- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes , a violation-guided minimum-change algorithm that directly controls the contextual curvature and improves strict higher-order robustness without obtaining safety through broad refusal. Experiment resutls show that reduces ASR across orders six through eight from 22.0% to 18.0% and SCFR from 9.5% to 7.5%, while changing benign helpfulness by only 0.3 points.
Critical strengths
- The problem is real and motivated. Solving compositional attacks is important and challenging.
- The curvature-based framing of compositional attack offers an interesting theoretical perspective in understanding the mechanism of compositional attack.
Critical weaknesses
- The paper is poorly written. Specifically, this paper contains
- A lot of undefined variables/terminologies. For example, the meaning of in line 148, in Line 149, LME in line 157. It also does not provide a detailed explanation of why in line 152, why in line 166 etc. Making it extremely hard to understand the theory in Section 3 and hard to understand the results in the later sections. I also don't understand what the meaning of "policy-preserving" in line 145, since policy can refer to both "language model as the policy model " and "the safety constraints that the model should follow.
- The threat model is never stated, and the theoretical contribution reads hollow. Computing the bound requires margins at all nodes and may be hard for large .
- The reported numbers seem weird. I don't understand why most of the numbers in Table 1, Table 3, Table 8, and Table 9 end with .0 and .5. Most importantly, in Table 8 and Table 9 (in the supplementary material), for almost every method, the gap between Llama, Qwen, and Gemma is the same. See the table below
Table 8: SCFR 6:8 (%)
| Method | Llama | Qwen | Gemma | Llama→Qwen | Qwen→Gemma |
|---|---|---|---|---|---|
| Base model | 18.5 | 21.5 | 25.1 | +3.0 | +3.6 |
| Single-factor tuning | 16.0 | 18.8 | 22.2 | +2.8 | +3.4 |
| Pairwise training | 13.5 | 16.0 | 18.5 | +2.5 | +2.5 |
| Random composition | 11.5 | 14.0 | 16.5 | +2.5 | +2.5 |
| Hard-negative training | 10.5 | 13.0 | 15.5 | +2.5 | +2.5 |
| ASCOT-inspired | 9.5 | 12.0 | 14.5 | +2.5 | +2.5 |
| Margin hinge | 10.5 | 13.0 | 15.5 | +2.5 | +2.5 |
| Active margin projection | 8.5 | 11.0 | 13.5 | +2.5 | +2.5 |
| Curvature hinge | 7.0 | 9.5 | 12.0 | +2.5 | +2.5 |
| C2SA | 5.0 | 7.5 | 10.0 | +2.5 | +2.5 |
Table 9: ASR 6:8 (%)
| Method | Llama | Qwen | Gemma | Llama→Qwen | Qwen→Gemma |
|---|---|---|---|---|---|
| Base model | 37.0 | 40.5 | 44.3 | +3.5 | +3.8 |
| Single-factor tuning | 34.5 | 38.0 | 41.5 | +3.5 | +3.5 |
| Pairwise training | 30.5 | 34.0 | 37.5 | +3.5 | +3.5 |
| Random composition | 27.5 | 31.0 | 34.5 | +3.5 | +3.5 |
| Hard-negative training | 25.5 | 29.0 | 32.5 | +3.5 | +3.5 |
| ASCOT-inspired | 24.0 | 27.5 | 31.0 | +3.5 | +3.5 |
| Margin hinge | 25.0 | 28.5 | 32.0 | +3.5 | +3.5 |
| Active margin projection | 22.0 | 25.5 | 29.0 | +3.5 | +3.5 |
| Curvature hinge | 18.5 | 22.0 | 25.5 | +3.5 | +3.5 |
| C2SA | 14.5 | 18.0 | 21.5 | +3.5 | +3.5 |
| Given this, I feel not confident in trusting the numbers reported in the paper and think the paper should be clearly rejected. |
Questions for authors
I have listed my questions in the weaknesses section.