acceptodds
Under review as a conference paper at ICLR 2027

Skill-Forest: Randomized Ensembles of Textual Skill Optimizers

Abstract

Agent skills adapt frozen large language model agents by encoding reusable procedural knowledge outside model parameters. Under limited data, noisy textual updates and repeated selection on small validation sets can make single-path optimization sensitive to early decisions and evaluation noise. We introduce Skill-Forest, an optimizer-level ensemble inspired by the randomize-and-aggregate principle of Random Forests. Each tree uses an independently sampled training–validation split and independently searches for its tree-local best Skill. Gain-prioritized expansion under a shared budget, validation-guided pruning, and early rejection enable efficient forest construction. The resulting tree-local best Skills are then aggregated into one deployable Skill through direct selection (Select) or hierarchical merge (Merge). This design diversifies both optimization paths and validation views to reduce the high variance associated with textual-gradient optimization. We evaluate Skill-Forest across five benchmarks spanning direct-chat tasks, tool use, and embodied tasks. Skill-Forest Select and Merge improve the mean held-out Hard score over No-skill by 12.82 and 12.87 percentage points (pp), respectively. Under the same reflection strategy, they outperform SkillOpt by 2.59 and 2.63 pp, respectively. Skill-Forest reduces training and validation rollouts during forest construction by 73.5% across the five benchmarks relative to complete tree construction, supporting efficient multi-path ensemble Skill exploration.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Agent Skills, Skill Optimization, Textual Gradients, Ensemble Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dLawOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Skill-Forest targets capability-matched Agent Skill optimization, where the same relatively weak backbone both executes tasks and optimizes Skills using only outcome-level supervision, without relying on a stronger optimizer. The method brings the randomize-and-aggregate principle of Random Forests into textual Skill optimization: starting from the same initial Skill, it runs K K independent Skill trees, each with its own sampled training-validation split; within each tree, patch-based branching preserves alternative update paths. At the forest level, a shared expansion budget, validation-gated pruning, gain-prioritized node expansion, early rejection, and Root validation reuse control construction cost. Tree-local winners are then aggregated into one deployable Skill via Select or Merge. Experiments cover five benchmarks spanning single-round QA, multi-round tool use, and persistent environment interaction. Skill-Forest outperforms baselines such as SkillOpt, GEPA, and Trace2Skill overall and reports substantial rollout savings.

Critical strengths

1.Reasonable and well-motivated direction: It extends Skill optimization from a single textual-gradient path to multi-tree search with independent data views, directly targeting the high variance of single-path optimization. The idea is natural and has practical engineering value.

2.Relatively complete efficiency design: The shared expansion budget, validation-gated pruning, gain-prioritized expansion, upper-bound early rejection, and Root evaluation reuse jointly reduce forest-construction rollouts. The paper also provides a fairly detailed cost breakdown.

3.Good experimental coverage: The five benchmarks span single-round direct chat, multi-round tool use, and persistent environment interaction. Under the capability-matched setting, Skill-Forest overall outperforms several representative baselines without relying on a stronger optimizer.

Critical weaknesses

1.Internal tension in the method: The paper acknowledges that tree-local validation gains are not comparable across trees, yet uses them to rank nodes in a global priority queue for allocating expansion budget. This may divert budget to trees with easier validation subsets or lower initial scores rather than to genuinely promising search directions.

2.Concerns about data usage and selection fairness: Each tree uses only about 50% of the training data, while baselines mostly use the full training set. Final Select/Merge also scores winners on the complete training split, without an independent selection set separate from the tree splits, raising the risk of selection overfitting.

3.Insufficient evidence strength: Main results lack cross-run standard deviations and significance tests. The efficiency comparison only counts forest construction and does not fully include aggregation/Merge rollouts and LLM call costs. It remains unclear whether the 2–3 pp gains are stable and how much is attributable to the forest structure itself.

Questions for authors

see Critical Weaknesses

Reviewer_C8oeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors introduce Skill-Forest, an optimization approach that uses multiple branching searches and merging to produce a deployable skill. They claim improved performance and validation-rollout efficiency without a stronger optimizer. The authors use GPT-5.4-mini and show results over five benchmarks against baselines of a naive baseline (no-skill), the base SkillOpt skill (Initial Skill), SkillOpt, Trace2Skill, and GEPA. Skill-Forest's best configuration improves average performance by 2.63 over SkillOpt and 1.5 over GEPA.

Critical strengths

Strongest positive is in the performance against SkillOpt, showing gains on five optimization runs over two benchmarks with a reduction in number of rollouts needed.

Critical weaknesses

Biggest weakness is that this approach is only evaluated on GPT-5.4-mini. I consider this largely 'sinking' this paper as we are unable to tell whether the improvements of SkillForest are actually generalizable or just related to a quirk of GPT-5.4-mini. The advantage over the strongest baseline (GEPA) is also minimal and the headline 73.5% rollout reduction is misleading as it compares against exhaustive branch construction and excludes aggreation.

Questions for authors

1.) Most important: do you have proof that this method would work for another model? Even if its another GPT model. 2.) How does Skill-Forest compare to GEPA across matched budgets? 3.) What is the efficiency improvement when all rollout, token, and aggregation costs are factored in?

Reviewer_1wJgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Skill-Forest, a framework for optimizing textual skills for frozen language-model agents. It constructs multiple skill-search trees using independently sampled training–validation splits, with trajectory-derived patches generating alternative updates. Forest construction combines parent-relative validation gates, gain-prioritized expansion, a shared expansion budget, and early rejection of candidates that cannot pass validation. Tree-local winners are subsequently selected or hierarchically merged into one deployable skill. Experiments use the same model for execution and optimization across five benchmarks. Seven-tree Merge achieves an average Hard score of 66.87%, exceeding SkillOpt by 2.63 percentage points and GEPA by 1.50 points. Additional experiments examine branching, randomized splits, and repeated optimization runs on LiveMath and ALFWorld.

Critical strengths
  1. A coherent and practically usable approach to exploring alternative skill updates. The framework combines branching and distinct validation views while ultimately producing a single deployable skill. It also operates without a stronger optimizer model. The early-rejection rule is a particularly sound component: it avoids completing evaluations that cannot change the acceptance decision. These properties make the approach practically relevant for improving existing skill optimizers.

  2. Encouraging empirical results with meaningful supporting controls. The evaluation spans direct reasoning, tool use, and embodied interaction. Seven-tree Merge improves over both SkillOpt and GEPA on four of five benchmarks. The five complete optimization runs, shared-split control, and SkillOpt Parallel×7 baseline on two benchmarks are valuable additions beyond repeated evaluation of a single learned skill. They provide credible evidence that broader optimization search can improve average downstream performance.

Critical weaknesses
  1. The comparisons do not adequately establish that the proposed search organization improves performance at comparable computational cost. Tables 3 and 6 reveal substantial budget differences: seven-tree forest construction uses 22,467 rollouts on SearchQA versus GEPA’s 4,000-call budget, and 2,089 on LiveMath versus 360. Across the five benchmarks, construction alone uses approximately 4.0–5.8 times GEPA’s configured budget, before final selection or merging. GEPA’s official guidance supports the chosen budget as a reasonable configuration, but this does not establish resource parity. Consequently, the 1.50-point average improvement over GEPA does not isolate an advantage of the proposed optimization strategy from an advantage of additional search. Likewise, the reported 73.5% saving is relative to exhaustive depth-three tree construction, whose corresponding performance is not reported; it does not establish a better performance–cost tradeoff than existing optimizers. The Parallel×7 experiments partially address this concern, but cover only two benchmarks and leave an unexplained cost discrepancy on LiveMath. Since efficient organization of optimization is a central contribution, controlled performance–budget comparisons are essential.

  2. The central variance-reduction claim is only partially supported, and the randomized-split ablation does not demonstrate the proposed explanation. Table 8 shows lower standard deviations than single-run SkillOpt, but the more relevant controls give a mixed picture. On LiveMath, the seven-tree forest has a standard deviation of 1.96 points, compared with 0.70 for SkillOpt Parallel×7. Relative to the shared-split forest, tree-specific splits increase the reported standard deviation on both LiveMath, from 1.42 to 1.96, and ALFWorld, from 1.09 to 3.13. On ALFWorld, this accompanies only a 0.25-point improvement in the mean. These observations do not establish that randomization generally increases variance, given the small number of runs, but they also do not substantiate the paper’s attribution of improved stability to diverse validation views. Select performs winner selection, while Merge performs nonlinear textual synthesis; neither inherits a variance-reduction argument merely from the Random Forest analogy. Moreover, the five-benchmark headline results average three executions of fixed learned skills, rather than complete optimization runs. The available evidence therefore supports promising mean performance more strongly than a general reduction in optimizer variance or sensitivity to validation splits.

Questions for authors
  1. Additional evidence needed for the efficiency claim: Under a common end-to-end rollout budget, does Skill-Forest retain its advantage over GEPA and repeated SkillOpt optimization? Please include final selection and merging in the budget, and separately report optimizer-model calls or token usage so that rollout savings can be interpreted as computational savings.

  2. Protocol clarification: Table 8 reports 212 rollouts for a single SkillOpt run on LiveMath but 2,450 for SkillOpt Parallel×7, whereas seven replicas would account for 1,484 rollouts before additional selection costs. What accounts for the remaining 966 rollouts? Please reconcile this accounting and explain whether the reported 10.1% saving remains under a consistent definition of optimization cost.

  3. Clarification and supporting evidence for the stability claim: Which results specifically demonstrate variance reduction attributable to tree-specific splits? Please reconcile this claim with the higher standard deviations than the shared-split control on both ablation benchmarks, and distinguish improvements in expected performance from improvements in stability.

  4. Additional evidence needed for the headline comparison: How uncertain are the reported gains over GEPA across complete optimization runs? Repeated test execution measures a different source of variability. Results across optimization seeds for the principal GEPA and Merge comparisons would materially affect my assessment of the reported 1.50-point average advantage.

These requests concern the submission’s existing claims and experimental interpretation, rather than extensions to new tasks or future applications.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.