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

MAPS: Multi-Agent Personality Shaping for Collaborative Reasoning

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

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.

Proposes MAPS, a multi-agent multimodal reasoning framework with five Big-Five-personality agents (four solver stages plus a Socratic Critic that triggers re-runs), built on GPT-4o and evaluated on MathVista, OlympiadBench and EMMA.

Problem

Existing multi-agent collaborative reasoning suffers from homogeneous agent behaviors and lacks reflective, iterative revision of intermediate steps on complex multimodal scientific problems.

Approach

Assigns Big Five personality traits to role-specialized agents (Interpreter, Aligner, Scholar, Solver) to diversify reasoning styles, and adds a Socratic Critic that scores each stage (0-5) and reruns the weakest stage when scores fall below a threshold.

Key results

• MAPS (GPT-4o base) reaches 31.14% average on the OlympiadBench subtasks and 58.00% on EMMA avg, versus 21.47% for GPT-4o direct on OlympiadBench.
• Reported overall average of 56.31% across 10 subtasks, exceeding the human expert average of 52.73%.
• Ablation on OlympiadBench: removing the Interpreter causes the largest drop (16.09 points), removing the Critic the smallest (7.05 points).

Surprising findings

• Removing the Critic causes the smallest performance loss among the ablated modules, despite being the component meant to provide reflection.
• The paper's Proposition 1 (monotonic free-energy descent) and Proposition 2 are argued informally, with no empirical verification that the Critic's scores track mutual information.

Limitations

• Critic scores are LLM self-assessments without gold labels, so reliability of stage-level confidence is not validated.
• Theoretical propositions rely on loose assumptions (e.g., f_k implicitly estimating mutual information) and are not rigorous.
• Evaluated mainly with GPT-4o; personality embeddings described as vectors but implemented via prompts.

Experimental setup

Datasets: MathVista OlympiadBench EMMA DiagramQG

Benchmarks: MathVista OlympiadBench EMMA DiagramQG

Metrics: accuracy performance gain over baseline

Baselines: Claude 3.5 Sonnet Gemini 2.0 Flash GPT-4o Qwen2.5-VL-72B InternVL2.5-8B-MPO LLaVA-Onevision-72B Human Expert Random Choice

Builds on: chain-of-thought Big Five personality theory Socratic questioning

Closest papers by meaning

• Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
• AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors (ICLR 2024 · similarity 0.86)
• Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation (ICLR 2026 · similarity 0.86)
• AgentCDM: Enhancing Multi-Agent Collaborative Decision-Making via ACH-Inspired Structured Reasoning (AAAI 2026 · similarity 0.86)
• Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.86)
• Group Perspective Matters: Regulating Debate Relationships Can Mitigate Blind Conformity in Multi-Agent Debate (NeurIPS 2026 · similarity 0.86)