REACT: Representational Early Analysis and Control of Chain-of-Thought Trajectories
원문 초록
Chain-of-thought (CoT) generation offers an opportunity to intervene before a reasoning model completes its answer. How does the safety effect of a single activation edit vary across early reasoning states, and what factors govern that variation? We study this question with REACT, a conditional activation intervention that learns a safety direction from paired safe and unsafe trajectories and applies it once to an early CoT hidden state. The model weights remain frozen, and all subsequent tokens are generated without further edits. To study timing, we train directions separately at each of the first five CoT positions and evaluate ungated edits with full post-edit regeneration. Across three DeepSeek-R1 distilled models, site-specific edits concentrate safety gains in the opening CoT tokens. Norm-matched controls on DeepSeek-R1-Distill-Llama-8B (8B) disentangle two factors behind this curve: holding an effective direction and edit budget fixed still yields much larger early ($t_1$-$t_5$) than later ($t_6$-$t_8$) gains, whereas the near-zero site-specific $t_3$ gain reflects higher optimisation variance despite rising risk-probe AUROC. Forced prefixes shift the high-gain region with the opening context, and a four-arm intervention shows that ~98% of the net gain at $t_2$ propagates through the changed next token. On the same 8B model, REACT raises the rate of fully safe CoT-answer outputs on StrongREJECT-273 from 29.06% to 82.05%, while keeping benign refusal and usefulness near baseline. An anonymized implementation is available at https://anonymous.4open.science/r/REACT_official2-333B.
제출 이력
확인된 버전만 표시합니다. 전체 제출 이력은 arXiv 원문에서 확인하세요.
- arXiv:2610.32113v2제출 시각 미확인
함께 살펴볼 논문
공통 태그를 근거로 연결했습니다. 실제 인용 관계나 논문의 우수성을 뜻하지 않습니다.
공통 태그로 연결된 공개 논문이 아직 없습니다.
토론
이 논문과 연결된 포럼 글이 아직 없습니다.