acceptodds
Withdrawn from ICLR 2027

EvoGap: Weakness-Graph-Guided Skill Evolution for Frozen LLM Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Frozen agents can improve through external skills, but what they learn is constrained by the tasks they encounter. Generating additional tasks can broaden practice, but may not sufficiently target diagnosed behavioral weaknesses. Consequently, agents may accumulate more experience without systematically repairing the weaknesses that actually limit their performance. To overcome this limitation, we propose EvoGap, a weakness-graph-guided framework that turns diagnosed failures into executable practice and uses validated outcomes to guide skill evolution and subsequent practice. An updatable weakness graph records behavioral gaps and repair evidence to select the next target. Depending on a weakness's repair state, EvoGap runs breadth-oriented or depth-oriented practice to generate executable tasks with verifiable outcomes. Environment feedback informs the extraction of environment knowledge and behavioral skills. Candidate skills are admitted only after targeted validation, and their measured effects are written back to the same graph. This feedback loop connects practice selection, skill updates, and the next round of practice. Across AppWorld, BFCL-v3, and the three -Bench domains, EvoGap consistently improves task success over prior memory- and skill-based baselines with a frozen backbone, by points on average and up to points.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, skill acquisition, self-improvement, continual learning, environment construction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_oBNxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes EvoGap, a framework for improving frozen LLM agents through weakness-guided skill evolution. It maintains a graph of diagnosed behavioral weaknesses, selects targeted practice through breadth- or depth-oriented task generation, and updates an external skill library using environment feedback and validated outcomes. Experiments across five benchmark domains and two frozen backbones demonstrate consistent improvements over existing memory- and skill-based methods.

Critical strengths
  • The paper addresses a practical problem of improving frozen LLM agents by explicitly identifying and targeting behavioral weaknesses rather than relying on passive experience accumulation.

  • The framework integrates weakness-guided practice generation, environment-grounded skill distillation, and validation-based skill updates into a coherent closed-loop process.

  • Experiments across multiple benchmarks and model backbones demonstrate promising performance gains, supported by ablations investigating several key design components.

Critical weaknesses
  • Sections 2 and 4.1–4.2 assume access to machine-checkable weakness triggers, error signatures, correct action sequences, and executable verifiers, but do not explain how these are obtained or what supervision they require.
  • Table 3 shows only a two-task improvement over simpler informed selection policies on 40 Telecom tasks, without statistical significance analysis. Moreover, the prerequisite edges introduced in Section 4.1 are never explicitly evaluated or ablated.
  • Equation (2) imposes HARM(ST)≤ϵHARM(S_T) \leq \epsilon, yet Section 4.3 treats unrelated canary tasks as warnings rather than hard admission constraints. No HARM results are reported, and improvements in aggregate success rates cannot rule out regressions on previously solved tasks.
Questions for authors

see wekaness.

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

The paper reframes skill self-evolution for frozen LLM agents as a sequential decision problem: which diagnosed behavioral weakness to practice next. An updatable weakness graph holds machine-checkable weakness nodes (trigger predicate + error-signature predicate, facets, repair tier, failure-prevalence prior). A linear selection score S(w) = αP(w) + γF(w) − λN(w) — which the paper shows subsumes random, failure-frequency, and cost-aware heuristics as special cases — selects the next weakness; depending on repair state, breadth- or depth-oriented practice generates executable, verifiable tasks ("solution-before-task": gold mechanism and verifier materialized before the surface task); two-arm execution (base vs candidate skill) yields evidence; an admission gate merges only skills showing net improvement without regression; measured effects are written back to the graph. Evaluation on BFCL-v3, AppWorld, and τ²-Bench (Retail/Airline/Telecom) with frozen Qwen3-Coder-Plus and Qwen-Flash reports +5.8 pp average (up to +12.5 pp) over A-Mem/ExpeL/AWM/SkillX, plus selection-policy, channel, and practice-mode ablations.

Critical strengths
  1. A conceptually appealing reframing. Treating self-evolution as weakness-conditioned practice selection — with random/failure-frequency/cost-aware heuristics falling out as coefficient settings of S(w) — is elegant, and the closed loop (SELECT–GENERATE–RUN–DIAGNOSE–PROPOSE–ACCEPT–SETTLE) with verifier-based evaluation is principled in design. I note this strength for completeness; as detailed below, it is not matched by the evidence actually delivered, and it does not weigh against rejection for this submission in its current form.
Critical weaknesses
  1. The submission is incomplete and its promised evidence is absent. Section 1 names five core metrics — held-out SR, target-error-signature disappearance, unseen-facet generalization, consistent-resolution, and HARM (regression on originally-correct tasks, itself a constraint in Eq. 2) — yet the experiments report only SR; the diagnostic signals were "tracked" but no numbers appear. Two ablations promised in the introduction (whether the admission gate matters; whether hierarchical retrieval is necessary) never appear. The appendix is empty (a LaTeX placeholder reading "You may include other additional sections here"), while the reproducibility statement claims details are "further elaborated in the Appendix." Two cross-references are broken ("deferred to ??"), Table 2's caption references "results not yet available," and an anomalous value (ExpeL/Flash/Telecom = 26.7, impossible for an Avg@4 result on a 40-task set) plus an unreconciled discrepancy between Table 6 (77.5%) and Table 2 (72.5%) for the same setting suggest the experimental pipeline itself was not audited before submission. A submission that does not deliver its own stated evaluation plan cannot be assessed as sound, and these are not defects addressable by clarification alone — they require the missing experiments to exist.
  2. The evidence that does exist does not support the claims, and leakage is unexcluded. The only experiment isolating the paper's stated contribution — the selection policy — beats a trivial failure-frequency heuristic by +2 tasks out of 40 (Table 3, conceded by the authors). The headline gains rest on 2 of 5 cells (+10.6 BFCL, +12.5 Telecom) with no variance reporting (Avg@4 with no std; Table 4 is single-trial), no gain at all on Airline, and a tie on Retail under the weak backbone. The four most relevant competitors named in the introduction (EvoSkill, Trace2Skill, SkillOpt, Skill Self-Play) are never run, so Table 2 conflates selection with the entire generation+gate pipeline. Most seriously, τ²-Bench has no training split, yet the provenance of D_seed, the "frozen calibration initial split," and the 40-task Telecom subset (the official Telecom split has 114 tasks) is never stated; generated practice tasks deliberately "reproduce the original trigger condition and keep the initial state and tool-chain comparable to the source case" — if the source failures were observed on evaluation tasks, this constitutes near-duplicate training on test failures regardless of the formal D_prac ∩ D_test = ∅ constraint. The claim "D_test never touched during evolution" is asserted, not demonstrated, and the information needed to verify it is not in the submission.
Questions for authors
  1. Exactly where do D_seed, the calibration "initial split," and the failure traces that seed the weakness graph come from, for each benchmark? Were failure traces ever collected on evaluation tasks? Please demonstrate (not assert) that this is not training on test failures.
  2. Where are the five promised core metrics (especially per-task HARM) and the two promised ablations (gate on/off; hierarchical retrieval on/off)? Were they run?
  3. Which LLMs perform weakness diagnosis, task generation, and step-evidence extraction, with what prompts? Which AppWorld split, which BFCL-v3 categories, and which model simulates the τ²-Bench user? What are the task counts per benchmark?
  4. Please reconcile Table 6 (77.5%) with Table 2 (72.5%) for the same backbone/benchmark, and explain the 26.7% value in Table 2.
Reviewer_Ly5tOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes weakness graph to point out the behavior weakness. The idea of waekness looks interesting and this weakness graph can help the skill evolution. Experiments are on 3 different benchmarks.

Critical strengths

This paper proposes weakness graph to point out the behavior weakness. The idea of waekness looks and this weakness graph can help the skill evolution. Experiments are on 3 different benchmarks.

Critical weaknesses

The liturature review is very insufficient, and it makes it very hard to position this paper and very hard to know the novelty of the paper. The overall quality of the paper is very bad: the method and experimental design looks very weak. The proposed comcept is very dense which is very hard to follow.

Questions for authors
  1. Weakness definition should be more carefully describe. For example, as described, C_w is “a predicate over the environment/task state for when the error may occur”, and how can only use a binary value without time indicator can define “when occur”. Another example is definnation of, \sigma_w(\tau), indicate that the error are on trajectory level, not on skill level, which is confusing, how do you define skill level weakness. The definition is a mismatch with paper’s motivation.
  2. It’s very unclear about how to connect the definition of weakness with the weakness graph, how do you define graph edge? That is very important because without this it cannot show the necessary of graph.
  3. There is no “rho” inside definition of PREF function in equation 1, how do you optimize that in equation 2.
  4. What is “??” in line 114.
  5. In experiment part, the efficiency on token cost for specifc preformance milestone should be reported. I’m also curious about the cross domain/tasks abilities.

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

Related papers

Discussion (0)

Sign in to comment.