Paper 2: Self-Improvement Memory

Xin Yu, Fei Xia, Jiaqi Pan, Youbao Tang, Chengzhi Liu, Adams Yu, Sicong Jiang, Yanfei Chen, Lingzhou Xue, Ting Yu (Google Cloud AI, Google DeepMind, Penn State), arXiv:2611.18856, submitted Nov 13, 2026.

Evidence on this page: paper-reported (checked) unless labeled. The main text, Tables 1 to 3, Limitations, Ethics, and Appendix F were read in full; Appendices B to E were skimmed, not line-read (see the flag at the end).

In one breath. Harness optimization usually happens in one fixed setting: one model, one benchmark. Real deployments break both. When optimization resumes in a new setting, the optimizer inherits an evolved harness but not the reasons behind it: why each edit was made, what it was supposed to fix, and what actually happened. The paper's fix is a structured memory, SIM. It records every attempted edit by root cause, with its rationale, the outcome predicted before evaluation, the observed outcome, and provenance. Carrying that record with the harness beat carrying the harness alone in every cross-dataset and cross-model setting tested. The strongest evidence is one repeated study: ten independent runs, in coding, for one transition (Terminal-Bench 2.1 to 4), which gave +2.36 points (95% CI [1.05, 3.68], p=0.003, 9 of 10 pairs). The other results are single runs per condition, including cross-model margins of +3.60 to +11.33 points. The paper is just as clear about the failure it targets. An evolved harness moved without its history can start below the plain seed. And in one of its own generational tests, carrying everything forward still lost to starting fresh.

The problem it names

Both motivating measurements are negative transfer: an artifact optimized in setting A starts below the unoptimized baseline in setting B.

The paper calls this setting Continual Recursive Self-Improvement (CRSI). It means continuing an existing harness-optimization method after a benchmark change (cross-dataset: policy model fixed) or a policy change (cross-model: benchmark and split fixed), with a compatible executable harness interface preserved. The question is precise: when the setting changes, what exactly should carry over, the artifact or the reasons behind it?

The mechanism

The base loop

The base method follows Meta-Harness (see the reading path). A setting is a policy model plus a task set, split into disjoint evolve and held-out evaluation sets. A run lasts N rounds. Each round works like this:

  1. The evaluator runs the current harness on the evolve set and returns scores, traces, and errors.
  2. The proposer, an LLM acting as harness engineer, generates k candidate harnesses from the harness, the feedback, and the run history.
  3. The selector screens proposals before execution, weighing performance, complexity, domain specificity, and regression risk. It then executes the survivors and picks the next harness.

If nothing is accepted, the harness stays the same. Only evolve-set feedback may influence proposals, selection, memory updates, or stopping. Held-out scores are computed before and after the run and never enter the loop.

SIM itself

The memory is a set of entries, one per root-cause hypothesis (for example, "premature completion"). Each entry holds an ordered list of attempts, and each attempt stores three things:

Operating rules that make it work:

The pre-registration discipline deserves emphasis. It is the same move as Paper 3's frozen anchors, in a different form. A prediction written before evaluation cannot be fitted to the outcome afterward. The memory is an evidence ledger, not a highlight reel, and Appendix F (below) shows it being used as one.

The four transfer conditions

All target runs share the same policy, tasks, and round and candidate budgets.

Problem-Only Memory is the crucial control. It tests whether the active ingredient is the diagnoses or the evidence attached to them. Note what Full SIM is: the artifact and its record. The paper does not test carrying the record without the artifact.

Experimental setup

Four domains:

Budgets are matched across conditions: 10 rounds, 5 candidate proposals per round, 5 rollouts per task, all rollouts counted. Tasks are split uniformly at random, unstratified, into roughly equal evolve and held-out halves. Cross-dataset runs use Gemini 3.8 Flash as the policy in both stages. Cross-model runs optimize with Claude Opus 5.5 and continue with Gemini 3.8 Flash, on the same benchmark and split. Proposer and selector are Gemini 3.8 Flash in the main comparisons. An ablation with Opus 5.5 as proposer and selector produced larger gains from both coding seeds. Policy weights are always frozen. The paper reports budgets as rounds, candidates, and rollouts; it does not report token or wall-clock costs.

The paper's own reporting caveat, which this site keeps: apart from the paired study below, each condition is one complete run. The five rollouts estimate execution variance, not optimization variance. Cross-domain margins are therefore descriptive, and all the inferential weight sits on the paired study.

Results

Cross-dataset (six settings, one run each)

The six settings are four domains, with coding and text classification each run from two seeds. Full SIM finishes highest in all six. Its margins:

Problem-Only Memory is unreliable. It ranges from -1.82 to +2.40 relative to Harness Transfer, so knowing the diagnosis without the evidence can be worse than knowing nothing.

Source optimization itself raised held-out source scores by 2.75 to 6.41 points in all six settings, and still produced the negative-transfer start above (FewShotAll, 72.92 against seed 75.72). From that start, harness-only continuation reached 74.12, still below Cold Start's 75.92, while Full SIM reached 77.92.

On the NoMemory seed, Cold Start gained more over the run than Full SIM (+3.40 against +2.20) but finished lower (72.00 against 73.60). The paper reports both endpoints and treats Harness Transfer, which shares Full SIM's starting harness, as the direct control. Because each cell is one run, "highest in all six" is a descriptive count, not a tested claim.

Cross-model (Opus 5.5 to Gemini 3.8 Flash, four domains, one run each)

On first contact, the Opus-evolved harness scores below Gemini's own seed in every domain. The deficits run from 0.60 points (text classification) to 5.48 (workspace, the 68.41 against 73.89 case). Harness-only continuation does not recover: it finishes below Cold Start in three of four domains. Full SIM finishes highest in all four, beating harness-only continuation by +3.60 to +11.33 points. The extreme case is workspace. Both start at 68.41. Full SIM reaches 84.75, and harness-only reaches 73.42. These are the largest margins in the paper, and they are single runs with no variance estimate.

Across model generations: the honest counterexample

The chain carries the harness plus SIM forward from generation to generation, scored on held-out Terminal-Bench 2.1. "Fresh RSI" restarts from the seed each time.

Policy Seed Fresh RSI CRSI (with SIM)
Gemini 2.5 Flash 17.8 21.4 22.4
Gemini 3.5 Flash 70.2 73.3 76.8
Gemini 3.8 Flash 87.2 88.9 90.5
Gemini 4 Argon 89.7 95.6 94.8

Continuation beats fresh restarts in two of the three later generations (+3.5, +1.6) and loses on the newest model (-0.8). The paper reports this without spin: inheritance is not uniformly better than starting over. Any doctrine built on this paper has to carry that sentence with it, and the synthesis does.

A second multi-model result points the other way. It is text classification, with Grok 4.20 Reasoning, Opus 5.5, and Gemini 3.8 Flash. A single harness built by unified synthesis, using feedback from all three models, scores 75.24 held-out. That is nearly identical to per-model harnesses (75.20); both are single runs and no significance test applies. Both are above the shared seed (73.28), and sequential CRSI reaches 74.20. Harnesses evolved under one policy and then evaluated across all three average only 66.84 to 72.24. Single-model evolution generalizes worst.

The paired study: the only repeated result

Source optimization was run ten independent times from the Terminus-2 seed on Terminal-Bench 2.1. From each resulting harness came two matched 10-round Terminal-Bench 4 continuations, one inheriting that run's SIM and one without. The unit of analysis is the independent source run.

Read the conditions with the number. This is one domain (coding), one transition (TB2.1 to TB4), ten runs, from a low base: the harnesses solve roughly one task in five, so the relative gain is about 12%. The CI's lower bound is 1.05 points. The result is established, and it is narrow.

The adversarial panel (one model, see the synthesis) claimed that removing the largest pair would collapse the result. It would not. Site arithmetic, from the reported summary statistics and not a paper result: a mean of +2.36 with t(9) = 4.06 implies a standard deviation of about 1.84. Dropping the largest delta (+5.46) leaves nine pairs with a mean of about +2.02, a standard deviation of about 1.57, t(8) of about 3.85, and p of about 0.005. This assumes the reported largest delta and should be re-checked against the paper's per-pair table. The effect survives its largest pair. What it does not have is breadth. The panel's proposed minimum replication is reasonable as a target: 30 or more pre-registered paired runs in coding, at least 5 runs per condition in the other settings, and Holm correction across claims.

Speed and retention (separate traces, not the paired study)

SIM reaches the strongest control's final score by round 5 (AHE seed) and round 2 (Terminus-2 seed). It is faster, not just better.

Source retention comes from the AHE-seed trace, where all conditions start at 90.22 on held-out TB2.1. After target optimization, Full SIM has lost 3 of 225 successful source rollouts, against 4 for Problem-Only Memory and 6 for Harness Transfer. That is one trace, and the differences are a few rollouts.

Appendix F: the mechanism evidence

These cases are the paper's real argument, because they show how the record changes behavior.

The frozen check

Held-out evaluation is structurally outside the loop. Only evolve-set scores, traces, and errors reach the proposer, selector, or memory. Held-out sets are scored before and after the run and never feed a decision. Inside the loop, the selector's pre-execution screen is the gate, and predictions go into memory before evaluation, so the record cannot be fitted to outcomes afterward.

The paper's ethics statement adds the operational corollary. Harnesses run inside benchmark environments with policy weights fixed, and scores come from frozen verifiers or a fixed judging procedure. The authors acknowledge the residual risk that a self-modifying harness could satisfy an evaluator without solving the intended task. That risk is exactly why held-out feedback is quarantined.

Notice what is frozen and what is not. The held-out sets are frozen. The selector, which makes every in-loop acceptance, is an LLM whose reliability the paper does not measure.

Failure modes the authors report

What a careful reader should NOT conclude

Limitations

The authors concede these:

Two more limits that the paper does not frame as limitations. Both come from the adversarial panel's valid points and are checked against the paper's own numbers:

[UNVERIFIED: Appendix B's exact per-domain task counts and Appendix E's full score tables were not read line by line for this site; every number on this page comes from the main text, Tables 1 to 3, and Appendix F.]

Next: Paper 3 turns the same skeleton on the measuring instrument itself.