Summary (Overview)

• PRAXIST is a lineage-centered generational system for autonomous R&D that converts reproducible artifacts and evaluator outcomes into a typed evidence graph of findings, lane-structured frontiers, and agendas, enabling cumulative knowledge inheritance across generations.
• On the standardized 75-task MLE-bench suite, PRAXIST achieves 60 medals (80.0%) with 49 gold, versus 55 medals (73.3%) and 34 gold for a Claude Code baseline on Claude Opus 4.8—at a recorded model spend of US$3,054 versus US$38,370, roughly a twelfth of the cost.
• Four open-ended case studies (quantitative trading, LiDAR-inertial-visual SLAM, tokamak magnetic control, rocket landing) demonstrate the system's generality, with each producing artifacts that improve on task-native baselines in headline accuracy, survival, or resource cost.
• The rocket case study achieves 100% landing success (12,288/12,288 trajectories) versus 17.12% for Weco (an autonomous code optimizer) and 4.03% for the starting artifact, at roughly one-fifth the cost.
• The key innovation is evidence inheritance: separating local artifact construction from cohort-level evidence synthesis so later attempts inherit validated mechanisms, unresolved claims, and useful constraints, with results attached to an inspectable lineage.

---

Introduction and Theoretical Foundation

Background and Motivation

Complex engineering and scientific breakthroughs are rarely found in a single attempt; they are constructed cumulatively from validated intermediate discoveries. Recent autonomous R&D systems solve problems by constructing executable artifacts, evaluating them with task-grounded feedback, and using results to guide later attempts. However, most systems treat each attempt as nearly self-contained—logs, memories, and search trees record what happened without establishing which design element produced an improvement, whether its evidence survived validation, or how it recombines with others. Long campaigns therefore keep re-learning the same lessons.

Theoretical Basis: Assembly Theory

The paper draws on assembly theory, which characterizes complex objects through the formation histories that build them from reusable substructures, with selection determining which persist. Because every step operates on parts already retained:

• Cost scales with the number of assembly operations rather than the size of the space
• Each element is amortized across later constructions
• Search need not foresee which variant will matter—only retain the right elements in the right roles and keep them recombinable

The authors call this conversion of evaluated outcomes into actionable, artifact-grounded state evidence inheritance.

Related Traditions

The work builds on several existing capabilities:

• Reflection and memory reuse across attempts
• Multi-agent systems that distribute planning and revision across roles
• Graph structures organizing agent computation, reasoning, and memory
• Quality-diversity search maintaining coverage across high-performing regions
• Provenance systems linking outputs to the processes producing them

---

Methodology

Formalization

PRAXIST runs autonomous R&D as an artifact-to-lineage process:

$$\text{Artifact} \rightarrow \text{Finding} \rightarrow \text{Frontier} \rightarrow \text{Agenda} \rightarrow \text{Lineage}$$

The state inherited into generation $g$ is:

$$\mathcal{S}_g = (\mathcal{F}_g, \mathcal{A}_g, \mathcal{G}_g, \mathcal{L}_g) \tag{1}$$

where $\mathcal{F}_g$ is the frontier, $\mathcal{A}_g$ is the agenda, $\mathcal{G}_g$ is the set of Gems (durable cross-generation lessons), and $\mathcal{L}_g$ is the lineage trace.

Key Components

1. Deep Innovation Gate (DIG): A pre-artifact gate requiring peers to inspect inherited state, identify a testable mechanism, and fix the intended intervention before construction begins. The resulting contract specifies mechanism family, intervention surface, parent lineage, evidence signature, validation hook, and forbidden changes.

2. Quantified Diversity (QD): A cohort-level allocation rule distributing contracts across distinct design cells:

$$c = (\text{mechanism family}, \text{intervention surface}, \text{intent}) \tag{2}$$

3. Findings: The reusable unit of knowledge, defined as:

$$\phi = (\text{intervention}, \text{outcome}, \text{evidence}, \tau, m, \alpha) \tag{3}$$

where $\tau$ is the finding type (positive, negative, diagnostic, uncertain, procedural), $m$ is evidence maturity, and $\alpha$ is the recommended inheritance action (reuse, validate, avoid, diagnose, preserve, archive).

4. Frontier Lanes: The frontier is partitioned into four lanes by operational role:

$$\mathcal{F} = \mathcal{F}^{\text{cf}} \cup \mathcal{F}^{\text{cd}} \cup \mathcal{F}^{\text{dg}} \cup \mathcal{F}^{\text{vl}} \tag{4}$$

• Confirmed ($\mathcal{F}^{\text{cf}}$): mature enough to serve as a parent or constraint
• Candidate ($\mathcal{F}^{\text{cd}}$): promising but immature evidence
• Diagnostic ($\mathcal{F}^{\text{dg}}$): failures, controls, and failure modes
• Validation ($\mathcal{F}^{\text{vl}}$): evidence scheduled for reproduction or ablation

5. PI/Chair Synthesis: A panel of Principal-Investigator roles (Builder, Skeptic, Portfolio, plus External-validity in high-stakes mode) independently interprets findings from complementary perspectives, after which a Chair merges the memos into a single agenda.

6. Memory Compression (Gems): Every $\rho$ generations, durable lessons are distilled into compact Gems—validated mechanisms, rejected assumptions, recurring failure modes, or procedural constraints—retained across reset boundaries.

7. Final Output:

$$a^{\star} = \arg\max_{a \in \text{art}(\mathcal{F}^{\text{cf}})} \text{SCORE}(a), \quad \text{output} = (a^{\star}, \mathcal{L}_G) \tag{5}$$

---

Empirical Validation / Results

MLE-bench Results

Table 2: MLE-bench Any Medal rate (%) by complexity tier and base LLM

Key findings:

• PRAXIST exceeds the Opus 4.8 sweep on every tier
• 49 of 60 medals (81.7%) are gold, versus 61.8% for Claude Code
• Recorded spend: ~US$3,054 (PRAXIST) vs US$38,370 (Claude Code)—a 12× cost reduction

Case Study: Rocket Landing

The discovered controller is deterministic (no neural networks), using rolling ZEM/ZEV guidance, a fuel-commit governor, and a closed-form box-constrained allocator. The 95th-percentile sink speed falls from 66.39 to 0.33 m/s, lateral speed from 1.41 to 0.05 m/s, and tilt from 4.56° to 0.35°.

Case Study: Quantitative Trading

• Discovered policy: recurrent LSTM actor-critic trained with PPO
• 53.07% CAGR versus 22.80% for the paired all-eligible equal-weight baseline (2.3× ratio)
• Positive in 26 of 28 quarters; beats baseline in every calendar year of the test period
• Robust to execution costs: at +50 bps per side, mean quarterly return declines from 12.28% to 7.09% but remains positive

Case Study: SLAM

• Mechanism: COVSCHED—a frame-level marginal-value scheduler and deduplicating map-admission policy for the sparse-direct visual update
• 72.4% average reduction in evaluator-captured visual-path (VIO) processing time across 14 sequences
• Mean APE RMSE: 0.0937 m (baseline) vs 0.0501 m (COVSCHED), but the paper deliberately does not claim this as an accuracy improvement due to a timestamp confound
• The supported claim: spending far less visual computation does not cost trajectory accuracy

Case Study: Fusion (Tokamak Magnetic Control)

The verdict is split: PRAXIST has higher aggregate survival and lower common-horizon tracking error, while the PCS-style baseline completes more episodes and holds the lower error on the benchmark's original full-horizon metric. The paper identifies a metric confound: the full-horizon percentile is not monotone in controller quality because longer-surviving controllers contribute more near-failure steps.

---

Theoretical and Practical Implications

Theoretical Contributions

1. Evidence inheritance as a systems requirement: The paper formulates the principle that prior evidence must be selectively retained and recombined rather than merely stored, drawing on assembly theory's description of objects built from reusable parts under selection.

2. Typed evidence over raw transcripts: PRAXIST's central design choice is that evidence is never inherited as a raw transcript or scalar score—it is assigned an explicit operational role (mature parent, fragile candidate, failure to avoid, or lesson to remember) before influencing future work.

3. Failures as first-class evidence: Negative and diagnostic findings are treated as inheritable constraints rather than discarded failures, generalizing verbal-reflection mechanisms from within-episode iteration to cross-generation inheritance.

Practical Implications

• Cost efficiency: An order-of-magnitude reduction in model spend while improving results suggests that structured evidence inheritance is more efficient than brute-force search
• Auditability: Every result arrives with a solution lineage—an inspectable account of mechanisms, controls, and failures—supporting reuse and extension by human scientists
• Generality: The system handles both competition-style ML engineering (MLE-bench) and open-ended engineering problems (rocket control, trading, SLAM, fusion)
• Applications: Model and algorithm development, controller synthesis for physical systems, simulation-driven engineering design, and quantitative strategy research

---

Conclusion

PRAXIST demonstrates that long-horizon R&D improves when a campaign inherits evidence, not just scores. The system realizes this as a generational artifact-to-lineage cycle: evaluated artifacts become typed findings, findings populate a lane-structured frontier, PI/Chair synthesis emits the next agenda, durable lessons are compressed into Gems, and the accumulated lineage both directs later attempts and documents why the final artifact is credible.

Two properties distinguish PRAXIST as a research collaborator:

1. High medal quality: 81.7% of its MLE-bench medals are gold
2. Solution lineages: Results arrive with an inspectable account of their formation

Future Directions

1. Inheritance across campaigns: Gems and lineages from one problem seeding the next
2. Extension to slower/noisier evaluators: Bringing the cycle closer to laboratory science
3. Deeper human-AI collaboration through lineages: Positioning PRAXIST as an instrument that compounds scientists' evidence rather than a replacement for them

The paper concludes: "Stronger artifacts at an order of magnitude less spend, each backed by an auditable lineage, are, to our knowledge, first brought together here: the operating profile production research requires, not the one a benchmark demonstration establishes."