acceptodds
Under review as a conference paper at ICLR 2027

SE-GoS: Self-Evolving Graph-of-Skills for Skill Library at Scale

Abstract

LLM agents use large libraries of reusable skills. At thousands of skill entries, retrieval becomes the bottleneck. Graph-of-Skills (GoS) retrieves dependency-aware bundles from a typed skill graph, and SkillDAG shows that such a graph can accumulate execution-backed structure online. Neither asks whether execution traces can be distilled into a better retrieval graph that generalizes to unseen tasks. We present Self-Evolving Graph-of-Skills (SE-GoS), which treats the retrieval graph as an index rather than a learned representation: the graph is maintained from execution traces while the retrieval pipeline, the skill library, and the model stay fixed. SE-GoS applies three updates: (1) topology, which induces relations from execution evidence and retracts an avoid edge only after repeated successful co-use; (2) edge-weight, which softly attenuates unsupported semantic edges and reinforces incoming edges to used skills; and (3) node-description, which updates retrieval-facing descriptions stored on graph nodes ranked too low. On SkillsBench, one evolution round lifts average reward from 52.4% to 59.4%, above full-library loading, vector retrieval, static GoS, and SkillDAG, and this ordering repeats on all three backbones. Retrieval over the evolved graph spends about two-thirds of the input tokens that loading the full library costs. Repeating the round does not help. The same graph improves a held-out split it never saw from 52.9% to 58.3%, so what it accumulates transfers rather than memorizes traces. Skill graphs can therefore be improved from execution experience without model training, retrieval-algorithm changes, skill-content modifications, or a model judging which skills are related.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, skill retrieval, self-evolution, Graph-of-Skills

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_S7DTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SE-GoS uses execution traces to update skill-graph relations, weights, and descriptions while keeping models and skills fixed. Benchmark evaluations, ablations, and held-out tests support improved retrieval utility.

Critical strengths

Encoding execution experience into inspectable and shareable retrieval state is practically useful. Keeping models and skill implementations fixed also helps isolate the effects of graph maintenance.

Held-out evaluation and complete component ablations provide encouraging evidence.

Critical weaknesses

The paper is motivated by missing prerequisite skills, but no dependency edges are generated in the reported trace analysis. The added navigation links mainly capture workflow. The results shows that execution feedback can improve retrieval, but the method’s ability to recover actual prerequisite relationships is untested.

It is unclear how a newly added skill would become connected to existing skills without execution history. The experiments keep the skill library fixed. Semantic initialization provides a starting point, but this scenario is not evaluated. In the experimental result, reward drops to 54.0% in the third round, raising concerns about the stability of repeated updates.

Questions for authors

How does ranking by skill usage or reward without graph diffusion perform using the same traces?

How would a newly added skill enter the existing graph without execution history, and how much usage evidence is needed to establish useful workflow connections?

1 comment
AuthorsOpenReview ↗

Response to Reviewer S7DT: Graph-Free Ranking and New-Skill Integration

Thank you for asking how graph-free ranking compares and how new skills enter the graph. Below, we compare reward-based ranking with SE-GoS and explain skill insertion and the execution evidence supporting workflow edges.

1. How does ranking by skill usage or reward without graph diffusion perform using the same traces?

With the same traces, SE-GoS exceeds graph-free reward ranking by 6.1–9.8% across the three SkillsBench backbones and by 9.3% on held-out tasks. Our MemRL* baseline adapts episodic-memory retrieval to rank skills by semantic relevance and reward utility, keeping source skills and model weights fixed. Tables 2 and 3(a) report:

SkillsBench protocolBackboneMemRL* R (%)SE-GoS R (%)SE-GoS − MemRL* (%)
Full-benchmarkdeepseek-v4-flash-073149.659.4+9.8
Full-benchmarkminimax-m2.720.928.5+7.6
Full-benchmarkgpt-5.2-codex32.038.1+6.1
Held-outdeepseek-v4-flash-073149.058.3+9.3

For this control, we considered four related approaches reviewed in Section 2: MemRL, SELM, TrackRL, and SkillBrew. Among them, we found public implementations for MemRL and SkillBrew. MemRL directly matches the requested graph-free ranking control with fixed skills and model weights, while SkillBrew curates the library. Appendix B.2 details the baseline choice and adaptation.

2. How would a newly added skill enter the existing graph without execution history, and how much usage evidence is needed to establish useful workflow connections?

a. How would a newly added skill enter the existing graph without execution history

A new skill can enter retrieval without execution history. Appendix C.2 specifies three steps:

  1. Add its node, retaining the existing graph and learned state.

  2. Connect it to relevant skills using deterministic text similarity.

  3. Include it in lexical seeding, making it retrievable immediately, even before it has graph neighbors.

Initialization needs no embedding model or LLM calls. Subsequent traces support the three updates, with only description rewriting calling the backbone.

b. how much usage evidence is needed to establish useful workflow connections?

One qualifying observation creates a workflow edge. Under Section 3.3's rule, a query seed A connects to a distinct skill B used on a task with reward ≥0.9. Further observations strengthen the edge. Table 5 reports query observations, which count qualifying retrieval queries and determine induced workflow weights, and distinct tasks, which count how many tasks supplied that evidence. Two qualifying queries from one task count as two observations but one supporting task. The table therefore shows both observation frequency and task coverage:

Evolution setCounting unitSupported onceTwiceThree timesTotal workflow edges
Full87, first roundQuery observations206211228
Full87, first roundDistinct tasks207201228
Train50 for held-outQuery observations111160127
Train50 for held-outDistinct tasks111160127

Table 4 shows that topology-only updating improves reward from 50.8% to 54.1%, even though most induced workflow edges have only one supporting observation.

Reviewer_BhSiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses skill retrieval at library scale (~1,000 skills): full-library loading does not scale and vector retrieval misses functional prerequisites (the "prerequisite gap"). Prior Graph-of-Skills (GoS) retrieves dependency-aware bundles via a typed graph + reverse-aware Personalized PageRank, but its graph is static. SE-GoS treats the retrieval graph as mutable, offline-maintained harness state — retrieval pipeline, skill library, and backbone fixed — updated once from execution traces via three training-free updates: (1) topology (induce workflow edges from successful co-use, dependency edges from ordered use + I/O-schema overlap, avoid edges from failed co-use), (2) edge-weight (soft-prune never-used surfaced skills; reinforce incoming edges of used skills), (3) node-description (ProTeGi-style rewrite of retrieval-facing text for poorly ranked used skills). Evaluation on SkillsBench (87 tasks, 1,000-skill library) and ALFWorld (140 episodes) with three backbones vs Vanilla/Vector/GoS/SkillDAG reports 52.4%→59.4% at ~68% of full-loading tokens, a held-out study (52.9%→58.3%), a 4-round evolution curve (plateau then reversal at round 3), and a full 2³ ablation. Rewards are unit-test-based, not LLM-judged.

Critical strengths
  1. A well-motivated problem with a clean framing and honest empirical protocol. "Retrieval graph as harness state" is a useful conceptual contribution; rewards come from unit tests/binary success rather than LLM judges; and the held-out transfer study (+5.4 over GoS, +4.1 over SkillDAG, tokens unchanged) directly addresses the obvious memorization criticism — this is the load-bearing evidence and it is genuinely encouraging.
  2. Exceptional transparency. A byte-level benchmark-fidelity audit, a complete hyperparameter table, a full 2³ factorial ablation, an honestly reported negative result (round-3 reversal 59.8→54.0), and the candid admission that the dependency channel never fires (Appendix B.3) — this is more rigorous protocol documentation than most agent-systems papers.
Critical weaknesses
  1. The headline numbers are in-domain with no statistical support. The main SkillsBench cells evolve on the same 87 tasks they are re-measured on (a "deployment-utility convention" inherited from SkillDAG), with 2 attempts/task, point estimates only, and significance explicitly deferred. The margins over the strongest baseline (SkillDAG) are +1.2/+1.3 on two of three backbones — plausibly within agentic-evaluation noise. Worse, for two of three backbone blocks, three of five baseline rows are quoted historical numbers from the GoS paper while SE-GoS/SkillDAG are fresh runs, stacking API drift on a possible reward-metric mismatch (the authors' fractional CTRF reward vs SkillsBench's upstream binary all-tests-pass reward for "source-reported" entries).
  2. The motivating mechanism never activates, and key design choices are untested. Dependency induction — the relation meant to close the "prerequisite gap" that motivates the paper — produces zero edges (max schema overlap 0.500 < ζ = 0.6), so all structural gains come from workflow co-occurrence and reweighting. The avoid channel (a single failed co-use creates an edge that shrinks bundles without backfill, θ_avoid = 1) is never ablated separately from workflow induction; the edge-reinforcement rule is admittedly non-causal (reinforcing all incoming edges to a used skill regardless of source), and the proposed safeguard against it (Eq. 9) is derived but never evaluated — while the round-3 reversal is direct evidence the raw update is not a stable fixed point. The headline +7.0 also mixes a 1.6-point substrate downgrade (semantic-only cold start 50.8 vs typed GoS 52.4) with the evolution gain, and evolution initialized from the original typed graph is never tested.
Questions for authors
  1. Can you report per-task paired bootstrap CIs (or a Wilcoxon signed-rank test) for SE-GoS vs SkillDAG on both the full-benchmark and held-out protocols?
  2. Which LLM performs the node-update rewriting (never identified in App. A.4)? If it belongs to an evaluated backbone family, how does that square with the "LLM-prior-free" claim in the abstract and Table 1?
  3. Do the quoted GoS/Vanilla/Vector numbers in the minimax and gpt-5.2-codex blocks use the same fractional CTRF reward as your fresh runs? And where do SE-GoS's ALFWorld evolution traces come from — the same 140 valid-seen episodes it is scored on, while SkillDAG trains on 420 train episodes?
  4. (Suggestions) A graph-free per-skill success-rate reweighting baseline and a random/shuffled-edge control would sharpen attribution; evolution from the original typed GoS graph would separate substrate effects from update effects; separate ablation of the avoid channel is needed given its aggressiveness.
4 comments
AuthorsOpenReview ↗

Response to Reviewer BhSi: Paired Statistics and Evaluation Protocols

Thank you for requesting paired tests and clarification of the evaluation protocols. Below, we report the paired analysis and clarify the reward definition, evolution traces, and LLM use. The companion response addresses the suggested baselines and structural controls.

1. Can you report per-task paired bootstrap CIs (or a Wilcoxon signed-rank test) for SE-GoS vs SkillDAG on both the full-benchmark and held-out protocols?

Both protocols favor SE-GoS, with 95% confidence intervals above zero and p<0.05. For deepseek-v4-flash-0731, we average each method's two attempts within task, then pair by task. Section 4.3 and Table 3(b) report:

ProtocolTasksW / T / LSE-GoS R (%)SkillDAG R (%)SE-GoS − SkillDAG (%)95% paired bootstrap CI (%)Two-sided Wilcoxon p
Full-benchmark8713 / 69 / 559.455.3+4.09[+0.24, +8.05]0.0419
Held-out376 / 30 / 158.354.2+4.16[+0.66, +8.61]0.0469

Appendix B.3 specifies 200,000 bootstrap resamples, including zero-difference pairs. The exact Wilcoxon test follows the standard convention of excluding zero differences and assigning average ranks to ties. Inference uses unrounded task means for the evaluated graphs.

2. Which LLM performs the node-update rewriting (never identified in App. A.4)? If it belongs to an evaluated backbone family, how does that square with the "LLM-prior-free" claim in the abstract and Table 1?

a. Which LLM performs the node-update rewriting (never identified in App. A.4)?

The corresponding execution backbone performs rewriting: deepseek-v4-flash-0731, minimax-m2.7, or gpt-5.2-codex. Rewriting uses training evidence offline, and descriptions are frozen before evaluation. Section 3.5 and Appendix A.5 now identify these models.

b. If it belongs to an evaluated backbone family, how does that square with the "LLM-prior-free" claim in the abstract and Table 1?

LLM-free applies to primary graph initialization. Topology and weight updates also use fixed rules, while description rewriting calls the backbone. The abstract, Table 1, and Section 3.1 replace the broader “LLM-prior-free” claim with this precise scope.

3. Do the quoted GoS/Vanilla/Vector numbers in the minimax and gpt-5.2-codex blocks use the same fractional CTRF reward as your fresh runs? And where do SE-GoS's ALFWorld evolution traces come from — the same 140 valid-seen episodes it is scored on, while SkillDAG trains on 420 train episodes?

a. Do the quoted GoS/Vanilla/Vector numbers in the minimax and gpt-5.2-codex blocks use the same fractional CTRF reward as your fresh runs?

Yes. Vanilla, Vector, MemRL*, GoS, SkillDAG, and SE-GoS all use the fraction of unit tests passed on SkillsBench. Our runs compute this fraction from CTRF outcomes, and the results taken from GoS use the same reward definition. Section 4.1 specifies the shared metric and result sources, with accounting details in Appendix B.1.

b. And where do SE-GoS's ALFWorld evolution traces come from — the same 140 valid-seen episodes it is scored on, while SkillDAG trains on 420 train episodes?

Evolution uses the separate 420 training episodes. The graph is then frozen and evaluated twice on all 140 valid_seen episodes, yielding 280 scored attempts. Section 4.1 and Appendix B.1 specify this separation.

AuthorsOpenReview ↗

Response to Reviewer BhSi: Graph-Free Ranking and Structural Controls

4. (Suggestions) A graph-free per-skill success-rate reweighting baseline and a random/shuffled-edge control would sharpen attribution; evolution from the original typed GoS graph would separate substrate effects from update effects; separate ablation of the avoid channel is needed given its aggressiveness.

a. A graph-free per-skill success-rate reweighting baseline

MemRL* adapts episodic-memory retrieval to rank skills by relevance and reward using SE-GoS's traces, with fixed skills and model weights. Tables 2 and 3(a) report:

Benchmark / protocolBackboneVector R (%)MemRL* R (%)SE-GoS R (%)SE-GoS − MemRL* (%)
SkillsBench fulldeepseek-v4-flash-073138.749.659.4+9.8
SkillsBench fullminimax-m2.710.420.928.5+7.6
SkillsBench fullgpt-5.2-codex21.532.038.1+6.1
SkillsBench held-outdeepseek-v4-flash-073138.149.058.3+9.3
ALFWorlddeepseek-v4-flash-073183.985.791.1+5.4
ALFWorldminimax-m2.750.755.068.6+13.6
ALFWorldgpt-5.2-codex92.992.993.6+0.7

SE-GoS outperforms graph-free reward ranking in all seven settings.

We considered four related approaches reviewed in Section 2: MemRL, SELM, TrackRL, and SkillBrew. We found public implementations for MemRL and SkillBrew. MemRL matches graph-free utility ranking, while SkillBrew curates the library. Appendix B.2 details the choice and adaptation.

b. a random/shuffled-edge control would sharpen attribution

Shuffling lowers reward from 59.4% to 42.4%, below the 50.8% cold start. We randomly reassign endpoints with seed 42 while preserving node and edge counts, relation types, and per-edge weights. This demonstrates the value of meaningful connectivity beyond graph size and accumulated weights. Section 4.4, Table 4, and Appendix C.1 report this control.

c. evolution from the original typed GoS graph would separate substrate effects from update effects

Evolution improves both initializations. Each uses its own execution trajectories. Section 4.5 and Table 6 report reward, node edits, and evolved edge counts:

InitializationInitial R (%)Evolved R (%)Gain (%)Node editsSemanticWorkflowDependencyAlternativeAvoidTotal edges
Semantic-only50.859.4+8.6686322800271,118
Typed from GoS52.460.8+8.43935242458231,253

Typed initialization reaches 60.8%, above SkillDAG's 55.3%. Both gains use each configuration's own initial reward; Section 4.2 distinguishes them from the +7.0% comparison with static GoS.

d. separate ablation of the avoid channel is needed given its aggressiveness.

Disabling only avoid filtering reduces reward in both protocols, with the graph otherwise unchanged. Table 4 reports the full-benchmark ablation, and Appendix C.1 reports held-out results and bundle diagnostics in Tables 12 and 11:

ProtocolWith avoid R (%)Without avoid R (%)Difference from disabling avoid (%)Avoid edgesTotal edgesAttempts with changed bundles (retrieval diagnostic)
Full-benchmark59.457.1−2.3271,1188/174
Held-out58.357.0−1.399991/74

Section 3.3 specifies the safeguards: pairs with an existing navigation relation are excluded from avoid induction, and any successful co-use retracts an exclusion. Appendix B.1 defines evidence within individual attempts.

Additional concern. The motivating mechanism never activates

Round 2 induces one dependency edge, retained in round 3, as Table 7 shows. Evolution from GoS's typed graph retains 45 dependency edges and improves reward from 52.4% to 60.8% in Table 6. Zero dependency edges occur only in the first round from semantic initialization.

Additional concern. “the edge-reinforcement rule is admittedly non-causal (reinforcing all incoming edges to a used skill regardless of source)”

Reinforcement encodes target-skill utility for retrieval. Task rewards do not identify individual edges' causal contributions. Updating all incoming non-avoid edges lets query-dependent diffusion use this utility along existing routes, as Section 3.4 explains. The edge-weight component alone yields +6.5% in Table 4.

Additional concern. “the round-3 reversal is direct evidence the raw update is not a stable fixed point.”

Repeated updates do not improve reward monotonically: Table 7 reports 59.4%, 59.8%, and 54.0% over three rounds. Section 4.5 recommends choosing the round on separate validation tasks.

Reviewer_BhSiOpenReview ↗

Response to Rebuttal

Thank you for the thorough rebuttal. The paired significance tests on DeepSeek (p < 0.05 on both full and held-out splits), the clarifications on reward consistency, ALFWorld split, and "LLM-free" scope, and all four new controls (MemRL*, shuffled edges, typed-GoS evolution reaching 60.8%, and the avoid-channel ablation) directly address my main concerns. In the final version, please also consider adding significance tests for the remaining backbones and highlighting validation-based checkpoint selection for round 3. I am raising my rating from 2 to 3.

AuthorsOpenReview ↗

Quick Response to Reviewer BhSi on further concerns

Dear Reviewer BhSi,

Thanks for the follow-up on this thread, we will follow your suggested investigations and prepare corresponding materials and results in the futural draft version.

Best regards,

Authors

Reviewer_Wge1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SE Graph of Skills, it treats a skill library’s retrieval graph as an index that can be improved offline from agent execution traces, using three training-free updates while the model, skills, and retrieval algorithm stay fixed. The paper shows the effectiveness of SE-GoS with improved rewards on SkillsBench with one evolution round increases reward from 52.4% to 59.4% with about 32% fewer tokens.

Critical strengths
  • The dynamic GoS evolving shows consistent gains and transfers across different backbones: SE-GoS beats Vanilla, Vector, static GoS, and the closest self-evolving baseline (SkillDAG) on SkillsBench across all three backbones.

  • The paper identifies a limitation of static structural skill retrieval: the graph is built once and never corrects itself. It then shows that the traces a single evaluation already produces (retrieved, used, reward) are enough to repair it.

Critical weaknesses
  • There is a conflict between paper framing and actual implementation details. The framing emphasizes induced workflow, dependency, and avoid relations. But Appendix B.3 admits that no dependency edge is induced.

  • The SkillDAG comparison does not isolate the LLM-prior-free claim. The paper says the two differ only on that axis, but SkillDAG is online and LLM-driven, SE-GoS is offline and batch, and they start from different graphs. SkillDAG was also re-run under the authors’ own harness, including a non-native train/held-out protocol. Any gap could reflect these differences rather than the absence of an LLM prior.

Questions for authors

How was SkillDAG run? Please specify its cold-start graph, whether it received the same traces or its native online protocol on SkillsBench, and what tuning it received compared with SE-GoS. Without this, the LLM-prior-free advantage is not isolated from protocol differences.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Wge1: Topology Updates and SkillDAG Comparison

Thank you for raising questions about topology induction and the SkillDAG comparison. Below, we clarify which relations are induced and specify each method's initialization, execution evidence, and update protocol.

1. How was SkillDAG run? Please specify its cold-start graph, whether it received the same traces or its native online protocol on SkillsBench, and what tuning it received compared with SE-GoS. Without this, the LLM-prior-free advantage is not isolated from protocol differences.

a. Please specify its cold-start graph

We used SkillDAG's official released 1,000-skill graph, containing 2,360 initial edges. Its official construction uses functionality/prerequisite candidates and an LLM relation classifier.

b. whether it received the same traces or its native online protocol on SkillsBench

SkillDAG generated its own traces using its native online graph-editing mechanism. SE-GoS collected separate traces and updated its graph offline. Appendix B.2 specifies both protocols, including our shared 50/37 held-out adaptation.

c. what tuning it received compared with SE-GoS

SkillDAG used the recommended top-k=5 and neighbor depth=2, without additional hyperparameter search. SE-GoS used GoS's released lightweight retriever defaults. Section 4.1 and Table 10 in Appendix B.2 specify the complete configurations.

d. Without this, the LLM-prior-free advantage is not isolated from protocol differences.

The LLM-free claim concerns primary graph initialization, which requires neither embeddings nor LLM calls. Description rewriting uses the backbone LLM. Section 3.1, Table 1, and Appendix A.5 make this distinction explicit. SE-GoS extends SkillDAG's unweighted relation editing with continuous edge-weight adaptation and retrieval-description updates. Table 4 shows gains from both across every tested component combination. SE-GoS outperforms SkillDAG on all three SkillsBench backbones and the held-out split, as Tables 2 and 3 show.

Additional concern. The framing emphasizes induced workflow, dependency, and avoid relations. But Appendix B.3 admits that no dependency edge is induced.

Table 7 shows one dependency edge induced in round 2 and retained in round 3. Table 6 reports 45 dependency edges retained when SE-GoS evolves GoS's typed graph, with reward rising from 52.4% to 60.8%, above 59.4% from semantic initialization and SkillDAG's 55.3%. The zero count is specific to the first round from semantic initialization. The revised results demonstrate both new dependency induction and improved retrieval with an existing dependency-aware graph.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Anonymous Reproducibility Materials

Dear Area Chair and Reviewers,

Thank you for the constructive feedback. We summarize the principal revisions to the manuscript submmited in September, with references to our detailed responses.

1. Innovation and gains over SkillDAG. SE-GoS extends SkillDAG's unweighted relation editing with continuous edge-weight adaptation and retrieval-description optimization. These components alone improve reward by 6.5 and 2.8 percentage points over the semantic cold start. SE-GoS outperforms SkillDAG on all three SkillsBench backbones and the held-out split (Tables 2–4; Wge1, Q1; BhSi, mechanism concerns).

2. Added paired statistical evidence. For the primary deepseek-v4-flash-0731 backbone, SE-GoS exceeds SkillDAG by 4.09 percentage points on full87 and 4.16 points on held-out tasks. The 95% paired bootstrap intervals are [0.24, 8.05] and [0.66, 8.61], with exact two-sided Wilcoxon p = 0.0419 and 0.0469, respectively. Held-out evaluation uses frozen graphs (Table 3; BhSi, Q1).

3. Demonstrated gains from graph structure. With the same traces, SE-GoS reaches 59.4% versus graph-free MemRL's 49.6%* in the primary full87 setting. Shuffling edges lowers reward to 42.4%, below the 50.8% cold start. Avoid filtering contributes 2.3 points on full87 and 1.3 points on held-out tasks (Tables 2–4 and 12; BhSi, Q4; S7DT, Q1).

4. Separated initialization and evolution gains. The abstract now reports 50.8% → 59.4%, using SE-GoS's semantic cold start already reported in the original ablation. The previous 52.4% was the static typed GoS baseline; that score and SE-GoS's 59.4% result are unchanged. The new typed-initialization experiment shows 52.4% → 60.8%, retaining 45 dependency edges. Round 2 also induces a new dependency edge from semantic initialization (Tables 6–7; Wge1/BhSi, dependency concerns; BhSi, Q4c).

5. Made execution protocols explicit. SkillDAG uses its official graph, own online traces and recommended settings. All SkillsBench methods share fractional unit-test reward; ALFWorld separates 420 training from 140 evaluation episodes. Primary initialization is deterministic and LLM-free, while description rewriting uses the execution backbone (Wge1, Q1; BhSi, Q2–3).

6. Explained new-skill entry and evolution duration. New skills enter through deterministic semantic connections and lexical seeding; one successful query-to-use observation establishes a workflow edge. Topology alone improves reward from 50.8% to 54.1%. Multi-round rewards of 59.4%, 59.8% and 54.0% motivate selecting the evolution checkpoint on separate validation tasks (Tables 4–5 and 7; S7DT, Q2 and stability concern).

Anonymous materials: Our repository includes code, experiment configurations, SkillsBench and ALFWorld task assets, skill archives, graph snapshots, selected per-attempt scores, evolution records, all 12 manuscript tables, and scripts for paired statistics and checksum verification. The paired comparisons can be recomputed from the included scores without a model API.

AuthorsOpenReview ↗

A kind reminder

Dear Reviewers,

Here is a gentle follow-up on our responses and the revised manuscript. We would appreciate your feedback on whether the added experiments and clarifications address your concerns. Please let us know if you have any further questions or concerns.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.