List your company

Text-based agent interaction on TextWorld Treasure

81AccuracyAgent-BRACE
65.469.4573.577.55May 12, 2026

Evaluation results

MethodLinks
Agent-BRACEBackbone=Qwen3-4B-Instruct, Independent runs=32026.058129.9
ReAct (RL)Backbone=Qwen3-4B-Instruct, Independent runs=32026.057416.5
Direct-Action (RL)Backbone=Qwen3-4B-Instruct, Independent runs=32026.0572.528.1
PABUBackbone=Qwen3-4B-Instruct, Independent runs=32026.0570.737.8
ReActBackbone=Qwen3-4B-Instruct, Independent runs=32026.0569.510
Base ModelBackbone=Qwen3-4B-Instruct, Independent runs=32026.056629.9