The Chain Holds, the Answer Folds: Trace–Answer Dissociation in Reasoning Models Under Adversarial Pressure
Yubo Li, Ramayya Krishnan, Rema Padman · 2026
Read The Chain Holds, the Answer Folds: Trace–Answer Dissociation in Reasoning Models Under Adversarial Pressure by Yubo Li, Ramayya Krishnan, Rema Padman (2026) 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-2605-29087.txt.
Description
Yubo Li, Ramayya Krishnan, and Rema Padman document unfaithful capitulation in reasoning models under adversarial pressure. The chain of thought holds up across every turn, yet the answer put out turns wrong. A latent-behavioral framework isolates the failure, which flip metrics and single-turn probes both miss. Across three datasets, correctness at the flip nears 50 percent in think mode and collapses to 11-15 percent under no-think. That paired contrast shows reasoning creates the gap. An independent judge corroborates 86 percent of labels.
Catalog record
| CREATOR | Yubo Li, Ramayya Krishnan, Rema Padman |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41801 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.29087