acceptodds
Withdrawn from ICLR 2027

TrajEdit: Safety Editing with Multi-Anchor Trajectories

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Rapidly patching vulnerabilities in Large Language Models (LLMs) against evolving jailbreak attacks remains a critical challenge. While Knowledge Editing offers a resource-efficient alternative to retraining, existing methods indiscriminately update parameters over the entire response sequence, often disrupting general factual knowledge irrelevant to the safe direction associated with the target correction. This inefficiency leads to severe over-refusal on benign queries, catastrophic forgetting of general capabilities, and limited generalization to broader safety-related tasks. In this paper, we propose TrajEdit, a solver-agnostic framework that reframes safety correction as sparse, multi-anchor trajectory editing. It constructs a safe trajectory for each selected anchor and jointly edits the corresponding divergence points, thereby reducing updates to generic content. In our implementation, the first chunk is retained as an anchor, while perplexity-based uncertainty serves as a simple and effective option for selecting additional anchors. Extensive experiments across LLMs demonstrate that TrajEdit achieves superior defense success while maintaining a high non-refusal rate comparable to leading baselines, effectively reconciling the trade-off between safety and locality.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model editing, safety and alignment, adversarial attacks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UPMZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TrajEdit is a framework for editing unsafe behavior in language models through sparse, multiple-anchor constraints. Instead of supervising an entire safe response, it selects the first response chunk and additional high-perplexity chunks from an unsafe generation. For each selected anchor, it constructs a safe continuation that shares the preceding harmful prefix, then applies an existing editing solver at the corresponding divergence point.

Critical strengths
  1. Empirical support for localization. The annotated harmful-onset evaluation goes beyond illustrative examples and reports quantitative evidence that perplexity carries localization information.
  2. Substantial capability preservation relative to the selected baseline.
Critical weaknesses
  1. The sequential evaluation is insufficiently specified. It is unclear which model checkpoint generates each reported measurement, how previous edits are retested, and whether generalization cases remain disjoint from later editing inputs.
  2. Capability preservation is incomplete and narrowly evaluated. Table 4 evaluates one model and one baseline configuration. TrajEdit still reduces MMLU from 0.4125 to 0.3345, an absolute drop of 7.8 percentage points.
  3. Uncertainty and cost reporting are incomplete. Results lack confidence intervals and variability across edit orders or random truncation choices. Editing-time units and whether timing includes anchor scoring, safe-target generation, and discriminator calls should also be specified.
Questions for authors

See weakness

Reviewer_LMo8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for safety-oriented knowledge editing that focuses on selected locations within harmful response trajectories rather than applying supervision to the entire target response. It divides harmful responses into fixed-length chunks and selects a sparse set of anchors consisting of the first chunk and additional high-perplexity chunks. For each anchor, the method constructs a target trajectory by retaining the preceding response prefix and generating a safe continuation. The authors additionally employ a dynamic classify procedure before editing to avoid unnecessary updates during sequential editing.

Critical strengths
  1. I think this paper represent a useful formulation of localized safety editing, which corrects harmful behaviors without unnecessarily affecting benign generation capabilities.
  2. the framework can be integrated with existing knowledge-editing algorithms rather than requiring a completely new procedure. This makes the approach potentially applicable across different editing backbones.
Critical weaknesses
  1. It is unclear to me why maximizing this conditional mutual information corresponds to learning a better safe continuation. If the modified hidden state already preserves information about the adversarial input, conditioning on that state may reduce the proposed quantity.
  2. The ablations do not convincingly distinguish the advantage of PPL-based localization from the general benefits of sparse supervision or including the first response chunk. In Table 3, First + Top-1 and First + Top-2 achieve identical success and generalization rates. The First Only baseline achieves an even higher non-refusal rate of 76.5%, kind of suggesting that additional anchors do not consistently improve the safety–locality trade-off. Appendix D.5 compares First + PPL against random anchor selection, but the random baseline does not fix the first chunk. I think the comparison may confound the effect of PPL ranking with the inclusion of the first anchor.
Questions for authors
  1. Could the authors clarify the information-theoretic derivation in Appendix B? What assumptions justify decomposing the global conditional mutual information into the proposed local terms? Why is maximizing I(A; R_safe | h') the appropriate objective, and how does it lead to Equation (3)?
  2. Can the authors provide controlled comparisons isolating PPL-based anchor selection?
  3. Could the authors report confidence intervals for the main results?
Reviewer_sCfhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper propose TrajEdit, a model editing method for LLM safety. Specifically, TrajEdit decomposes the initial harmful response to multiple chunks, and constructs the safe response for each chunk. Then, these multiple triplet of (prefix, harmful suffix, safe suffix) are selectively utilized to extract editing direction. The proposed method has been demonstrated on several benchmarks with Qwen3-8B and Llama-3.1-8B.

Critical strengths
  • The proposed idea is valid and its empirical effectiveness over the baseline has been demonstrated.
Critical weaknesses
  • Some important information and results are omitted. As the proposed method significantly relies on the constructed safe refusals for each chunk (lines 210-212), the way to construct and their quality is important; but, this information, details of P and the validity of R^(i)_safe, is not presented now. Also, in Table 4, there is only comparison with UnKE; it should be extended to include the comparison with other baselines similar to Table 1. Lastly, in D.3 and D.4, only the success rates are presented but other metrics (Gen. and Loc.) should be presented for these experiments.
  • Presentation quality. The current draft has many rooms to be improved. First, in my opinion, the space of main pages is not utilized effectively. Currently, some figures occupy lots of space (e.g., Figure 2 and Figure 4), but I don't know whether they are necessary. This is critical because many important information or results are currently in Appendix, which makes readability poor. For example, all the definitions to understand Section 2 are in Appendix A (line 96). By definition, Appendix is not mandatory to read and should be used for supplementing the draft.
Questions for authors
  • What are PPL-based locations in Section 2? Do they mean top-k high-likelihood chunks as in Section 4?
  • What is hh in Eq.2? Does it mean the hidden feature at the last token of the common prefix as defined in line 244? Then, which layer is used? No such information now.
  • In Table 2, the proposed method requires shorter time for editing; but, the proposed method requires to generate the safe refusal per chunk, so it's natural to require more time than others. So, does time for the proposed method in Table 2 include generation time for safe refusals?
  • In Figure 5, performances at step 0 are different across the methods, which is unnatural as step 0 indicates no editing. How does it happen?
  • Table 3: No reason to write ID column. Also, for the column of Flu., the current bold is wrong.
  • It's hard to follow the intention of section 4.5. Why this experiment is important?
  • Table 5: currently, k is omitted for columns of TrajEdit and TrajEdit*

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.