Method
Amitava Das, Vinija Jain, Aman Chadha · 2025
Read Method by Amitava Das, Vinija Jain, Aman Chadha (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-2508-02063.txt.
Introduction
Amitava Das, Vinija Jain, and Aman Chadha trace unsafe model answers back to training data. Their 2025 paper presents TraceAlign for language models gone off-rails. Fine-tuned models can drift under tricky prompts and paraphrased jailbreaks. Knowing which training texts planted the bad beliefs is the missing piece.
A Belief Conflict Index scores how far an answer strays from aligned policy. The score rests on matching text spans against the training corpus. Three defenses follow: filter bad answers live, retrain against conflict, and veto risky decoding paths. Together they cut drift by up to eighty-five percent on a custom benchmark. Normal task skill barely moved. Safety work gains a root-cause tool, not just a patch.
Catalog record
| CREATOR | Amitava Das, Vinija Jain, Aman Chadha |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 145 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-02063