The paper introduces REACT, a method that intervenes early in a chain‑of‑thought (CoT) generation by applying a learned safety direction to an early hidden state, without altering model weights. Experiments on distilled reasoning models show that the most effective control occurs within the first two to three tokens, with diminishing returns thereafter. On an 8B model, REACT increases the proportion of fully safe CoT answers from 29.06% to 82.05% while maintaining baseline performance on benign refusal and usefulness.
arXiv:2606. 25013v1 Announce Type: new Abstract: Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts.
The paper introduces TRUST, a token‑level reward model that predicts the correctness of partial chain‑of‑thought (CoT) traces in vision‑language‑action (VLA) policies, enabling monitoring and selective steering of reasoning. On driving and manipulation VLA benchmarks, TRUST improves reasoning accuracy and reduces collision rates and trajectory errors, though its impact on overall task performance varies across tasks. The study defines two evaluation axes—correctability and actionability—to assess when CoT can serve as a runtime safety interface.
The paper investigates why chain-of-thought (CoT) reasoning in language models is hard to control, attributing the issue to an entropy collapse in the first few tokens after the opening thinking tag. By applying two inference-time interventions—prefilling a single token and banning frequent opening tokens with in-context examples—the authors significantly improve CoT controllability, especially for addition and style tasks. The study also distinguishes between addition/style and suppression controllability, revealing different underlying mechanisms and correlations with monitorability across models.
The paper introduces GUARD, a method for natural forgetting in large reasoning models that transforms unsafe disclosures into safe-exit trajectories using guided answer‑reasoning distillation. It aligns a frozen model with guidance tokens and distills this behavior into the parameters, aiming for a coherent, non‑disclosing chain of thought followed by a refusal‑style answer. The authors also propose the Natural Forgetting Reasoning Score (NFRS) to evaluate structural stability, fluency, and unsupported substitutes, and demonstrate GUARD’s effectiveness on R‑TOFU and a STAR‑1‑derived harmful‑intent setting.
arXiv:2608. 03842v1 Announce Type: cross Abstract: When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate.
arXiv:2511. 05852v4 Announce Type: replace-cross Abstract: Knowledge editing (KE) offers a lightweight alternative to retraining for updating large language models (LLMs).
arXiv:2605. 24396v2 Announce Type: replace Abstract: Long chains of thought (CoT) from current language models frequently contain logical gaps and unjustified leaps, limiting the gains from additional test-time compute.
arXiv:2609.13737v1 Announce Type: new Abstract: As large language models (LLMs) are increasingly deployed, the generation of harmful content has become a critical safety concern. Existing safeguards...
arXiv:2607. 06503v1 Announce Type: new Abstract: Large language model (LLM) agents solving multi-step tasks frequently commit to trajectories that are doomed to fail, yet continue to consume substantial inference compute before the failure becomes observable.
arXiv:2610.15365v1 Announce Type: new Abstract: Large language models (LLMs) can produce harmful content while generating a response. A useful safeguard should not only detect this risk during genera...
arXiv:2610.22640v1 Announce Type: new Abstract: Model-editing backdoors aim to induce harmful responses under a trigger while preserving ordinary behavior on untriggered input. Recent model-editing a...
EOPSA (Efficient On-Policy Self-Distilled Safety Alignment) addresses inefficiencies in On-Policy Self-Distillation (OPSD) for safety alignment by focusing training on safety-critical tokens. It introduces Adaptive Rollout Scheduling, which limits generation length based on a Teacher Rescue Rate metric, and Selective Distillation, which filters out safety-neutral tokens to concentrate gradient updates on safety-pivotal transitions. Experiments on models up to 32B parameters show that EOPSA reduces rollout computation by about 50% and backpropagates through only roughly 2% of tokens, outperforming full-token distillation baselines in safety compliance and reasoning retention.