← News

Models & toolsSources: arXiv · 19 Nov 2026

BaseRSI lets a vision-language model internalize its own prompt-and-tool harness

The arXiv paper BaseRSI proposes a training-free loop in which a base model improves its surrounding harness and writes those gains into its own weights.

In plain words

A vision-language model is one that processes both images and text. The paper describes a way for such a model to improve the scaffolding of prompts and tools around it, then fold those improvements directly into its own parameters without extra training rounds.

What was announced

The authors frame recursive self-improvement as shifting from gains that stay in the external harness toward changes written into the model itself. They identify three problems with existing approaches: improvements that remain harness-bound, supervision that reads only the model's outputs, and edits whose site and strength are fixed in advance. BaseRSI addresses this by having the base model improve its harness and then internalize what the harness adds through training-free edits, applied repeatedly to its own responses.

The supporting mechanism, ReGap, supervises each transformer block using the activation difference the harness introduces at that block, updating weights along this signal in a low-dimensional subspace. The ReGap-based algorithm ReKey writes this signal by regression into a few key-value neurons per block and measures each edit's site and strength on the model's own responses. The authors state the model fills every role of the recursion with forward passes alone, and that edits at different sites compose in one checkpoint without refitting. On ten visual reasoning benchmarks, they report consistent improvements both in distribution and on held-out benchmarks, with code released on GitHub.

What it does not show

  • Evidence is limited to the paper's abstract; no independent replication or third-party evaluation is available.
  • The abstract does not specify which vision-language model was tested, the size of improvements, or how the method compares to specific baselines.
  • Claims of consistent improvement rest on the authors' own reported benchmark results, which the abstract summarizes without detail.

For security and IT teams

Scope
The source is the paper's abstract only, covering the method's stated design and headline benchmark claim rather than full experimental detail.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.