A new arXiv preprint proposes locating the specific neural components that matter most for reasoning tasks, then modifying model activations to steer outputs accordingly. The authors argue this approach can make inference on hard problems more dependable without extra post-training or costly sampling. The work is presented as a way to improve reliability and efficiency during deployment.
Researchers introduced PhysMent, a benchmark designed to test how well large language models reason about physical systems by running experiments rather than answering static questions. The work argues that strong scores on existing science benchmarks do not show whether models can actively probe the physical world. The abstract notes that this ability remains poorly understood.
A new arXiv paper introduces HISPO (Hierarchical Importance-Sampling Policy Optimization with Entropy-Derived Segments), a reinforcement learning approach aimed at reinforcement learning with verifiable rewards. The method addresses the difficulty of assigning credit across long solution traces by splitting them into segments derived from entropy, so that different parts of a model's mathematical reasoning receive appropriate weight during training. The work is a cross-listed submission on arXiv's machine learning category.
A new arXiv preprint examines the kinds of mistakes large language models make while working through mathematics problems, grouping them into distinct error categories. The authors use that taxonomy of reasoning failures to target improvements in the models' mathematical problem-solving. The work aims to give a clearer picture of where current LLM reasoning breaks down and how to address it.
Researchers present Bellman Policy Optimization (BPO), a reinforcement learning approach for training large language models with verifiable rewards that does not require a separate critic network. The method is derived from Policy Mirror Descent and targets autoregressive generation. It aims to improve reasoning performance in LLMs while simplifying the training setup.
A revised arXiv paper examines a failure mode in which large language models lose reasoning ability on out-of-distribution tasks as model depth grows. The authors report that giving models access to tools mitigates this depth-related collapse, and they note that OOD generalization remains hard to measure reliably.
A new arXiv paper introduces Hindsight-Anchored Policy Optimization, a method for training large language models with verifiable rewards. It uses hindsight learning combined with a Thompson sampling-inspired adaptive gate to address cold-start problems in sparse-reward, on-policy training. The approach builds on mixed-policy methods that blend off-policy and on-policy data.
A new arXiv preprint describes a test-time adaptation technique for improving large language model reasoning on downstream tasks without expensive post-training. The method builds on predictive entropy as a model-derived signal but adds a stability-aware component to guide adaptation. The abstract presents the approach as a lightweight alternative to retraining or fine-tuning.