- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper investigates whether the order in which a set of benign downstream tasks is presented during sequential fine-tuning affects a model's final safety. Furthermore, it explores whether accurately assessing the immediate impact of a specific task on safety can help identify training sequences that yield better long-term safety outcomes.
Then this paper proposes P-DC-SCTR, a method that encodes the current state, the local response to a candidate task, and the set of remaining tasks into a 16-dimensional feature vector, which is then fed into a small MLP to predict the maximum and minimum potential final-state risks resulting from the selection of that task.
Critical strengths
-
The paper further refines the phenomenon of "safety degradation resulting from benign fine-tuning" into the concept of "sequential effects under a fixed set of tasks."
-
It demonstrates that, even when the set of tasks, the training data, and the training budget for each task remain constant, the order in which tasks are performed can significantly alter the final safety outcome.
-
This paper introduces P-DC-SCTR, which combines information about the current state, the candidate transition, and the remaining-task context to estimate these path values.
Critical weaknesses
-
The introduction contains little discussion of the authors' own proposed method, and the explanation of how the P-DC-SCTR method addresses the challenges mentioned in the introduction is not sufficiently clear.
-
The experimental results presented in the main text are somewhat limited. Given the critical importance of ablation studies, the results from these experiments should be included in the main text rather than relegated to an appendix.
-
As shown in Table 4, removing , results in minimal performance change, with some metrics even showing improvement, while the remaining task metrics contribute the most significantly. Consequently, the current results fail to clarify the following: Does the performance gain stem from effectively modeling the long-term consequences of state-dependent conditions, or from learning the fixed structural composition of the four tasks?
-
The primary training and testing for P-DC-SCTR are based on a single Qwen3-0.6B model, a fixed set of four tasks, and identical task data; variations in experimental results stem primarily from the inherent randomness of the training process. This approach falls significantly short of addressing scenarios that involve novel task combinations, different models, or long-term safety forecasting for sequences of tasks. If the number of tasks is increased—and larger or more diverse models are employed—would the existing experimental conclusions still hold? Furthermore, are there inherent limitations to an exhaustive experimental methodology?
Questions for authors
- The precise formulas for B/Q/C, as well as and , remain insufficiently clear.
- This safety anchor employs a LoRA with a rank of 8; however, it remains to be determined which specific parameters should be updated during the subsequent fine-tuning phase, and whether the same adapter should continue to be used.
7 comments
Response to Weakness 2
We thank the reviewer for emphasizing the importance of the ablation studies. We agree that the key numerical comparisons should appear in the main text alongside their interpretation. We will include the feature-group ablations and grouped-input comparison in Section 5.4. The submission summarizes the findings at p. 8, lines 427–431, with detailed values in Table 4 and Appendix E.2. The key comparisons are shown below; each entry reports and Spearman correlations, respectively.
| Removed feature group | seed43 | seed44 |
|---|---|---|
| None (Full model) | 0.684/0.721 | 0.758/0.759 |
| , | 0.673/0.730 | 0.740/0.777 |
| 0.698/0.734 | 0.756/0.756 | |
| Remaining-task indicators | 0.549/0.590 | 0.543/0.587 |
In these feature-group ablations, removing remaining-task indicators produces the largest correlation decrease. Removing the two R signals or has a smaller effect, with some metrics improving. Because each deletion retains other state and preview information, the grouped-input comparison below further separates the contributions of task context, state statistics, and preview measurements.
Response to Weakness 3
To examine the roles of task context and state-conditioned information, we compare complete input groups, their candidate-selection decisions, and terminal outcomes under matched training histories.
All configurations use the same development tree and fixed hyperparameters. Context-only uses depth, candidate identity, and the post-candidate remaining-task mask (9 dimensions). No-preview adds parent-state statistics (11 dimensions); local-only uses parent-state statistics and the five preview features (7 dimensions); Full uses all 16 inputs. Structural lookup predicts the mean development-tree target for the same depth, candidate, and remaining-task set. All predictors are evaluated without fitting to the held-out labels. Results are shown below; entries are Spearman correlations.
| Held-out tree | Full | Context-only | No-preview | Local-only | Structural lookup |
|---|---|---|---|---|---|
| seed43 | 0.684/0.721 | 0.742/0.728 | 0.721/0.715 | 0.514/0.541 | 0.819/0.851 |
| seed44 | 0.758/0.759 | 0.807/0.764 | 0.738/0.763 | 0.439/0.542 | 0.884/0.871 |
| seed45 | 0.714/0.691 | 0.810/0.760 | 0.774/0.749 | 0.252/0.325 | 0.910/0.874 |
| seed46 | 0.780/0.763 | 0.847/0.801 | 0.795/0.791 | 0.441/0.487 | 0.942/0.870 |
Context-only has higher correlation point estimates than Full for both targets on all four holdouts, and structural lookup is higher still. These results identify task context as the strongest source of terminal-value predictability among the evaluated representations.
To examine whether the input comparison also matters for candidate selection, we compare action regret within the same parent state. For selected action , high-risk regret is , and low-risk regret is . Lower is better. Mean regret over the 17 decision parents per tree is shown below; entries are high/low risk in percentage points.
| Held-out tree | Full | Context-only |
|---|---|---|
| seed43 | 1.18/0.88 | 0.91/0.88 |
| seed44 | 0.85/1.21 | 0.21/1.21 |
| seed45 | 0.65/0.59 | 0.35/0.47 |
| seed46 | 0.26/0.47 | 0.21/0.47 |
Context-only has equal or lower mean regret throughout, and both configurations select the same low-risk complete order on all four holdouts. Full and structural lookups each obtain better low-risk complete-order outcomes on different trees; pooled prediction correlation alone therefore does not determine the quality of the selected complete order.
To check whether task context omits terminal differences associated with training history, we compare matched state–action pairs with identical structural inputs and reversed first-two-task orders. Across five trees, including the development tree, the 60 pairs have the same final two tasks; their mean absolute terminal-ASR difference is 2.97 pp, and the maximum is 9.0 pp. The matched pairs describe terminal differences associated with the preceding training history. On the holdouts, the zero-difference baseline predicts these pairwise differences more accurately than Full, distinguishing observed history-related variation from the estimator’s ability to predict it.
Together, these comparisons show the central role of task context in the evaluated terminal-value predictions and characterize outcome variation across different training histories. The realized one-step oracle in Section 5.3 separately motivates terminal-value estimation by quantifying the mismatch between immediate safety effects and complete-order outcomes.
Response to Weakness 1
We thank the reviewer for the helpful suggestion to clarify the introduction. We agree that the connection between the local-to-global safety gap and P-DC-SCTR should be more explicit. To address this, we will expand the method overview in the introduction to connect the challenge, terminal-value objective, and estimator design.
The challenge is to evaluate a candidate task by the safety outcome after remaining tasks have been completed. We therefore formulate the prediction targets as Q+(s,a) and Q−(s,a): the highest and lowest terminal ASRs reachable after choosing task a and completing the remaining tasks in different orders.
DC-SCTR uses a 20-step preview from the current checkpoint to measure the candidate’s immediate behavioral response. P-DC-SCTR combines these measurements with parent-state behavioral statistics and remaining-task context to predict the terminal targets. Its supervision is computed from reachable terminal outcomes in the development tree. Thus, the preview describes a local transition, while the learning objective concerns the completed adaptation trajectory.
These components connect the research challenge to local transition measurement and terminal-value estimation. Section 5.4 evaluates terminal-value prediction and complete-order selection separately.