Active Minimax Decoder: Reducing Hallucinations in Small Language Models via Binary Adversarial Verification

We introduce the Active Minimax Decoder (AMD), a training-free, inference-time framework that reduces hallucinations in Small Language Models through binary adversarial verification. AMD employs a three-agent architecture: a Generator produces candidate responses, an Adversary identifies factual inconsistencies, and an Arbiter renders binary accept/abstain decisions. On TruthfulQA (817 questions), AMD enables SmolLM2-360M (0.36B parameters) to achieve a 0.61% hallucination rate ,17x lower than Gemini-2.0-Flash in vanilla mode and 96% reduction versus its own baseline. When AMD elects to answer, it achieves 97.5% accuracy. Our adversary ablation study reveals that adversary capability, rather than size, determines verification quality: LFM2-350M outperforms Qwen-1.5B as an adversary despite being 4x smaller.