Research map / Large language models / Agents, code and math
NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning
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 NeSyPr, which compiles symbolic planner plans into a vector-quantized procedural memory inside an LM so that embodied reasoning runs in single-step inference without online symbolic tools, evaluated on PDDLGym, VirtualHome, and ALFWorld.
Problem
LM-based embodied agents need structured, adaptive, and timely reasoning in dynamic environments, but online symbolic planners and large reasoning models are too slow or resource-heavy for latency- and connectivity-constrained settings.
Approach
Symbolic planner plans are encoded into a vector-quantized procedure-book inside a memory-augmented LM decoder, so composed procedures are used at inference; test-time contrastive planning uses success and failure procedure banks with contrastive decoding to adapt without symbolic tools.
Key results
- On PDDLGym Minecraft test, NeSyPr reaches 65.2% CSR versus 53.3% for LongMem with 1.3B parameters.
- On the ALFWorld unseen split, NeSyPr reaches 59.7% CSR, compared with 49.1% for Optimus-2.
- In the efficiency analysis with LLaMA3.1-8B, NeSyPr has about 5.2 s latency versus 59.5 s for BoT and 111.0 s for DeepSeek-R1-8B, with 89.0% CSR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply NeSyPr instead of LongMem | PDDLGym Minecraft test · PDDLGym Minecraft test | CSR: 65.2 -> 53.3 | Helped |
| apply NeSyPr instead of LongMem | PDDLGym Rearrangement test · PDDLGym Rearrangement test | CSR: 73.5 -> 58.2 | Helped |
| use NeSyPr rather than the strongest listed baseline | VirtualHome and ALFWorld · VirtualHome and ALFWorld train, seen, and unseen sets | CSR: 12.5%, 9.6%, and 12.4%; 9.8%, 8.3%, and 10.6% | Helped |
| use NeSyPr instead of BoT with the same LLaMA 3.1-8B backbone | Minecraft continual planning · Minecraft continual planning | latency: 54.3 seconds longer | Helped |
| add EMA updates to the procedure-book (ablation) | NeSyPr · Minecraft | CSR: 2.2% | Helped |
| add the procedure-book (ablation) | NeSyPr · Minecraft | CSR: 17.3% | Helped |
| use both successful and failure procedures rather than successful procedures only (ablation) | NeSyPr · Minecraft | CSR: 2.0% | Helped |
| use both successful and failure procedures rather than failure procedures only (ablation) | NeSyPr · Minecraft | CSR: 1.9% | Helped |
| enable contrastive planning (ablation) | NeSyPr · Minecraft | CSR: 5.3% | Helped |
| enable contrastive planning rather than following positive procedures alone (ablation) | NeSyPr · Minecraft | CSR: 4.3% | Helped |
| enable contrastive planning rather than following failure procedures alone (ablation) | NeSyPr · Minecraft | CSR: 5.5% | Helped |
Surprising findings
- NeSyPr with a 3B LM (83.6% CSR) outperforms PlaSma with an 8B LM (80.5% CSR) in the efficiency comparison.
Limitations
- Performance partially depends on the pretrained knowledge of the underlying LM, as the authors acknowledge.
- Requires symbolic planner-generated plans for training and symbolic observation interfaces (PDDL), limiting direct applicability to raw sensory inputs.
Experimental setup
Datasets: PDDLGym (Minecraft, Rearrangement, GlibRearrangement) VirtualHome ALFWorld
Benchmarks: PDDLGym VirtualHome ALFWorld
Metrics: CSR CGC Executability SPL latency FWT BWT FR RR
Baselines: ZSP RAP LLM-Planner CoT ToT GoT ReAct Reflexion LongMem LM2 BUTLER Optimus-2
Builds on: Adaptive Control of Thought (ACT) BUTLER LM2 LongMem Optimus-2 BoT
Closest papers by meaning
- NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains (ICLR 2025 · similarity 0.90)
- Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning (NeurIPS 2025 · similarity 0.89)
- Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models (NeurIPS 2026 · similarity 0.88)
- R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner (CVPR 2025 · similarity 0.88)
- OmniEVA: Embodied Versatile Planner via Task-Adaptive 3D-Grounded and Embodiment-aware Reasoning (ICLR 2026 · similarity 0.88)
- Embodied CoT Distillation From LLM To Off-the-shelf Agents (ICML 2024 · 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.