An agent can outgrow a verifier's tests—or learn to exploit them. Keeping the judge fixed while the system changes leaves the improvement loop measuring against a moving mismatch.
Unified Co-Evolution, posted October 13, evolves executable verifier programs together with harness state. Candidate verifiers are tested on response pairs that differ in a declared property, with selection emphasising their weakest contrast family. Read the research paper.
Improvement needs an evaluator that remains useful
Open-ended agent tasks rarely offer the clean pass-or-fail checks available in a unit test. A model may have to judge whether a draft is better, decide what to repair and determine which experience should influence future tasks. Errors in that judgement have two effects. They can spoil the current response, and they can enter persistent harness state that shapes subsequent work. A weak verifier therefore creates a route by which a local mistake becomes a repeated habit.
Unified Co-Evolution treats the verifier as another evolving component. All model parameters remain fixed; the changes occur in executable verifier programs and in the agent's surrounding harness state.
Contrast pairs give the judge an examination
Candidate verifier programs are tested on response contrasts: pairs that differ in one declared property and have a known preferred side. The system selects using the verifier's weakest contrast family, a measure the authors call driving capacity. This discourages a judge from looking strong merely because it handles abundant easy examples while missing a consequential category.
The selected verifier diagnoses one weakness in a draft, guides one repair, chooses the response to serve and controls whether the accepted repair can enter harness memory. As tasks finish, their diagnostics help generate new verifier candidates. A candidate must beat the current judge on the same contrasts before replacing it.
The safeguard sits inside the feedback loop
This design links persistent improvement to the quality of the process approving it. That is a more specific claim than saying the agent critiques itself. The critique mechanism has an explicit comparison test, and its weakest measured area influences selection. A repair that looks better to an untested model is not automatically allowed to become future guidance.
Across six open-ended benchmarks, the authors report average preference rates of 59.00 percent for Qwen3-8B and 57.19 percent for Qwen3.8-27B against matched direct generation. On three verifiable benchmarks, the reported mean rises from 36.06 to 38.48 percent. These are bounded experimental gains, with different evaluation meanings for preference and verifiable success.
Preference rates are model-judged, not human votes
The open-ended comparisons use a separate model judge, with both response presentation orders and half credit for ties. Fifty percent is parity against matched direct generation. The actor, task order and one-revision protocol are shared across the compared methods. These controls make the reported preference margins interpretable without implying that fifty-nine percent represents a human survey.
The verifier’s internal contrast test and the external outcome comparison are also different measurements. A candidate can discriminate its contrasts better yet fail to improve delivered responses. Retaining both levels of evaluation is essential because the system is explicitly changing its own feedback mechanism. The separate verifiable benchmarks provide another check, although their smaller gain should not be collapsed into the open-ended preference measure.
Known contrasts cannot cover unknown weaknesses
The contrast families define what the system knows how to test. If important defects never appear in those contrasts, improving the weakest measured family can still leave a serious unmeasured blind spot. Independent held-out contrasts and examination of served answers are necessary to determine whether a better verifier produces the intended real improvement.
The feedback infrastructure deserves the same scrutiny as the agent. Persistent self-improvement combines proposals, judgements and retention rules. Keeping weights fixed does not remove the risk of accumulating errors; it relocates that risk into the programs and memories that decide how the model will act next.
Questions, news tips or corrections? Contact [email protected].
