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.
- An Opus-evolved harness starts at 68.41 on Gemini (agentic workspace domain), below the plain seed harness's 73.89 on the same model. Moved without its history, the inherited artifact is worse than the vendor seed.
- A source-evolved text-classification harness (seed FewShotAll) starts target optimization at 72.92, against the seed's 75.72.
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:
- The evaluator runs the current harness on the evolve set and returns scores, traces, and errors.
- The proposer, an LLM acting as harness engineer, generates k candidate harnesses from the harness, the feedback, and the run history.
- 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:
- The modification and its rationale.
- Predicted and observed outcomes. Which tasks the edit was predicted to fix, recorded before evaluation, and which it actually fixed or missed, plus pass counts.
- Provenance: domain, benchmark, policy model, round.
Operating rules that make it work:
- Proposer and selector both receive the complete memory in context every round. There is no retrieval and no truncation. Each new proposal must cite an existing root cause or name a new one.
- After selection, only the chosen candidate is recorded, with its measured outcome, even when the edit regresses or fails to fix a predicted task. Failed predictions stay marked unresolved in the entry.
- Entries with identical root-cause text are merged without dropping histories. If no edit is accepted, the memory is unchanged.
- In CRSI, the target run starts from the source harness and its memory, and keeps appending under the same schema. The proposer and selector can retain, revise, replace, or remove inherited modifications while the records persist. So the memory outlives the edits it describes. When a later round replaces a mechanism, the earlier attempt and its unresolved predictions stay on record.
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.
- Cold Start: seed harness, no memory.
- Harness Transfer: source harness, no memory.
- Problem-Only Memory: source harness plus root causes only, with no attempts or outcomes.
- Full SIM: source harness plus the complete record.
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:
- Coding: Terminal-Bench 2.1 to Terminal-Bench 4 (containerized shell tasks, scored by their own tests, pass@1).
- Text classification: a 3-dataset source suite to an 8-dataset target suite absent from source optimization (pooled accuracy).
- Agentic workspace: Harvey LAB to the equal-weight mean of JobBench, GDPval, and APEX-Agents, scored by benchmark rubrics and win-rates with Gemini 3.8 Flash as LLM judge.
- Engineering design: EngDesign to Frontier-Eng, frozen verifiers.
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:
- +2.43 to +4.35 points over Harness Transfer.
- +0.20 to +6.06 over Problem-Only Memory.
- 0.20 to 4.35 over the strongest alternative in each setting.
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.
- Scores. The mean start was 18.36 (held-out TB4). The final was 21.70 with SIM against 19.33 without.
- Test statistics. Mean paired advantage +2.36 points, 95% CI [1.05, 3.68], paired t(9) = 4.06, p = 0.003. The signed-rank check gives p = 0.006, and the standardized effect is d_z = 1.28.
- Pairs. SIM wins 9 of 10 pairs. The single loss is -0.61; the largest win is +5.46.
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 completion gate. Source optimization from Terminus-2 built an execution-gating interlock: block completion when commands were discarded, unexecuted, or errored. With SIM, the target proposer reads the gate's rationale, retains it, and extends it with a target-specific insight: passing a feasibility checker is not the same as optimality. The example is a photonic-waveguide routing task whose checker passes any feasible route while the score is route cost. Without SIM, the same inherited gate is edited four times, twice to repair completion deadlocks the run itself created, and its round-3 rationale credits the inherited gate to the run's own first round. Same artifact, same budget. The only difference is whether the optimizer knows why the gate exists.
- Failed predictions are kept. Two accepted target edits aimed at the feasibility-versus-optimality problem. Each predicted fixes for two tasks (photonic-waveguide-routing, production-planning). Neither fixed either task: 0 of 4 predictions came true. Both attempts stay in memory with their predictions unresolved. Note who accepted them: the selector, on aggregate effect.
- Diagnoses accumulate evidence. One root cause, "deliverable internally correct but violates the verifier's external contract," carries three attempts under one entry: a source attempt (fixed 1 of 4 predicted tasks) and two target attempts (0 of 2, then 1 of 2). Later rounds can see which remedies were tried, where, and what remains unresolved.
- The transfer taxonomy.
- Infrastructure transfers directly: the shell and process layer, indexing, query sanitization.
- Mechanisms transfer with recalibration: interception heuristics, retrieval budgets, intervention thresholds. Recorded zero-score rates differ by policy, 0.09% for Opus against 4.49% for Gemini, so thresholds are policy-specific.
- Skeletons transfer and get extended: the completion contract, the routing skeleton.
- Some components depend purely on the setting. Examples are parser repairs tuned to one model's output errors, a SMILES normalization that never fires on the target suite, and refusal handling carried forward with no tasks that trigger it. The authors note that the presence of such a component alone shows neither benefit nor loss.
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
- Negative transfer of bare artifacts. Evolved harnesses start below seed in a new setting: in every cross-model domain tested, and in at least one cross-dataset case.
- Diagnosis without evidence misleads. Problem-Only Memory underperforms even Harness Transfer in some settings (-1.82 in the worst case). A root-cause label carried without attempts and outcomes can anchor the optimizer on the wrong fix.
- Self-inflicted repair loops without memory. The no-memory continuation keeps re-editing the inherited gate and repairing deadlocks of its own making, while misattributing provenance. The cost is wasted rounds and corrupted rationale.
- Accepted edits that do not work. The Appendix F edits above fixed none of their predicted tasks. Selection validates a candidate's aggregate effect, not its diagnosis.
- Inheritance can lose to restarting. In the Gemini 4 Argon generation, CRSI scored 94.8 against 95.6 fresh. On the NoMemory seed, Cold Start gained more than Full SIM even while finishing lower.
- Judge bias. Workspace scores come from an LLM judge (Gemini 3.8 Flash), and the authors' ethics statement says they may reflect the judge's biases.
What a careful reader should NOT conclude
- Not "+2.36 points is what SIM buys you." That figure comes from one paired study, in coding, Terminal-Bench 2.1 to 4, from low starts (about 18 to 22 points). Cross-model margins were much larger, and they were single runs. The honest summary: positive everywhere tested, magnitude depends on the setting, and statistically established only in coding, for one transition.
- Not "always carry the harness forward." The paper's own tables contain three counterexamples: the Argon generation, the FewShotAll start, and the NoMemory gain. The claim is that carrying the record with the artifact beats carrying the artifact alone, as a continuation procedure. It is not a guarantee for every transition.
- Not "the record replaces the artifact." Full SIM carries both. The paper never tests the record without the harness.
- Not "memory helps because it is more context." Problem-Only Memory is also more context, and it sometimes hurts. The active ingredient is attempt-level evidence: rationale plus pre-registered prediction plus observed outcome plus provenance.
- Not "the memory replaces evaluation." SIM shapes proposal and selection. It does not replace held-out evaluation, which stays outside the loop. Remove the quarantine and SIM becomes a channel for overfitting with better documentation.
Limitations
The authors concede these:
- Independent repetitions exist only for coding. The other three domains have one run per condition.
- All settings require compatible harness interfaces. CRSI does not address a harness rewrite that changes the interface itself.
- The optimizer's planning and search procedures are held fixed. SIM is evaluated as a complete procedure, so carrying memory and continuing to update it are confounded by design.
- The transfer sequences studied are short: one transition, and one four-generation chain.
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:
- Negative transfer is documented but unbounded. The paper shows transplanted harnesses starting 0.60 to 5.48 points below seed cross-model, and 2.80 points below seed cross-dataset (FewShotAll). It shows inheritance losing to a fresh restart on Argon (-0.8). It offers no detector for when an inherited artifact is about to hurt, no bound on how far below seed a transplant can start, and no stated condition under which a practitioner should not inherit. SIM recovers from every bad start in the tested settings. The paper does not show when it would fail to. In practice, measure the start before trusting the inheritance.
- Selector reliability is never tested. The selector is the in-loop gate. It is an LLM (Gemini 3.8 Flash in the main runs) that reads the full memory and decides what is accepted. The paper does not test whether its judgments are reliable, and it runs no adversarial arm with proposals designed to flatter it. The one direct exhibit is Appendix F's pair of accepted edits that met 0 of 4 predictions. A biased selector could accept regressions, and SIM would record them faithfully. TRUST addresses the statistical side of this problem (noise-driven false accepts). Nobody in the family has tested the adversarial side.
[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.