AtManRL introduces differentiable attention manipulation to train salient chain-of-thought reasoning in large language models. By optimizing an additive attention mask that identifies the tokens necessary to recover the model’s answer, the method derives a saliency reward combined with an outcome reward in a GRPO framework with decoupled advantages. Across three models and four benchmarks, AtManRL improves saliency and faithfulness metrics while maintaining comparable accuracy to length‑penalty baselines.
arXiv:2510. 13554v2 Announce Type: replace-cross Abstract: The reasoning pattern of Large language models (LLMs) remains opaque, and reinforcement learning (RL) typically applies uniform credit across an entire generation, blurring the distinction between pivotal and routine steps.
arXiv:2608.31066v1 Announce Type: new Abstract: Chain-of-thought (CoT) reasoning improves multi-step problem solving, but long reasoning traces inflate inference cost. Token-level CoT compression red...
The paper introduces RECAP, a training method that enhances vision‑language models by incorporating temporally localized reactivation of image evidence—called visual revisit—into the reinforcement learning process. Visual revisit peaks are shown to predict correctness and causally improve performance; masking them hurts accuracy more than masking non‑peaks, while adding revisit peaks to failed reasoning partially restores correct answers. RECAP applies revisit‑conditioned advantage estimation, dynamic discounting, and gated value prediction to propagate final rewards through visually important reasoning steps, achieving significant gains on HallusionBench and other benchmarks without extra rewards or inference‑time changes.
The paper introduces Perception and Reasoning Attribution (PARA), a method that quantifies how much of a vision‑language model’s errors can be fixed by better perception versus stronger reasoning. PARA reveals that reinforcement learning (RL) benefits more from improving reasoning than perception, and motivates a new group‑relative attention distillation objective called GRAD. GRAD directly guides weaker rollouts toward the visual attention patterns of the best rollouts, consistently boosting performance across multiple model scales and benchmarks.
arXiv:2605.31378v2 Announce Type: replace Abstract: Large Reasoning Models (LRMs) still struggle with fine-grained translation quality estimation (QE), even with long reasoning chains. We argue that...
arXiv:2607. 09786v1 Announce Type: new Abstract: Length-penalized reinforcement learning can shorten chain-of-thought reasoning while hiding an influence that drives the model's answer.
The paper introduces Declarative Attention (DA), a protocol that lets language models explicitly declare which parts of their context to focus on during generation. By partitioning decoding into full-context, region-specific, and recent-output-only modes, the inference engine can skip large portions of the KV cache, dramatically reducing attended tokens. Experiments on 15 long-context tasks with off-the-shelf models show significant savings (52.0% and 31.1% reductions) with only modest accuracy drops that diminish as model size increases.
The paper introduces Stepwise Marginal Information Gain (MIG), an intrinsic process reward that evaluates how each reasoning step of a large language model (LLM) or vision-language model (VLM) improves the likelihood of the reference answer. MIG rewards only new likelihood maxima, preventing duplicate credit, and is combined with outcome, format, and self‑distillation objectives to guide training. Experiments on eight benchmarks show that this method outperforms outcome‑only reinforcement learning and improves accuracy by up to 4.8 points over binary‑reward training, including a 12.6‑point gain on MathVerse and a 12.9‑point advantage on vision‑language transfer at 7B parameters.
The paper investigates how training large language models to use fewer tokens in chain-of-thought (CoT) reasoning impacts the faithfulness and monitorability of the generated explanations. Three efficiency methods—fixed generation budget, per-example length target, and group-relative length reward—were applied during fine‑tuning, and the resulting models were evaluated on how well their CoT reflects decision processes and whether it signals changes due to input interventions. Results show that while faithfulness generally decreases because models become less consistent, monitorability remains relatively robust, with models still indicating the influence of input changes even when CoT is shortened.
arXiv:2608. 03550v1 Announce Type: new Abstract: Chain-of-Thought (CoT) prompting remains the standard baseline for evaluating models' reasoning abilities.
arXiv:2605.11467v2 Announce Type: replace-cross Abstract: Reasoning models post-hoc rationalize answers they have already committed to internally, producing chains of *reasoning theater*: deliberativ...
The paper introduces RECAP, a redundancy-aware credit assignment method that improves reasoning efficiency in large language models by assigning credit to each reasoning step based on its downstream role and contribution to the correct answer. RECAP uses a semantic dependency graph to measure structural responsibility and evaluates step efficacy via changes in gold-answer log-likelihood, enabling step-specific updates without requiring a separate reward model or concise trajectories. Experiments on two 7B models across four mathematical reasoning benchmarks show that RECAP enhances the accuracy-efficiency trade-off, boosting pass@1 by 2.0–3.7 percentage points while cutting reasoning tokens by 8–31% compared to GRPO.