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

Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning

ICLR 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 HILA, a multi-agent framework where LLM agents learn via GRPO with cost-aware rewards when to defer to a human (proxy GPT-4o-mini/GPT-4o) and use expert demonstrations for SFT continual learning (DLPO).

Problem

Purely autonomous multi-agent LLM systems are closed-world and fail on tasks needing knowledge beyond their training; existing human-in-the-loop systems use heuristic deferral and treat expert feedback as one-time fixes.

Approach

A metacognitive policy (EVAL/CREATE/DEFER) trained with GRPO and cost-aware rewards decides when to defer; an outer loop converts expert demonstrations from deferrals into SFT data so the backbone's reasoning improves over time (Dual-Loop Policy Optimization).

Key results

• On Llama3-8B, HILA reaches 89.86 on GSM8K, 35.83 on AMC, 9.37 on AIME, 72.15 on HumanEval, 73.62 on MMLU, outperforming the best autonomous baseline on each benchmark.
• On GSM8K across four backbones, HILA achieves 94.72 (Qwen2.5-7B), 91.17 (Qwen2.5-3B), 89.86 (Llama3-8B), 83.85 (Llama3-3B), the best in each column.
• DLPO-updated Llama3-8B backbone improves vanilla GSM8K from 72.76 to 82.11 without deferral at inference time.

Surprising findings

• Backbone updated with DLPO transfers gains to other inference frameworks (e.g., DyLAN and Debate) that do not use deferral, suggesting the expert demonstrations improve the base model itself.
• Deferral rate drops substantially after DLPO (e.g., MMLU defer 19% to 5%), while accuracy rises.

Limitations

• Human expert is simulated by GPT-4o-mini/GPT-4o rather than real humans.
• Gains depend on expert quality; stronger proxies give better results.
• Performance peaks at intermediate rounds and cost rises sharply with number of agents.

Experimental setup

Datasets: GSM8K MMLU HumanEval AIME AMC

Benchmarks: GSM8K AMC AIME HumanEval MMLU

Metrics: accuracy solve rate action distribution

Baselines: Vanilla CoT Self-Consistency PHP LLM-Debate G-Debate DyLAN GPTSwarm AgentPrune AFlow

Builds on: GRPO LLM-Debate DyLAN GPTSwarm AFlow

Closest papers by meaning

• Learning to Deliberate: Meta-policy Collaboration for Agentic LLMs with Multi-agent Reinforcement Learning (AAAI 2026 · similarity 0.92)
• ReMA: Learning to Meta-Think for LLMs with Multi-agent Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
• Beyond Single Stationary Policies: Meta-Task Players as Naturally Superior Collaborators (NeurIPS 2024 · similarity 0.90)
• NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming (NeurIPS 2026 · similarity 0.89)
• Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.89)
• Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.89)