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

MARS: Multi-Agent Adaptive Reasoning with Socratic Guidance for Automated Prompt Optimization

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 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 MARS, a five-agent framework (Planner, Teacher-Critic-Student Socratic dialogue, Target) that casts automated prompt optimization as a POMDP and is evaluated on 17 BBH/MMLU/C-Eval/GSM8K/LSAT-AR tasks using a single training example per task.

Problem

Automated prompt optimization suffers from rigid meta-prompt templates and inefficient exploration of the prompt space, with generate-and-search methods only doing local search around candidates.

Approach

A Planner generates task-specific sub-goal trajectories, and a Teacher-Critic-Student Socratic dialogue iteratively refines the prompt in a pseudo-gradient fashion, with the Target model's accuracy as reward, all modeled as a POMDP whose hidden state is the Student's reasoning state.

Key results

• On 12 general tasks with DeepSeek-V2.5, MARS reaches 85.11% average accuracy versus 79.07% for the best baseline (OPRO).
• On domain-specific tasks (Chinese, math, law), MARS averages 75.81% versus 69.39% for PE2.
• With GPT-4o as the base model, MARS averages 76.58% across five task groups versus 73.08% for PE2.

Surprising findings

• Using only one training example per task, MARS outperforms baselines that use 20 to 100 training examples.
• Even with zero training examples, MARS (90.76 on Boolean Expressions, 0-shot row) already exceeds all listed baselines on several tasks.

Limitations

• Reported results rely on a single primary backbone (DeepSeek-V2.5) and a single run with no variance reported.
• Proposition-based theoretical claims rest on assumed quantities (positive expected advantage, bounded variance, Lipschitz reward) that are not verified empirically.

Experimental setup

Datasets: BIG-Bench Hard MMLU C-Eval GSM8K LSAT-AR

Benchmarks: BIG-Bench Hard MMLU C-Eval GSM8K LSAT-AR

Metrics: accuracy

Baselines: CoT (zero-shot) CoT (few-shot) APE ProTeGi OPRO PE2

Builds on: APE ProTeGi OPRO PE2

Closest papers by meaning

• AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems (NeurIPS 2026 · similarity 0.90)
• A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.90)
• How to Auto-optimize Prompts for Domain Tasks? Adaptive Prompting and Reasoning through Evolutionary Domain Knowledge Adaptation (NeurIPS 2025 · similarity 0.89)
• SePO: Self-Evolving Prompt Agent for System Prompt Optimization (NeurIPS 2026 · similarity 0.88)
• Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models (ICML 2024 · similarity 0.88)
• Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.88)