acceptodds
Under review as a conference paper at ICLR 2027

SkillGATE: Gate-Aware Monte Carlo Tree Search for Skill Retrieval

Abstract

Skill Retrieval (SR) aims to identify the most relevant skills from external skill libraries, and becomes increasingly challenging as libraries grow in scale and diversity. Existing methods either rank skills independently or rely on predefined graph propagation and hierarchical routing, making them vulnerable to semantic distractors, local trapping, and early routing errors. We formulate SR as an adaptive information-foraging process that coordinates region-level navigation with skill-level selection according to the utility and uncertainty observed during search. Based on this formulation, we propose SkillGATE, a graph-guided hierarchical retrieval framework with Gate-Aware Monte Carlo Tree Search (MCTS). SkillGATE constructs a graph-preserving hierarchical index and performs adaptive retrieval through selection, expansion, simulation, and backpropagation. G-PUCT guides action selection, expansion explores new regions, simulation evaluates candidate skills, and backpropagation updates search statistics. Experiments on six SR benchmarks show that SkillGATE consistently improves diverse retrieval and reranking backbones, achieving a 16.3% improvement in overall R@1 over the strongest retriever-based baseline. Our code is available at https://anonymous.4open.science/r/SkillGate-EFD0/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Skill Retrieval, Monte Carlo Tree Search, Graph-guided Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_ucezOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SkillGATE, a skill retrieval framework that combines a skill graph, hierarchical indexing, and gate-aware Monte Carlo Tree Search (MCTS). The method introduces four search actions: node-aware descent, skill-aware descent, horizontal search, and pruning. It uses historical rewards and entropy to guide action selection and employs a dual-path backpropagation mechanism to update search statistics. Experiments on six benchmarks demonstrate performance improvements over various retrieval and reranking methods.

Critical strengths
  1. The combination of skill graphs and hierarchical indexing enables cross-region exploration and helps recover from incorrect hierarchical routing decisions.

  2. The paper provides extensive experiments, including component ablations, comparisons of different search strategies, hyperparameter sensitivity analyses, and scalability evaluations.

Critical weaknesses

W1. The proposed method introduces four search actions, a heuristic gating function, a specialized reward function, and a dual-path backpropagation mechanism. Although the authors conduct component ablations and compare their method with Greedy, UCT, and PUCT, these experiments do not sufficiently isolate the benefits of the complex MCTS framework itself. In particular, Table 2 shows that removing skill-aware descent leads to performance even worse than the original Ada-002 retriever in Table 1. Meanwhile, Table 7 lacks a simple baseline that combines hierarchical indexing and skill-aware graph propagation without MCTS-based search feedback.

W2. Several structured retrieval baselines perform poorly, even compared with simple retrieval methods. However, the original objectives of these methods may differ from the single-skill ranking task considered in this paper. The paper does not sufficiently clarify how these baselines were adapted to the current evaluation setting or whether their original configurations were faithfully preserved.

Questions for authors
  1. Why is such a complex MCTS framework necessary? Could the authors include a simpler baseline that combines hierarchical indexing and skill-aware graph propagation without MCTS-based search feedback? Furthermore, what motivates the specific formulations of the gating functions for the four actions, and have alternative mappings been explored?

  2. Do the search strategies compared in Table 7 use equivalent computational budgets? Could the authors provide recall–latency curves for different numbers of search iterations and compare the methods under matched computational budgets?

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

The paper treats skill retrieval over large libraries as adaptive search. Offline, it builds a skill graph whose edges combine embedding similarity, TF-IDF similarity, and Jaccard overlap of LLM-extracted input/output/API labels. It then recursively partitions that graph with Leiden into a hierarchy. Online, MCTS runs over the hierarchy with four actions: child descent, "skill-aware" descent via intra-subgraph PPR, horizontal moves to sibling subgraphs via global PPR, and pruning. Actions are chosen by G-PUCT, which mixes historical return with a gate built from the node's return and the normalized entropy of its visit distribution. Visit counts are backpropagated along the trajectory and rewards along the hierarchical ancestor path. The final ranking is top-K by a hybrid relevance score over the accumulated candidate pool. On SRA-Bench (26,262 skills, six query sets), the method plugged into three retrievers and three rerankers improves overall R@1 by 16.3 points over the best retriever baseline and by 10.8 over the best reranker baseline. Ablations, per-level routing-failure/rescue statistics, a search-policy comparison, and a library-size scaling study are included.

Critical strengths
  1. The per-level diagnosis in Table 3 shows where hierarchical routing fails: node-descent failures of 30.6% at L1 and 20.2% at L4. It also shows how often the graph actions recover them (skill-aware descent 42–75%, horizontal 35–65%). This is concrete mechanistic evidence for the core design choice, coupling hierarchy with graph moves, not only an end-metric gain.
  2. Gains hold across three retrieval and three reranking backbones. Ablations remove each action and component, and Appendix E compares G-PUCT with RW, HKD, PPR, Greedy, UCT and PUCT on the same index. Together these show the improvement is not tied to a single backbone.
Critical weaknesses
  1. The headline comparison mixes the effect of the search with the effect of a stronger relevance function. SkillGATE's relevance r(q,x) is a hybrid of dense similarity (0.50), TF-IDF (0.25) and exact match on LLM-extracted labels (0.25). The final output is top-K of r over the MCTS candidate pool (Alg. 2, line 33). The Table 1 baselines are mostly single-signal retrievers. The only flat-ranking control with the same weighted similarity is the "Vanilla" curve in Fig. 3(b): at the full library it reaches 39.22% R@1 vs 47.99% for SkillGATE. That ~8.8-point gap is encouraging, but it is roughly half of the +16.3 headline, and it is reported only for R@1 and on the 10% sampled subset (the 47.99 matches the Ada-002 subset result in Table 6). Please add flat ranking by the same r, and a standard hybrid baseline such as BM25+dense with RRF, to Table 1 for every backbone, both metrics and all six datasets, so the search's own contribution is clear.
  2. Robustness and cost are under-characterized. The method has more than a dozen hyperparameter groups (Table 4), plus the prior weight λ and a hard-coded 0.4 constant in the expansion rule (Alg. 2). They were grid-searched on SkillRet. Sensitivity is reported only for the relevance and reward weights, not for τκ, τH, λ, the iteration budget, Knbr, lmax or the Leiden resolution. Main results have no variance even though the simulation is stochastic. Retrieval takes about 3.1 s/query with 128 iterations (Table 9), and there is no recall-vs-budget curve or comparison with simply reranking a deeper dense shortlist under the same compute. R@10 also drops on ToolQA (−5.0/−5.5) and MedCalc; the ToolQA drop is explained only qualitatively (similar candidate skills), without supporting analysis.
Questions for authors
  1. Given that the final ranking uses r(q,x), what fraction of R@1 gains come from queries where a higher-r distractor exists in the full library but is excluded from S*(q)? Equivalently, what are flat-r R@1/R@10 on the full benchmark for each backbone (extending the Fig. 3(b) "Vanilla" point)?
  2. In the reranker setting, where exactly does the reranker enter? Does it replace r inside the reward, rerank the final pool, or both?
  3. Were hyperparameters tuned with the SkillRet-Emb backbone, which is itself trained on SkillRet data? If so, is there any overlap between the tuning and evaluation signals?
  4. Main-table variance over MCTS seeds?
  5. Code is stated as "available at here" but no link is given. Will it be released? Suggestions (not decisive): report latency and accuracy as a function of iterations, and quantify the ToolQA R@10 drop (e.g., how often the gold branch is pruned or never visited). The optimal-foraging framing is motivational only, since no patch-leaving rule is derived; either connect it more formally or tone it down. Typos: "Hierachy statics", "Updation Statics", "For eanch", "Tabel 6", "Ahorizon" vs "Ahoriz".

Comments on the paper

2 comments

Thank you for your response, I still have some concerns unresolved:

  1. The newly added Non-MCTS baseline provides evidence for the contribution of adaptive search, but its implementation details remain unclear, particularly how the search actions are performed without MCTS and whether the computational budgets are comparable.

  2. You should explain why removing skill-aware descent causes performance to fall below the original Ada-002 retriever, do you investigate its interactions with other components?

  3. You report overall retrieval latency and iteration-based Hit rates, but do not provide Recall–Latency curves or matched-budget comparisons across search strategies, leaving the computational fairness of the comparisons unclear.

Area_Chair_GAneOpenReview ↗

Please take a look at the response!

Dear reviewers,

As the rebuttal period is about to end, please take a look at the authors' responses.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.