Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation

Siddhant Bhambri, Upasana Biswas, Subbarao Kambhampati · 2025

Read Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation by Siddhant Bhambri, Upasana Biswas, Subbarao Kambhampati (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-13792.txt.

Description

Siddhant Bhambri, Upasana Biswas, and Subbarao Kambhampati test whether the reasoning traces of large language models are sound and clear, in work from 2025. Such traces are the steps a model writes before its final answer. They guide the model and are used to train smaller models, on the belief that they are correct and readable.

The team built question sets where each problem pairs with a trace that is checked as right or wrong, step by step. The final answer stays correct in both cases. That split isolates the effect of trace quality. They also tuned models on varied traces to test how well people can read and trust them.

Catalog record

CREATORSiddhant Bhambri, Upasana Biswas, Subbarao Kambhampati
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE32 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-13792

Browse the full library · About the Tower of Babel