Agents no longer accept self-edits on a higher score alone: TRUST cuts accepted bad updates from 45.59% to 2.78%

The acceptance step of a self-improving agent moves from accepting any candidate with a higher validation score to a per-round test with statistical control on the share of accepted updates that are inferior. The greedy gate only checks whether a candidate beats the incumbent on a fixed validation set, and finite-sample noise plus overfitting from reusing validation data can make inferior updates look beneficial and pass. Averaged over 30 rounds and 5 seeds in six settings spanning HotpotQA, DROP and APPS with DeepSeek-V4-Flash and Qwen-3.8-Flash, TRUST lowers the mean ratio of accepted inferior updates from the greedy gate's 45.59% to 2.78% and mean maximum drawdown from 4.963 to 0.047 percentage points.

Each round's candidate acceptance becomes a hypothesis test whose testing level is set jointly by the remaining error budget and an LLM assessment of historical context. Earlier statistical gates either control false acceptance within a single round or control the familywise error rate across rounds, and the latter proved conservative in practice: SGM-B accepted no update in any run and SGM-H accepted only three updates. The test statistic is an e-value, an evidence measure accumulated by betting, with overlap correction for validation samples reused across rounds; in all six settings TRUST reaches a higher final score and higher true evolution rate than PACE, with the accepted-inferior-update ratio lower in one setting, equal in three and higher in two.

The statistical guarantees tie the acceptance rule to final performance gain, not only to a bound on false acceptance. Earlier acceptance work focused on controlling false acceptance within a round or across rounds, without connecting the testing level to a lower bound on the final gain of the whole self-improvement process. Under explicit assumptions, TRUST proves that the proportion of accepted updates causing unacceptable performance loss is bounded and derives a lower bound on final performance gain; the mean final gain across the six settings is 24.84 percentage points, the highest among the compared gates.

The demonstrated scope is HotpotQA, DROP and APPS with DeepSeek-V4-Flash and Qwen-3.8-Flash, run for 30 rounds and 5 seeds inside an editable agent harness with reasoning, routing, memory and workflow components and 20 operators, using a 200-example validation pool and 300 held-out test examples; the statistical guarantees hold under the explicit assumptions stated in the work.

The work does not address how to identify the overlap correction strength from data or how to optimally tune the other parameters; the noninferiority margin gives an interpretable loss tolerance, but the theoretical guarantee can be conservative and only partly explains practical behavior; the LLM allocator can occasionally misjudge candidate updates and choose poor testing levels.