NoisyCoconut: Counterfactual Consensus via Latent Space Reasoning

TL;DR

NoisyCoconut enhances LLM reliability by injecting noise into latent representations during inference, enabling confidence estimation and abstention without retraining, significantly reducing errors on reasoning benchmarks.

Contribution

Introduces a novel inference-time method that manipulates internal model representations with noise to improve reliability without retraining or model modification.

Findings

Reduces error rates from 40-70% to below 15% through noise-based agreement.

Achieves over 95% accuracy on mathematical reasoning tasks with selective abstention.

Operates effectively across multiple reasoning benchmarks without access to training data.

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.