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

COMPLLLM: Fine-tuning LLMs to Discover Complementary Signals for Decision-making

NeurIPS 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 ComplLLM, a post-training framework that fine-tunes Qwen3-8B (SFT then GRPO) to extract text signals that complement an upstream agent's decision, using decision-theoretic complementary value as reward.

Problem

Multi-agent decision pipelines benefit only when agents provide complementary information; existing interpretability and hypothesis/topic methods surface frequent or predictive signals rather than ones that add decision value beyond an existing recommendation.

Approach

Define complementary value as the improvement in best-attainable decision payoff from a text-extracted signal beyond the agent's recommendation; estimate the data-generating process with an LLM reference, generate SFT labels from complementary signals, and use this value (normalized) as the GRPO reward with a support check against hallucination.

Key results

• On the synthetic dataset, ComplLLM reaches 0.98 surface similarity and 0.67 F1 versus 0.42 F1 for zero-shot and 0.38 for HypotheSAE.
• On MIMIC-CXR, ComplLLM signals raise accuracy from 0.819 (agent) to 0.839 with 12 significant signals, more than HypotheSAE (0.831, 6 signals) and zero-shot (0.825, 5 signals).
• On Review5K, Gemini 2.0 Flash accuracy with human review plus ComplLLM signals is 79.7% versus 78.8% without them.

Surprising findings

• Few-shot prompting gets higher surface similarity than zero-shot on synthetic data yet lower F1, meaning it names similar signals at the wrong instances.
• On DICES and Review5K, zero-shot and few-shot baselines extract no significant complementary signals.

Limitations

• Rare but important signals below the frequency threshold N_tau may be dropped.
• When the recommending agent and supervisor are the same person, complementary value may not hold in hindsight as beliefs change.
• Physician feedback found some signals redundant with or only partly relevant to the diagnosis; validation was on only six cases with two physicians.

Experimental setup

Datasets: MIMIC-CXR MIMIC-IV DICES Review5K CheXpert labels

Benchmarks: MIMIC-CXR (synthetic complementarity) MIMIC-CXR DICES Review5K

Metrics: surface similarity F1 similarity complementary information value breadth accuracy

Baselines: Zero-shot Few-shot BERTopic HypotheSAE

Builds on: GRPO HypotheSAE BERTopic

Closest papers by meaning

• AgentCDM: Enhancing Multi-Agent Collaborative Decision-Making via ACH-Inspired Structured Reasoning (AAAI 2026 · similarity 0.85)
• Think Twice, Act Once: A Co-Evolution Framework of LLM and RL for Large-Scale Decision Making (ICML 2025 · similarity 0.85)
• Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation (ICLR 2026 · similarity 0.85)
• RAGEN-2: Reasoning Collapse in Agentic RL (ICML 2026 · similarity 0.85)
• Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs (ICML 2026 · similarity 0.85)
• Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.84)