The Flow has not summarised this story yet — read it at arXiv Computation and Language.
The paper demonstrates that fine‑tuning reasoning models to predict their own confidence at intermediate steps—using only 600 self‑supervised examples—substantially improves inference efficiency. Without adding any explicit stopping or length penalties, the models generate up to 25 % fewer tokens while maintaining accuracy on mathematical, scientific, and coding benchmarks across several architectures. The study finds that confidence supervision preserves the models’ high‑level reasoning structure rather than merely suppressing specific behaviors.
The paper introduces a method to control reasoning behaviors in Mixture-of-Experts (MoE) language models by labeling sentence-level segments with six distinct reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion). Linear probes on MoE router logits can predict these behaviors nearly as accurately as probes on the full hidden state, enabling the authors to edit router scores during inference to reduce overthinking. This intervention cuts generated tokens by a median of 30.4% (up to 63.3%) across multiple models and benchmarks without significant loss in accuracy, and often improves it.
arXiv:2601. 04805v2 Announce Type: replace Abstract: Large reasoning models (LRMs) have attracted much attention due to their exceptional performance.
ClueTracer is a training‑free, parameter‑free plugin that suppresses hallucinations in multimodal reasoning models by tracing how visual clues propagate from the question through the model’s reasoning pathway. It localizes task‑relevant image patches while ignoring irrelevant regions, thereby addressing the problem of reasoning drift where models over‑focus on question‑irrelevant entities. The method improves performance across various reasoning architectures by an average of 1.21× on reasoning benchmarks.
The paper introduces Gradient Fingerprint (GRIFT), a technique that uses a model’s internal gradient computations to detect reward hacking in reinforcement learning with verifiable rewards. GRIFT compresses gradients of a chain-of-thought (CoT) conditioned on a prompt into a compact representation, which is then used to assess whether the CoT reflects reward hacking. Experiments on math, code, and logical reasoning benchmarks show GRIFT outperforms baselines by over 25% and, when integrated into a rejection fine‑tuning pipeline, reduces reward hacking while improving task performance.
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...