Aspect Research

An independent lab working on how language models store and update what they know. Our published work tunes the context a coding agent is given. The current work looks at where else that knowledge could live, and what each option costs. Aimed at open models.

Email GitHub

Current work

In progress

Probe-and-refine compiles what a coding agent needs to know about a repository into an artifact it can reuse. Repositories are only one case. Whatever a model needs to know has to live somewhere, and where it lives decides what the model can do with it and what it costs to keep current.

We are studying this on open models. The first preprint is in preparation.

In preparation

Expected Aug 2026

[working title TBD] — first preprint of the current program

The current line of work: where a model's knowledge lives, and what it costs to keep current. Preprint in preparation.

Draft available on request →

Published

Peer-checkable

On SWE-bench Verified, repository guidance that is iteratively refined outperforms static context. The current work starts from this result.

PreprintarXivCOLM 2026 WAB

Jan–May 2026

Probe-and-Refine Tuning of Repository Guidance for Coding Agents

A methodology for refining repo-level guidance via synthetic bug-fix probes; raises SWE-bench Verified resolve rate to 33.0% from a 25.5% unguided baseline over 4 trials, with gains driven by coverage rather than precision, plus evidence of cross-model degradation.

Accepted at the Workshop on Agent Behavior, COLM 2026. The accepted revision is not yet on arXiv.

Read →

Other preprints →