Semantic Gravity Wells: Why Negative Constraints Backfire
Shailesh Rana · 2026
Read Semantic Gravity Wells: Why Negative Constraints Backfire by Shailesh Rana (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-2601-08070.txt.
Description
Shailesh Rana studies a simple sounding test that large language models keep failing. His 2026 arXiv paper looks at negative constraints, such as telling a model not to use a given word. Models break such rules with striking regularity, and the reasons have stayed poorly understood.
Rana presents a broad mechanistic study of why negative instructions fail. The title names the effect a semantic gravity well, where the banned word pulls the model toward itself. For readers who build or test language models, the paper gives a clearer picture of a basic instruction-following flaw. It treats the failure as a mechanism to explain, not just an error rate to report.
Catalog record
| CREATOR | Shailesh Rana |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-08070