acceptodds
Under review as a conference paper at ICLR 2027

Benign Tasks, Unsafe Orders: The Local-to-Global Safety Gap in Continual LLM Fine-Tuning

Abstract

Continual fine-tuning can alter the safety of large language models even when all downstream tasks are benign. We show that this risk is inherently path dependent: applying the same tasks in different orders can lead to substantially different terminal safety outcomes. More importantly, we identify a local-to-global safety gap: state-conditioned measurements can accurately characterize the immediate safety effect of a candidate task, yet even the realized one-step effect can be a poor indicator of safety after the remaining tasks are completed. Thus, reliable local monitoring does not necessarily support reliable long-horizon task-order decisions. Motivated by this observation, we formulate continual task ordering as a terminal path-value estimation problem and introduce P-DC-SCTR, which jointly models the current state, local transition effects, and remaining adaptation path. Experiments across multiple continual fine-tuning settings consistently reveal strong order-dependent safety variation and confirm the local-to-global gap. Path-aware modeling further provides more informative estimates of terminal safety consequences than local signals, while reliable end-to-end ordering remains an open challenge. These findings suggest that safety under continual adaptation should be modeled over training trajectories rather than isolated updates.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Safety, Continual Fine-Tuning, Task Ordering, Local-to-Global Safety Gap, Path-Value Estimation, Safety Alignment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KkG8OpenReview ↗
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
  1. 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."

  2. 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.

  3. 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
  1. 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.

  2. 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.

  3. As shown in Table 4, removing RadvR_{adv}, RdirR_{dir} 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?

  4. 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
  1. The precise formulas for B/Q/C, as well as RadvR_{adv} and RdirR_{dir}, remain insufficiently clear.
  2. 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
AuthorsOpenReview ↗

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)Q^{+}(s,a) and Q−(s,a)Q^{-}(s,a): the highest and lowest terminal ASRs reachable after choosing task aa 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.

AuthorsOpenReview ↗

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 Q+Q^{+} and Q−Q^{-} Spearman correlations, respectively.

Removed feature groupseed43seed44
None (Full model)0.684/0.7210.758/0.759
RadvR_{\mathrm{adv}}, RdirR_{\mathrm{dir}}0.673/0.7300.740/0.777
B/Q/CB/Q/C0.698/0.7340.756/0.756
Remaining-task indicators0.549/0.5900.543/0.587

In these feature-group ablations, removing remaining-task indicators produces the largest correlation decrease. Removing the two R signals or B/Q/CB/Q/C 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.

AuthorsOpenReview ↗

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 Q+/Q−Q^{+}/Q^{-} Spearman correlations.

Held-out treeFullContext-onlyNo-previewLocal-onlyStructural lookup
seed430.684/0.7210.742/0.7280.721/0.7150.514/0.5410.819/0.851
seed440.758/0.7590.807/0.7640.738/0.7630.439/0.5420.884/0.871
seed450.714/0.6910.810/0.7600.774/0.7490.252/0.3250.910/0.874
seed460.780/0.7630.847/0.8010.795/0.7910.441/0.4870.942/0.870

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 a^\hat a, high-risk regret is max⁡aQ+(s,a)−Q+(s,a^)\max_{a} Q^{+}(s,a)-Q^{+}(s,\hat a), and low-risk regret is Q−(s,a^)−min⁡aQ−(s,a)Q^{-}(s,\hat a)-\min_{a} Q^{-}(s,a). 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 treeFullContext-only
seed431.18/0.880.91/0.88
seed440.85/1.210.21/1.21
seed450.65/0.590.35/0.47
seed460.26/0.470.21/0.47

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.

Reviewer_iLCaOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how the order of benign downstream tasks affects safety in continual fine-tuning. The authors show that the same set of tasks, learned in different orders, can lead to very different safety outcomes. They also argue that knowing the immediate safety effect of a task is not enough to predict the final safety after all tasks are learned. They call this the local-to-global safety gap. To address it, they propose a path-aware estimator that predicts the best and worst final safety reachable after each task choice.

Critical strengths
  1. The question is interesting and relevant. Safety under sequential adaptation is less studied than single-stage fine-tuning.
  2. The experimental design is controlled. Only the task order changes, so the order effect is clearly isolated.
  3. The paper is honest about its limits. The authors openly report that their estimator does not lead to better task orders.
Critical weaknesses
  1. The setting does not match continual learning. In continual learning, tasks arrive over time and we cannot choose what comes next. The method assumes we know all future tasks and can reorder them. This is closer to curriculum design than continual learning, which greatly limits its practical use.

  2. The main findings are not surprising. Prior work has already shown that safety is forgotten during sequential fine-tuning and that the order of tuning stages affects this forgetting [1]. Safety degradation under continual adaptation has also been framed as catastrophic forgetting and studied across task sequences [2], and alignment drift across continual domain tasks has been analyzed and mitigated [3]. The idea that safety should be studied over the training trajectory is therefore not new. That a greedy one-step choice fails on a path-dependent problem is also expected, since each update changes the starting point of the next one.

  3. The method does not solve the problem it targets. The goal is to find safer task orders, but the proposed estimator does not do this better than simple baselines. Better prediction without better decisions makes the method contribution weak.

  4. Generalization is unclear. The estimator is trained and tested on the same few tasks, and it relies on task identity features. It likely learns patterns specific to these tasks rather than a general rule. It also needs all possible orders to be trained first, so it offers no saving over simply trying them.

References

[1] Chained Tuning Leads to Biased Forgetting. ICML 2024 Workshop.

[2] Unforgotten Safety: Preserving Safety Alignment of Large Language Models with Continual Learning. arXiv 2025.

[3] Continual Safety Alignment via Gradient-Based Sample Selection. Findings of ACL 2026.

Questions for authors

Please refer to the Weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Weakness 1

We appreciate the reviewer’s thoughtful observation about task availability in continual learning. We agree that many online settings follow an externally determined task sequence. To clarify the practical setting addressed by our work, we study sequential fine-tuning of several available downstream datasets whose training order can be scheduled. Each task starts from the checkpoint produced by the preceding updates, and we examine how this scheduling decision affects terminal safety.

Task scheduling overlaps with task-level curriculum design and is also studied in continual-learning research: Bell and Lawrence’s The Effect of Task Ordering in Continual Learning, cited in our manuscript, examines how task reordering affects forgetting. Section 3.1 specifies our fixed-task setting and represents the training state through the current model parameters and remaining tasks.

We fix the task data, safety initialization, and per-task training budgets to isolate the influence of order on terminal safety and examine whether immediate safety measurements can guide the scheduling decision.

AuthorsOpenReview ↗

Response to Weakness 2

We thank the reviewer for the insightful discussion of prior safety-forgetting research and path-dependent decisions. We agree that forgetting, task-order effects, and the limitations of one-step greedy selection are important established considerations. To clarify our contribution, we examine whether accurately measured immediate safety changes reliably guide complete-order selection and relate this evaluation to the candidate-conditioned terminal path-value target.

To examine this, we enumerate all 24 orders of the same four benign tasks, keeping initialization, data, and per-task budgets fixed. On the enumerated trees, we retrospectively select the next task at each step using measured one-step ASR changes after full candidate-task training. Using fully measured candidate outcomes allows us to evaluate the immediate-safety decision criterion separately from errors in short-preview predictions.

The relationship between these studies and our focus is summarized below.

Related workRelevant contributionComplementary focus of our study
[1] Chained Tuning Leads to Biased ForgettingStudies safety/capability tuning orders and group-dependent forgetting and relates forgetting to first-task loss curvature.Safety initialization is fixed before ordering; only benign downstream tasks are permuted.
[2] Unforgotten SafetyFrames safety preservation during downstream fine-tuning as a continual-learning problem and evaluates safety-preserving methods across downstream datasets.We evaluate complete sequential adaptations of the same benign task set and compare their permutations.
[3] Continual Safety Alignment via Gradient-Based Sample SelectionStudies safety drift across domain adaptation, proposes sample selection, and evaluates alternative task orders.We hold samples and updates fixed and evaluate task-level decisions against exhaustive terminal references.

Across the five main-line trees, the results are shown below.

Selection ruleTerminal ASR (%)Distance from the corresponding observed terminal extreme (pp)
Minimize realized one-step ΔASR\Delta\mathrm{ASR}24.5–30.014.0–17.5 above the minimum
Maximize realized one-step ΔASR\Delta\mathrm{ASR}12.0–15.018.0–28.0 below the maximum

These complete trajectories are reported in Appendix D.2. In all five trees, minimizing immediate risk produces higher terminal ASR than choosing in the opposite direction. Relative to the mean over all 24 orders, the minimizing rule ends approximately 1.6–6.9 pp higher, while the maximizing rule ends approximately 8.1–10.9 pp lower.

In the additional 7B trees, greedy selection improves on the random-order mean in its intended direction, with remaining gaps to the observed extrema. These results show that the terminal effects of one-step selection differ across the evaluated model settings.

These results quantify the mismatch between immediate safety measurements and terminal-order decisions for the evaluated one-step greedy rule, even after local prediction error is removed. This motivates the candidate-conditioned terminal values defined in Section 3.3, which describe the lowest and highest terminal ASR reachable after a choice and completion of the remaining tasks. This decision-focused diagnosis complements the cited forgetting studies.

AuthorsOpenReview ↗

Response to Weakness 3

We appreciate the reviewer’s important emphasis on actual task-order outcomes. We agree that prediction accuracy should be assessed alongside the safety of the resulting complete orders. Accordingly, we compare P-DC-SCTR with realized one-step greedy, the random-order mean, copy-dev, and simplified feature representations, while also examining terminal-value prediction and within-state candidate selection.

The terminal-value target is directly connected to safer ordering. By Eqs. (5)–(7), Q−(s,a)Q^{-}(s,a) is the lowest terminal ASR reachable after choosing task aa and completing the remaining tasks, with V−(s)=min⁡a∈Rs  Q−(s,a)V^{-}(s)=\underset{a\in R_s}{\min}\;Q^{-}(s,a).

P-DC-SCTR estimates this terminal target using the current state, candidate transition, and remaining-task context.

We evaluate the estimator after fitting and hyperparameter selection on seed42 and freeze it for seed43–46. Complete-order outcomes are compared directly with realized one-step greedy, the mean over all 24 orders, and copy-dev, which replays the development tree’s best order. The low-risk results are shown below; ASR is in percent, and lower is better.

Held-out treeP-DC-SCTRRealized one-step greedycopy-devRandom-order mean
seed4315.028.012.022.19
seed4414.024.510.522.88
seed4512.530.015.023.12
seed4610.026.513.521.67

Relative to realized one-step greedy, the terminal-ASR reductions are 13.0, 10.5, 17.5, and 16.5 pp. Compared with copy-dev, the Full-minus-baseline differences are +3.0, +3.5, −2.5, and −3.5 pp, and all four paired 95% intervals in the submission’s Table 6 include zero.

The structural baselines help clarify the contribution of the full representation. Context-only and No-preview select the same low-risk order as Full in every holdout. For the high-risk stress test, where higher ASR is the objective, full yields terminal ASRs of 27.5%, 31.0%, 30.5%, and 35.0% on seed 43–46, respectively, while both simpler models reach 27.5%, 40.0%, 33.0%, and 35.0%. They match Full on two trees and better meet the maximization objective in point estimates on the other two.

We separately evaluate terminal-value prediction and within-state candidate comparison. On the four holdouts, the Spearman correlations are 0.684–0.780 for Q+Q^{+} and 0.691–0.763 for Q−Q^{-}, and low-risk terminal-optimal action agreement is 10/17, 11/17, 13/17, and 13/17 (Table 3). On seed43, this agreement is 10/17 compared with 4/17 for the realized one-step greedy. These predictions and within-state comparisons provide supporting evidence for terminal-value modeling and are interpreted alongside the complete-order outcomes above.

Together, the results support terminal-value modeling and show improved low-risk rollout ASR point estimates over realized one-step greedy across all four holdouts. The copy-dev and context-only comparisons show that stronger terminal-value prediction does not consistently translate into better complete-order selection.

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

This paper studies an interesting problem -- how the order of a fixed set of benign finetuning tasks changes terminal safety. The author conducted exhaustive four-task tress experiment to reveal substantial order effects and show that accure, one-step risk changes do not reliably identify terminal extremal continuations. The paper frames ordering as terminal path-value estimation and proposes P-DC-SCTR, an MLP on 16 state, transition and remaining-task features. Trained on one tree, it predicts path values on four held-out seed trees at ρ = 0.68–0.78, but greedy rollouts do not beat replaying the development tree's best order.

Critical strengths
  1. Clean controlled design: task set, data, budget and anchor are fixed, only the order varies, and all 24 orders are enumerated.
  2. The oracle experiment, which uses realized one-step ASR after full training, cleanly separates estimation error from target mismatch.
  3. The paper is well-written and the presentation is easy understanding.
Critical weaknesses
  1. External validity. The "aligned" model is a 0.6B model with a rank-8 LoRA anchor trained on 1,800 examples. That is far shallower than production alignment, and order effects could be smaller or different for large-scale-aligned instruct models. Could you please try larger model?
  2. In seed42, the lowest-ASR order (qa → code → math → summarization) is the exact reverse of the highest (summarization → math → code → qa). The two oracle-greedy orders end in summarization (low ASR) and code (high ASR). This suggests the last tasks dominate terminal safety. In that case the gap is largely ordinary forgetting: early safety changes get overwritten by later tasks. I think it could be a case the recency dominate the the safety.
Questions for authors

as listed above

4 comments
AuthorsOpenReview ↗

Response to Weakness 1

We thank the reviewer for the valuable suggestion to examine a larger instruction-tuned model. We agree that model scale and safety initialization are important considerations when assessing the scope of task-order effects. To address this concern, we conduct a controlled evaluation of all 24 task orders on Qwen2.5-7B-Instruct using its publicly released checkpoint.

We start directly from the public Qwen/Qwen2.5-7B-Instruct checkpoint (approximately 7.6B parameters), without the additional 1,800-example safety anchor used in the main Qwen3-0.6B experiments. Within each tree, the same four benign tasks, training data, per-task budgets, and optimization configuration are fixed; only task order varies. The results are shown below.

Training seedMin. terminal ASR (%)Max. terminal ASR (%)Max. − min. (pp)Prompt-paired bootstrap 95% interval for range (pp)
4228.560.532.0[25.0, 40.5]
4734.059.525.5[20.5, 34.5]

Both seeds share the same starting checkpoint, with initial HarmBench ASR of 6.5% and XSTest over-refusal of 3.6%. Prompt-paired permutation reference tests also support terminal-ASR differences across orders(10,000 permutations per tree; raw p≈0.0001p \approx 0.0001 in each seed).These results provide further empirical support for substantial task-order effects on terminal safety in a larger, publicly released instruction-tuned model with a different safety initialization.

The main experiment exhaustively evaluates task orders using a shared safety-aligned anchor, giving every order the same measured starting point (2.5% HarmBench ASR; 9.6% XSTest over-refusal). With task data and budgets fixed, this supports a controlled study of safety during benign sequential adaptation. The public 7B instruct checkpoint extends the comparison to a different model scale and safety initialization.

The original paper additionally evaluates all 24 orders on Qwen3-1.7B and SmolLM2-1.7B without the custom anchor, with ranges of 26.5 and 20.5 pp, respectively (Figure 3; Appendix C.1). Qwen3-1.7B provides additional supporting evidence; the SmolLM2 result has weaker statistical evidence after multiplicity correction.

AuthorsOpenReview ↗

Response to Weakness 2 (Part I): Last-task effects

We appreciate the reviewer’s insightful observation about last-task effects. We agree that the final task can strongly influence terminal safety and that recency-related forgetting is a plausible contributor. To examine this explanation, we compare orders with the same final task or final two tasks, and evaluate whether realized one-step safety changes reliably guide complete-order selection.

First, we group all 24 orders in each tree in the main experiment by the final task, with six permutations of the preceding three tasks per group. The results are shown below; each cell reports mean terminal ASR (%) / within-group range (pp).

TreeLast = codeLast = mathLast = qaLast = summarizationBetween-group / total sum of squares
seed4230.4 / 9.019.4 / 1.526.2 / 22.016.2 / 8.00.63
seed4329.5 / 4.020.9 / 10.522.8 / 22.015.6 / 11.00.55
seed4429.7 / 10.021.3 / 4.525.7 / 24.514.9 / 10.00.52
seed4530.3 / 4.021.6 / 5.523.9 / 18.016.7 / 8.50.65
seed4629.0 / 8.019.7 / 7.024.1 / 20.513.8 / 6.50.60

Code-ending groups have the highest mean ASR and summarization-ending groups the lowest in all five trees. Final-task grouping accounts for 52%–65% of the total sum of squares, indicating a strong association with terminal ASR. Within-group variation also remains: qa-ending orders span 18.0–24.5 pp. The two extreme orders noted by the reviewer differ in both the final task and the preceding order. Comparing orders with the same final task lets us examine the contribution of their preceding task orders.

We assess residual variation by fixing either the final task or the final two tasks. The latter includes all 12 matched pairs with reversed first two tasks. Prompt-paired permutation reference tests are jointly Holm-adjusted across the ten main-line tests. The results are shown below.

TreeFinal task fixed: within-group variation, adjusted ppFinal two tasks fixed: mean absolute difference (pp)Final two tasks fixed: adjusted pp
seed420.00102.710.228
seed430.00103.960.0035
seed440.00102.580.270
seed450.00103.080.151
seed460.00102.500.270

The tests support variation among orders with the same final task in all five trees. With the final two tasks fixed, mean absolute differences are 2.50–3.96 pp, with statistical support after correction only in seed43. Together, these results show a strong last-task association and variation related to earlier ordering.

In both 7B seeds, the ranking of the final-task groups differs from that in the main experiment: summarization-ending orders have the highest mean ASR and code-ending orders the lowest, indicating setting-dependent final-task risk rankings. Between-group / total sum of squares is 0.533 / 0.557 for seeds42 / 47, and within-final-task variation is supported in both (raw p≈0.0001p \approx 0.0001 each). With the final two tasks fixed, mean absolute differences are 4.50 / 3.58 pp (raw p=0.0007p = 0.0007 / 0.08130.0813), resolved in seed42 but not seed47. These supplementary tests are reported separately from the main-line correction family.

AuthorsOpenReview ↗

Response to Weakness 2 (Part II): Oracle analysis and the local-to-global gap

Later updates may change the safety effects measured after earlier tasks, which helps explain why immediate and terminal outcomes can differ. To assess the implications for ordering decisions, Appendix D.2 applies greedy selection using realized one-step ASR changes after fully training each candidate task. This removes local proxy-prediction error from the comparison and directly tests whether immediate safety changes reliably guide terminal-order selection.

Across seed42–46, minimizing realized one-step ΔASR\Delta\mathrm{ASR} selects qa → summarization → math → code, yielding 24.5%–30.0% terminal ASR, 14–17.5 pp above the observed minima. Maximizing it selects code → qa → math → summarization, yielding 12.0%–15.0%, 18–28 pp below the maxima. In all five trees, the rule favoring immediate safety therefore yields higher terminal ASR than the rule favoring the opposite direction. For their respective objectives, both rules also have less favorable terminal-ASR point estimates than the mean over uniformly random complete orders.

We additionally evaluate the same oracle rules in the two 7B seeds. Low-risk gap means greedy ASR minus the observed minimum; high-risk gap means the observed maximum minus greedy ASR. The results are shown below.

Training seedLow-risk greedy ASR (%)Low-risk gap (pp)High-risk greedy ASR (%)High-risk gap (pp)
4238.09.555.55.0
4742.58.546.513.0

The low-risk gaps repeat across seeds (9.5 and 8.5 pp). For seed42, joint prompt resampling that reselects both the path and terminal extrema gives a median gap of 9.5 pp and a 95% resampling interval of [3.5, 18.0] pp, assessing sensitivity to evaluation prompts. For the same seed, the high-risk gap is not statistically resolved by the fixed-path paired interval (greedy minus maximum: −5.0 pp, 95% interval [−13.0, 3.0]). Both rules yield more favorable terminal-ASR point estimates than the random-order mean for their respective objectives in both 7B seeds. Thus, local guidance can improve on random ordering while still leaving a gap to the safest observed complete order.

These results distinguish accurately measured immediate safety changes from reliable terminal-order decisions. This motivates Q+(s,a)Q^{+}(s,a) and Q−(s,a)Q^{-}(s,a) in Section 3.3, which evaluate candidates through the highest and lowest terminal ASRs reachable over all orders of the remaining tasks, incorporating subsequent adaptation and forgetting.

Dear Reviewer,

We hope our response has addressed your questions. As the discussion phase is coming to a close, we are looking forward to your feedback and would like to know if you have any remaining concerns we can address. We are grateful if you find our revisions satisfactory and consider raising your score for our paper.

Thank you once again for the time and effort you have dedicated to reviewing our paper.

Best regards!

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.