Investigating the Interplay between Contextual and Parametric Chain-of-Thought Faithfulness under Optimization
arXiv:2605.24960v1 Announce Type: new Abstract: Chain-of-Thought (CoT) faithfulness, i.e., whether CoTs genuinely reflect large language models' (LLM) underlying behavior, is typically evaluated under two disjoint paradigms: contextual faithfulness, measured by perturbing the input or CoT trace, and parametric faithfulness, assessed by intervening on a model's parametric knowledge. Yet prior work compares them only descriptively. We fill this gap by proposing FaithMate, a unified preference-alignment interface for optimizing models towards either faithfulness paradigm. It enables us to investi
The proliferation of Large Language Models (LLMs) and their integration into critical applications necessitates deeper understanding and control over their reasoning processes and faithfulness. This research addresses a current gap in systematically evaluating different facets of CoT faithfulness.
This research provides a framework for optimizing LLMs not just for performance, but for the fidelity and transparency of their underlying reasoning, which is crucial for safety, reliability, and trustworthiness in AI systems.
The introduction of FaithMate and a unified approach to contextual and parametric faithfulness enables more targeted development and evaluation of LLMs, potentially leading to more robust and explainable AI systems.
- · AI developers
- · AI safety researchers
- · Developers of explainable AI
- · Black box AI solutions
- · Users distrustful of AI reasoning
Improved methods for evaluating and enhancing Chain-of-Thought (CoT) faithfulness in LLMs are developed and adopted.
More reliable and transparent AI systems emerge, increasing trust and broader adoption in sensitive domains.
Regulatory frameworks begin to incorporate requirements for demonstrable AI faithfulness and explainability, driven by these advancements.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL