Research map / Multi-agent systems and games / LLM agents
AgentOdyssey: Open-Ended Long-Horizon Text Game Generation for Test-Time Continual Learning Agents
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.
Introduces AgentOdyssey, a procedurally generated open-ended long-horizon text game framework with multifaceted diagnostics for evaluating test-time continual learning agents across memory, exploration, knowledge, skills and planning.
Problem
Existing benchmarks assume no learning at test time and lack scalable, uncontaminated, long-horizon environments to evaluate continual learning agents that acquire knowledge and skills during deployment.
Approach
An LLM-driven program-synthesis game generator (Aider-based entity, rule, and quest generators with automated run-time testing) produces open-ended text RPG games with world graphs, action/step rules, and main/side quests, paired with diagnostic QA and exploration/action-diversity metrics.
Key results
- Long Context agent with GPT-5 achieves the highest main quest reward (3) among agents, while human reaches 9 in the same game; Claude-Opus-4.6 is the best frontier model in additional tests but still below humans.
- SFT agent with Qwen3-4B and short-term memory reaches main quest reward 7 in Experiment 2 game, the best in that group (vs. 0 for SFT baseline).
- Long Context agent with GPT-5 achieves 34.8% increase in World Knowledge QA accuracy after gameplay, while its token cost grows quadratically with steps.
Surprising findings
- Reflection and summarization do not help reasoning-model agents and in several cases hurt SFT and long-context agents.
- Long Context agents with Qwen3-4B collapse to zero world knowledge and episodic memory accuracy after gameplay, converging on a single action.
Limitations
- Environment supports only text observations and a single agent.
- Turn-based design assigns fixed duration per action, simplifying temporal dynamics.
- Qwen3-4B is insufficiently capable for the main game, requiring a simpler game for test-time training experiments.
Experimental setup
Datasets: procedurally generated AgentOdyssey games
Benchmarks: AgentOdyssey
Metrics: main quest reward normalized combined reward world knowledge QA accuracy episodic memory QA accuracy object exploration action exploration action diversity token cost invalid-action rate
Baselines: Vanilla RAG Mem0 Raptor Voyager MEM1 MemoryLLM MPlus Random agent No Memory agent
Builds on: ReAct Aider PPO verl-agent ALFWorld
Closest papers by meaning
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.88)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.88)
- Learning to Continually Learn via Meta-learning Agentic Memory Designs (NeurIPS 2026 · similarity 0.87)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.87)
- AdaMEM: Test-Time Adaptive Memory for Language Agents (ICML 2026 · similarity 0.87)
- Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks (ICML 2026 · 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.