Organon : From Evolutionary Search to Recursive Self-Improving Agents for Scientific Discovery

1 Introduction

• Accumulated experience can expose how search strategies, selection, memory, and resource allocation shape later experimentation, yet effective strategies may stall as search progresses.A strategy that produces rapid early gains can concentrate effort in a narrow solution family.Source: §1 · p. 2
• Prior frameworks adapt designated components or fixed update procedures, leaving their adaptive scope bounded at design time.When bottlenecks lie outside those levers or within revision procedures, these systems cannot diagnose or restructure the underlying operation.Source: §1 · p. 2
• Organon uses one unified agent to revise its search, diagnostics, and self-improvement procedures from accumulated experience.Its editable operational stack includes the evolutionary loop, selection, variation, population management, memory, tools, resource allocation, and worker coordination.Source: §1 · p. 3
• Organon’s state-preserving runtime revisions retain the population and search history, and the paper reports evaluation and analysis across 21 tasks.Capabilities can be added, replaced, or withdrawn during search while preserving accumulated state.Source: §1 · p. 3

2 Preliminaries

• The task is to maximize the best score among at most B candidates generated from a baseline program, under an official evaluator where larger scores are better.The evaluator and candidate-generation budget remain fixed throughout the investigation.Source: §2 · p. 3
• Evolutionary search selects previously evaluated programs as parents, varies them into candidates, and manages which results remain in the population.These procedures distribute effort across competing approaches and use experimental feedback to develop new solutions.Source: §2 · p. 3
• The search history records evaluated candidates, scores, feedback, and parent–offspring relationships, supporting assessment of quality and search behavior.Accumulated experience provides a basis for examining and developing the procedures that govern search.Source: §2 · pp. 3–4

3.1.1 Agent Composition and Persistent State

• The editable composition spans the evolutionary search capability and the tools, memory, worker coordination, and execution loop governing how the agent researches and improves itself.The persistent history preserves the active population, candidate records, worker findings, and revision history across search calls and revisions.Source: §3.1.1 · pp. 4–5

3.1.2 Evolutionary Search as an Agent Capability

• Each candidate attempt is a worker subagent’s development task; Organon evaluates the resulting candidate with the fixed task evaluator and updates the population and history under its survival policy.The worker receives a prompt, role, tools, skills, and relevant history context; evaluation may also include auxiliary tests or diagnostics.Source: §3.1.2 · p. 5

3.2 Recursive Self-Improvement During Search

• Organon couples candidate evolution, which updates persistent search state, with capability evolution, which revises the agent’s composition using accumulated evidence.The problem definition remains fixed across these changes.Source: §3.1.1 · p. 4, §3.2 · p. 5

3.2.1 Research Horizons

• At each research horizon, the agent allocates up to k candidate attempts before reassessing accumulated evidence and deciding whether to continue, revise capabilities, or conclude.Each allocated attempt consumes budget, including development or evaluation failures; the run returns the highest-scoring officially evaluated candidate or stops when the agent judges it complete.Source: §3.2.1 · pp. 5–6

3.2.2 From Evidence to Capability

• The agent can diagnose its operation from accumulated history and revise any part of its composition, including search, diagnostics, memory, worker coordination, and its own execution loop.For example, lineage diagnostics can inform parent selection, while recurring failures can prompt worker-instruction, checking, repair, or memory changes.Source: §3.2.2 · p. 6

3.2.3 Self-Improvement Through Composition

• Capability revisions integrate components into the live runtime through dependency-aware activation and tracked, reversible effects, with persistent history, remaining budget, and the official evaluator invariant across changes.The agent evaluates each revision over subsequent research horizons and may retain, refine, or withdraw it.Source: §3.2.3 · p. 6

3.3 Implementation Details

• The implementation exposes replaceable evolutionary-search providers within a dynamically composed runtime, and changes providers between bounded search calls while preserving search state.Candidate records retain parent identifiers, fitness scores, and survival status; the official evaluator supplies fitness, and each new candidate consumes one unit of fixed budget.Source: §3.3 · pp. 6–7

4.1 Experimental Setup

• The evaluation covers 21 tasks across algorithm design, kernel engineering, mathematical construction, and system optimization, with at most 100 candidate-generation attempts per method and task.The comparison aligns the language-model backbone and maximum candidate count; worker subagents cannot compile, execute, or test their candidates.Source: §4.1 · p. 7

4 Experiments

• ALE-Bench reports mean private performance after selection using public cases, KernelBench reports speedups relative to PyTorch eager, and mathematical results report best-found objective values.System results are means over three evaluations of each selected program.Source: §4.1 · pp. 7–8, §4.2 · p. 8

4.2 ALE-Bench Results

• Organon records the highest mean private ALE-Bench performance at 2478.6, exceeding EvoX by 20.5%, and scores highest on all ten tasks.The mean is averaged across tasks; search and program selection use public cases, while the reported performance is private.Source: Figure 3 · p. 8, Table 1 · p. 8, §4.2 · p. 8

4.3 KernelBench Results

• Organon achieves the highest KernelBench speedup on three of four workloads, with execution-time reductions up to 40.1% versus the strongest comparator on the same workload.EvoX leads on the remaining workload.Source: Table 2 · p. 9, §4.3 · p. 9

4.4 Mathematical Construction Results

• Organon leads the compared methods on third-order autocorrelation, but EvoX and ShinkaEvolve achieve stronger results on circle packing and Heilbronn.Organon comes closest to the AlphaEvolve reference on autocorrelation; its geometric-task constructions remain near the reference values.Source: §4.4 · p. 9, Table 3 · p. 9

4.5 System Optimization Results

• All compared methods improve on the cited human-best references across the four system tasks, while Organon’s clearest comparative gain is on LLM-SQL.PRISM results are tied at 26.2560 for all methods; EPLB and Cloudcast results are closely spaced.Source: §4.5 · pp. 9–10, Table 4 · p. 10

• Organon revised its composition in all 21 runs, first after a median of four candidate attempts, and no two runs changed the same capability subset.Runs used 2–22 activated revisions (median 11); variation, memory, and self-correction appeared in every run.Source: §5 · p. 10, Figure 4 · p. 10
• Search policies coevolved variation, selection, and survival, shifting from unconstrained rewrites toward specialized operators and portfolios of bounded refinement and structural redesign.Selection and survival also adapted to measured yield, candidate size, behavioral niches, ancestry, repairable failures, and archived successes.Source: §5 · pp. 10–11
• On 19 tasks, custom measurement tools informed revisions, while workflow changes on 16 tasks added such practices as compilation checks, regression repair, and concurrency tuning.Memory evolved on all 21 tasks from passive digests into active control; in circle packing, storing and verifying a numerical arrangement ended a 17-attempt plateau.Source: §5 · p. 11
• Organon corrected faulty components and revision diagnostics, and withdrew changes on 12 tasks when evidence did not support keeping them.In Cloudcast, it restored a prior component version while retaining population and diagnostic history accumulated under the replacement.Source: §5 · p. 11

• LLM-guided evolutionary methods use generated code, empirical evaluation, and reuse of candidates; later approaches adapt sampling, exploration, resources, or selection strategies.EvoX evolves parent-selection, variation, and inspiration-context strategies, while Dream-RSI replays discovery trees to refine policies for a fixed discovery agent.Source: §6 · p. 11
• Self-evolving-agent work ranges from revising prompts, harnesses, tools, and memory to searching over agent variants or recursively modifying agent code.The cited approaches include systems that make both task execution and the mechanism generating future modifications editable.Source: §6 · pp. 11–12
• Several related approaches evaluate or constrain revisions, including pairwise judgments, regularization against benchmark-specific edits, and downstream research performance under fixed budgets.These approaches differ in whether they revise an inner agent, modify their own source, or search an archive of self-modifying agents.Source: §6 · p. 12

• A single persistent agent revises search policies, diagnostic instruments, and its procedures for further revision while retaining its population and accumulated experience.Source: §7 · p. 12
• Across 21 tasks, Organon achieves strong results in algorithm design, kernel engineering, and system optimization, and remains competitive in mathematical construction.Its trajectories show accumulated evidence becoming task-specific strategies, tools, and memory.Source: §7 · p. 12
• Future work should use repeated runs and controlled ablations to isolate individual revision contributions and carry capabilities from one task into the next.The stated aim is for self-improvement to compound across problems as well as within them.Source: §7 · p. 12

A.1 Records and Analysis Protocol

• Section A.1 describes the records and how revisions are identified.Source: A Detailed Anatomy of Agent… · p. 16

A.2–A.3 Timing and Inventory

• Sections A.2–A.3 cover the timing of self-modification, search progress, and a per-run inventory.Source: A Detailed Anatomy of Agent… · p. 16

A.4–A.12 Capability Changes and Trajectories

• Sections A.4–A.12 examine capability categories, self-correction and withdrawal, and representative trajectories.Source: A Detailed Anatomy of Agent… · p. 16

A.1 Records and Analysis Protocol

• Revision analysis distinguishes proposals, definitions, activation, subsequent use, and observed consequences; only activated and used revisions count as changes.Some defined revisions were superseded or replaced before activation or use.Source: §A.1 · p. 16
• Each of 21 tasks contributes one complete Organon run, including component histories, candidate records, evaluations, tools, memory, and worker sessions.Source: §A.1 · p. 16
• Temporal association between a revision and later progress does not identify its causal contribution, so the analysis characterizes capability development rather than revision effects.Source: §A.1 · p. 16

A.2 Timing of Self-Modification and Search Progress

• In 7 runs, no candidate under the initial composition improved on the starting program, so all progress followed self-modification.The initial failures included regressions, workers missing output contracts, and kernels misrouted by the evaluation harness.Source: §A.2 · p. 16
• In a third regime, early candidates closed most of the gap and later search refined strong configurations or broke plateaus.Circle packing improved from a radius sum of 0.96 to 2.60 within three attempts.Source: §A.2 · p. 16
• Median 4 attempts preceded the first revision, and runs had a median of 11 activated revisions.Table 5 reports revision timing, activated revisions, and best recorded scores for every run.Source: §A.2 · p. 16, Table 5 · p. 17
• Heilbronn’s record held for 35 attempts before compute windows were tied to parent budgets; third-order autocorrelation’s held for 41 before workers received the strongest other candidates’ source.Source: §A.2 · p. 17
• AHC027 broke a 20-attempt plateau after selection spread beyond a saturated leader, while circle packing advanced by less than 0.0001 over 17 attempts before workers received its best arrangement as verified data.Source: §A.2 · p. 17

A.3 Cross-Task Inventory

• Every run revised candidate variation and maintained persistent memory; the inventory omits these categories from its table.Source: §A.3 · p. 17
• No two runs revised the same set of categories: AHC016 revised every category, while EPLB focused on worker directions and loop concurrency and left selection and survival unchanged.Source: §A.3 · p. 17
• LayerNorm left the loop unchanged while revising selection three times and worker guidance eight times; categories stayed unchanged when diagnoses found no limitation.Source: §A.3 · p. 17

Calibrating the scale of variation

• For AHC008, bounded edits kept the next six children within 0.55–1.35× of their parent’s score, with two setting new bests; edits over 140 lines improved in 1 of 11 attempts (mean score ratio 0.68).In LayerNorm, after large rewrites scored 17.09× and 17.63×, the agent instead required exactly one change per candidate.Source: §A.4 · p. 18

Task-specific operators

• Task-specific operators targeted observed bottlenecks: AHC011 focused on completed trees, AHC015 robustified worst-case runtime, and EPLB changed one parameter or decision rule at a time.Heilbronn added an extend contract that preserves the parent’s computation before adding work, plus an audit mode after a swallowed exception blocked a search move.Source: §A.4 · p. 18

Recombination

• Recombination varied what was transferred, from complete programs to isolated donor changes, change lists, measured-gain lists, and design reports.In AHC027, the first two attempts transplanted the same exactly scored 2-opt move into another branch and improved; in AHC026, two neutral report-based outcomes led to full-source transfer.Source: §A.4 · p. 19, Table 7 · p. 19

Validation and repair before recording

• In selected runs, AHC011, AHC026, and AHC039 compiled candidates before official evaluation and allowed one or two bounded repair attempts; in AHC039, judge-matched compiler and flags followed two zero scores traced to incompatibility.AHC026, AHC027, and AHC039 also tested candidates on small case sets, requesting repairs for illegal or crashing candidates, or those more than 5% (AHC027) or 3% (AHC039) below their parent.Source: §A.4 · pp. 19–20

Allocation from measured yield

• The agent reallocated attempts using measured operator outcomes, including increasing AHC015 hybridization from none to five of twelve planned attempts and removing VGG16 redesign after its lower mean score.In EPLB, it revised a direction-weight formula after diagnosing that the original favored directions that beat weak parents early.Source: §A.4 · p. 20

Sharpening selection pressure

• Fitness-proportional sampling was often replaced because scores differed little relative to their magnitude, with alternatives including tournaments, rank weights, elite subsets, and explicit rank plans.In AHC027, rank weighting restored the best candidate’s selection frequency from about 27% under a three-way tournament to about 58%.Source: §A.5 · p. 20

Spreading selection when concentration fails

• When five of seven AHC025 attempts went to one parent before any children were recorded and none improved it, selection began accounting for in-flight work and spreading attempts across parents.Other runs penalized recently selected parents or made broader exploration conditional on the leader stalling; in AHC008, a later breakthrough came from the leader’s own child.Source: §A.5 · pp. 20–21

Coupling parents with operations

• Parent choice was tied to the intended operation: compact programs could receive redesign, the current best small refinements, and candidates with high worst-case runtime robustification.Conversely, some runs assigned operations based on parent history or rank, so selection depended on what the next attempt was meant to do.Source: §A.5 · p. 21

Selection on attributes beyond fitness

• Selection also used attributes beyond fitness, including compactness, repeated-evaluation means, behavioral niches, simplicity, and diagnostics that explained apparent zero scores.VGG16 and LayerNorm kept some failed candidates selectable when the agent diagnosed that they contained the real designs.Source: §A.5 · p. 21

A.6 Population Management

• The agent revised population survival to preserve task-relevant diversity, ancestry, compact programs, repairable failures, reliable measurements, and near-champions.It also revived historically culled candidates when repeated measurements placed them among the top set, and re-bred a dataset-specific historical champion when the current best lagged.Source: §A.6 · pp. 21–22

A.7 Evolutionary-Loop Restructuring

• Organon treated worker coordination as a measured design choice, using concurrent batches, pipelines, streaming refill, validation, and repairs within the search loop.It also adjusted concurrency to avoid duplicated work and stale parent information, and hardened loops against setup and evaluation failures.Source: §A.7 · pp. 22–23
• VGG16 recorded a 2.4-fold increase in throughput after its first pipelined loop, while search calls lengthened as productive compositions emerged.Across 21 tasks, the agent made a median of 13 calls (range 2–18), and 196 of 244 calls after the first followed at least one revision.Source: §A.7 · p. 22

A.8 Worker Specialization and Model Allocation

• Worker specialization used exploration and engineering personas, implementation-family awareness, and selective model routing, which sometimes produced records but could also add cost without improving scores.Third-order autocorrelation routed 25 of 100 workers to DeepSeek-V4-Pro, and its final best came from such a worker.Source: §A.8 · pp. 23–24

A.9 Memory

• Memory evolved from stored reports into task-specific context and control signals for selection, assignment, recombination, and preserving solutions.The agent also recorded its own component behavior: VGG16 used timestamped loop logs to identify idle time and motivate streaming refill, while AHC026 added worker timing logs to inform concurrency decisions.Source: §A.9 · pp. 24–26

A.10 Measurement and Diagnostic Capabilities

• The agent built diagnostics to inspect evaluation context, estimate noisy scores, test feasibility and generalization, and detect when objectives were near saturation.For example, it found import-format errors that caused KernelBench candidates to fail, and redirected search when PRISM found identical scores across 50 evaluation instances and 250 additional instances.Source: §A.10 · pp. 26–27

A.11 Self-Correction, Withdrawal, and Recursive Revision

• Every run included at least one correction to the agent’s own instruments or revisions, and twelve runs explicitly withdrew, deactivated, or reverted an earlier change.Corrections to diagnostic instruments changed later decisions, including a revised AHC011 regression check that detected a child losing twelve completed trees relative to its parent.Source: §A.11 · pp. 27–28

A.12 Representative Trajectories

• Across five representative trajectories, Organon changed its selection, memory, and search procedures in response to observed failures and progress.AHC008 advanced from 3.67 to 8.88 million (2.42× the starting score); circle packing reached 2.635314 after storing and reusing a verified arrangement; AHC026 finished at 9,488.82 against 9,258.80; AHC027 reduced recorded cost by 14.9%.Source: §A.12 · pp. 28, 30
• In LLM-SQL, the recorded score rose from 0.676 to 0.737 as the agent introduced behavioral niches, opportunity-based roles, and archive rebreeding.It used per-dataset profiles and gaps to select donors, then coarsened niches when near-duplicates persisted; it corrected repair to target the largest recoverable gap and recovered a culled starting program. A sign error in comparing speed opportunity went undetected, though speed-oriented recombination remained in the portfolio.Source: §A.12 · pp. 30–31
• Heilbronn’s recorded minimum triangle area reached 0.0365035 after the agent revised computation budgets, inheritance, and its budget reader.Escalating computation windows brought evaluations near the time limit without a record, so the agent reduced them; after correcting budget-reader errors, each correction preceded a new record, and later directions targeted inert operators and overly coarse search steps.Source: §A.12 · p. 31
• ConvDiv’s recorded speedup reached 5.93× after progressing from 1.13× through 3.29×, while LayerNorm produced a 15.98× kernel after harness-aware changes.The agent addressed invalid evaluations, fallback execution, tuner duplication and timing, and regression from tuner rewrites; it also ranked ConvDiv candidates by repeated means after measuring a ±5% spread between identical evaluations.Source: §A.12 · p. 31
• Cloudcast cost fell from 1035.1 to 618.1 within 24 attempts, while PRISM’s feasibility analysis led the agent to conclude its objective was saturated.After building per-configuration analysis, held-out configurations, bounds, and a bottleneck analyzer, the agent tried twelve optimality-verification attempts and reverted when they all tied the best; for PRISM, it refined an implausible feasibility result by bisection before redirecting search toward simpler, safe programs preserving the score.Source: §A.12 · p. 31

A.13 Scope of Interpretation

• Across one run per task, revisions were not isolated, so improvements are temporal associations rather than evidence that particular changes caused them.Some revisions contained undetected errors, such as the LLM-SQL sign error and arithmetic inconsistencies in Cloudcast instructions; others were never activated or used. The agent nevertheless repeatedly revised capabilities from measured evidence, monitored behavior, and corrected or withdrew revisions, including diagnostic procedures.Source: §A.13 · pp. 31–32