Game Theory Driven Multi-Agent Framework Mitigates Language Model Hallucination

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

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

G-Frame is a game-theoretic multi-agent framework that orchestrates hierarchical LLM agents — governed by team-game and Bayesian-game principles — to drive an end-to-end automated pipeline for chemical data synthesis and lightweight model training. Applied to a 7B model (OmniChem), it achieves a 79.46% hallucination reduction relative to the base model and performance parity with GPT-4o mini on chemistry benchmarks.

Problem

Lightweight LLMs fail in rigorous scientific domains (e.g., chemistry) because autoregressive generation mimics linguistic surface patterns rather than internalizing axiomatic physical constraints, producing plausible but factually wrong outputs. Existing mitigations (tool integration, supervised fine-tuning on scarce domain data, closed-source API dependence) do not close this structured-reasoning gap.

Method

G-Frame implements a three-layer hierarchical multi-agent architecture:

• Macro-level (Bayesian game): A decisional agent treats the entire pipeline as a Partially Observable Markov Decision Process. It forms a prior policy over workflow and hyperparameter settings, then updates to a posterior decision \(d_{\text{posterior}} = h(d_{\text{prior}}, \Gamma)\) using real-time feedback \(\Gamma\) (e.g., KV-cache occupancy, queue length, ChemJudge checkpoint scores).

• Micro-level (team game): Each task agent (data cleaning, augmentation, CoT synthesis, adaptive training) spawns 2–3 executive agents with granular, prompt-engineered roles. Sequential execution restricts per-step information entropy, minimizing KL divergence between generated and true distributions.

• Communication: A decoupled async multi-queue architecture (Multi-Producer, Single-Consumer via asyncio) merges hardware state vectors (\(\text{DataPoint}_V\)) and task progress vectors (\(\text{DataPoint}_T\)) into a unified System State Vector consumed by the decisional agent.

The pipeline ran on ~500k chemistry articles/books → 5B-token pre-training corpus (8:2 domain/general mix), then synthesized 363,045 CoT chains and 199,589 QA pairs for supervised fine-tuning of Qwen2.5-7B-Instruct. Adaptive training uses Bayesian-controlled hyperparameter adjustment guided by ChemJudge (LLM-as-referee) payoff signals. Inference is served via vLLM; the decisional agent uses DeepSeek-R1 API.

Key Contributions

• G-Frame framework: A reusable meta-harness coupling team-game collaborative synthesis with Bayesian adaptive control for automated domain model construction.
• OmniChem 7B: A chemistry-specialized LLM with 79.46% fewer hallucinations vs. base model and performance comparable to GPT-4o mini on ThChem and ChemBench.
• Synthetic corpus at scale: 363,045 chemical CoT chains + 199,589 QA pairs generated via multi-agent pipelines, to be openly released.
• Adaptive concurrency control: Bayesian-governed asynchronous concurrency achieves ~22× throughput vs. synchronous mode and outperforms fixed concurrencies of 50 and 100 by 25% and 6.5%, respectively.
• Downstream integration: OmniChem deployed inside a Deep Research framework (academic API + GraphRAG) producing reports scoring ~90% of GPT-o3 and Gemini on LLM-as-Judge evaluations.

Results

• ChemJudge (471-question hallucination test, judged by Gemini 3.1 Pro): OmniChem beats GPT-4o mini; 79.46% hallucination reduction vs. base Qwen2.5-7B-Instruct.
• ThChem 1.0: 79.45% accuracy (80%–97% of GPT-o3 level).
• ThChem 2.0: 62.08% accuracy.
• ChemBench: 49.82% accuracy, comparable to GPT-4o mini.
• SQuAD 2.0 QA quality: G-Frame multi-agent approach raises F1 by 20%–40% over single-LLM generation across lightweight model scales.
• Deep Research report (TADF materials): 90% of GPT-o3/Gemini score on four-dimension 100-point LLM-as-Judge rubric.

Limitations

• Decisional agent relies on DeepSeek-R1 via external API (Volcano Engine), introducing external dependency and potential data sovereignty concerns despite the local-deployment motivation.
• Evaluation benchmarks (ThChem, ChemJudge) appear to be custom or lightly adopted; ChemBench is the only widely recognized external benchmark.
• ChemJudge scores are produced by an LLM referee (Gemini 3.1 Pro), introducing judge-model bias.
• Results demonstrated exclusively in chemistry; generalizability to other rigorous scientific domains is asserted but not empirically validated.
• Ablation data (full table) is relegated to Supplementary Information, limiting reproducibility verification from the main text.
• The 5B-token corpus construction pipeline is compute-intensive (A100/H100 GPUs), potentially limiting accessibility.

Relevance to Harnesses / Meta-Harnesses

G-Frame is a concrete instantiation of a meta-harness: it programmatically orchestrates heterogeneous agents (decisional, task, executive) through command-line invocations of independent Python modules, controls the full ML lifecycle (data cleaning → augmentation → CoT synthesis → pre-training → SFT → adaptive training), and dynamically reconfigures itself via Bayesian feedback loops. The decoupled async communication architecture and prior-posterior update cycle directly parallel harness control-plane patterns (state observation → policy update → dispatch). For researchers tracking harnesses, G-Frame demonstrates how game-theoretic control policies can replace hand-authored orchestration scripts for closed-loop, self-adapting pipelines — a design transferable to any domain requiring structured multi-stage agent coordination.