SR-TTT: Surprisal-Aware Residual Test-Time Training
Swamynathan V P · 2026
Read SR-TTT: Surprisal-Aware Residual Test-Time Training by Swamynathan V P (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-2603-06642.txt.
Description
V P revisits SR-TTT, routing high-surprisal tokens to a Residual Cache, credited with Needle-in-a-Haystack gains. Those gains were artifacts: the metric read logits at answer positions rather than one position earlier. The cache attended non-causally over tokens including the answer. Correction refutes the SR-TTT hypothesis. Failure splits into storage and addressing: biased surprisal gating stores early needles at minimal rates, while addressing cannot resolve ordered slots with near-interchangeable contents. Exact match stays 0% in all 2,250 trials, and v1 claims are retracted.
Catalog record
| CREATOR | Swamynathan V P |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 13245 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.06642