NoisyCoconut: Enhancing LLM Reliability via Latent Space Reasoning

A new inference-time method called NoisyCoconut improves large language model (LLM) reliability by manipulating internal representations without retraining. The approach injects controlled noise into latent trajectories to generate diverse reasoning paths; unanimous agreement among these paths serves as a confidence signal, allowing models to abstain when uncertain. Experiments demonstrate effective coverage-accuracy tradeoffs across multiple reasoning benchmarks, requiring no access to training data or parameter modification. This provides a practical pathway to improving LLM output reliability while maintaining compatibility with existing models.