NoisyCoconut: Counterfactual Consensus via Latent Space Reasoning

NoisyCoconut is an inference-time method that improves LLM reliability by injecting controlled noise into internal representations to generate diverse reasoning paths, enabling models to abstain when uncertain without requiring retraining. The technique reduces error rates from 40-70% to below 15% on mathematical reasoning tasks through unanimous agreement among noise-perturbed paths, offering practical reliability improvements compatible with existing models.

NoisyCoconut addresses a critical limitation in large language models: their tendency to produce confident but incorrect outputs. Rather than pursuing expensive fine-tuning approaches, this method operates at inference time by manipulating the model's internal latent representations. The core innovation involves injecting structured noise into reasoning trajectories to create multiple diverse paths through the model's computation, then using consensus among these paths as a confidence signal.

This research builds on growing recognition that LLMs struggle with calibration and reliability, particularly on structured tasks like mathematical reasoning. Previous approaches required retraining models or accessing training data, creating practical barriers to deployment. NoisyCoconut's parameter-free approach sidesteps these constraints entirely, making it immediately applicable to existing closed-source and open-source models alike.

The technique's practical impact is substantial for production systems where reliability matters. Reducing error rates from 40-70% to below 15% through selective abstention represents meaningful progress toward more trustworthy AI systems. By enabling models to decline answering when uncertain rather than hallucinating, the method improves overall output quality from the user's perspective.

For the broader AI field, NoisyCoconut demonstrates that significant reliability gains don't require architectural changes or extensive computational overhead. The unanimous-agreement threshold provides an elegant mechanism for trading coverage against accuracy—organizations can choose their preferred accuracy-abstention tradeoff. Future research will likely explore whether similar latent-space perturbation strategies apply to other LLM failure modes beyond reasoning tasks.

• →NoisyCoconut improves LLM reliability through inference-time noise injection without requiring model retraining or parameter modification
• →Unanimous agreement among noise-perturbed reasoning paths reduces mathematical reasoning errors from 40-70% to below 15%
• →The method enables selective abstention, allowing models to decline uncertain predictions rather than hallucinate answers
• →Approach maintains compatibility with existing models including closed-source systems, enabling immediate practical deployment
• →Demonstrates that significant reliability improvements can be achieved through latent space manipulation rather than architectural changes