PaperlayerView exact revision on arXiv

Source-linked AI summary

SkillVine: Agent Skill Evolution via Branching Exploration

Kaiwei Liu, Jiqian Dong, Liran Dong, Shuai Mao, Mingming Zhao, Bufang Yang, Jie Chuai, Zhitang Chen, Guoliang Xing, Zhenyu Yan

arXiv:2609.32731v1cs.AI

TL;DR

Linear skill evolution applies updates sequentially to the latest library, leaving alternative paths unexplored. SkillVine treats evolution as branching graph search with collaborative trunk and branch updates, a parent-node selector, and adaptive update granularity. It outperforms six baselines in nine of ten model–benchmark combinations and finds every best final library through a branch rather than the trunk.

  • Problem

    Linear skill evolution applies updates sequentially to the latest library, risking overwritten useful content and leaving alternative evolutionary paths unexplored.

    Source: §1 · pp. 1–2

  • Method

    SkillVine preserves skill-library versions in a graph and combines trunk updates with branches, complementarity-aware parent selection, and adaptive branch-update granularity.

    Source: §1 · p. 2

  • Results

    Nine of ten model–benchmark combinations favor SkillVine over six baselines, and every best final skill library is discovered through a branch rather than the trunk.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    Branching exploration offers an alternative approach to agent self-evolution, with the best final skill libraries found on branches rather than the linear trunk.

    Source: §1 · p. 2, §5 · p. 10

Abstract

from arXiv · show

Agent skills encapsulate reusable procedural knowledge that enables LLM agents to perform tasks, and they can be improved automatically using trajectories from interactions with the environment. This is the classic problem of skill evolution. Existing approaches predominately follow a linear evolution paradigm, in which updates are sequentially applied to the latest skill-library version. As a result, they inevitably fall into local optima, leaving many promising evolution paths unexplored. We propose SkillVine, an automatic skill-evolution framework that formulates skill evolution as a graph search problem and employs a branching exploration strategy. Equipped with a trunk-branch collaborative searching mechanism, an intelligent parent-node selector, and an adaptive-granularity update rule, SkillVine achieves a balance between exploration and exploitation. We evaluate SkillVine on 5 benchmarks with two LLMs. Results show that SkillVine discovers better skill-library versions along branches than along the linear trunk and achieves the best test performance in nine of ten benchmark-model combinations.

1 INTRODUCTION

SkillVine addresses the limits of linear skill evolution by preserving historical library versions and exploring alternative paths. Across five benchmarks and two base LLMs, it outperforms six baselines in nine of ten model–benchmark combinations, with every best final library found on a branch.

1 INTRODUCTION

  • Adding skills alone may not improve performance: expanding a library to 202 skills can reduce performance by up to 21%, with skill shadowing accounting for up to 68% of that degradation.The passage describes skill shadowing as agents increasingly selecting inappropriate skills as the library grows.Source: §1 · p. 1
  • Linear evolution applies each update to the latest library, risking overwritten useful content, local optima, and the loss of alternative paths for comparison.A helpful rule for current tasks can limit later performance, and linear evolution cannot return to the pre-update version to explore another path.Source: §1 · pp. 1–2
  • SkillVine preserves complete skill-library versions in a graph and combines continuous trunk updates with branches that explore capabilities the trunk lacks.Nodes store library versions and validation results; directed edges record parent–child relationships and allow returns to earlier versions.Source: §1 · p. 2
  • SkillVine outperforms six baselines in nine of ten model–benchmark combinations, and every best final skill library is discovered through a branch rather than the trunk.With DeepSeek-V4-flash, it leads on all five benchmarks.Source: §1 · p. 2
  • A complementarity-aware selector chooses branch starting points using validation accuracy, prior child count, and capability complementarity with the trunk.Adaptive branch-update granularity responds to whether the trunk accepts its update, helping reduce harmful updates from noisy or misleading batches.Source: §1 · p. 2

2 RELATED WORK

Related work covers lifelong skill management and branching search across decision-making and LLM-agent tasks. SkillVine differs by branching over complete skill-library versions rather than action, reasoning, or tool-use trajectories.

2.1 LIFELONG EVOLUTION OF SKILLS

  • Prior work on lifelong skill evolution spans automatic skill curation, test-time optimization, and continuous management, alongside taxonomies for lifecycle stages and skill operators.The cited survey organizes the domain into an eight-stage lifecycle and a ten-operator taxonomy.Source: §2.1 · p. 2

2.2 BRANCHING SEARCH AND EXPLORATION

  • Branching methods include value-guided tree search, archiving and revisiting promising states, and copying high-performing training states before perturbing them.Examples include MCTS and UCT, Go-Explore, and Population-Based Training.Source: §2.2 · p. 3
  • Unlike methods that branch over action, reasoning, or tool-use trajectories, SkillVine branches over skill-library versions, with each node a complete snapshot.Edges represent parent–child relationships between those snapshots.Source: §2.2 · p. 3

3 SKILLVINE

SkillVine formulates skill evolution as budget-constrained optimization and searches a recoverable graph of skill-library versions. It combines a continuous trunk with branches from historical checkpoints, selecting and updating branches according to validation evidence and the trunk’s update outcome.

3.1 PROBLEM FORMULATION

  • SkillVine seeks the skill-library version with highest expected validation accuracy among versions explored within a fixed training budget.Each candidate update is either discarded or applied to a specific version, allowing alternative evolutionary paths to be explored.Source: §3.1 · p. 3

3.2 SKILLVINE CORE CONCEPT

  • The trunk updates on every batch, while branches start from selected historical checkpoints and explore paths using only a subset of batches.The trunk accumulates long-horizon information and provides a reference for branch selection and updates.Source: §3.2 · pp. 3–4

3.3 SKILLVINE METHODOLOGY

  • SkillVine records each skill-library version as a recoverable graph node, with edges linking each child version to its parent.Stored metadata includes parent and node state, validation accuracy and solved or failed samples, and expansion count.Source: §3.3 · p. 4
  • The selector ranks historical versions using validation accuracy, underexploration, and complementarity with the current trunk.Complementarity credits validation samples a trunk parent misses and penalizes samples the candidate loses that the trunk solves.Source: §3.3 · p. 5
  • Branches use coarse-grained updates when the trunk accepts its update and fine-grained edits when the trunk rejects it.Coarse updates require one evaluation; fine-grained updates evaluate edits individually and return to the latest accepted version after a rejected edit, trading evaluation cost against update reliability.Source: §3.3 · pp. 5–6

4 EXPERIMENTS

SkillVine is evaluated against six baselines on five benchmarks with two base LLMs; results compare its branching search, component ablations, evolution paths, and parameter settings.

4.1 EXPERIMENTAL SETUP

  • The evaluation covers five benchmarks spanning spreadsheet editing, mathematical reasoning, document question answering, open-domain answering, and table-based tasks, against six baselines.The baselines include linear evolution, candidate-update selection, random parent selection, Hermes-SE, SkillOpt, and SkillOpt-Lite.Source: §4.1 · p. 6

4.2 MAIN RESULTS

  • SkillVine scores highest on all five benchmarks with DeepSeek-V4-flash and four of five with Qwen3.8-Flash, and beats Best-of-3 and Random in all ten combinations.On SpreadsheetBench with DeepSeek-V4-flash, its test score rises from 48.04 for Linear to 71.53, a 23.49-point increase.Source: §4.2 · p. 7, Table 1 · p. 8

4.3 ABLATION STUDY

  • SkillVine generally outperforms ablations, while the w/o Selector setting uses slightly more rollouts yet performs worse.Branch exploration and AGBU require additional rollouts but improve performance; removing complementarity changes parent selection in five of ten combinations.Source: §4.3 · p. 8, Table 2 · p. 8

4.4.1 SKILL EVOLUTION DYNAMICS

  • Across all ten benchmark–model combinations, the final best skill-library nodes are reached through branch paths, and AGBU rejects at least one harmful edit in eight paths.Discovery-only selectors, refinement-only selectors, and a mix of both appear in the best paths.Source: §4.4.1 · p. 9, Figure 4 · p. 9

4.4.2 IMPACT OF SYSTEM PARAMETERS

  • Performance generally improves and then levels off with more branch attempts; shorter branch intervals and more training steps generally improve performance.More steps provide additional branch-exploration rounds and divide evolution into more stages, allowing updates to be evaluated gradually.Source: §4.4.2 · p. 9, Figure 5 · p. 9

5 CONCLUSION

SkillVine enables parallel branching exploration for automatic agent-skill evolution. It outperforms six baselines in nine of ten model–benchmark combinations and finds better skill-library versions through branches than along the trunk.

5 CONCLUSION

  • SkillVine enables parallel search through branching exploration for automatic skill evolution.Its core components are trunk-branch collaborative search, complementarity-aware node selection, and adaptive-granularity updates.Source: §5 · p. 10
  • SkillVine outperforms six baselines in nine of ten model–benchmark combinations.Source: §5 · p. 10
  • SkillVine finds better skill-library versions through branching than along the linear trunk.Source: §5 · p. 10

AI USE STATEMENT

The authors disclose generative-AI assistance in several research and writing activities, while distinguishing activities for which they did not use it.

AI USE STATEMENT

  • Generative AI assisted with writing, language polishing, literature retrieval and discovery, research ideation and execution, and drafting sections.Source: AI USE STATEMENT · p. 10
  • The authors state that generative AI was not used to generate synthetic datasets or prove mathematical claims.Source: AI USE STATEMENT · p. 10
  • The authors reviewed all AI-assisted work and take responsibility for the paper’s final content, claims, and artifacts.Source: AI USE STATEMENT · p. 10

A.1 SKILLVINE EVOLUTION PATHS ACROSS 10 BENCHMARK–MODEL COMBINATIONS

Across all ten benchmark–model combinations, the final best nodes are reached through paths that diverge from the trunk. The paths use both discovery and refinement selectors, and eight include an update in which AGBU rejects at least one edit.

A.1 SKILLVINE EVOLUTION PATHS ACROSS 10 BENCHMARK–MODEL COMBINATIONS

  • All ten final best nodes are reached through paths that diverge from the trunk, rather than remaining on the linear trunk.Table 3 traces version numbers from each initial node to its final best node.Source: §A.1 · p. 14, Table 3 · p. 14
  • Six paths use both discovery and refinement selectors, while two use only discovery and two only refinement.D:vXXXX denotes discovery-selector branch nodes, and R:vXXXX denotes refinement-selector branch nodes.Source: §A.1 · p. 14, Table 3 · p. 14
  • Eight of the ten paths include an update in which AGBU rejects at least one harmful edit.Table 3 marks such updates with †.Source: §A.1 · p. 14, Table 3 · p. 14

A.2 EXAMPLES OF COMPLEMENTARY NODES

Two examples show that complementary branch nodes can retain or add validation capabilities, and evolution from either selected branch yields a final node that outperforms the best trunk node on validation and test sets.

A.2 EXAMPLES OF COMPLEMENTARY NODES

  • In Case 1, R:v0007 solves seven validation samples missed by M:v0002 while retaining all 21 samples the trunk node solves.The selected refinement node therefore adds coverage without losing the trunk node’s solved samples in this example.Source: §A.2 · p. 14
  • In Case 2, D:v0010 solves nine samples missed by M:v0031 but loses five samples that the trunk node solves.This example illustrates complementarity despite the branch node not retaining every trunk capability.Source: §A.2 · p. 14
  • In both cases, evolution from the selected branch produces a final node that outperforms the best trunk node on validation and test sets.The examples show the selector’s complementarity-aware choice can open better evolution paths.Source: §A.2 · p. 14

A.3 EXAMPLES OF HARMFUL EDITS REJECTED BY AGBU

The examples show AGBU rejecting harmful edits while retaining useful edits, allowing branch updates to improve even when corresponding trunk updates are rejected.

A.3 EXAMPLES OF HARMFUL EDITS REJECTED BY AGBU

  • Across both cases, AGBU separates harmful edits from useful ones, preserving beneficial edits from otherwise rejected updates.The examples concern two trunk-update batches filtered during branch updates.Source: §A.3 · pp. 14–15, Table 5 · p. 16
  • In Case 1, the trunk’s overgeneralized full-cell-text rule is rejected, while the branch retains useful edits and improves from 47/70 to 49/70.The harmful rule loses more validation samples than it gains; the retained edits select the correct column and compute durations.Source: §A.3 · pp. 14–15, Table 5 · p. 16
  • In Case 2, the trunk’s broad name-shortening rule lowers its score from 152/200 to 148/200, while the branch rejects it and improves from 147/200 to 151/200.The branch retains two useful edits for selecting the answer type and formatting paired answers.Source: §A.3 · p. 15, Table 5 · p. 16

A.4 AN EXAMPLE OF EVOLUTION PROCESS

In the SpreadsheetBench example, a branch reaches a higher score than the trunk after later trunk updates are mostly rejected. Across validation results, SkillVine attains the highest score, including ties, in nine of ten model–benchmark combinations.

A.4 AN EXAMPLE OF EVOLUTION PROCESS

  • On SpreadsheetBench, the branch reaches 30/39, while the trunk’s best score is 22/39 after most later trunk updates are rejected.The path uses both discovery and refinement selectors; at R:v0007 and D:v0026, AGBU filters harmful edits while keeping useful ones.Source: §A.4 · p. 16

A.5 VALIDATION RESULTS

  • SkillVine attains the highest validation score, including ties, in nine of ten model–benchmark combinations; the exception is LiveMath with Qwen3.8-Flash.The authors attribute the exception to a small validation set and stochastic agent rollouts, which make the final score more sensitive to individual samples.Source: §A.5 · p. 17, Table 6 · p. 17

Related papers

Loading 2609.32731v1…