The paper introduces Hybrid Decision Models (HDM), a "zero-or-T thinking" paradigm where a single frozen backbone either answers instantly via implicit reasoning in one forward pass or invokes full autoregressive reasoning costing T tokens. A lightweight self-routing block estimates the "value of thinking"—the expected utility gain of slow over fast mode—from the backbone's implicit representations, triggering slow-mode generation only when this value exceeds a request-specific price λ. For episodic agentic tasks lacking per-decision labels, the authors propose Calibrated Decision Policy Optimisation (CDPO), a reinforcement-learning method training the router from episode-level returns. Evaluated on MMLU-Pro, AGIEval, GPQA-Diamond, JevBench, BANKING77, and MiniWoB, HDM-4B achieves 0.857 accuracy on JevBench versus Jev's 0.851 (+0.7%) at a median latency of 0.179s against 0.652s (-72.5%). On MiniWoB, HDM-27B reduces token cost by 80.1% while maintaining performance, demonstrating efficient adaptive computation allocation.
Read original
huggingface/daily-papers