Research map / Multi-agent systems and games / LLM agents

Embodied Task Planning via Graph-Informed Action Generation with Large Language Models

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

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 GiG, a Graph-in-Graph memory for LLM embodied planning that pairs a GNN-encoded scene-graph embedding memory with an outer state-transition graph, retrieval of past successful traces, and a bounded 1-step lookahead, reporting large Pass@1 gains on Robotouille and ALFWorld.

Problem

LLM planners for long-horizon embodied tasks lose strategy coherence due to context window limits and hallucinate state transitions that violate environment constraints; tree-based memories also block parallel sub-goals.

Approach

A two-tier graph memory: a GAT-encoded scene graph per step as node embeddings in an outer state-transition graph, enabling similarity-based retrieval of past successful sub-trajectories, loop detection, and a bounded 1-step lookahead using known PDDL transition logic injected into the prompt.

Key results

• Pass@1 of 93% (GiG) vs 71% for ReCAP on Robotouille Synchronous with Qwen3-235B; GiG+Exp reaches 97%.
• On Robotouille Asynchronous with Qwen3-235B, GiG+Exp reaches 82% vs 35% for ReCAP.
• On ALFWorld GiG reaches 97% Pass@1 with Qwen3-235B and DeepSeek-R1 vs 89% for ReCAP and 91% for ExpeL.

Surprising findings

• Experience memory from a large model improves smaller models (Qwen3-30B 27% to 42%) without fine-tuning.
• Ablation shows experience retrieval alone (95%) outperforms bounded lookahead alone (80%).

Limitations

• Zero-shot out-of-domain transfer to unseen task distributions is not addressed.
• Inference latency of LLM reasoning limits real-time use in dynamic environments.
• Bounded lookahead requires known transition logic; not applicable where dynamics are unknown.

Experimental setup

Datasets: Robotouille Synchronous Robotouille Asynchronous ALFWorld

Benchmarks: Robotouille Synchronous Robotouille Asynchronous ALFWorld

Metrics: Pass@1 steps FLOPs token consumption

Baselines: ReCAP ReAct ExpeL CoT

Builds on: ReAct ReCAP ExpeL Graph Attention Network

Closest papers by meaning

• R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner (CVPR 2025 · similarity 0.90)
• From Assumptions to Actions: Turning LLM Reasoning into Uncertainty-Aware Planning for Embodied Agents (ICLR 2026 · similarity 0.90)
• Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.89)
• Graph-enhanced Large Language Models in Asynchronous Plan Reasoning (ICML 2024 · similarity 0.89)
• SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning (NeurIPS 2026 · similarity 0.89)
• Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making (NeurIPS 2024 · similarity 0.89)