Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
VERITAS verbalizes layer-wise computation to detect unfaithful chain-of-thought, beating LLM judges by 28.4% balanced accuracy.
VERITAS produces a verbalized, layer-wise account of how chain-of-thought reasoning evolves using Jacobian projections at each generated token. A classifier trained on those features detects unfaithful reasoning on four FaithCoT-Bench and BonaFide datasets. It improves balanced accuracy by 28.4% over LLM-as-Judge and averages a 16.0% gain over activation probing, with reported generalization across datasets and architectures.
- VERITAS verbalizes layer-wise Jacobian features of CoT generation.
- It tracks how reasoning anchors emerge, transfer, and reach the output.
- Balanced accuracy beats LLM-as-Judge by 28.4% on four datasets.
- Average gain over activation probing is 16.0%.
- Authors report generalization across datasets and model architectures.
From the original article
1 min read at arxiv.orgCurrent 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. We introduce VERITAS to verbalise the model's computation: how different reasoning anchors emerge, transfer across layers, and surface in the output.
By Yujing Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.