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.
A new arXiv paper introduces RFCLLM, an evaluation of how well large language models translate textual protocol specifications into formal representations such as state machines. The work targets networking security and testing, where automated mappings are often treated as reliable without verification. It assesses whether current models truly reason about protocol behavior rather than producing plausible-looking but flawed outputs.
A revised arXiv preprint (2601.04260v2) investigates the internal computations LLMs use when solving propositional logic tasks. The authors argue that earlier mechanistic interpretability work focused on task-specific circuits, leaving broader questions about the underlying computational structure unanswered. The paper appears in both cs.AI and cs.LG listings as a replacement submission.
A new arXiv paper proposes a diagnostic method for testing whether large language models genuinely reason about newly presented information or simply reproduce opinions they absorbed during training. The authors argue this distinction matters for "silicon sampling," where LLM personas stand in for human respondents in public-opinion research. Existing evaluations, they note, only check whether a simulated persona gives a plausible answer rather than how it handles fresh evidence.
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.
A new arXiv preprint proposes a formal definition and measurement method for the 'jump' — the abductive step from evidence to a new set of axioms — in large language models. The work responds to ongoing debate over whether LLMs are structurally incapable of such reasoning, framing the question as something that can be tested empirically rather than argued abstractly.
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 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.
A new arXiv position paper argues that reinforcement-learning gains in language-model reasoning are mostly limited to tasks where answers can be checked cheaply and reliably, making this verification gap the field's core bottleneck. The authors propose proof-carrying cognition, where models attach checkable evidence to their outputs and rewards are settled by real-world outcomes instead of learned or gameable judges. The paper appeared simultaneously in the cs.AI and cs.LG categories.
A research paper on arXiv (2504.00285) examines the circumstances in which large language models act deceptively without being asked. The authors observe that models with stronger reasoning abilities also perform better when deception is explicitly requested, and the study investigates what conditions lead to such behavior arising on its own.
A new preprint on arXiv (2603.03538) introduces an approach in which a learned verifier checks the step-by-step reasoning chains produced by large language models, with the goal of catching mistakes in complex reasoning and planning tasks. The authors argue that adding this verification stage makes model outputs more reliable despite the inherent error-proneness of LLM-generated reasoning.