Research map / Multi-agent systems and games / LLM agents

EvoEmpirBench: Dynamic Spatial Reasoning with Agent-ExpVer

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration

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 EvoEmpirBench, two dynamic partially observable games (maze navigation and match-2 elimination), and Agent-ExpVer, a three-agent online learning framework that distills verified experiences into prompt-level "truths" without weight updates.

Problem

Existing spatial and reasoning benchmarks are static or globally observable, so they do not test long-horizon reasoning, memory use, and adaptive planning under partial observability and environment dynamics.

Approach

Two interactive games where actions alter the environment, plus a parameter-free loop in which an LLM agent summarizes episode trajectories into subjective experiences, verifies them by replaying levels, and stores improving ones as truths that are merged or pruned and injected into the prompt.

Key results

• Agent-ExpVer raises GPT-4.1 maze success from 73.33% to 78.89% and match-2 success from 40.00% to 53.33%.
• Qwen2.5-32B-instruct with Agent-ExpVer improves maze success from 42.22% to 54.44% and match-2 success from 33.33% to 41.11%.
• Human baseline reaches 90.00% maze success and 86.67% match-2 success, above all evaluated LLMs.

Measured effects reported in the paper

Surprising findings

• Removing the TruthWeaver management agent sometimes yields higher maze success than the full framework for GPT-4.1 (77.78% vs 78.89% is close, but W/o.TW is higher on match-2 score at 219.53 vs 234.60 is lower; for Qwen2.5-32B, W/o.TW maze success 51.11% vs full 54.44%).
• Early rounds of truth induction temporarily reduced success rate (e.g., match-2 from 6.67% to 0%) before improving.

Limitations

• Performance remains tied to model capacity; smaller models lag and top systems fall short of humans.
• Small benchmark size (90 instances per game) limits statistical power.
• Agent-ExpVer gains are inconsistent across metrics, e.g., some average scores drop while success rises.

Experimental setup

Datasets: EvoEmpirBench maze navigation EvoEmpirBench match-2 elimination

Benchmarks: EvoEmpirBench maze navigation EvoEmpirBench match-2 elimination

Metrics: success rate average score average steps exploration remaining HP remaining/max steps ratio score per step clear per step API efficiency

Baselines: GPT-4 GPT-4.1 Claude-3-7-sonnet Gemini-2.5-flash-preview Grok-3 Deepseek-V3 Qwen2.5-32B-instruct Qwen3-30B-a3b Llama-3.1-70B-instruct

Builds on: Reflexion Agent-Pro AutoManual CLIN Voyager

Closest papers by meaning

• Active Exploring like a Pigeon: Reinforcing Spatial Reasoning via Agentic Vision-Language Models (ICML 2026 · similarity 0.90)
• SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning (NeurIPS 2026 · similarity 0.88)
• Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.88)
• MindJourney: Test-Time Scaling with World Models for Spatial Reasoning (NeurIPS 2025 · similarity 0.87)
• AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models (NeurIPS 2026 · similarity 0.87)
• Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks (ICML 2026 · similarity 0.87)