The BaseResearcharXiv
Values Shape What Self-Improving Coding Agents Build
According to its abstract, a preprint by Jason Liang, Dylan Waldner and Risto Miikkulainen tests whether a short value statement in a self-improving coding agent's instruction file changes what it builds. Using four Schwartz value types, the authors report that values affect performance and solution design, and that human oversight is still needed.
Why it matters
The same thing, for different readers.
What it means for you
Researchers report that a short statement of values in an AI coding agent's instructions changes what it builds, not just how well it performs.
For researchers
Liang, Waldner and Miikkulainen report that openness values raised held-out performance, conservation values lowered it, and letting agents revise their own values narrowed the differences.
The question it raises
The authors report that human oversight is still needed to design an agent's values, as letting the agent rewrite its own values reduced the effect.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Self-improving coding agents given a fixed short value statement in a persistent instruction file, from four Schwartz higher-order value types, across diverse coding tasks and multiple self-improvement iterations.
- What it found
- The authors report that openness values increase held-out performance while conservation values score below an agent with no values; values change solution size, complexity and approach; allowing the agent to revise its own values narrows the differences.
- What it does not show
- The excerpt was cut off and does not give models, tasks or effect sizes.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- What happens when AI helps build more capable AI? · What values should AI have, and could it have a mind of its own?
- People
- Jason Liang (Wrote it) · Dylan Waldner (Wrote it) · Risto Miikkulainen (Wrote it)
- What we read
- We read the abstract.
- Found by
- The Radar, run of November 18, 2026, via
arxiv-ai-safety-query(tier 1: the source’s own channel or an official record) - Cite this
- arXiv, “Values Shape What Self-Improving Coding Agents Build”, November 17, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-17-values-self-improving-coding-agents