Score $\times$ Decoder: A Unified View of Unsupervised Inference-Time Scaling for Hallucination Mitigation
For practitioners needing to reduce hallucinations without supervision, this work provides a systematic comparison of unsupervised methods, though findings are incremental and limited to a single model and dataset.
The paper investigates unsupervised inference-time scaling for hallucination mitigation in LLMs, evaluating a grid of four intrinsic scores and three decoding families on MATH500 with Qwen3-1.7B. They find that self-verification with a virtual-thinking prefix works well, but no single score is universally best; performance depends on the decoder and model capability.
Large language models hallucinate even when the answer lies within their parameters. While inference-time scaling can surface this latent knowledge, the most effective methods require supervision: a trained verifier or reward model. We ask what can be done with only a base language model: which intrinsic signal best identifies correct outputs, and how should it be decoded? We cast this as a score~$\times$~decoder grid pairing four scores (perplexity, contrastive, power-distribution likelihood, and self-verification) with three decoding families (optimization, sampling, consensus), and evaluate every cell on MATH500 with the base and instruction-tuned Qwen3-1.7B. While self-verification, which prompts the model to judge its own answer and is sharpened by a training-free virtual-thinking prefix, works well in most settings, no score has a fixed quality: its value depends on the decoder that consumes it and on model capability. When no supervision is available, the score and the decoding family must be chosen together.