SYNTHESIS NOTE
Topics›Evolution›this note

Can an AI agent reliably improve itself through hidden evaluation?

AIDE2 rewrites its own code and selects improvements based on hidden evaluations. But what are these evaluations hidden from, and does the partition actually prevent gaming or circularity?

Synthesis note · 2026-09-24 · sourced from Evolution

The abstract states the system in one sentence: AIDE2 "proposes changes to its own code, benchmarks modified versions of itself on a suite of AI R&D tasks, and keeps the changes that perform best on hidden evaluations." It is the implementation of the loop defined in How does an AI agent improve its own research code?, applied to "a frontier AI research agent."

The result reported for it: "In an autonomous 8-day run, AIDE2 discovered seven successive improvements." The discussion adds that the loop "accepted seven rewrites, each under a fixed evaluation budget." Successive matters, since each accepted rewrite becomes the agent the next round edits, so the seven are a chain and not seven independent candidates.

The selection rule is where the vault's interest sits. Keeping what scores best on evaluations the agent is not scored on in the open is the "partition nothing that proposes changes can see" in Can deterministic checks protect LLM judges from failure?. It is also an external anchor in the sense of Can models reliably improve themselves without external feedback?: the loop rewrites the agent, and the scoring stays outside it. My reading, not the paper's: the excerpt says "hidden" and does not say hidden from the proposing agent, so the fit to the guardrail move is a reading (What exactly does hidden mean in AIDE2's evaluation system?).

Two things the excerpt leaves out. It says "keeps the changes that perform best", which does not say whether variants that lost are retained as stepping stones the way the Darwin Gödel Machine's archive does (Can AI systems improve themselves through trial and error?). And "a suite of AI R&D tasks" is not enumerated, so what the selection was tuned against is not visible.

Inquiring lines that read this note 17

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

Do AI capability benchmarks accurately measure reasoning ability or just surface patterns? How can we verify agent claims against their actual capabilities and actions? Why does self-revision fail to improve and instead amplify confidence? How prevalent is reward hacking in frontier models? How should reasoning time be allocated across search iterations? How can evaluation criteria remain robust against agent gaming? Can AI systems safely improve themselves recursively? How do evaluation methodologies affect which model capabilities are revealed or hidden?

Related concepts in this collection 5

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
14 direct connections · 99 in 2-hop network ·medium cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

AIDE2 proposes changes to its own code, benchmarks the modified versions on AI R&D tasks and keeps the ones that perform best on hidden evaluations — seven successive improvements in an autonomous 8-day run