Ingest: Meta-n: Recursive Self-Improvement through Emergent Depth

Type: types/ingest-report.md

Classification

An empirical research preprint proposing an agent architecture and reporting benchmark comparisons, component removals, and analysis of generated code. Authors Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa, and Dongyeop Kang are affiliated with the University of Minnesota and Seoul National University. The retained observation is the full August 2026 paper; its results are author-reported, with no independent reproduction in this ingest.

Summary

Meta-n repeatedly applies a fixed LLM procedure, Ω, to prior solver code and execution evidence, producing another wrapper containing task guidance and callable helpers. An evolutionary archive retains alternative wrapper chains; search stops by convergence or configured limits. Across eight benchmark families and two backbones, the paper reports gains within supplied evaluator, wrapper, evidence-formatting, and archive interfaces. For example, GPT-5.2 CO-Bench held-out scores reach 0.870 ± 0.011 when selecting the best chain per task, versus 0.806 ± 0.017 for one best chain and 0.702 ± 0.025 for OpenEvolve. A Gemma CO-Bench component removal reduces validation score from 0.845 to 0.751 when context passed between executing layers is removed. The useful contribution is an inspectable account of accumulated guidance, reusable helpers, and selective rollback under fixed improvement machinery. Higher wrapper depth, archive selection, and improvement of the procedure that produces future improvements remain different claims.

Quotes

No source quotes have been retained yet.

Connections Found

The source is a comparison case for machinery persisting by warrant: useful adaptation can occur while the proposer stays fixed, but success does not establish that proposer's generality or justify every protected choice. It also compares with learning inside a fixed decomposition. The experiments vary wrappers and selected execution channels while retaining the evaluator, layer interface, and evidence formatter; the reported gains concern that supplied space.

The LawBench example compares with diagnostic richness: a later layer identifies an over-prescriptive directive and retains a useful helper while reversing the directive. This is an illustration of diagnosis from code and traces, not an isolated test of diagnostic access. The large context-removal effect concerns communication between executing layers. Finally, compounding measurement limits what increasing depth and archive scores establish: the paper does not separately measure whether a retained gain makes a later improvement episode more productive.

Learning Claims (our opinion)

On the source's account, Ω reads traces and the previous wrappers, explains failures, and writes another pre-process plus helper library. At execution time, outer guidance passes through inner layers, while helpers are merged with name-based overrides. Retained chains therefore change future solver behavior without updating model weights or Ω's template. The archive also selects by scores; selection and diagnosis coexist rather than constituting one mechanism.

Against the theory-builder definition, conditions 1 and 2 are met: wrapper guidance and helpers are stated units that change solver behavior. The LawBench sequence shows condition 3 in one episode. An operative prescription for exhaustive analysis is criticized as over-constraining; a successor reverses it while preserving a label-reconciliation helper, and the reported score recovers from 0.773 to 0.833. This is criticism aimed at what a unit says, beyond ranking whole candidates by outcome, and it illustrates addressable theory. Condition 4 is met within a search: Ω writes each new layer from traces of the previous chain and its wrappers, so the reversed prescription is a new conjecture shaped by the criticism, and it is scored again. Meta-n is therefore a theory builder at the grade of rounds within one run, at the strength of the reported sequence. Each search grows wrapper chains for one benchmark family until convergence or a limit; held-out scoring reuses the selected chains, and the paper does not carry retained wrappers into a later episode, which Extractable Value 4 below proposes testing. As a separate learning claim, the benchmark gains concern the bounded search. The trace and score sequence do not isolate diagnosis as their cause or show that every generated rationale drives its associated edit.

The effective update space is broad within the wrapper contract: generated programs can branch on task content, supply libraries, and alter guidance. It does not include the benchmark's task definitions and evaluators, Ω's template, the wrapper interface, or the trace formatter. Evidence is bounded too: at most 20 traces, a failure-biased sample, truncated outputs, structured summaries from depth three, and removal of oldest code on overflow. Consequently, the paper's strict information-growth argument is stronger than the implemented input policy supports. Fixedness alone does not establish a mistake, but improvements within these boundaries do not compare their alternatives.

The source adds a useful example of criticism implemented through a new overriding layer rather than an in-place edit. It does not require changing the theory-builder definition. Its proposed numerical ranking of earlier systems at roughly 2.5 meta-levels should remain source vocabulary: nesting another wrapper and changing the procedure that constructs wrappers are different properties, and Ω remains fixed here as well.

Extractable Value

1. Selective repair can preserve a useful component while reversing its accompanying instruction. The LawBench example makes the addressability mechanism concrete within a fixed wrapper interface. It is a worked case, not a controlled estimate of addressability's benefit. [quick-win]
2. Separate build-time diagnostic access from run-time guidance. Removing inter-layer context costs 0.094 validation points in the Gemma CO-Bench setup, whereas removing libraries costs 0.020. These are channel-removal results within the supplied stack; they do not show that giving the proposer more evidence caused the gain. The distinction is directly useful when designing KB learning experiments. [quick-win]
3. Report retained per-task performance separately from one deployable procedure. The archive preserves specialists even when deeper chains regress other tasks. Its consolidation guard inherits frozen best traces for non-focus tasks, so its monotonicity concerns retained per-task results; it does not guarantee that a newly selected single chain will preserve every capability. [quick-win]
4. Test later improvement productivity separately from accumulated task competence. A matched later episode with and without retained wrappers would help determine whether their reuse makes improvement cheaper or more effective. The current depth and archive comparisons leave that causal question open. [experiment]

Limitations (our opinion)

The paper's strongest causal descriptions exceed some of its comparisons. Appendix C's GPT-5.2 depth-one condition removes context, helper injection, and stacking together, uses four runs per seed and approximately 870K tokens, and reports validation only. Its difference from the full stack therefore does not isolate pure recursion at matched compute. Likewise, separate channel removals do not establish an additive partition of gains into conditioning, code transfer, and residual machinery when interactions are possible. OpenEvolve and Gödel Agent differ from Meta-n along several design axes; their gaps cannot uniquely identify cross-task transfer or an architectural ceiling. The increased-budget Gödel Agent study stops below Meta-n's token spend.

Estimator and task-pool differences constrain the headline comparisons. Archive-best uses different winners for different tasks; not every result includes a best-single-chain estimate. AlphaEvolve Math aggregates compare differing problem pools, and the AlgoTune single-shot and agentic aggregates initially average seven and eight tasks, respectively. The matched-seven AlgoTune comparison still favors single-shot, but has a smaller gap. TerminalBench compares against Meta-n's own seeds rather than the two external systems; GPT-5.2 TerminalBench is single-seed, with variation across categories rather than repeated-run uncertainty. The prominent ARC-AGI-2 figure of 0.331 is a development archive-best score. The captured paper asserts an above-zero held-out advantage but does not supply a corresponding numerical held-out result.

Layer-role analysis is descriptive and partly scaffolded: the evidence formatter switches attention toward stack-level summaries at depth three. LLM raters have weaker agreement on abstract roles, with mean Cohen's κ of 0.59. Neither these labels nor deeper task wins establish a hierarchy of increasingly capable improvers. A simpler account remains compatible with much of the evidence: repeated guided program search, reusable task helpers, and per-task selection improve a bounded solver family.

Holding Ω fixed and checking generated code prevent some forms of corruption; they do not establish unconditional stability or prevent harmful guidance. The paper itself reports layer interference and regressions. No implementation was inspected or executed for this ingest, so safeguards, mechanisms, and outcomes remain supported by the paper rather than independent code verification.

Recommended Next Action

Review Diagnostic richness constrains outer-loop learning quality for one bounded addition distinguishing proposer evidence from guidance passed between executing layers, using Meta-n's rollback example and context-removal result without treating the latter as a diagnostic-access ablation.