AI Paper Atlas

Research map / Large language models / Agents, code and math

Large Language Models as Commonsense Knowledge for Large-Scale Task Planning

Zirui Zhao, Wee Sun Lee, David Hsu

NeurIPS 2023 · Large language models › Agents, code and math · approach: LLM planning with symbolic tools · 40 citations (OpenAlex)

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 changedWhereMeasuredEffect
add LLM-induced commonsense world model and heuristic policy with MCTS vs UCT aloneMCTS · VirtualHome Seen Simplesuccess rate (%): 0.0 -> 91.4Improved
use LLM as heuristic in MCTS rather than as policy onlyGPT3.5 Policy · VirtualHome Seen Comp.success rate (%): 47.0 -> 71.2Improved
use LLM as heuristic in MCTS rather than as policy onlyGPT3.5 Policy · VirtualHome Seen NovelComp.(3)success rate (%): 5.4 -> 33.6Improved
GPT3.5-MCTS vs GPT3.5 policy on unseen apartmentGPT3.5 Policy · VirtualHome Unseen NovelComp.(3)success rate (%): 6.9 -> 38.8Improved
GPT3.5-MCTS vs GPT3.5 policy on NovelComp.(2) unseenGPT3.5 Policy · VirtualHome Unseen NovelComp.(2)success rate (%): 54.0 -> 70.4Improved
add LLM heuristic policy (PUCT) to MCTS (ablation)GPT3.5-MCTS · VirtualHome Seen Simplesuccess rate (%): 0.0 -> 91.4Improved
use LLM-generated commonsense state prior instead of uniform prior (ablation)GPT3.5-MCTS · VirtualHome Seen Simplesuccess rate (%): 3.2 -> 91.4Improved
use LLM-generated commonsense state prior instead of uniform prior (ablation)GPT3.5-MCTS · VirtualHome Seen Comp.success rate (%): 0.0 -> 71.2Improved
give full observability instead of LLM commonsense belief (ablation)GPT3.5-MCTS · VirtualHome Seen Simplesuccess rate (%): 91.4 -> 94.0Improved
give full observability instead of LLM commonsense belief (ablation)GPT3.5-MCTS · VirtualHome Seen Comp.success rate (%): 71.2 -> 80.7Improved
fine-tune GPT-2 policy vs few-shot GPT-3.5 policypolicy baseline · VirtualHome Seen NovelSimplesuccess rate (%): 74.3 -> 41.2Hurt
use LLM as world model with search instead of LLM as policy for flight planningGPT-3.5 flight path prediction · Flight route planningpath accuracyImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy