acceptodds
Under review as a conference paper at ICLR 2027

SkillAdam: Stable and Efficient Skill Evolution for Agents

Abstract

Agent skills provide a lightweight mechanism for equipping frozen language-model agents with domain knowledge and procedural guidance. However, building high-quality skills remains costly and difficult to scale. As expert-written skills require substantial human effort, recent skill self-evolution methods automate skill optimization using execution feedback. Yet these methods still rely on heuristic update strategies that often lead to unstable optimization and low iteration efficiency. To address these limitations, we identify two key principles for stable and efficient skill optimization. (1) Direction Stability requires effective corrections to accumulate across iterations rather than be overwritten by iteration-local feedback. (2) Update Adaptivity requires the scope of each revision to reflect the consistency of recent case-level improvements. We introduce SkillAdam, an Adam-inspired framework that applies these principles to the optimization of discrete and non-differentiable skill documents. Inspired by Adam's first moment, SkillAdam maintains an optimization memory that accumulates identified problems and prior solution attempts to stabilize the update direction. Inspired by Adam's second moment, it maintains a volatility-driven edit budget that tracks the history-weighted variation of recent case-level improvements to adaptively control the update magnitude. Extensive experiments on seven benchmarks spanning both short- and long-horizon tasks show that SkillAdam achieves state-of-the-art performance with substantially more stable optimization, while requiring substantially fewer optimization iterations and lower optimization cost than prior methods. Code is available in the supplementary materials.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Agent Skills, Self-Evolution, Textual Optimization

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_1BFVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SkillAdam optimizes natural-language skills for frozen LLM agents. It maintains an Evolving Issue Tracker (EIT) to record problems and previous repair attempts, and adjusts the edit budget using the variance of case-level score improvements. Candidate skills are evaluated before being accepted. The paper reports results on seven benchmarks, with larger gains on long-horizon planning, lower optimization costs, and improved cross-model transfer compared with SkillOpt.

Critical strengths

S1. The reduction in optimization cost is substantial. On DeepPlanning, token usage drops from 226.6M to 74.0M while DP-Avg improves from 21.7% to 28.3%.

S2. The evaluation covers seven benchmarks and includes cross-model transfer. On GPT-5.4-mini, the transferred skills outperform SkillOpt on five of six benchmarks, with the average score increasing from 63.1% to 67.8%.

Critical weaknesses

W1. I find EIT an incremental change to the memory mechanisms already used by SkillOpt. Tracking issues and repair histories is a reasonable design, but I do not see clear evidence that this organization preserves useful corrections better than existing optimizer memory. The ablation removes EIT and conversation history together, so it does not establish EIT's distinct contribution. I also find the Adam framing overstated: the connection is mainly conceptual and makes the method appear more novel than the demonstrated advance warrants.

W2. The variance-based edit rule is not sufficiently justified. The comparison with a fixed budget shows that the proposed scheme can help, but does not explain why variance is the right signal. No comparison with other improvement signals is reported. It is also unclear whether the number of requested diff hunks reflects the scale of changes in the skill's strategy or behavior. Without establishing this connection, I am not convinced that the budget provides the claimed control over update magnitude, particularly when several parts of a skill need to change together.

Questions for authors

Q1. What evidence separates EIT's contribution from that of conversation history and shows that issue-level tracking helps preserve successful repairs across revisions? Does it improve subsequent editing decisions compared with SkillOpt's optimizer memory?

Q2. Using improvement variance to control the edit budget seems heuristic. Why was this signal chosen, and how does it compare with other improvement signals?

Q3. Does changing the hunk budget reliably change the extent of revisions to the skill's strategy or behavior? How does the method handle improvements that require coordinated changes across the skill, including substantial restructuring?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 1BFV (1/2)

Thanks for your careful and constructive comments, and for recognizing the substantial cost reduction and the cross-model transfer of SkillAdam. We have revised the paper following your suggestions, with all changes marked in blue.

 

Q1: Distinct contribution of the EIT beyond the conversation history, and whether issue-level tracking preserves repairs and improves editing decisions compared with SkillOpt's optimizer memory.

A1: Thank you for this critical question. We agree that the original ablation removes the EIT only together with the fixed budget. The conversation history is a global component kept in all variants, so we ran an ablation on DeepPlanning that removes only the EIT while keeping the conversation history, the adaptive budget, and all other settings unchanged:

ConfigurationShopping Level 1Shopping Level 2Shopping Level 3Shopping (all levels)TravelAverage
SkillAdam (EIT + optimizer history + adaptive budget)52.036.050.045.011.728.3
Remove the EIT only, conversation history kept, −M-\mathsf{M} (new)44.024.040.035.00.017.5
Remove the adaptive budget only, −B-\mathsf{B} (Table 3)48.036.040.041.71.721.7
Remove the EIT and adaptive budget, −B,−M-\mathsf{B},-\mathsf{M} (Table 3)56.020.030.036.71.719.2

Removing only the EIT leads to the largest drop among all variants, even though the conversation history is still available, so the history cannot substitute for the EIT. This variant even falls below the one that removes both the EIT and the adaptive budget. Without the EIT, the patch generator still sees earlier attempts in the history but has no record of which issues are already resolved, so the adaptive budget narrows each revision without knowing which corrections must be preserved. Issue-level tracking is therefore what allows successive revisions to build on earlier repairs.

The EIT also differs from SkillOpt's optimizer memory, which records the outcomes of rejected edits. The EIT is organized by issue, tracking each problem's status and all attempts, so the optimizer knows which problems are solved and avoids overwriting them. The comparison with SkillOpt is consistent with this difference: SkillAdam reaches higher accuracy at a lower optimization cost (Tables 2 and 5), its accepted skills stay above the initial accuracy while SkillOpt's fluctuate (Figure 4), and its skills are rated higher in the blind skill-quality evaluation (Figure 3). We have added this ablation to Table 3 and Section 4.4 of the revised paper (marked in blue).

 

Q2: Rationale for using improvement variance as the budget signal, and comparison with other signals.

A2: Thank you for this question. The choice of variance is inspired by Adam, which scales its step size by the variability of recent gradients. In skill revision, we measure this variability by the variance of per-case improvements. A high variance means a revision helps some cases but hurts others, so the next edit should be narrower; a low variance supports a broader edit. This complements the acceptance gate, which uses the mean improvement.

Table 3 shows that this signal is effective: replacing it with a fixed budget clearly lowers accuracy, most of all on DeepPlanning Travel. We have not compared other signals, since our goal is a principled and effective design rather than a search over signals.

AuthorsOpenReview ↗

Response to Reviewer 1BFV (2/2)

Q3: Whether the hunk budget controls the extent of strategy and behavior changes, and how coordinated revisions are handled.

A3: Thank you for this insightful question. Each hunk is intended to be one logical change, so the hunk count reflects how many parts of the strategy a revision modifies, and the extent of revisions follows the budget:

Requested σt\sigma_t#PatchesAvg. realized hunksMax realized hunksWithin budget
1331.33372.7%
2221.64395.5%
3211.673100%
4122.083100%

Coordinated changes are not blocked. Related edits within a section fit into one hunk, the budget returns to bbaseb_{base} once improvements become consistent, and since the budget is soft, a necessary change is never truncated. A large restructuring may take several iterations under inconsistent evidence, which we now note as a limitation in Section 6 of the revised paper (marked in blue).

 

Q4: Clarifying the connection between SkillAdam and Adam.

A4: Thank you for this comment. We agree that the connection to Adam is a conceptual analogy, and Adam serves as the inspiration for our design.

Skill self-evolution faces two problems similar to those in stochastic optimization. Successive revisions often overwrite earlier corrections, causing a seesaw effect, and revisions without guidance on their scope keep adding rules that overfit a few cases. Following Adam's principles of aggregating history and scaling the step by signal variability, we designed the EIT and the volatility-driven edit budget to address these two problems.

Our experiments confirm that the design is effective: SkillAdam achieves higher accuracy at lower cost (Tables 2 and 5), keeps its accepted skills stable (Figure 4), relies on the adaptive budget (Table 3), and produces higher-quality skills that transfer better (Figure 3 and Table 4).

 

Thanks again for your careful and valuable comments, hope our response addresses your concerns. If our responses answer your questions well and reassure your concerns, we would appreciate if you could reassess our work and increase the rating.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer 1BFV,

Thank you again for your detailed and constructive review. As the discussion period is coming to an end, we would like to kindly check whether our response has addressed your concerns. In particular, we added an ablation that removes only the EIT while keeping the conversation history, which shows that the history cannot substitute for the EIT, explained the rationale for the variance-based budget signal, showed that the extent of revisions follows the hunk budget, and clarified that the connection to Adam is a design inspiration rather than a mathematical correspondence. All revisions are marked in blue in the updated paper.

If any question remains, we would be glad to clarify it before the discussion closes. If our response has resolved your concerns, we would greatly appreciate it if you could consider updating your rating.

Best regards, The Authors

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

This paper proposes SkillAdam, an Adam-inspired framework for stable and efficient agent skill evolution. It introduces an Evolving Issue Tracker to preserve optimization history and a volatility-driven edit budget to adaptively control skill revisions. Experiments across seven benchmarks demonstrate improved performance, optimization stability, and efficiency compared to existing methods.

Critical strengths
  • The paper identifies important limitations of existing skill evolution methods, particularly unstable updates and inefficient optimization.
  • The two proposed mechanisms are well-motivated, and the connection to Adam provides a clear conceptual framework.
  • Experiments cover seven benchmarks involving both short- and long-horizon tasks, with comparisons against several competitive baselines.
  • Clear presentation. The paper is well organized and makes it easy for the reader to quickly grasp the main idea and method.
Critical weaknesses
  • Several closely related studies have explored similar ideas, including SkillProx [1], which uses feedback-aware, validation-gated skill refinement; SkillAdaptor [2], which performs targeted skill updates based on trajectory-level failure analysis; and SkillRise [3], which studies efficient cross-task skill evolution through reinforcement learning.

  • The cumulative ablations do not independently isolate the contributions of the Evolving Issue Tracker, optimizer conversation history, and acceptance mechanism.

  • SkillAdam and SkillOpt use different data partitioning and optimization protocols. It need to be explained further.

  • The edit budget is only a soft constraint based on the number of diff hunks, which may not accurately reflect the magnitude of semantic changes. Sensitivity analysis of its hyperparameters is also missing.

[1] Zheng, Mingxuan, et al. "Skillprox: Self-evolving agent skills via proximal textual gradient descent." arXiv preprint arXiv:2608.07449 (2026).

[2] Yu, Zhuoyun, et al. "Skilladaptor: Self-adapting skills for llm agents from trajectories." arXiv preprint arXiv:2606.01311 (2026).

[3] Yao, Zhiyuan, et al. "Skillrise: Agentic reinforcement learning for cross-task skill evolution." arXiv preprint arXiv:2607.26784 (2026).

Questions for authors

See weaknesses above

7 comments
AuthorsOpenReview ↗

Response to Reviewer ukCy (1/3)

Thanks for recognizing the importance of the problem, the motivation of our two mechanisms, the coverage of our experiments, and the clarity of our presentation. We have revised the paper following your suggestions, with all changes marked in blue.

 

Q1: Differences from recent related work (SkillProx, SkillAdaptor, and SkillRise).

A1: Thank you for pointing out these works. SkillRise and SkillAdaptor address different settings or problems, while SkillProx shares part of our motivation but takes a different approach.

SkillRise trains the policy weights with reinforcement learning, whereas SkillAdam keeps the agent frozen and requires no training. SkillAdaptor decides where to edit within a single failed trajectory, but does not track issues across iterations or control how much each revision changes, which is exactly what SkillAdam addresses.

SkillProx, a concurrent work, also accepts candidates on the same batch and also observes that iterative patching accumulates redundant content. However, it lets the skill grow and prunes it afterwards using a separate validation split. SkillAdam instead controls growth during revision: the edit budget bounds each revision, and the EIT refines guidance for known issues instead of appending new rules. This also protects earlier corrections from being overwritten, which pruning cannot recover, so the two approaches are complementary. As SkillProx is the closest to our setting, we have added it to the discussion of skill self-evolution methods in Section 2 of the revised paper (marked in blue).

 

Q2: Individual contributions of the EIT, the optimizer conversation history, and the acceptance mechanism.

A2: Thank you for this suggestion, which we fully agree with. We ran an additional ablation on DeepPlanning that removes only the EIT while keeping the optimizer conversation history and the adaptive budget, so that the EIT and the budget can now each be removed independently:

ConfigurationShopping Level 1Shopping Level 2Shopping Level 3Shopping (all levels)TravelAverage
SkillAdam (EIT + optimizer history + adaptive budget)52.036.050.045.011.728.3
Remove the EIT only, −M-\mathsf{M} (new)44.024.040.035.00.017.5
Remove the adaptive budget only, −B-\mathsf{B} (Table 3)48.036.040.041.71.721.7
Remove the EIT and adaptive budget, −B,−M-\mathsf{B},-\mathsf{M} (Table 3)56.020.030.036.71.719.2

Removing the EIT alone causes the largest drop, larger than removing the budget, and the adaptive budget brings no gain without the EIT, indicating that the two mechanisms are complementary. Since the conversation history is kept in this and all other variants, the result isolates the contribution of the EIT beyond the history. The history is a global component of the optimizer that serves as lightweight context for the patch generator, so we did not remove it, and its own contribution is not isolated.

We do not ablate the acceptance mechanism separately, because both other mechanisms depend on it. The paired evaluation of each candidate on the batch produces the outcome that the EIT records for every attempt, as well as the per-case improvements from which the volatility and the edit budget are computed. Removing the gate would therefore also remove the inputs of the EIT and the budget, rather than isolating the gate itself. We have added this variant to Table 3, together with this explanation, in Section 4.4 and Appendix B.3 of the revised paper (marked in blue).

AuthorsOpenReview ↗

Response to Reviewer ukCy (2/3)

Q3: Further explanation of the data partitioning and optimization protocol relative to SkillOpt.

A3: Thank you for this suggestion. The data partitions of the two methods are in fact consistent, and there is no data leakage. On the six benchmarks outside DeepPlanning, we directly use the splits released in SkillOpt's official repository. SkillOpt divides the non-test examples into train and selection, while SkillAdam uses their union as one optimization pool, so both methods see exactly the same non-test examples and are evaluated on the same test cases, which are never used for optimization or skill selection. On DeepPlanning, which SkillOpt does not cover, both methods use the odd-numbered cases for optimization and the even-numbered cases for testing, and SkillOpt further divides its optimization cases into train and selection. We have added this configuration to Appendix B.2 of the revised paper (marked in blue).

SliceSkillOpt TrainSkillOpt SelectionSkillAdam PoolTest (shared)
Shopping Level 11692525
Shopping Level 21692525
Shopping Level 3641010
Travel (English)40206060

The difference in optimization protocol reflects the difference between the two methods rather than a difference in data. SkillOpt decides acceptance on its held-out selection split, whereas SkillAdam compares each candidate with the current skill on the same batch, because the EIT records the outcome of every attempt and the edit budget is computed from the per-case improvements of this paired comparison. This protocol gives SkillAdam no additional data or optimization signal, since outside DeepPlanning it traverses the pool once while SkillOpt runs four epochs. We have stated this explicitly in Section 4.2 of the revised paper (marked in blue).

 

Q4: Whether the hunk-based budget reflects the magnitude of semantic changes.

A4: Thank you for this insightful comment. We agree that a hunk count is not a direct measure of semantic change. It is designed to control the scope of a revision, namely how many separate parts of the skill are touched, since each hunk is intended to be one logical change. The realized patches confirm this control:

Requested σt\sigma_t#PatchesAvg. realized hunksMax realized hunksWithin budget
1331.33372.7%
2221.64395.5%
3211.673100%
4122.083100%

The number of modified locations grows with the budget, and Table 3 shows that replacing the adaptive budget with a fixed one clearly lowers accuracy. We have added this analysis to Appendix B.3 of the revised paper (marked in blue).

AuthorsOpenReview ↗

Response to Reviewer ukCy (3/3)

Q5: Sensitivity to the hyperparameters of the edit budget.

A5: Thank you for pointing this out. The edit budget has four hyperparameters, listed in Table 7 of the paper:

Benchmarkβ2\beta_2bbaseb_{base}bminb_{min}VmaxV_{max}Volatility metric
SearchQA0.9410.05Exact Match
SpreadsheetBench0.9420.10Hard task success
OfficeQA0.9410.05Exact Match
DocVQA0.9410.04ANLS-hard
LiveMathematicianBench0.9410.05Exact Match
ALFWorld0.9410.10Episode success
DeepPlanning Shopping Levels 1–30.9410.05Score
DeepPlanning Travel (English)0.9410.10Composite score

Only VmaxV_{max} varies across benchmarks, and within a narrow range, while the other values are shared; all were set heuristically without a search. Our logs show that these values keep the budget adaptive:

Benchmark#PatchesAvg. requested σt\sigma_tAvg. realized hunksWithin budget
SearchQA351.291.3477.1%
SpreadsheetBench82.382.12100%
OfficeQA82.382.3875.0%
DocVQA102.201.30100%
LiveMathematicianBench52.601.40100%
ALFWorld63.332.17100%
DeepPlanning Shopping Level 132.671.67100%
DeepPlanning Shopping Level 253.001.20100%
DeepPlanning Shopping Level 343.251.75100%
DeepPlanning Travel (English)43.501.50100%
All882.141.5988.6%

On every benchmark, the average requested budget lies between bminb_{min} and bbaseb_{base}, so the budget follows the observed volatility rather than staying constant. We have added this discussion to Appendix B.3 of the revised paper (marked in blue).

 

Thanks again for your careful and valuable comments, hope our response addresses your concerns. If our responses answer your questions well and reassure your concerns, we would appreciate if you could reassess our work and increase the rating.

Reviewer_yh46OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SKILLADAM, an Adam-inspired framework for optimizing discrete natural-language agent skill documents. Functioning as analogues to first and second moments, an Evolving Issue Tracker accumulates identified failure patterns and prior solution outcomes to stabilize update directions, while a volatility-driven edit budget tracks case-level improvement variance to adaptively constrain the number of unified-diff hunks per revision. Evaluated across seven short-horizon and long-horizon benchmarks, SKILLADAM achieves state-of-the-art accuracy and stronger cross-model transfer to smaller models while reducing optimization token consumption by over 67 percent compared to SkillOpt on DeepPlanning.

Critical strengths
  • Coupling the Evolving Issue Tracker with a volatility-driven edit budget is consequential because it prevents conflicting revisions across iterations and improves long-horizon planning accuracy by 6.6 percentage points on DeepPlanning while cutting optimization token cost by two thirds.

  • The step-by-step optimization trajectory analysis and blind skill-quality evaluation are consequential because they demonstrate that SKILLADAM reaches peak accuracy at one fifth of the token cost while avoiding the severe performance regressions seen in baseline updates.

Critical weaknesses
  • Merging train and selection splits to evaluate and accept candidate patches on the exact same mini-batch used to generate them is consequential because it risks batch-level overfitting, alters the data-partition protocol relative to SkillOpt, and yields marginal gains of 0.0 to 1.1 points on five short-horizon benchmarks.

  • Relying on a soft prompt-level hunk request that is neither enforced nor truncated alongside heavily hand-tuned per-benchmark acceptance and stopping thresholds is consequential because it weakens the algorithmic guarantee of adaptive step sizing and requires domain-specific tuning.

Questions for authors
  • Since SKILLADAM merges train and selection splits and evaluates candidate patches on the same mini-batch BtB_t used to generate them, how does it perform under SkillOpt's exact held-out selection gating protocol, and what is the isolated ablation impact of removing only the Evolving Issue Tracker while keeping adaptive budgeting enabled in Table 3?

  • Because the edit budget σt\sigma_t is only a soft prompt instruction without truncation, how often does the patch generator exceed σt\sigma_t in practice, and how sensitive is SKILLADAM to replacing the heavily hand-tuned per-benchmark acceptance thresholds in Table 8 with a unified acceptance rule?

4 comments
AuthorsOpenReview ↗

Response to Reviewer yh46 (1/3)

Thanks for your careful reading and for recognizing the effectiveness and efficiency of SkillAdam, as well as our optimization trajectory analysis and skill quality evaluation. We have revised the paper following your suggestions, with all changes marked in blue.

 

Q1: Performance under SkillOpt's held-out selection protocol, and the risk of batch-level overfitting.

A1: Thank you for this important question. We did not rerun SkillAdam under SkillOpt's held-out gate, because the same-batch paired comparison is part of the method rather than an evaluation choice. The outcome of each candidate on its own batch is what the EIT records for every attempt, and the per-case improvements on that batch are what the volatility-driven budget is computed from, so a held-out gate would remove these inputs and change the method itself. The concurrent work SkillProx independently adopts the same kind of same-batch acceptance.

The comparison remains fair, and same-batch acceptance does not overfit the batch. SkillAdam only merges SkillOpt's train and selection splits into one pool, so both methods use the same non-test examples and test cases, and the test split is never used for optimization or selection. SkillAdam also uses less optimization, with a single pass over the pool outside DeepPlanning versus SkillOpt's four epochs. On DeepPlanning, which SkillOpt does not cover, we use the following split, which we have added to Appendix B.2 of the revised paper (marked in blue):

SliceSkillOpt TrainSkillOpt SelectionSkillAdam PoolTest (shared)
Shopping Level 11692525
Shopping Level 21692525
Shopping Level 3641010
Travel (English)40206060

As for overfitting, the gate accepts a candidate only if it solves at least one more case than the current skill on the batch (A4), and Figure 4(a) shows that all skills accepted by SkillAdam stay above the initial test accuracy, although test scores never enter the acceptance decision. Revisions accepted on the same batch therefore generalize to unseen cases. We have clarified this rationale in Section 4.2 of the revised paper (marked in blue).

 

Q2: Ablation that removes only the EIT while keeping the adaptive budget.

A2: Thank you for this suggestion. Following it, we ran the requested ablation on DeepPlanning, which removes only the EIT while keeping the optimizer history, the adaptive budget, and all other settings unchanged:

ConfigurationShopping Level 1Shopping Level 2Shopping Level 3Shopping (all levels)TravelAverage
SkillAdam (EIT + optimizer history + adaptive budget)52.036.050.045.011.728.3
Remove the EIT only, −M-\mathsf{M} (new)44.024.040.035.00.017.5
Remove the adaptive budget only, −B-\mathsf{B} (Table 3)48.036.040.041.71.721.7
Remove the EIT and adaptive budget, −B,−M-\mathsf{B},-\mathsf{M} (Table 3)56.020.030.036.71.719.2

Removing the EIT causes a clear drop even when the adaptive budget is enabled, and this drop is larger than that caused by removing the budget. Without the EIT, the adaptive budget does not help and performs even below the variant that removes both mechanisms, because a narrower revision scope only pays off when the revisions are directed by the tracked issues. The two mechanisms are therefore complementary, and the gains of SkillAdam come from their combination. We have added this variant to Table 3 and its discussion to Section 4.4 of the revised paper (marked in blue).

AuthorsOpenReview ↗

Response to Reviewer yh46 (2/3)

Q3: How often do the generated patches follow the hunk budget σt\sigma_t in practice?

A3: Thank you for this question. We counted the realized hunks of all 88 non-empty patches in our main runs and compared them with the budget requested in each iteration:

Benchmark#PatchesAvg. requested σt\sigma_tAvg. realized hunksWithin budget
SearchQA351.291.3477.1%
SpreadsheetBench82.382.12100%
OfficeQA82.382.3875.0%
DocVQA102.201.30100%
LiveMathematicianBench52.601.40100%
ALFWorld63.332.17100%
DeepPlanning Shopping Level 132.671.67100%
DeepPlanning Shopping Level 253.001.20100%
DeepPlanning Shopping Level 343.251.75100%
DeepPlanning Travel (English)43.501.50100%
All882.141.5988.6%

The generator follows the budget in the large majority of iterations, and nearly all violations occur when the budget is a single hunk. We keep the budget soft on purpose, because truncating a patch could break a coordinated change, and every patch must still pass the acceptance gate. We have added these statistics to Appendix B.3 of the revised paper (marked in blue).

AuthorsOpenReview ↗

Response to Reviewer yh46 (3/3)

Q4: How are the acceptance thresholds in Table 8 set, and how sensitive is SkillAdam to a unified acceptance rule?

A4: Thank you for raising this point. Most thresholds in Table 8 are the same rule expressed on different batch sizes. Except for DeepPlanning Travel, the primary metrics are per-case 0/1 scores, so each threshold falls between two adjacent values of k/∣B∣k/|B|:

BenchmarkBatchPrimary metricGain thresholdEquivalent gain ruleProtection thresholdEquivalent protection rule
SearchQA15Exact match≥ 0.04≥ 1 more case solved> −0.04no net loss
SpreadsheetBench15Hard task success≥ 0.06≥ 1 more case solved> −0.08at most 1 case lost
OfficeQA10Exact match≥ 0.08≥ 1 more case solved> −0.08no net loss
DocVQA15Hard ANLS≥ 0.01≥ 1 more case solved> −0.05no net loss
LiveMathematicianBench10Exact match≥ 0.08≥ 1 more case solved> −0.08no net loss
DeepPlanning Shopping5Case accuracy> 0.15≥ 1 more case solved> −0.15no net loss
ALFWorld10Episode success≥ 0no fewer cases solved—no net loss

The primary rule is thus shared: solve at least one more case than the current skill on the same batch. The remaining differences only concern ties. A continuous auxiliary metric breaks ties when the 0/1 primary metric carries no information, and it rarely changes a decision:

Role of the auxiliary metricCount
Accepted revisions relying on an auxiliary metric because the primary metric was tied4 of 57
Decisions vetoed by an auxiliary protection condition although the primary gain held2 of 88

ALFWorld has no auxiliary metric, so it accepts ties to avoid rejecting nearly every revision on its small batches, and DeepPlanning Travel uses a continuous primary metric. A unified primary rule would thus change only these few tie-related decisions. We have added this unified form to Appendix B.3 of the revised paper (marked in blue).

 

Q5: Performance gains on the short-horizon benchmarks.

A5: Thank you for this observation. We agree that the accuracy gains on short-horizon benchmarks are modest. SkillAdam targets preserving corrections across iterations, which matters most in long-horizon tasks, where its advantage is clear (Table 2). Short-horizon tasks involve fewer such conflicts, and SkillOpt already performs strongly on them.

Even where accuracy is close, SkillAdam is much cheaper to optimize. On DeepPlanning it uses about one third of SkillOpt's tokens (Table 5). For the other benchmarks, SkillOpt's reported token counts include test evaluations and cannot be aligned, so we compare the rollouts required by each protocol:

BenchmarkSkillOpt rolloutsSkillAdam rolloutsRatio
SearchQA10,5201,1259.4×
SpreadsheetBench9203602.6×
OfficeQA6081663.7×
DocVQA1,3963204.4×
LiveMathematicianBench4041033.9×
ALFWorld4201143.7×

SkillOpt requires several times more rollouts on every benchmark. As both methods run the same target model on the same tasks with skills of comparable length, a single rollout costs a similar number of tokens, so the gap in rollouts translates into a comparable gap in optimization cost. SkillAdam's skills are also rated higher in quality (Figure 3) and transfer better (Table 4). We have clarified this in Section 4.3 and added the rollout comparison to Section 5.2 of the revised paper (marked in blue).

 

Thanks again for your careful and valuable comments, hope our response addresses your concerns. If our responses answer your questions well and reassure your concerns, we would appreciate if you could reassess our work and increase the rating.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer yh46,

Thank you again for your careful review and positive assessment of SkillAdam. As the discussion period is coming to an end, we would like to kindly check whether our response has addressed your concerns. In particular, we clarified that SkillAdam uses exactly the same data as SkillOpt and explained why same-batch acceptance does not overfit, added the requested ablation that removes only the EIT while keeping the adaptive budget, reported how often patches follow the edit budget, and showed that most thresholds in Table 8 reduce to one shared rule. All revisions are marked in blue in the updated paper.

If any question remains, we would be glad to clarify it before the discussion closes. If our response has resolved your concerns, we would greatly appreciate it if you could consider updating your rating.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.