Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection

Key points

• Current Chain-of-Thought (CoT) verification methods detect reasoning faithfulness from model outputs or static activations, but are limited in explaining how unfaithfulness develops during generation.1
• VERITAS adopts the layer-wise Jacobian projection to produce a verbalised account of the model's internal reasoning evolution at each generated CoT token.1
• By training a classifier on the extracted features along the traces, we show that these structured features provide insightful perspectives on how and where reasoning diverges toward unfaithfulness.1
• Our work shows that verbalising model computation provides a richer and more interpretable lens on CoT faithfulness, moving beyond detection toward a mechanistic understanding of how reasoning fails.1

Sources (1)

• [1]Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 10:44 AM

  We introduce VERITAS to verbalise the model's computation: how different reasoning anchors emerge, transfer across layers, and surface in the output.

  Current Chain-of-Thought (CoT) verification methods detect reasoning faithfulness from model outputs or static activations, but are limited in explaining how unfaithfulness develops during generation.

Extractive summary: sentences quoted from the sources.