Research map / Large language models / Agents, code and math
REI-Bench: Can Embodied Agents Understand Vague Human Instructions in Task Planning?
Chenxi Jiang, Chuhao Zhou, Jianfei Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-turn standard dialogue context to the instruction | SayCan · REI-Bench (Explicit REs) | success rate (%): 57.7 -> 46.90 | Hurt |
| increase implicit referring expressions (Explicit to Mixed to Implicit) (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench | success rate (%): -16.8% / -15.5% / -21.5% at Mixed, further -8.0% / -8.5% / -8.0% at Implicit (Standard/Noised/Short) | Hurt |
| increase implicit REs: object omission rate rises (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench (Standard Context) | object omission rate (%): 22.6 -> 53.9 | Hurt |
| add naming-ambiguity noise and short context (ablation) | SayCan · REI-Bench | success rate (%) | No clear effect |
| add TOCC (decoupled RE resolution before planning) (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench (Implicit REs, Standard Context) | overall error rate (%): 77.9 -> 70.7 | Improved |
| add TOCC (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench (Implicit REs, Standard Context) | object omission rate (%): 53.9 -> 40.1 | Improved |
| add aware prompt (AP) (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench (Mixed REs, Standard Context) | overall error rate (%): 69.9 -> 71.0 (object omission 38.8 -> 31.3) | Mixed |
| add chain-of-thought prompting (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench (Implicit REs, Standard Context) | overall error rate (%): 77.9 -> 77.9 | No clear effect |
| add 2-shot in-context learning (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench (Explicit REs, Standard Context) | overall error rate (%): 53.1 -> 60.8 | Hurt |
| provide dialogue context to TOCC (vs instruction only) (ablation) | SayCan + TOCC (LLaMA3.1-8B) · REI-Bench (Implicit REs, Standard Context) | overall error rate (%): 90.6 -> 70.7 | Improved |
| use TOCC rather than CoT for token cost (ablation) | SayCan (LLaMA3.1-8B) · REI-Bench | avg total tokens: 3613 -> 1956 | Improved |
| add TOCC vs no prompting method (ablation) | SayCan (DeepSeekMath-7B) · REI-Bench | success rate | Hurt |
Closest papers by meaning
- LoTa-Bench: Benchmarking Language-oriented Task Planners for Embodied Agents (ICLR 2024 · similarity 0.89)
- PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks (ICLR 2025 · similarity 0.88)
- OntoPlan: An Ontology-Grounded Scene Representation and Agentic Framework for Scalable Robot Task Planning (NeurIPS 2026 · similarity 0.88)
- R2C: Mapping Room to Chessboard to Unlock LLM As Low-Level Action Planner (CVPR 2025 · similarity 0.88)
- Towards Reliable LLM-based Robots Planning via Combined Uncertainty Estimation (NeurIPS 2025 · similarity 0.87)
- TaskBench: Benchmarking Large Language Models for Task Automation (NeurIPS 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.
