Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Learning Dynamic Belief Graphs to Generalize on Text-Based Games
Ashutosh Adhikari, Xingdi Yuan, Marc-Alexandre Côté, Mikuláš Zelinka, Marc-Antoine Rondeau, Romain Laroche, Pascal Poupart, Jian Tang, Adam Trischler, Will Hamilton
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 text observations as input alongside the OG belief graph (ablation) | GATA (OG) · TextWorld, avg across levels | relative improvement (%): +8.0 -> +20.4 | Improved |
| add text observations as input alongside the COC belief graph (ablation) | GATA (COC) · TextWorld, avg across levels | relative improvement (%): +20.2 -> +24.2 | Improved |
| add text observations to ground-truth-pretrained discrete graph agent (ablation) | GATA-GTP · TextWorld, avg across levels | relative improvement (%): -1.2 -> +14.9 | Improved |
| add learned belief graph (COC pre-trained) to text-only Tr-DQN | Tr-DQN · TextWorld, avg across levels | relative improvement (%): +20.2 (COC, graph only) vs Tr-DQN baseline | Improved |
| use COC instead of OG self-supervised pre-training for graph updater (graph only) (ablation) | GATA · TextWorld, avg across levels | relative improvement (%): +8.0 -> +20.2 | Improved |
| use COC instead of OG pre-training (with text obs) (ablation) | GATA · TextWorld, avg across levels | relative improvement (%): +20.4 -> +24.2 | Improved |
| add recurrent memory (DRQN) to transformer DQN | Tr-DQN · TextWorld, avg across levels | relative improvement (%): +3.9 avg; 20 games +10.3, 100 games -2.6 | Mixed |
| add episodic counting bonus to Tr-DRQN | Tr-DRQN · TextWorld, avg across levels | relative improvement (%): +3.9 -> +3.6 | No clear effect |
| replace learned continuous belief graph with ground-truth-pretrained discrete graph updater | GATA · TextWorld, avg across levels | relative improvement (%): +24.2 -> +14.9 (both with text obs) | Hurt |
| give agent the ground-truth full graph instead of an inferred belief graph | GATA · TextWorld, avg across levels | relative improvement (%): +24.2 -> +81.6 | Improved |
| train on 100 instead of 20 games | GATA (OG, graph only) · TextWorld, relative improvement | relative improvement (%): -0.2 -> +16.1 | Improved |
Closest papers by meaning
- Monte Carlo Planning with Large Language Model for Text-Based Game Agents (ICLR 2025 · similarity 0.87)
- Strategist: Self-improvement of LLM Decision Making via Bi-Level Tree Search (ICLR 2025 · similarity 0.87)
- Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision (NeurIPS 2025 · similarity 0.87)
- Can Graph Learning Improve Planning in LLM-based Agents? (NeurIPS 2024 · similarity 0.87)
- BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games (ICLR 2025 · similarity 0.86)
- AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents (NeurIPS 2026 · similarity 0.86)
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.
