Research map / Large language models / Agents, code and math
NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains
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 NeSyC, a neuro-symbolic continual learner that uses LLM hypothesis generation with ASP-based contrastive validation and memory-based error monitoring to generalize actionable knowledge for embodied agents in open-domain settings.
Problem
Embodied agents relying on fixed symbolic or LLM knowledge generalize poorly to open-domain environments with incomplete, inconsistent, or changing action preconditions and effects.
Approach
Combines LLM-generated ILP-style hypotheses (with subsumption chain-of-thought) and contrastive TPR/FPR scoring validated by an ASP solver (clingo) in a rule-reformulation loop, and re-enters that loop when a memory-based error handler detects action failures during planning.
Key results
- NeSyC reaches 70.7% SR on ALFWorld High Dynamic versus 30.9% for AutoGen, the strongest multi-agent baseline.
- In the Static setting CLMASP outperforms NeSyC on ALFWorld (88.5 vs 82.9 SR) because it uses expert-level knowledge; NeSyC stays robust as dynamics increase.
- NeSyC reports 91.1% SR on Minecraft Low Dynamic and 81.5% SR on RLBench Low Dynamic, while CLMASP drops to 0% SR on RLBench Low Dynamic.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeSyC rather than the strongest baseline CLMASP | embodied task planning · ALFWorld Low Dynamic and High Dynamic | SR (%): 58.8{\pm 4.4} -> 78.9{\pm 3.7}; 23.8{\pm 3.9} -> 70.7{\pm 4.1} | Helped |
| Use NeSyC rather than CLMASP | embodied task planning · VirtualHome Low Dynamic and High Dynamic | SR (%): 28.9{\pm 0.4} -> 79.6{\pm1.9}; 0.0{\pm 0.0} -> 77.5{\pm1.3} | Helped |
| Use NeSyC rather than CLMASP | embodied task planning · Minecraft Low Dynamic and High Dynamic | SR (%): 52.2{\pm5.7} -> 91.1{\pm1.6}; 48.9{\pm3.1} -> 87.8{\pm5.7} | Helped |
| Use NeSyC rather than CLMASP | embodied task planning · RLBench Low Dynamic and High Dynamic | SR (%): 0.0{\pm0.0} -> 81.5{\pm4.5}; 0.0{\pm0.0} -> 79.0{\pm6.6} | Helped |
| Replace the subsumption CoT prompt with a simple LLM prompt (ablation) | NeSyC · ALFWorld Static, Low Dynamic, and High Dynamic | SR (%): 82.9{\pm 3.4} -> 41.5{\pm 4.4}; 78.9{\pm 3.7} -> 39.0{\pm 4.4}; 70.7{\pm 4.1} -> 36.6{\pm4.3} | Hurt |
| Replace the ASP solver with LLM symbolic CoT planning (ablation) | NeSyC task planner · ALFWorld | SR (%): 82.9{\pm 3.4} -> 71.5{\pm 4.1}; 78.9{\pm 3.7} -> 60.2{\pm4.4}; 70.7{\pm 4.1} -> 27.6{\pm4.0} | Hurt |
| Replace the ASP solver with standard CoT planning (ablation) | NeSyC task planner · ALFWorld | SR (%): 82.9{\pm 3.4} -> 69.1{\pm 4.2}; 78.9{\pm 3.7} -> 58.5{\pm 4.4}; 70.7{\pm 4.1} -> 24.4{\pm3.9} | Hurt |
| Replace ASP planning with naive LLM prompting (ablation) | NeSyC task planner · ALFWorld | SR (%): 82.9{\pm 3.4} -> 48.8{\pm 4.5}; 78.9{\pm 3.7} -> 35.8{\pm 4.3}; 70.7{\pm 4.1} -> 17.1{\pm 3.4} | Hurt |
| Replace both the subsumption CoT prompt and ASP solver with simple LLM prompting (ablation) | NeSyC · ALFWorld | SR (%): 82.9{\pm 3.4} -> 25.2 {\pm 3.9}; 78.9{\pm 3.7} -> 9.8{\pm 2.7}; 70.7{\pm 4.1} -> 3.3{\pm1.6} | Hurt |
Surprising findings
- CLMASP, the neuro-symbolic baseline, falls to 0% SR on RLBench and VirtualHome High Dynamic, below LLM-based baselines.
- Replacing the ASP solver with LLM reasoning (SymbCoT or CoT) improves GC and Step in some settings but lowers SR.
Limitations
- Performance degrades or fails to improve with smaller LLMs such as Llama-3-8B due to rule conflicts during reformulation.
- Relies on a small set of expert-level episodic experiences (about 5-15 per environment) plus ChatGPT-generated data, and LLM-generated guidance can add errors.
Experimental setup
Datasets: ALFWorld VirtualHome Minecraft (PDDLGym) RLBench
Benchmarks: ALFWorld VirtualHome Minecraft RLBench real-world desk rearrangement
Metrics: SR GC Step
Baselines: LLM-planner ReAct Reflexion AutoGen ProgPrompt CLMASP
Builds on: inductive logic programming answer set programming CLMASP SymbCoT VoxPoser
Closest papers by meaning
- Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning (NeurIPS 2025 · similarity 0.90)
- NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning (NeurIPS 2025 · similarity 0.90)
- VisualPredicator: Learning Abstract World Models with Neuro-Symbolic Predicates for Robot Planning (ICLR 2025 · similarity 0.88)
- Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks (ICML 2026 · similarity 0.88)
- R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner (CVPR 2025 · similarity 0.88)
- Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models (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.