GeoFaith: A Spatio-Temporal Dual View of Faithful Chain-of-Thought

Researchers introduce GeoFaith, a framework for detecting and improving faithfulness in chain-of-thought reasoning by LLMs, addressing the problem of plausible-sounding but inaccurate explanations. The method combines geometric latent structures with entropy analysis and includes a reinforcement learning approach that achieves superior performance on faithfulness detection while maintaining accuracy.

Chain-of-thought prompting has become fundamental to LLM capabilities, yet a critical vulnerability persists: models generate reasoning that sounds correct while masking flawed logic. GeoFaith directly tackles this faithfulness-accuracy tradeoff that has frustrated AI safety researchers and practitioners. The framework operates on two dimensions—spatial (geometric structure of model activations) and temporal (entropy changes across reasoning steps)—enabling detection of when models confabulate versus genuinely reason.

This work builds on growing recognition that outcome supervision alone corrupts reasoning quality. Prior faithfulness assessment methods relied on human evaluation (expensive), external tools (domain-specific), or simple heuristics (unreliable). GeoFaith's bootstrapping pipeline expands annotated data from 1,000 to 20,000 samples across four domains, creating practical scale. The resulting 8B faithfulness detector outperforming GPT-4 on benchmarks signals that faithfulness can be systematically measured and optimized.

The reinforcement learning component addresses a real deployment challenge: how to improve reasoning without sacrificing accuracy. By jointly optimizing correctness, faithfulness, and consistency, the approach produces shorter, more interpretable chains—directly beneficial for applications requiring explainability like healthcare, legal analysis, or financial advising. Organizations using LLMs for critical decisions gain better visibility into model confidence and reasoning validity.

Longer-term implications center on trustworthiness standards. As AI adoption deepens in regulated industries, demonstrable faithfulness becomes competitive advantage and compliance necessity. This research contributes methodology and benchmarks for measuring what many claim to offer but few can prove.

• →GeoFaith detects unfaithful reasoning by analyzing latent geometric structure and entropy dynamics rather than relying solely on human evaluation
• →The bootstrapped faithfulness detector with 20k samples outperforms GPT-4 on standard benchmarks despite being based on 8B parameters
• →Joint optimization of accuracy, faithfulness, and consistency produces shorter reasoning chains without degrading task performance
• →Scalable faithfulness assessment addresses a critical gap for deploying LLMs in regulated industries requiring explainability
• →Public code release enables downstream research on measuring and enforcing reliable reasoning across AI applications