Research map / Multi-agent systems and games / LLM agents
EvoC2F: Compiling Tool Orchestration for Efficient and Evolvable LLM Agents
Lei Wei, Qi Liu, Ruiyang Huang, Xiao Peng, Sicong Xie, Lanbo Lin, Chenhao Jiang, Yuanwu Xu, Tianyuan Yang, Jiayao Liu, Li Cai, Zhaolu Kang and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EvoC2F, a method that compiles LLM agent tool orchestration into constrained, evolvable structures to balance flexibility, parallelism, and reliability.
Problem
Tool-augmented LLM agents must balance planning flexibility with the reliability needed for production; existing approaches either run tools sequentially without parallelism or generate unconstrained code, which hinders optimization and verification.
Approach
Compiles tool orchestration into a constrained representation that supports parallel execution and verification, and evolves it from experience; the abstract is truncated so the precise mechanism is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic plan compiler (parallel scheduling) (ablation) | EvoC2F · StableToolBench | SoPR / latency: 64.9 -> 72.2 SoPR; 11.2s -> 5.5s | Improved |
| add effect annotations (ablation) | EvoC2F · StableToolBench | SoPR: 64.8 -> 72.2 | Improved |
| add verification gates for skill admission (ablation) | EvoC2F · StableToolBench | SoPR / regression rate: SoPR 70.5 -> 72.2; regression rate 7.2% -> 0.8% | Improved |
| add macro-skills (ablation) | EvoC2F · StableToolBench | SoPR / tool calls: 66.1 -> 72.2 SoPR; 21.5 -> 16.8 tool calls | Improved |
| add learned skill retrieval router (ablation) | EvoC2F · StableToolBench | SoPR: 70.8 -> 72.2 | Improved |
| EvoC2F vs ReAct on latency | Claude-4-Sonnet · StableToolBench | latency (s): 16.9 -> 5.5 | Improved |
| grow skill library over sequential tasks | EvoC2F · ToolSeq-500 | success rate: 68.2 -> 75.0 | Improved |
| use composed skills (>=2) in a task | EvoC2F · ToolSeq-500 | success rate: 70.6 -> 78.4 | Improved |
| increase compiler weight lambda toward reliability (ablation) | EvoC2F · StableToolBench validation | SR vs latency: SR 68.5 -> 74.6; latency 5.1 -> 7.8 | Mixed |
| transfer learned skills zero-shot to ShortcutsBench | EvoC2F · ShortcutsBench | ASA: 46.8 -> 49.3 | Improved |
| add safe fallback disabling harmful skills (ablation) | EvoC2F transfer · ShortcutsBench | ASA: 49.3 -> 52.5 | Improved |
Closest papers by meaning
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.90)
- EvoMAS: Evolutionary Generation of Multi-Agent Systems (ICML 2026 · similarity 0.89)
- PACE: Two-Timescale Self-Evolution for Small Language Model Agents (NeurIPS 2026 · similarity 0.89)
- CATArena: Evaluating Evolutionary Capabilities of Code Agents via Iterative Tournaments (ICML 2026 · similarity 0.88)
- AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.88)
- AutoTool: Efficient Tool Selection for Large Language Model Agents (AAAI 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
