- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
In this paper, the authors study how a reasoning model can recover after harmful reasoning has entered its chain of thought. For that during training, they insert harmful and persuasive drift into the model’s own reasoning and train it to continue safely, using a frozen copy of the model as the teacher. The teacher judges each continuation twice: once after a note saying the reasoning was manipulated (Recognition), and once with the corrupted reasoning removed and a note stating the right final behavior (Response). On Qwen3-4B, 55.5% of safe responses stay safe after H-CoT hijacking (vs. at most 30.2% for baselines), HarmBench attack success falls from 30.9% to 6.2%, and over-refusal stays below 4%.
Critical strengths
Prior work has shown that safety alignment often depends on the first few output tokens [1]. Attack such as H-CoT [2] corrupt the reasoning midway, and post-training approaches that start from clean reasoning, such as STAR-1 [3], fail against them. The proposed method ImmuneCoT, trains the model directly on corrupted reasoning so that it learns to recover after the reasoning has gone wrong. This targets a gap that current safety training does not cover.
[1] Safety Alignment Should Be Made More Than Just a Few Tokens Deep, 2025. [2] H-CoT: Hijacking the Chain-of-Thought Safety Reasoning Mechanism, 2025. [3] STAR-1: Safer Alignment of Reasoning LLMs with 1K Data, 2026.
Critical weaknesses
The ablation results does not clearly show that combining the two signals is what helps. Specifically, the authors hypothesize that Recognition and Response are complementary. However, combining them without the adjustment against the corrupted-context score gives 27.5% HarmBench ASR, worse than Response alone (20.9%). The authors should test Response alone with the same adjustment.
Questions for authors
- How does a Response-only variant perform when it uses the same adjustment against the corrupted-context score as the full method?
- Just curious, does recovery hold when the harmful drift comes from the model itself, as in self-jailbreak [1], rather than being inserted?
[1] Mao, Y., Zhang, C., Wang, J., Guan, X., Cao, B., Lu, Y., Lin, H., Han, X. and Sun, L., 2026, July. When models outthink their safety: Unveiling and mitigating self-jailbreak in large reasoning models. In Findings of the Association for Computational Linguistics: ACL 2026 (pp. 22274-22302).
2 comments
Follow-up to Reviewer bDvD
Dear Reviewer bDvD,
Thank you again for suggesting the Response-only control with explicit Base adjustment and the evaluation on model-generated drift. With the discussion period closing soon, we would appreciate hearing whether these two additions address your questions. If anything remains unclear about the control or the recovery results, we would be happy to clarify.
Dear Reviewer bDvD,
Thank you for your careful review of this paper and for identifying the key comparison missing from our validation of the complementarity of the two signals. Your question about harmful drift produced by the model itself also prompted us to extend the recovery experiments to naturally generated drift prefixes.
W1&Q1: Response-Only with Explicit Base Adjustment
Thank you for requesting this control. Response-only with inverse-Base weighting yields 28.4% HarmBench ASR versus ImmuneCoT's 6.2%; Base division alone does not reproduce the gains.
Here, N normalizes over vocabulary, qB is the floored corrupted-context Base score, and q~rec,q~res are softly anchored scores:
P_{\mathrm{full}}&=\mathcal N[\tilde q_{\mathrm{rec}}\tilde q_{\mathrm{res}}/q_B],\\ P_{\mathrm{PoE\ w/o\ Base}}&=\mathcal N[\tilde q_{\mathrm{rec}}\tilde q_{\mathrm{res}}],\\ P_{\mathrm{Response\text{-}only}}&=\mathcal N[\tilde q_{\mathrm{res}}],\\ P_{\mathrm{Response+explicit\ Base}}^{\mathrm{ideal}} &=\mathcal N[\tilde q_{\mathrm{res}}/q_B]. \end{aligned}$$ Original Response-only sets the Recognition weight to zero, giving $\tilde q_{\mathrm{rec}}=q_B$ and cancelling the denominator. The new control omits this factor; its normalized score uses $\min\{\log\tilde q_{\mathrm{res}}-\log q_B,5\}$ and per-token loss clip 10. | Method | HarmBench ASR ↓ | WJ-H ASR ↓ | [email protected] ↓ | | --- | --- | --- | --- | | Response-only | 20.9% | 65.1% | 15.7% | | PoE w/o Base adjustment | 27.5% | 77.3% | 15.3% | | Response-only + explicit Base adjustment (new) | 28.4% | 73.8% | 24.0% | | **ImmuneCoT** | **6.2%** | **47.8%** | **0.6%** | **The full target outperforms both single-signal targets and unadjusted PoE on all three metrics.** This supports the complete fusion design, rather than unadjusted combination or inverse-Base weighting alone. The stabilized ratio control is not a strictly matched one-factor comparison. **Revision note:** We add this control, the formulas, and numerical settings. Thank you for helping us clarify the source of the gains. ### Q2: Recovery from Model-Generated Harmful Drift Thank you for this question. It prompted us to test whether learned recovery also holds when the harmful reasoning originates from the model itself. **ImmuneCoT achieves 35.72% SRR (514/1,439), with 176 explicit R³ safe outcomes (12.23%), the highest R³ rate among the ten methods.** Using RQ3's paired settings and safety threshold, we test the 500 longest drift prefixes from 572 automatically screened Base trajectories on WildJailbreak, without rewriting. Ten models receive identical requests/prefixes, with five continuations and paired Clean outputs each (2,500 per condition). Missing scores and empty final answers are excluded. SRR = S→S/Clean-safe; S→S means safe under both conditions. DeepSeek-v4-pro applies the frozen RQ3 stage prompt, withholding model identities and final answers. Ordered Rollback requires Recognition followed by explicit withdrawal; R³ additionally requires final safety. All 8,811 eligible continued-reasoning trajectories received annotations with validated exact evidence spans and ordering. | Method | Valid pairs | Clean-safe | S→S | Recognition | Ordered rollback | Explicit R³ safe | SRR ↑ | | --- | --- | --- | --- | --- | --- | --- | --- | | Base | 2486 | 176 | 38 | 10 | 5 | 4 | 21.59% | | SafeChain | 2474 | 805 | 423 | 71 | 41 | 33 | 52.55% | | UnsafeChain | 2483 | 1052 | 116 | 38 | 25 | 9 | 11.03% | | STAR-1 | 2495 | 1668 | 181 | 123 | 85 | 33 | 10.85% | | AdvChain | 2447 | 1340 | 87 | 553 | 277 | 37 | 6.49% | | SFT | 2491 | 512 | 96 | 85 | 67 | 27 | 18.75% | | DPO | 2471 | 536 | 101 | 78 | 65 | 29 | 18.84% | | OPSD | 2497 | 1048 | 272 | 216 | 171 | 106 | 25.95% | | GRPO | 2486 | 236 | 59 | 19 | 11 | 6 | 25.00% | | **ImmuneCoT** | **2496** | **1439** | **514** | **275** | **216** | **176** | **35.72%** | Models have different Clean-safe subsets; on common pairs, ImmuneCoT exceeds STAR-1 and OPSD by 25.65 and 17.84 points. OPSD and SafeChain reach 10.11% and 4.10% R³, respectively. SafeChain has higher SRR (52.55%); UnsafeChain reaches 11.03%. On 619 common Clean-safe pairs, SafeChain also leads: 59.61% versus 39.10%. Both train on WildJailbreak-derived data; UnsafeChain additionally learns corrected safe CoT and answers for unsafe Base cases. **Under shared harmful prefixes generated by Base itself, ImmuneCoT improves safe recovery and exhibits complete visible R³ trajectories.** **Revision note:** Appendix B.6 adds prefix construction, settings, recovery results, and process analysis; RQ3 retains H-CoT and references this supplement. Thank you for motivating this additional test. --- Thank you again for your in-depth analysis and suggestions. If the added control and supplementary experiment adequately address your concerns, we sincerely hope you will consider re-evaluating this paper.