NoisyCoconut: Counterfactual Consensus via Latent Space Reasoning

2026-05-06

AI Summary

This work proposes a training-free, adaptive abstention mechanism to enhance the reliability of large language models during inference. By injecting controlled noise into the model’s latent representations at inference time, the method generates diverse reasoning paths and evaluates output confidence based on inter-path consistency, enabling the model to abstain when uncertain. Notably, this approach is the first to leverage latent-space perturbations for consensus-based confidence estimation without modifying model parameters or requiring additional training data. Experimental results demonstrate that the method reduces error rates from 40–70% to below 15% across multiple reasoning benchmarks and achieves over 95% accuracy on mathematical tasks through selective abstention.

Abstract

This paper presents NoisyCoconut, a novel inference-time method that enhances large language model (LLM) reliability by manipulating internal representations. Unlike fine-tuning methods that require extensive retraining, NoisyCoconut operates directly on model representations during inference and requires no retraining. Rather than training models to reason in latent space, we inject controlled noise into latent trajectories to generate diverse reasoning paths. Agreement among these paths provides a confidence signal, enabling models to abstain when uncertain. We demonstrate that this approach achieves effective coverage-accuracy tradeoffs across multiple reasoning benchmarks without requiring access to training data or modification of model parameters. This approach provides a practical pathway to improving the reliability of LLM outputs while maintaining compatibility with existing models. Our experiments show that unanimous agreement among noise-perturbed paths reduces error rates from 40-70% to below 15%, enabling models to exceed 95% accuracy on mathematical reasoning tasks through selective abstention.

Problem

• LLM reliability
• inference-time method
• selective abstention
• reasoning uncertainty
• latent space

Innovation

• latent space reasoning
• controlled noise injection
• inference-time intervention
• selective abstention
• counterfactual consensus