source

Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization

Researchers have introduced OR-Clarify, a novel benchmark designed to evaluate how Large Language Models (LLMs) handle incomplete natural-language problem descriptions during the optimization modeling process. Unlike pre…

→ View original source

Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference

This research investigates the revitalization of layer dropout (stochastic depth) in Large Language Model (LLM) training, addressing its recent omission from standard pre-training recipes. Through over 2,400 experiments …

→ View original source

Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems

This paper introduces a formal game-theoretic framework to model orchestrator-worker interactions in multi-agent LLM systems as a bilevel coordination game. The researchers demonstrate that under bounded coupling, worker…

→ View original source

Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction

To address the issue of geometric collapse in online 3D reconstruction caused by cumulative pose drift in long videos, the authors propose Scal3R, a framework that reformulates reconstruction as a multi-reference relativ…

→ View original source

A Common Measure of Communication for Speech Brain-Computer Interfaces

To address the lack of standardized benchmarking in speech brain-computer interfaces (BCIs), researchers Jayalath, Ballyk, and Jones propose a new information-theoretic metric called open-vocabulary mutual information (O…

→ View original source

VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement

The researchers introduce VeriPhy, an auditable physical-verification system designed to evaluate and refine video generation models by moving beyond simple scalar quality scores. Unlike monolithic prompting methods, Ver…

→ View original source

Select, Compress, Reinvest: A Controlled Study of Visual-Token Allocation in Long-Video MLLMs

This study investigates optimal visual-token allocation strategies in long-video Multimodal Large Language Models (MLLMs) by isolating three variables: frame selection, spatial compression, and token reinvestment. To ens…

→ View original source

DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training

The DRACO (Distributing Rubric-based Advantage for Credit Optimization) framework addresses the challenge of credit assignment in long-horizon agent training within outcome-blind settings where programmatic verifiers are…

→ View original source

Rethinking On-Policy Distillation of Large Language Models II: One Training Example

This research investigates the data efficiency of On-Policy Distillation (OPD) by analyzing its performance in a data-minimal setting where the student model is trained on a single query. The authors demonstrate that one…

→ View original source

Compile by Training: Turning Natural-Language Specifications into Local Neural Functions

The "compile by training" framework introduces a method for transforming natural-language specifications into efficient, reusable, and local neural functions. To address the latency, cost, and dependency issues inherent …

→ View original source

Exploring Collaboration between a language and a non-language agent

Researchers have introduced LLAMIA-Bench, a novel benchmark comprising six diverse collaborative chess tasks designed to evaluate the integration of Large Language Models (LLMs) with non-language agents. These tasks spec…

→ View original source

Debias-SparseGPT: Bias-Aware Pruning for Large Language Models

Researchers have introduced Debias-SparseGPT, a post-training pruning framework designed to mitigate the bias amplification often caused by weight sparsification in Large Language Models (LLMs). While existing methods li…

→ View original source

Language Models Can Control Their Own Attention

Researchers have introduced Declarative Attention (DA), a novel protocol designed to optimize inference efficiency in long-context language models by enabling them to manage their own attention mechanisms. Rather than re…

→ View original source

Ignorance or Incompetence? Constructing Knowledge-Gated, Verifiable Tasks for LLM Agents

Researchers introduce a novel knowledge-gated task-construction protocol designed to distinguish between an LLM agent's inherent competence and its access to private, domain-specific conventions. By decoupling task instr…

→ View original source

Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs

Researchers propose a novel framework called "control-data flow separation" to address the instability inherent in prompt optimization within multi-agent Large Language Model (LLM) systems. Currently, prompt optimization…

→ View original source

DramaChain Bench: An End-to-End Benchmark for Short-Drama Generation

Researchers have introduced DramaChain Bench, a comprehensive end-to-end evaluation framework designed to assess the multi-stage production pipeline of commercial short dramas, including scriptwriting, storyboarding, key…

→ View original source

From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

To mitigate GPU fragmentation caused by maintaining multiple legacy models, researchers developed a consolidated self-hosted LLM designed to handle diverse corporate workloads. By analyzing production error patterns acro…

→ View original source

Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving

The researchers introduce Qwen-Drive-1.0, a vision-language foundation model designed to unify 3D perception, visual question answering, and motion planning for autonomous driving. Building upon a pretrained vision-langu…

→ View original source

Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions

Researchers have introduced a method to transform the deterministic Aardvark Weather model into a probabilistic forecasting system capable of disentangling different sources of uncertainty. By attaching specific stochast…

→ View original source

Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement

This research investigates the efficacy of On-Policy Distillation (OPD), questioning whether student improvement stems from teacher supervision or inherent self-improvement mechanisms. The authors demonstrate that teache…

→ View original source

ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models

Researchers have introduced ContextBias, a controlled evaluation framework designed to investigate whether stereotypical associations between professional roles and visual attributes persist when the semantic context is …

→ View original source

Super Library Agent: Joint Generation and Maintenance of Multiple Applications Beyond the Single Codebase

Researchers have introduced the "Super Library Agent" problem, which addresses the inefficiencies inherent in using LLM coding agents to manage portfolios of related applications. Traditional application-by-application w…

→ View original source

Acquire, Repair, Preserve: A Diagnosis-Guided Post-Training Recipe for Small-Model Dialogue Game Agents

This study addresses the limitations of small-scale language models in interactive dialogue games, specifically focusing on their inability to maintain state, interpret feedback, and execute valid actions. Using a 2B ope…

→ View original source

ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL

The researchers introduce ContextPilot, a proactive context management framework designed to optimize long-horizon agentic reasoning in Large Language Models (LLMs). Recognizing that growing interaction histories create …

→ View original source

GGSS: Geodesic-Gated Spherical Steering for Inference-Time Debiasing of Generative Vision-Language Models

Researchers have introduced GGSS (Geodesic-Gated Spherical Steering), a novel inference-time intervention designed to mitigate demographic bias in generative vision-language models (VLMs). Unlike previous debiasing metho…

→ View original source

Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning

Researchers have introduced the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task to address the limitations of current generative AI, which struggles with long-form video editing due to entity fragmentation a…

→ View original source

TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback

TacForcing introduces a novel streaming action-generation framework designed to address the limitations of chunk-based vision-language-action (VLA) models in contact-rich manipulation tasks. Traditional models often suff…

→ View original source

Procedura: Agentic 3D Modeling with Procedural Control

Procedura introduces a novel agentic framework for 3D modeling that treats shapes as parametric, procedural programs rather than static meshes. Moving away from traditional dense mesh generation, which often lacks part d…

→ View original source

GameWAM: A World Action Model for Video Games

Researchers have introduced GameWAM, the first World Action Model (WAM) designed for closed-loop gameplay and GUI control in video games. Unlike traditional agents that map visual context directly to actions, or predicti…

→ View original source

Aphanta: Diagnosing Task-Aligned Image-Edited Intermediates for Multimodal Reasoning

The researchers introduce Aphanta, an automated diagnostic framework designed to evaluate the effectiveness of the MLLM-to-image-editor-to-MLLM pipeline for multimodal reasoning. The framework assesses three distinct rea…

→ View original source

CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes

Researchers have introduced CritICL, a novel inference-time framework designed to facilitate weak-to-strong generalization in Large Language Models (LLMs) by leveraging the structured failure patterns of smaller models w…

→ View original source

Agentic Game Development as a Verifiable Trajectory Data Engine for Scaling World Models

Researchers propose a novel approach to scaling world models by utilizing agentic game development as a verifiable trajectory data engine. The authors argue that current scaling strategies, which rely heavily on massive …

→ View original source

TTPO: Test-Time Policy Optimization

Researchers have introduced Test-Time Policy Optimization (TTPO), a novel framework designed to enable test-time training (TTT) for large language models without relying on ground-truth labels. While existing methods lik…

→ View original source

RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval

Researchers have introduced RetrievalRouter, a lightweight, query-aware routing mechanism designed to optimize the trade-off between retrieval accuracy and computational latency. Recognizing that modern retrieval pipelin…

→ View original source

LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale

Researchers have introduced LibriBrain100, a massive magnetencephalography (MEG) dataset engineered to advance non-invasive neural speech decoding. The dataset provides over 100 hours of high-quality MEG recordings captu…

→ View original source

When "Must" Becomes "Maybe": Constraint Weakening in LLM Agent Workflows

Researchers have investigated a critical failure mode in multi-stage LLM agent workflows where mandatory constraints are inadvertently weakened during state transformations. By focusing on "safety blockers" as a proxy fo…

→ View original source

Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training

Researchers have introduced Game2World Engine (G2WEngine), a comprehensive framework designed to convert raw gameplay videos into high-quality training data for video world models by decoupling game environments from scr…

→ View original source

LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training

Researchers have introduced LAION-BVD, a massive-scale open-access video dataset designed to facilitate multimodal pre-training across video, audio, and image modalities. The dataset was constructed by harvesting 1.3 bil…

→ View original source

ClawProBench: Trace-Aware Evaluation of AI Agents with Runtime Coverage and Frozen Workplace-Style Holdouts

Researchers have introduced ClawProBench, a novel trace-aware evaluation framework designed to address the limitations of traditional agent benchmarks that rely solely on final-answer correctness. By shifting the focus f…

→ View original source

One Polluted Page Is Enough: Evaluating Web Content Pollution in LLM Recommenders

Researchers Minghao Luo and Liang Chen introduced FORGE (Fake Online Recommendations in Generative Environments), a novel benchmark designed to evaluate how Generative Engine Optimization (GEO) can manipulate Large Langu…

→ View original source

The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search

Researchers address two primary bottlenecks in Retrieval-Augmented Generation (RAG): inaccurate measurement of evidence utilization and inefficient context budget allocation. The study identifies a "diagnostic illusion" …

→ View original source

Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization

Researchers have introduced Environment-Regularized Policy Optimization (ERPO), a novel framework designed to resolve the stability-exploration trade-off inherent in Large Language Model (LLM) policy optimization. Tradit…

→ View original source

Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models

Researchers have introduced SparsePR, a training-free block-sparse attention mechanism designed to accelerate video transformers and world models without sacrificing generative quality. Addressing the limitations of stan…

→ View original source

Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts

To address the prohibitive computational costs of hyperparameter sweeps in large-scale Mixture-of-Experts (MoE) models, researchers propose a two-step transfer framework designed to estimate optimal learning rates. The m…

→ View original source

CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment

To address the inherent tension between safety alignment and model utility in large language models (LLMs), researchers introduced CLEAR (Continuous Latent Adapter Routing). Unlike standard Supervised Fine-Tuning (SFT) o…

→ View original source

Peer-Voted LLM-Agent Stress Tests Find Feed-Induced Lexical Convergence but No Reliable Matched-Exposure Advantage for Distributed Sources

Researchers introduced PV-SST, a peer-voted social-platform testbed, to evaluate population-level behaviors in large language model (LLM) agents through a preregistered matched-exposure experiment. The study conducted 44…

→ View original source

τ_0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation

Researchers have introduced $\tau_0$-VLA, a hierarchical vision-language-action (VLA) foundation model designed to address the complexities of long-horizon robot manipulation. Unlike traditional VLA models that rely on a…

→ View original source

The Embedder's Dilemma: LLMs Are Better, but at What Cost?

This study investigates the trade-offs between using large language models (LLMs) and dedicated embedding models for downstream NLP tasks. Researchers conducted a cost-aware comparison involving ten LLMs and 26 embedding…

→ View original source

Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

The paper introduces Hierarchical Self-Improvement (HSI), a framework designed to evolve task-specific agent harnesses rather than relying on manual prompt or workflow adjustments. Using a single frozen LLM as a backbone…

→ View original source

FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills

FlowEvo is a training-free framework designed to enable large language model (LLM) agents to achieve self-evolution through the simultaneous co-evolution of workflows and executable skills during inference. Unlike tradit…

→ View original source

GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation

Researchers have introduced GOAG (Generative and Object-Agnostic Grasp Planner), a novel deep generative model designed to improve dexterous robotic manipulation by overcoming the limitations of object-specific training.…

→ View original source

CoToGrasp: Contact-Topology-Conditioned Dexterous Grasp Synthesis via Canonical Workspace Learning

Researchers have introduced CoToGrasp, a generative framework designed for dexterous grasp synthesis that conditions grasp generation on specific contact topologies rather than just physical stability. Unlike traditional…

→ View original source

Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

This research investigates the reasoning capabilities of three frontier mixture-of-experts models from Alibaba, OpenAI, and NVIDIA (3.6–4.0B active parameters each) when fine-tuned for reasoning in Greek, a low-resource …

→ View original source

FlashPrefill V2: Block-Sparse Prefill Attention for Long-Context LLM Serving

FlashPrefill V2 advances long-context Large Language Model (LLM) serving by transitioning a previous algorithmic prototype into a production-ready sparse attention framework. To address the quadratic complexity of the pr…

→ View original source

Listening Forward: Next Patch Embedding Prediction Enables Scalable Audio Learners

Researchers have introduced NAPE (Next-Audio-Patch-Embedding prediction), a minimalist self-supervised learning framework designed to shift audio representation learning toward a causal, autoregressive paradigm. Unlike t…

→ View original source

Towards Quantifying Benchmark Optimization in ASR Models

This research introduces a methodology for quantifying benchmark optimization in Automatic Speech Recognition (ASR) models, specifically addressing scenarios where audio content underdetermines the reference transcript. …

→ View original source

VA-Judger: Reward Modeling from Human Preference Feedback for Joint Video-Audio Generation

To address the limitations of separate metrics for audio quality, visual fidelity, and synchronization in joint video-audio generation, researchers introduced VA-Judger, a chain-of-thought omni-reward model designed to a…

→ View original source

SoftVTBench: A Deformation-Aware Visuo-Tactile Dataset and Benchmark for Deformable-Object Manipulation

Researchers have introduced SoftVTBench, a novel visuo-tactile dataset and benchmark designed to improve the physical interaction quality during deformable-object manipulation. Addressing the limitations of success-only …

→ View original source

Decision-Metric Alignment in Latent World Models: Diagnostics and Action-Conditioned Objectives for MPC Planning

This research addresses the "decision-metric alignment" problem in JEPA-style latent world models, where Euclidean distance to a goal latent in the embedding space may fail to accurately rank action sequences by real-wor…

→ View original source

SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents

The research introduces SkillGate, a novel training framework designed to improve skill selection in long-horizon agents that rely on external instruction libraries. The authors identify a phenomenon termed "selector cre…

→ View original source

DiSCO: Defending text-to-image generation through distribution-guided contrastive prompt optimization

Researchers have introduced DiSCO, a zero-shot, black-box defense mechanism designed to prevent text-to-image models from generating Not-Safe-For-Work (NSFW) content, specifically addressing the "benign adversarial" prob…

→ View original source

PTXBench: Benchmark and Adapt LLMs for GPU Kernel Optimization with Architecture-specific PTX

The researchers introduced PTXBench, a specialized benchmark designed to evaluate and adapt large language models (LLMs) for optimizing GPU kernels using architecture-specific PTX instructions. The framework assesses mod…

→ View original source

MoE-ViE: Mixture of Experts Vision Encoder for Efficient Image and Video Understanding

Researchers have introduced MoE-ViE, a series of Mixture-of-Experts Vision Encoders designed to scale vision-language model capacity without the proportional increase in computational cost associated with dense architect…

→ View original source

Agent Lightning v1.0: Towards Harnessed Agentic RL

Agent Lightning v1.0 introduces a lightweight framework designed for "harnessed agentic RL," a paradigm where the deployment-time harness—which manages tools, context, and control flow—directly participates in model post…

→ View original source

Advancing Open and Reproducible Relational Learning: RelArena-α, TabPFN-Rel and RPI

Prior Labs has released an alpha-version suite of open-source tools designed to standardize and accelerate research in relational learning. The release introduces RelArena-$\alpha$, a unified benchmarking framework model…

→ View original source

Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization

Researchers have introduced Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization (SA-MRPO) to address inefficiencies in multi-objective reinforcement learning for language models. Traditional metho…

→ View original source

AnyTalk: Speech Animation for Arbitrary Characters Leveraging a Video Generation Model

AnyTalk introduces a novel framework for generating 3D speech animations for arbitrary characters without necessitating character-specific animation data or manual rigging. The method leverages the motion priors of pre-t…

→ View original source

StreamOPD: A Post-Training Recipe with Spatio-Temporal Cue Gating for Streaming Video Understanding

The paper introduces StreamOPD, a post-training methodology designed to enhance streaming video understanding using a memory-free, recent-window protocol. To overcome the limitations of reinforcement learning—which often…

→ View original source

Nanbeige4.2-3B on Apple Silicon: Fixing Deployment Bugs and Decreasing Looped Transformer Memory Overhead

This research addresses deployment failures and memory inefficiencies encountered when running the Nanbeige4.2-3B agentic model on Apple Silicon (MPS). The model utilizes a Looped Transformer (LT) architecture, which reu…

→ View original source

A Pathway to General-Purpose Scientific AI: Multimodal Comprehension of Scientific Images

This paper outlines a strategic framework for developing general-purpose scientific AI by addressing the challenges of interpreting multimodal scientific data. To facilitate this, the authors introduced the ALD/E-ImageMi…

→ View original source

SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning

The researchers introduce SimpleOPD, a tokenizer-agnostic on-policy distillation (OPD) framework designed to transfer long-context reasoning capabilities from powerful teachers, such as the SU-01 model, to short-context …

→ View original source

CPI-Bench: A Comprehensive,Practical and Intelligent Benchmark for Real-World Image Editing

Researchers have introduced CPI-Bench, a multidimensional benchmark designed to evaluate image editing models within complex, real-world contexts, addressing the limitations of existing benchmarks that focus solely on si…

→ View original source

Mitigating Gender Bias in English to Romanian Machine Translation

Researchers Ioana Grigore and Sergiu Nisioi have introduced a novel hybrid pipeline designed to mitigate gender bias in English-to-Romanian machine translation, addressing the challenge of translating from a gender-neutr…

→ View original source

PixSDS: Why Latent SDS Makes Noisy Pixels

The paper introduces PixSDS, a method designed to mitigate the structured color artifacts and high-frequency texture noise commonly produced by latent Score Distillation Sampling (SDS) in text-to-3D generation. The autho…

→ View original source

Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing

Researchers have introduced Hybrid-Policy Self-Editing (HPSE), a novel framework designed to address the lack of "composability" in Unstructured Knowledge Editing (UKE). While current editors can inject free-form passage…

→ View original source

TailBooster: A Dual-Layer Generative Framework for Extreme Value Augmentation with Operational Validity Enforcement

To address the scarcity of extreme event data in air transport modeling, researchers introduced TailBooster, a dual-layer generative framework designed for extreme value augmentation in mixed-type tabular records. The me…

→ View original source

From Inaudible Inputs to Model Failures: Low-Frequency Safety Risks in LALMs

This research investigates a critical security vulnerability in Large Audio-Language Models (LALMs) where inaudible, low-frequency signals can trigger significant model failures. The authors introduce Intermittent Low-Fr…

→ View original source

From Atomic Evidence to Logical Composition: Structured Compositional Reasoning over Compound Answer Options

Researchers Obed Junias and Maria Leonor Pacheco address a specific failure mode in Large Language Models (LLMs) where models struggle to process compound answer options containing logical operators like AND, OR, and NEI…

→ View original source

Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity

This research investigates how instruction tuning influences the relationship between model confidence and the lexical diversity of generated rationales. By evaluating three matched pairs of base and instruction-tuned mo…

→ View original source

DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation

Researchers have introduced DreamX-Phi 1.0, an action-conditioned video world model designed to predict future visual observations for robotic manipulation based on input frames, language instructions, and action sequenc…

→ View original source

Specification-first convergence with an AI coding agent: a case study of dismantling a core architectural invariant across 189 files in a 717k-line codebase with no test oracle and no human code review

This case study demonstrates a large-scale architectural refactoring performed by an AI coding agent using a "specification-first" protocol on a 717,725-line TypeScript production application. The objective was to disman…

→ View original source

From Synthesis to Removal: Physics-Grounded Reflection Simulation and Diffusion-Based Video Dereflection

Researchers have introduced a closed-loop framework designed to address the challenges of video reflection removal, a task often hindered by the lack of paired datasets and temporal coherence. The proposed solution begin…

→ View original source

ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-based Agents

Researchers have introduced ToolHazard, a scalable framework designed to synthesize adversarial environments for evaluating and aligning LLM-based agents. Recognizing that existing security research is often limited by m…

→ View original source

Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence

Researchers have introduced Mechanist, an agentic system designed to automate the discovery of the underlying mechanisms of artificial intelligence. To facilitate autonomous scientific inquiry, the authors constructed a …

→ View original source

Ready Cohorts: Bounding GPU Opportunity and Avoiding Host Round Trips in LLM-Agent Control

The "Ready Cohorts" framework optimizes LLM-agent execution by offloading control-path routing decisions directly to the GPU. By utilizing four-byte, GPU-computed routing, the system avoids costly round-trips to host mem…

→ View original source

DistilVDR: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation

DistilVDR introduces a compact, end-to-end visual document retrieval (VDR) system designed to overcome the latency and computational costs associated with multi-billion-parameter models. The researchers developed a 524M …

→ View original source

Reference-Free Post-Training of Open Large Language Models for Multilingual Machine Translation

Researchers have introduced a novel reference-free post-training methodology to enhance multilingual machine translation (MT) in open large language models. Building upon the supervised fine-tuned (SFT) MiLMMT-46-v0.1 mo…

→ View original source

VibeLifeBench: Can Your Life Agent Be Proactive and Persistent in a Living World?

Researchers from Xiaohongshu Inc have introduced VibeLifeBench, a novel benchmark designed to evaluate the proactivity and persistence of Large Language Model (LLM) agents within complex, long-horizon lifelike scenarios.…

→ View original source

Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA

The Macaron-V1 framework introduces a co-designed system for open continual learning, utilizing a Mixture-of-LoRA (MoL) architecture to enable experiential intelligence in real-world environments. Instead of modifying th…

→ View original source

Gaming Without an Attacker: Benchmark Fingerprinting in LLM-Driven Search Under Selection Pressure

This research investigates how Large Language Models (LLMs) optimize for evaluation signals rather than true task performance, a phenomenon termed "benchmark fingerprinting." The study utilized two GPU-kernel-optimizatio…

→ View original source

Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory

Researchers have introduced Agent Memory Distillation (AMD), a training-free framework designed to enhance the performance of small language model (SLM) agents by transferring structured knowledge from a large teacher mo…

→ View original source

Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control

Researchers have introduced Enfold, a method designed to enhance embodied control efficiency by distilling the internal computational processes of world generative models into predictive representations. Rather than rely…

→ View original source

Zero Gap Is Not Restoration: Stratified Per-Question Probability Evaluation and Step-wise Mitigation of Benchmark Contamination

This paper critiques the current methodologies for evaluating and mitigating benchmark contamination in large language models, arguing that the standard Gap of Aggregate Performance (G-AP) metric provides a flawed repres…

→ View original source

Beyond Simply Environment Scaling: Designing Effective Environment Distributions for Multimodal Agent Learning

Researchers Kejian Zhu et al. investigate the limitations of scaling multimodal environment pools for agent training, discovering that simply increasing the quantity of environments does not guarantee improved agent perf…

→ View original source

Do AI Personas Grow? Analyzing and Benchmarking Personality Evolution in LLM Agents After Life Events

Researchers investigated whether personality-conditioned Large Language Model (LLM) agents undergo plausible evolution following significant life events. By utilizing the Big Five traits as a psychometric framework, the …

→ View original source

DataSpace: Benchmarking Data Agents for Verifiable Analytics over Heterogeneous Workspaces

Researchers have introduced DataSpace, a comprehensive benchmark designed to evaluate data agents capable of performing verifiable analytics across heterogeneous workspaces. Unlike existing benchmarks that focus on isola…

→ View original source

Teaching Nemotron Greek: Mining a Corpus, Adapting Retrieval, and Grounding Generation for Modern Greek across Specialist Domains

Researchers have developed an end-to-end adaptation of NVIDIA's Nemotron retrieval stack specifically for Modern Greek, addressing the language's absence in major multilingual retrieval benchmarks. The methodology involv…

→ View original source

Weights or Skills? A Survey of Robot-Learning Techniques: from Action-Predicting Weights to Robots that Write their Own Skills

This survey categorizes the evolving landscape of robot learning by distinguishing between two primary paradigms: vision-language-action (VLA) models that embed competence within frozen weights, and agents that generate …

→ View original source

FactorJEPA: Factorizing Monolithic Futures into Layout-Agent-Interaction Channels for Crowded and Chaotic Global South Urban Worlds

Researchers have introduced FactorJEPA, a novel predictive architecture designed to model the complex, high-density dynamics of "DENSEWORLD" environments—urban settings in the Global South characterized by heterogeneous …

→ View original source

OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models

Researchers have introduced OSReward, a new benchmark designed to evaluate the reliability of Vision-Language Models (VLMs) acting as judges for Computer-Using Agent (CUA) trajectories. As human verification is unscalabl…

→ View original source

Task-Conditional Flow Matching for Balanced Multilingual Text Embedding Adaptation

Researchers have introduced Task-Conditional Flow Matching (TCFM), a novel framework designed for the balanced adaptation of multilingual text embedding models. Recognizing that a uniform training objective often fails t…

→ View original source

From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models

In the paper "From Economic Agents to Agentic Economies: A Systems Blueprint for Economic World Models," researchers propose a comprehensive implementation roadmap for developing Economic World Models (EWMs). Unlike trad…

→ View original source

ContextMaster: Interactive Multi-Shot Video Creation via Fixed-Budget Sparse Context Routing

The researchers introduce ContextMaster, a unified framework designed for Interactive Multi-Shot Video Creation (IMVC), a setting that requires seamless transitions between text-to-video generation, reference conditionin…

→ View original source

Interpretable MEG Decoding of Perceived Speech: Cortical Sources and the Stimulus Features That Drive Retrieval

Researchers have developed a more interpretable magnetoencephalography (MEG) decoding architecture designed to retrieve perceived speech from non-invasive recordings by targeting wav2vec 2.0 audio embeddings. To improve …

→ View original source

Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay

Researchers have introduced Activity Frames, a deterministic, zero-model compilation pipeline designed to transform passive screen-capture streams into structured memory for computer-use agents. Unlike traditional method…

→ View original source

MameLoshnLM: Yiddish Language Model and Evaluation Benchmark

Researchers have introduced MameLoshnLM, the first open-source 8B-parameter language model specifically engineered for the Yiddish language. To overcome the limitations of noisy, machine-translated multilingual datasets,…

→ View original source

AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning

The researchers introduce AgentOPSD, a critic-free, recursive framework designed to improve turn-level credit assignment in long-horizon, multi-turn agentic reinforcement learning tasks. Addressing the limitations of tra…

→ View original source

EnvACE: Internalizing Environment Dynamics via World Rehearsal for Agentic Reinforcement Learning

EnvACE introduces a novel agentic reinforcement learning framework designed to mitigate the high costs and scaling limitations associated with external environment interactions during LLM agent training. Instead of relyi…

→ View original source

NOLLI: A Difficulty-Calibrated Puzzle Benchmark for Diagnosing the English-Korean Performance Gap

Researchers have introduced NOLLI, a procedurally generated puzzle benchmark designed to diagnose the performance disparities between English and Korean in large language models. The benchmark consists of 7,500 items acr…

→ View original source

Consistency-Driven Co-Evolution for Self-Supervised Cross-Representation Learning

Researchers have introduced CoCoEvolve, a novel framework designed to facilitate self-supervised cross-representation learning across chart images, tabular data, and visualization code. Addressing the inherent one-to-man…

→ View original source

ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

The researchers introduce ABSeeker, a framework designed to train long-horizon search agents by addressing the limitations of uniform credit assignment in supervised fine-tuning (SFT) and reinforcement learning (RL). To …

→ View original source

Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning

Researchers have introduced a new framework to address the challenges of long-horizon reasoning in Large Language Models (LLMs), specifically focusing on "cross-skill" tasks that require seamless transitions between diff…

→ View original source

LLaDA MoE v2: Scaling Mixture-of-Experts Diffusion Language Models

Researchers have introduced LLaDA MoE v2, a study that systematically characterizes the scaling laws and architectural requirements for Mixture-of-Experts (MoE) diffusion language models (dLLMs). By conducting an IsoFLOP…

→ View original source

When Many Answers Are Valid, Voting Fails: Symbolic Verification for Best-of-K Causal Reasoning in LLMs

Researchers have introduced CALVER (Causal Axiom-Level VERification), a training-free symbolic verification method designed to improve causal reasoning in Large Language Models (LLMs). Traditional self-consistency method…

→ View original source

When Agents Learn to Be You: Benchmarking Privacy Leakage, Impersonation Risk, and Defenses in Persona Skills

Researchers have introduced AntiSkillBench, a comprehensive end-to-end benchmark designed to evaluate privacy leakage, attribute disclosure, and behavioral impersonation within the persona-skill pipeline. This pipeline i…

→ View original source

OmniPack: Unified Token Compression for Efficient Omni-modal Large Language Models

OmniPack is a training-free framework designed to optimize the efficiency of Omni-modal Large Language Models (Omni-LLMs) by addressing the computational overhead caused by redundant audio and visual token sequences. To …

→ View original source

ContinualSkillBench: Can LLM Agents Truly Evolve Their Capabilities?

Researchers have introduced ContinualSkillBench, a dynamic evaluation framework designed to assess the ability of Large Language Model (LLM) agents to evolve their capabilities through in-context continual skill learning…

→ View original source

JoyAI-Video-Edit: Real-Time Open-Ended Video Editing with Autoregressive Diffusion

Researchers have introduced JoyAI-Video-Edit, a 16B-parameter autoregressive diffusion framework designed for real-time, open-ended video editing. Unlike traditional methods, this system enables causal generation without…

→ View original source

Wnuan: Staged Post-Training for Question Answering over Proprietary Enterprise Knowledge

The Wnuan framework introduces a three-stage post-training pipeline designed to integrate proprietary enterprise knowledge into large language models while maintaining general capabilities. The methodology involves const…

→ View original source

GEOID-Flood: A Large-Scale Multi-Modal Benchmark Dataset for Flood Segmentation

Researchers have introduced GEOID-Flood, a large-scale multi-modal benchmark designed to evaluate geospatial foundation models for flood segmentation tasks. Derived from Copernicus Emergency Management Service activation…

→ View original source

WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity

Researchers have introduced WorldExam, a hierarchical diagnostic benchmark designed to evaluate controllable video generation models as functional world models. Moving beyond simple visual fidelity, the benchmark assesse…

→ View original source

EMBL AI Librarian: Life-Sciences Knowledge Layer for AI Agents

Researchers have introduced the EMBL AI Librarian, a specialized knowledge layer designed to optimize life-sciences information retrieval for AI agents. While existing databases like Europe PMC contain over 40 million in…

→ View original source

SULAND v2: A Refined RGB Dataset and Deep Learning Object Detection Benchmark for UAV/UGV-Based SUrface LANDmine Detection Under Domain Shift

Researchers have introduced SULAND_v2, a refined RGB dataset and benchmarking framework designed to improve surface landmine detection (specifically PFM-1 and PMA-2) for UAV and UGV platforms under domain shift. The orig…

→ View original source

Not All Tokens Deserve Equal Credit: Counterfactual Sensitivity Credit Reallocation for Long-CoT Reasoning

This paper addresses the inefficiency of credit assignment in Reinforcement Learning with Verifiable Rewards (RLVR) for long Chain-of-Thought (CoT) reasoning. Current methods like Group Relative Policy Optimization (GRPO…

→ View original source

Safeguards Based on Copyable Context Cannot Provide Reliable Safety for LLMs

Researchers Pingyu Wu et al. investigate the fundamental limitations of Large Language Model (LLM) safeguards that rely on copyable context to determine safety. The authors argue that because current safeguards decide wh…

→ View original source

AISPA: User-Centric System Prompt Auditing for Large Language Model Applications

Researchers have introduced Artificial Intelligence System Prompt Assurance (AISPA), a user-centric auditing framework designed to evaluate the instructions developers use to govern Large Language Model (LLM) behaviors. …

→ View original source

Beyond Borrowed Histories: Person-Aligned User Simulation for Interactive Role-Playing Evaluation

Researchers have introduced PALATE (Person-Aligned LLM-Simulated-User Assessment with Tailored Evaluation), a scalable benchmarking framework designed to address the limitations of existing role-playing agent (RPA) evalu…

→ View original source

LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger

LedgerMind introduces a novel framework for multimodal agentic reasoning designed to address the limitations of aggregate accuracy metrics by enforcing strict provenance constraints. Rather than relying on simple final-a…

→ View original source

OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models

Researchers have introduced OmniScope, a training-free token compression framework designed to optimize inference for omnimodal large language models (OmniLLMs) by addressing cross-modal salience mismatch. Unlike existin…

→ View original source

Σ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

Researchers have introduced Σ-Mem, an online reliability memory framework designed to enhance coordination in LLM-based multi-agent systems by modeling agent competence rather than just interaction history. Unlike tradit…

→ View original source

Beyond Geometric Complementarity: Coherent Overlap in Sparse Mixture-of-Experts Routing

This research investigates the mechanisms driving sparse Mixture-of-Experts (MoE) routing, challenging the assumption that expert benefits stem solely from geometric complementarity. By introducing the Expert Subspace Se…

→ View original source

Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

This research investigates the implications of lossy verification in Speculative Decoding (SD), a technique designed to accelerate large language model inference by utilizing a lightweight draft model to propose tokens f…

→ View original source

Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

This research presents the first systematic investigation into filesystem-based memory for Large Language Model (LLM) agents, treating a directory tree of Markdown files as a unified storage medium for declarative memory…

→ View original source

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

The Qwen-UI-Agent technical report introduces a foundation GUI agent designed for real-world digital environments, including mobile, computer-use, web, and DeepSearch platforms. To bridge the gap between simulation and r…

→ View original source

Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations

Researchers have introduced Fairness Pruning, a lightweight structural intervention designed to localize demographic bias within the Gated Linear Unit (GLU) MLP layers of large language models. By employing minimally con…

→ View original source

β-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

The researchers introduce $\beta$-OPSD, a novel framework designed to improve the reliability and performance of on-policy self-distillation (OPSD) for reasoning language models. The authors identify that standard OPSD i…

→ View original source

ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

Researchers have introduced ReToken, a lightweight method designed to enhance the performance of vision-language models (VLMs) when navigating long visual contexts. To address the computational burden of processing exces…

→ View original source

SecRespond: Benchmarking AI Agents for Real-World Post-Compromise Incident Response

Researchers have introduced SecRespond, a novel benchmark designed to evaluate the efficacy of Large Language Model (LLM) agents in post-compromise incident response workflows. Unlike traditional cybersecurity benchmarks…

→ View original source

OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding

Researchers have introduced OmegaUse-OfficeVal, a novel benchmark designed to evaluate Large Language Model (LLM) agents performing long-horizon office-suite workflows through the lens of economic grounding. The dataset …

→ View original source

SpecFirst: Behavioral Specification Elicitation as a First-Class Step in Agent-Based Program Synthesis from Scratch

SpecFirst introduces a novel two-stage framework designed to improve agent-based program synthesis from scratch, specifically addressing the limitations of single-pass LLM approaches that struggle with ambiguous document…

→ View original source

TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

Researchers have introduced TurboVLA, a novel Vision-Language-Action (VLA) paradigm designed to overcome the high computational and memory overhead inherent in traditional LLM-centric $V \to L \to A$ architectures. Unlik…

→ View original source

Temporal-Distance JEPA: Plan-Aware Representation Learning for Latent World Model Predictive Control

Researchers have introduced Temporal-Distance-JEPA, a novel representation learning framework designed to enhance latent Model Predictive Control (MPC) within Joint-Embedding Predictive Architectures (JEPAs). Traditional…

→ View original source

Visual prompt engineering for video models

This research introduces Visual Prompt Engineering (VIPE), a novel technique designed to enhance the performance of video foundation models in visual reasoning tasks. Unlike traditional text-based prompt engineering, VIP…

→ View original source

Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion

Researchers Cédric Bonhomme and Alexandre Dulaunoy introduced a reproducible pipeline designed to map Common Vulnerabilities and Exposures (CVEs) to MITRE ATT&amp;CK Enterprise techniques using free-text descriptions. Mo…

→ View original source

Parallel Decoding Distillation for Fast Image and Video Generation

To address the high computational costs associated with the iterative sampling processes in video diffusion and flow models, researchers have introduced Parallel Decoding Distillation (PDD). Unlike existing acceleration …

→ View original source

Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification

To address the computational bottlenecks caused by long token sequences in diffusion transformers during video generation, the researchers introduce Sol-Attn (Sparsifying Online Attention), a training-free method for dyn…

→ View original source

The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation

This research investigates the mechanics of multi-turn long-horizon planning in foundation model agents through a controlled, unified environment. The authors analyze planning evolution across three distinct stages: acqu…

→ View original source

Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation

This research investigates the limitations of applying Classifier-Free Guidance (CFG) within On-Policy Distillation (OPD) for diffusion models. Standard OPD approaches typically employ velocity matching on the CFG-compos…

→ View original source

DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and Benchmark for Driving Style Identification

Researchers have introduced DriveDNA, a large-scale multimodal naturalistic dataset designed to benchmark personalized driving-style identification. The dataset comprises 975 hours of human-controlled driving, captured a…

→ View original source

Spectral Prior for Reducing Exposure Bias in Diffusion Models

To address the issue of exposure bias in diffusion models—caused by error accumulation and frequency-dependent SNR discrepancies during iterative sampling—the researchers propose Spectral Alignment (SPA). The authors ide…

→ View original source

Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems

The paper introduces Agentic Context Management (ACM) as a framework to address the failure modes of production AI agents, which often stem from inefficient management of conversation histories, tool definitions, and exp…

→ View original source

Closing the Loop: Training-Free Revisit Consistency for Autoregressive Generative Rendering

To address the challenge of revisit inconsistency in autoregressive generative rendering, researchers have introduced a training-free method designed to maintain 3D world persistence during long-horizon video synthesis. …

→ View original source