Research map / Large language models / Agents, code and math
Embodied CoT Distillation From LLM To Off-the-shelf Agents
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 DeDer, which distills LLM (PaLM) embodied reasoning into a two-tier hierarchy of small LMs (T5 reasoning-policy with embodied KG and contrastive attention, GPT-2 planning-policy) for ALFRED task planning on off-the-shelf devices.
Problem
LLM-based embodied task planners are too computationally heavy for capacity-limited, off-the-shelf devices, and direct end-to-end distillation of LLM planning into small models performs poorly on complex, long-horizon embodied tasks.
Approach
Decompose the policy into a reasoning-policy that produces multiple MDP-featured rationales (distilled from PaLM via MDP-featured retrieval-augmented in-context learning and LLM self-critic filtering) and a planning-policy conditioned on those rationales; the reasoning-policy uses an embodied knowledge graph, prompt pools, causal and gated attention, and a contrastive behavior-based loss for singl
Key results
- DeDer-GPT2 reaches 40.3% SR in the Unseen Environment split versus 17.2% for LLM-planner-PaLM, and 100% SR on Train.
- DeDer achieves 21.6% higher SR than LLM-planner-PaLM on average, while using a model roughly 2700x smaller.
- Ablations show MDP-featured in-context learning with PaLM improves over few-shot by about 6.9 SR points in Unseen Environment (40.3 vs 29.9 in the detailed table); the reasoning-policy with KG and contrastive loss adds further gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeDer instead of the strongest baseline | ALFRED embodied task planning · ALFRED Seen | SR: 66.8{\pm0.0} -> 81.8{\pm0.5} | Helped |
| Use DeDer instead of the strongest baseline | ALFRED embodied task planning · ALFRED Unseen Spatial | SR: 32.7{\pm2.1} -> 52.7{\pm1.0} | Helped |
| Use DeDer instead of the strongest baseline | ALFRED embodied task planning · ALFRED Unseen Environment | SR: 27.6{\pm1.7} -> 40.3{\pm0.9} | Helped |
| Use MDP-featured in-context learning and self-verification instead of fixed few-shot examples (ablation) | DeDer rationale construction with PaLM · ALFRED Unseen Environment | SR: 29.9{\pm1.0} -> 40.3{\pm0.9} | Helped |
| Add contrastive learning (ablation) | DeDer reasoning policy · ALFRED Unseen Environment | SR: 33.9{\pm2.0} -> 40.3{\pm0.9} | Helped |
| Add the embodied knowledge graph (ablation) | DeDer reasoning policy · ALFRED Seen | SR: 71.0{\pm6.3} -> 81.8{\pm0.5} | Helped |
| Add the embodied knowledge graph (ablation) | DeDer reasoning policy · ALFRED Unseen Environment | SR: 35.6{\pm5.6} -> 40.3{\pm0.9} | Helped |
| Add causal attention Ψc (ablation) | DeDer reasoning policy · ALFRED Seen | SR: 75.9{\pm0.3} -> 81.8{\pm0.5} | Helped |
| Add gated attention Ψg (ablation) | DeDer reasoning policy · ALFRED Unseen Environment | SR: 30.5{\pm2.6} -> 40.3{\pm0.9} | Helped |
| Replace causal and gated attention with basic attention (ablation) | DeDer reasoning policy · ALFRED Seen | SR: 81.8{\pm0.5} -> 74.7{\pm0.4} | Hurt |
| Generate rationales iteratively without attention (ablation) | DeDer reasoning policy · ALFRED Seen | SR: 81.8{\pm0.5} -> 76.0{\pm0.7} | Hurt |
| Increase reasoning-policy size from t5-small to t5-large (ablation) | DeDer · ALFRED Unseen Environment | SR: 40.3{\pm0.9} -> 52.5{\pm0.7} | Helped |
Limitations
- Dependence on the pre-trained knowledge embedded in the small LMs.
- Reduced sLM capacity hurts zero-shot performance under significant domain shift.
- Evaluated only on ALFRED with a rule-based low-level executor.
Experimental setup
Datasets: ALFRED expert trajectories (312, generated with PDDL planner)
Benchmarks: ALFRED
Metrics: task success rate goal-conditioned success rate inference time
Baselines: SayCan ZSP LLM-Planner End2End SCoTD SCOTT
Builds on: SayCan LLM-Planner ZSP SCoTD SCOTT End2End
Closest papers by meaning
- Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making (NeurIPS 2024 · similarity 0.91)
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.89)
- Embodied Task Planning via Graph-Informed Action Generation with Large Language Models (ICML 2026 · similarity 0.89)
- From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.88)
- R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner (CVPR 2025 · similarity 0.88)
- True Knowledge Comes from Practice: Aligning Large Language Models with Embodied Environments via Reinforcement Learning (ICLR 2024 · 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.