Large-small model collaboration for cross-stage hallucination suppression in trustworthy language models

Large language models (LLMs) can generate fluent and informative responses, but they often produce factually incorrect content, especially when backbone retraining is expensive and deployment domains evolve over time. To address this issue, we propose a large–small model collaboration framework for hallucination suppression based on dual-proxy logit steering and cross-stage proxy refinement. Specifically, two lightweight proxy models are trained from paired positive and negative supervision: a Factual Alignment Proxy (FAP), which captures fact-supporting tendencies, and a Hallucination Suppression Proxy (HSP), which captures anti-factual tendencies. During inference, the two proxies guide a frozen backbone model by reweighting next-token logits, thereby promoting factual continuations while discouraging hallucination-prone ones. To support post-deployment adaptation, we further introduce a feedback loop that updates only the factual proxy using inference-time error signals and corrective evidence, without modifying the backbone model. Experiments on TruthfulQA, BioGEN, and SubjQA show that the proposed framework improves factuality across both contextual and non-contextual settings while preserving informativeness. Additional analyses on ablation and iterative refinement demonstrate the contribution of dual-proxy guidance and the value of lightweight proxy-only updates. Overall, the framework provides a practical parameter-efficient alternative to repeated backbone retraining for factuality-sensitive language generation. • Large-small model framework with dual-proxy logit steering suppresses hallucinations. • Lightweight factual proxy updated from errors and evidence adapts post-deployment. • On TruthfulQA, BioGEN, SubjQA, factuality gains without informativeness loss.