NoisyCoconut: Enhancing LLM Reliability via Latent Space Noise Injection
Researchers Michael Jerge and David Evans have introduced NoisyCoconut, a novel inference-time method designed to improve the reliability of large language models (LLMs) without requiring retraining or access to training data. Published on arXiv in May 2026, this technique manipulates internal model representations by injecting controlled noise into latent trajectories. This process generates diverse reasoning paths, where unanimous agreement among these paths serves as a confidence signal. Consequently, the model can abstain from answering when uncertain, effectively managing coverage-accuracy tradeoffs. Experimental results demonstrate significant performance improvements, reducing error rates from 40-70% to below 15%. Specifically, the method enables LLMs to exceed 95% accuracy on mathematical reasoning tasks through selective abstention. By operating directly during inference and maintaining compatibility with existing models, NoisyCoconut offers a practical pathway for enhancing output reliability in AI systems. This development represents a significant advancement in machine learning techniques focused on robustness and trustworthiness in artificial intelligence applications.
Wire timeline
NoisyCoconut: Enhancing LLM Reliability via Latent Space Noise Injection
Researchers Michael Jerge and David Evans have introduced NoisyCoconut, a novel inference-time method designed to improve the reliability of large language models (LLMs) without requiring retraining or access to training data. Published on arXiv in May 2026, this technique manipulates internal model representations by injecting controlled noise into latent trajectories. This process generates diverse reasoning paths, where unanimous agreement among these paths serves as a confidence signal. Consequently, the model can abstain from answering when uncertain, effectively managing coverage-accuracy tradeoffs. Experimental results demonstrate significant performance improvements, reducing error rates from 40-70% to below 15%. Specifically, the method enables LLMs to exceed 95% accuracy on mathematical reasoning tasks through selective abstention. By operating directly during inference and maintaining compatibility with existing models, NoisyCoconut offers a practical pathway for enhancing output reliability in AI systems. This development represents a significant advancement in machine learning techniques focused on robustness and trustworthiness in artificial intelligence applications.
cs.AI updates on arXiv.org