SID AI released SID-2, a reinforcement-learning-trained agentic search model with a 262k-token context window that replaces fixed RAG pipelines by iteratively invoking developer-implemented keyword and vector search tools before returning ranked document IDs via a report_helpful_ids call. The model offers three reasoning-effort tiers—Low, Medium, High—created by merging separately trained experts via multi-teacher on-policy distillation, plus Metric Steering to target NDCG, F1, or Recall@k. A 12-agent SID-2 Swarm variant forks contexts to explore corpus slices in parallel, achieving 0.720 NDCG on SID's 506-question benchmark across 19 corpora (HotpotQA, Enron, SEC filings, PubMed, BrowseComp-Plus), surpassing GPT-6 Astra's 0.704 and Claude Opus 5.5's 0.577, while a single Medium agent scores 0.585 NDCG in 3.11 seconds versus Opus 5.5's 25.1 seconds. Swarm latency reaches 17.7 seconds for six agents. Training used a two-stage RL loop: stage one optimizes retrieval quality; stage two adds a multiplicative latency penalty on positive rewards, with a proposer-solver self-play loop generating hard queries against a frozen SID-1 solver, yielding a claimed 6x throughput gain over SLIME. Deployment is available via hosted API (platform.sid.ai), Baseten, and SageMaker with OpenAI-compatible SDKs; pricing starts at $0.25/$1.00 per million input/output tokens for SID-2 and $0.50/$2.00 for Swarm. All results are self-reported; model size and weights remain undisclosed.
Read original
marktechpost