Skip to content

Game Theory Driven Multi-Agent Framework Mitigates Language Model Hallucination

🕒 Published (v1): 2026-07-09 12:28 UTC · Source: Arxiv · link

Why this paper was selected

Game-theoretic multi-agent framework reduces hallucination in lightweight LLMs — practical for agent reliability

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

G-Frame is a multi-agent framework that applies team game and Bayesian game theory to automate high-quality chemistry data synthesis and adaptive model training, eliminating the closed-loop between data generation and fine-tuning. The resulting 7B model OmniChem achieves a 79.46% reduction in hallucinations relative to its base model and reaches parity with GPT-4o mini on chemistry benchmarks.

Problem

Lightweight LLMs applied to rule-based scientific domains (chemistry) hallucinate because autoregressive generation mimics surface linguistic patterns rather than internalizing axiomatic physical constraints and structured causal reasoning. Existing mitigations—tool augmentation and supervised fine-tuning—are constrained by scarcity of high-quality domain data and dependence on closed-source APIs, which compromise data sovereignty and scalability.

Method

G-Frame is a three-layer hierarchical architecture (decisional → task → executive agents) governed by two interlocking game-theoretic mechanisms:

Team Game (micro-level): Macroscopic tasks are decomposed into short subtasks, each executed by a task agent comprising 2–3 executive agents with granular, prompt-engineered utility functions. Sequential execution confines per-step information entropy, formally acting as a distribution refinement that minimizes \(D_{\mathrm{KL}}(P_{\mathrm{generated}} \| P_{\mathrm{true}})\).

Bayesian Game (macro-level): A decisional agent models global system control as a Partially Observable Markov Decision Process (POMDP). It establishes a prior policy \(d_{\mathrm{prior}}\) from hardware/data state, then iteratively updates to a posterior \(d_{\mathrm{posterior}} = h(d_{\mathrm{prior}}, \Gamma)\) using real-time feedback \(\Gamma\) (throughput, KV cache occupancy, ChemJudge checkpoint scores), maximizing a predefined global payoff.

This closed loop drives three pipeline stages: (1) corpus cleaning (5B-token OCR-heavy chemistry literature via MinerU + Qwen2.5-7B-Instruct), (2) synthetic data generation (363,045 CoT chains distilled via teacher-student-supervisor agent teams; 199,589 QA pairs with generate-evaluate-correct cycles), and (3) adaptive training of OmniChem (Qwen2.5-7B-Instruct base, mixed 8:2 domain/general corpus to prevent catastrophic forgetting, hyperparameters tuned by Bayesian decisional agent guided by ChemJudge payoff signals). The decisional agent uses DeepSeek-R1 via Volcano Engine API; executive/task agents run via vLLM-served local models.

Key Contributions

  • G-Frame: a reusable, domain-agnostic multi-agent framework coupling team game (entropy minimization) with Bayesian game (POMDP adaptive control) for automated data synthesis and training
  • Automated closed-loop pipeline producing 363,045 CoT chains and 199,589 QA pairs from ~500,000 chemistry articles and ~1,000 textbooks
  • OmniChem, a 7B chemistry reasoning LLM with demonstrated molecular design, knowledge graph construction, and retrosynthetic planning capabilities
  • Adaptive asynchronous concurrency control achieving ~22× throughput improvement over synchronous synthesis mode
  • ThChem benchmark (custom chemistry reasoning evaluation) and ChemJudge (LLM-as-a-judge hallucination scoring, 471 questions, judged by Gemini 3.1 Pro)
  • Planned public release of CoT and QA datasets

Results

  • Hallucination (ChemJudge, 471 Qs, Gemini 3.1 Pro judge): OmniChem outperforms GPT-4o mini; 79.46% reduction in hallucination rate vs. Qwen2.5-7B-Instruct base
  • ThChem 1.0 / ThChem 2.0 / ChemBench accuracy: 79.45% / 62.08% / 49.82%; comparable to GPT-4o mini, reaching 80–97% of GPT-o3 level
  • Deep Research report quality: ~90% of GPT-o3 and Gemini 2.5 Pro scores on a 4-dimension LLM-as-a-Judge rubric (information accuracy, depth/logic, structure, novelty)
  • QA generation quality (SQuAD 2.0): G-Frame multi-agent approach raises F1 by 20–40% over single-LLM generation across lightweight model scales
  • Concurrency optimization: adaptive control achieves ~22× throughput vs. synchronous mode; 25% and 6.5% better than fixed concurrency at 50 and 100, respectively
  • Molecular design validation: DFT-confirmed BODIPY deep-red absorber at 709 nm; logP reduction from 1.82 → −4.44 with solubilizing substituents; lidocaine retrosynthesis matches commercial route

Limitations

  • Domain coverage is limited to chemistry; cross-domain generalizability of G-Frame is asserted but not demonstrated
  • The decisional agent relies on DeepSeek-R1 via an external commercial API, reintroducing the closed-source dependency the paper aims to avoid
  • ChemBench score (49.82%) is substantially below GPT-o3 levels, indicating a significant capability gap on that benchmark
  • ThChem is a novel benchmark introduced by the same group; independent validation is absent
  • Ablation is qualitative/summarized; the full ablation table is deferred to supplementary material not included in the provided text
  • Training process details (learning rates, epochs, LoRA vs. full fine-tuning) are truncated in the provided text

Relevance to Agentic AI / LLM Agents

G-Frame is a concrete instantiation of hierarchical multi-agent orchestration where agent collaboration is governed by formal game-theoretic contracts (utility functions, POMDP belief updates) rather than ad-hoc prompting—directly advancing the agentic AI design space. The Bayesian decisional agent that issues and revises workflow policies based on real-time environment feedback is an operative closed-loop agent architecture applicable well beyond chemistry. The team game decomposition—restricting per-step entropy by chaining agents with narrow scoped utility functions—offers a principled account of why multi-agent pipelines outperform monolithic LLM calls on complex tasks. This work also demonstrates that multi-agent frameworks can be self-improving training engines, not just inference-time orchestrators, which is a growing direction in agentic AI research.