Search

4 results for “MATH500”

1. Research · Language models

   Entropy signals guide efficient reasoning repair in LLMs

   REWIND: Uncertainty-Guided Trajectory Repair for Test-Time Reasoning

   Ding et al. introduce REWIND, a training-free method that uses token entropy dynamics to select optimal points for regenerating suffixes, improving reasoning accuracy across multiple models and benchmarks.

   MATH500: 52% accuracy (as reported)

   15 Oct 2026The Hong Kong University of Science and Technology (Guangzhou), University of Edinburgh, Huawei

2. Research · Evaluation & benchmarks

   LLMs fail multi-step math despite solving each step correctly

   Solving Every Step Is Not Enough: Milestone Oracles Reveal a Composition Gap in LLM Math Reasoning

   OracleLadder diagnostic reveals a composition gap that accounts for 33–48% of failures across six models.

   MATH500: 3 families solved directly (as reported)

   3 Oct 2026Rutgers University, Harvard University, The Hong Kong University of Science and TechnologyCode

3. Research · Language models

   Counterfactual prompts let beam search self-distill language models at inference

   Beam Search as Test-Time Self-Distillation via Counterfactual Contexts

   The method contrasts answer probabilities under excellent and poor reasoning prompts, then uses that signal to steer decoding without training.

   MATH500: 0.64 pass@1 (as reported)

   4 Oct 2026Huawei Noah’s Ark Lab, UCL Centre for AI

4. Research · Safety & security

   Learned latent links can undermine safety in multi-agent systems

   Safety of Latent Communication in Multi-Agent Systems

   The authors show that modifying representation-level communication links can increase harmful compliance without changing the safety-aligned agents themselves.

   four safety benchmarks: 76.9% mean harmful-compliance score (as reported)

   2 Oct 2026CISPA Helmholtz Center for Information Security