When the Model Said ‘No Comment’, We Knew Helpfulness Was Dead, Honesty Was Alive, and Safety Was Terrified
Gautam Siddharth Kashyap, Mark Dras, Usman Naseem · 2026
Read When the Model Said ‘No Comment’, We Knew Helpfulness Was Dead, Honesty Was Alive, and Safety Was Terrified by Gautam Siddharth Kashyap, Mark Dras, Usman Naseem (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-2602-07381.txt.
Description
Aligning a language model to be helpful, harmless, and honest at once is hard: fine-tuning creates interference between goals, and mixture-of-experts routing misfires. Gautam Siddharth Kashyap and two co-authors name that failure Axis Collapse and propose AlignX to fix it.
Stage one uses prompt-injected fine-tuning to extract axis-specific features without catastrophic forgetting from disjoint feature spaces. Stage two calibrates expert routing with a module built on fractal and natural shape. Gains on helpfulness, honesty, and safety benchmarks follow. Alignment teams get a two-stage design that treats multi-objective match as a routing and representation problem, not just a training trick.
Catalog record
| CREATOR | Gautam Siddharth Kashyap, Mark Dras, Usman Naseem |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-07381