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

Summary (Overview)

• Novel framework: Introduces Organon, a recursively self-improving evolutionary agent that autonomously revises its own search strategies, diagnostic tools, memory, worker coordination, and even the procedures governing its own self-improvement during live execution.
• State-preserving revision: Leverages DeepSeek Harness and Cordis's spatiotemporal composability to add, replace, or withdraw capabilities at runtime without losing the accumulated population or search history.
• Strong empirical results: Achieves best performance on all 10 ALE-Bench Lite tasks (20.5% improvement over EvoX), 3 of 4 KernelBench workloads (up to 1.67× speedup over strongest competitor), and competitive results across mathematical and system optimization tasks.
• Demonstrated self-improvement: Analysis reveals the agent develops task-specific search strategies, builds new diagnostic capabilities, and corrects or reverses its own earlier revisions across all 21 tasks.
• Key insight: The agent's accumulated experience drives improvements to both the mechanisms governing search AND the procedures governing its subsequent self-improvement, enabling recursive capability evolution.

Introduction and Theoretical Foundation

The paper addresses a fundamental challenge in automated scientific discovery: while LLM-guided evolutionary search systems accumulate rich empirical histories (candidate solutions, execution profiles, failed trials), the operational mechanisms governing how they conduct research remain largely static.

The core problem: Existing evolutionary frameworks (FunSearch, AlphaEvolve, EvoX) use fixed search procedures or restrict adaptation to predetermined components. When progress stalls due to bottlenecks outside those designated levers, the system cannot diagnose the underlying cause or restructure its operation.

Limitations of current approaches:

1. Fixed-component adaptation (ShinkaEvolve, PACEvolve, AdaEvolve, EvoX): Adaptation is confined to a predetermined subset of components (population sampling, context management, exploration schedules), bounding adaptive scope at design time.
2. Unconstrained in-context micro-decisions (Autoresearch, CORAL, SwarmResearch): Ad-hoc local choices lead to cognitive biases toward familiar/leading candidates, leader-following, premature consensus, and near-greedy behavior.

Theoretical foundation: The paper invokes Francis Bacon's Novum Organum (1620): "the art of discovery may advance as discoveries advance." This motivates the central question: Can an autonomous agent's experience drive the continual self-improvement of how it discovers?

Methodology

Agent Composition and Persistent State

Organon conducts each investigation as a single continuing agent whose capabilities evolve with task experience. After $r \geq 0$ adopted capability revisions, the agent $\mathcal{A}_r$ has an active editable composition:

$$\mathcal{C}_r = \left(L_r, \mathcal{T}_r, \mathcal{M}_r, \mathcal{W}_r, \Phi_r\right)$$

where:

• $L_r$ = evolutionary search capability
• $\mathcal{T}_r$ = available tools and skills (including internal diagnostics)
• $\mathcal{M}_r$ = memory and context management
• $\mathcal{W}_r$ = worker subagent coordination
• $\Phi_r$ = agent's execution loop and instructions

The agent maintains a persistent history:

$$\mathcal{H} = \left(\mathcal{P}, \mathcal{R}_{\text{cand}}, \mathcal{F}_{\text{work}}, \mathcal{R}_{\text{rev}}\right)$$

comprising the active population $\mathcal{P}$, full candidate records, worker findings, and revision history.

Evolutionary Search Loop

The search capability $L_r$ coordinates four adaptable procedures:

$$L_r = \left(\operatorname{Sel}_r, \operatorname{Var}_r, \operatorname{Eval}_r, \operatorname{Surv}_r\right)$$

corresponding to parent selection, candidate variation, internal evaluation protocol, and population survival. The search proceeds as:

$$\begin{array}{ll} x_{\text{par}} \sim \operatorname{Sel}_r(\mathcal{H}), & x_i \sim \operatorname{Var}_r(x_{\text{par}}, \mathcal{H}; \mathcal{W}_r, \mathcal{T}_r), \\ y_i = \operatorname{Eval}_r(x_i; f, \mathcal{T}_r), & \mathcal{H} \leftarrow \operatorname{Surv}_r\big(\mathcal{H}, (x_i, x_{\text{par}}, y_i)\big). \end{array}$$

Recursive Self-Improvement

The agent operates in research horizons: at each decision point with remaining budget $B_{\text{remaining}}$, it selects a horizon $k$ (number of candidate attempts) before reassessing. After each horizon:

$$\mathcal{C}_{r+1} = \begin{cases} \mathcal{C}_r, & \text{if } \mathcal{A}_r \text{ continues under the current composition}, \\ \operatorname{Revise}(\mathcal{C}_r, \mathcal{H}'), & \text{if } \mathcal{A}_r \text{ adopts a capability revision}. \end{cases}$$

Implementation via Spatiotemporal Composability

Capability revisions use reactive coeffects (coordinating cross-component dependencies) and revertible effects (isolating component contributions):

$$e_p(\Gamma_r) = (\Gamma_{r+1}, u_p), \qquad u_p(\Gamma_{r+1}) \simeq \Gamma_r$$

This enables clean addition, replacement, or withdrawal of any component while preserving the population and history. For example, activating a new parent-selection policy $\pi_p$:

$$\operatorname{Sel}(\Gamma_{r+1}) = \pi_p, \quad \operatorname{Sel}(u_p(\Gamma_{r+1})) = \operatorname{Sel}(\Gamma_r)$$

Empirical Validation / Results

Experimental Setup

• 21 tasks across 4 categories: algorithm design (10 ALE-Bench tasks), kernel engineering (4 KernelBench workloads), mathematical construction (3 tasks), system optimization (4 tasks)
• Comparison methods: OpenEvolve, ShinkaEvolve, EvoX (run through SkyDiscover)
• Common budget: B = 100 candidate-generation attempts per task, all using DeepSeek-V4.1-Flash with high reasoning effort

ALE-Bench Results

Table 1: Private performance on ALE-Bench Lite (mean over 3 evaluations, higher is better)

Organon achieves highest score on all 10 tasks, with mean performance 20.5% above EvoX.

KernelBench Results

Table 2: KernelBench speedups relative to PyTorch eager (higher is better)

Organon wins 3 of 4 workloads with up to 40.1% execution-time reduction vs. strongest comparator.

Mathematical Construction Results

Table 3: Mathematical construction results (arrows indicate improvement direction)

Organon wins on autocorrelation but trails on geometric tasks, suggesting difficulty escaping strong local optima.

System Optimization Results

Table 4: System-optimization objectives (mean over 3 evaluations; bold = best among compared methods)

Analysis of Self-Improvement Trajectories

Key findings from trajectory analysis:

• All 21 tasks saw composition revisions (median 11 per run, first after median of 4 attempts)
• 14 of 21 tasks: ≥70% of total improvement followed the first revision
• No two runs modified the same subset of capabilities (task-specific adaptation)

Three perspectives on self-improvement:

1. Revising search policy: Coevolution of variation, selection, and survival—e.g., AHC024 raised crossover from 1/3 to 2/3 of attempts after observing +4.3 vs. +0.5 parent-relative gains; AHC026 discovered size-aware selection after finding redesign gains came only from parents <27 KB.

2. Developing investigation means: Custom diagnostic tools on 19 tasks, workflow revisions on 16 tasks, memory evolution on all 21 tasks—e.g., LayerNorm traced 7 of 8 initial failures to an indented Triton import rule (unlocking 15.98× kernel); circle packing ended a 17-attempt plateau by storing the best arrangement as numerical data (2.6287 → 2.6353).

3. Assessing own changes: Withdrawal of ineffective revisions on 12 tasks; correction of inverted survival rule in AHC025 that had retained the 5 weakest candidates; state-preserving reversal in Cloudcast restored a prior component while retaining accumulated experience.

Theoretical and Practical Implications

Theoretical contributions:

1. Recursive self-improvement paradigm: Demonstrates that an agent's accumulated experience can drive durable improvements to its own investigative process—not just to its knowledge of the target problem. The agent revises its own diagnostics, which then expose new bottlenecks, prompting further revisions.

2. State-preserving capability evolution: Shows that spatiotemporal composability (reactive coeffects + revertible effects) enables clean addition/replacement/withdrawal of capabilities during live execution without severing the link between historical evidence and subsequent evaluation.

3. Beyond fixed adaptation: Establishes that unconstrained, evidence-driven self-modification outperforms both fixed-component adaptation (EvoX-style) and unconstrained in-context micro-decisions (agent-style), avoiding both bounded adaptive scope and greedy local behavior.

Practical implications:

• Automated scientific discovery: Organon's approach could accelerate discovery across algorithm design, kernel optimization, and systems engineering by automatically developing task-specific strategies.
• Generalizable self-improvement: The recursive mechanism (improving the improvement procedures) suggests a path toward agents that compound their capabilities across problems, not just within them.
• Diagnostic-driven search: The finding that custom measurement tools (on 19 of 21 tasks) determined what to revise next highlights the value of building diagnostic instruments as first-class research capabilities.

Conclusion

Organon demonstrates that recursive self-improvement during search is both feasible and effective. A single persistent agent that can revise its search policies, diagnostic instruments, and revision procedures—while preserving its population and accumulated experience—achieves strong results across diverse computational discovery tasks.

Key takeaways:

• Self-improvement is broad and task-specific: no two runs modified the same capability subset
• Evidence-driven revision (not ad-hoc) enables the agent to diagnose limitations and correct its own errors
• State-preserving reversal allows rejecting operational hypotheses without discarding accumulated experience

Future directions:

1. Isolate contribution of individual revisions through repeated runs and controlled ablations
2. Carry capabilities across tasks so that self-improvement compounds across problems, not just within them
3. Address remaining gaps in mathematical construction (escaping strong local optima)