What Survives When You Compress a Recursive Reasoner for the Edge?
Pearse Jim, Steven Kolawole, Opegbemi M. Busoye, Glory Bagai, Virginia Smith · 2026
Read What Survives When You Compress a Recursive Reasoner for the Edge? by Pearse Jim, Steven Kolawole, Opegbemi M. Busoye, Glory Bagai, Virginia Smith (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-2606-26488.txt.
Description
Pearse Jim and four colleagues ask what survives when a recursive reasoning model is compressed for small devices. Such models solve hard structured puzzles with few parameters. They update an internal state again and again instead of writing long outputs. Shrinking them for edge hardware is necessary, but compression errors compound across their reasoning cycles. Standard intuition from ordinary models does not apply.
One pattern held across every precision level, three tasks, and two designs. Aggressive compression kept single-cell predictions accurate. Whole-puzzle accuracy collapsed to zero. Quantization, pruning, distillation, and linear attention all failed the same way. The collapse hit one design family but not attention-based recursion on the same task. A per-channel calibrated low-precision format reversed the damage without retraining. The team also offers a label-free signal that tracks the reasoning path and predicts damage early.
Catalog record
| CREATOR | Pearse Jim, Steven Kolawole, Opegbemi M. Busoye, Glory Bagai, Virginia Smith |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.26488