Research map / Multi-agent systems and games / LLM agents
Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks
Read the paperCheck your idea against 108,967 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 Neuro-Symbolic Skill Induction (NSI), which lifts agent interaction traces into modular logic-grounded programs with control flow and variable binding, refined offline and online via reflective planning.
Problem
Prompting-based LLM agents struggle with long-horizon planning, and existing programmatic skill induction produces state-blind linear scripts that cannot encode conditional logic needed in dynamic environments.
Approach
Skills are represented as graph-structured programs with neural perception (LLM maps observations to predicate facts) and symbolic execution (CheckOp, DataOp, LoopOp, PrimitiveOp nodes), induced by intra-trajectory consolidation and inter-trajectory merging with structural operators (branching, crossover, variable lifting, loop folding), then grown online by grafting recovery trajectories onto fai
Key results
- NSI reaches 98.0% success on ALFWorld vs 91.3% for the strongest baseline AWM.
- NSI achieves 76.5% score and 44.5% success on WebShop, versus 51.8 score for AWM offline.
- On TextCraft, NSI reaches 95.2% success vs 92.5% for AWM; the ablation without online honing gets 78.5%.
Surprising findings
- ASI, which induces parameterized programmatic skills, performs far below text-based AWM (e.g., 70.6% vs 91.3% on ALFWorld), which the authors attribute to an expressiveness gap in linear scripts.
Limitations
- Experiments rely solely on GPT-4o as backbone; generalization to other LLMs is not shown.
- Full prompts and code are deferred to a future release, limiting reproducibility at present.
- Empirical consistency is checked against recorded traces rather than online environment verification, which may not capture all state variations.
Experimental setup
Datasets: ALFWorld WebShop TextCraft ALFRED
Benchmarks: ALFWorld WebShop TextCraft
Metrics: success rate reward score average steps per skill invocation survival analysis over horizon length
Baselines: ReAct Reflexion ADaPT StateAct AWM ASI
Builds on: ASI AWM ReAct Reflexion options framework
Closest papers by meaning
- Efficient Skill Grounding via Code Refactoring with Small Language Models (ICML 2026 · similarity 0.89)
- NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains (ICLR 2025 · similarity 0.88)
- Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning (NeurIPS 2025 · similarity 0.88)
- Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents (ICML 2026 · similarity 0.87)
- Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.87)
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills (NeurIPS 2026 · similarity 0.87)
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.