Most machine learning memory optimizes compression of observations—autoencoders, predictive coding, latent state descriptions. Scientific reasoning in unknown interactive worlds requires a different substrate: transition logs, exploration graphs, and experiment records optimized for intervention, not reconstruction. This concept paper contrasts compressive observation memory with transition-centric experiment memory, explains why causal structure emerges through interaction (Pearl), situates ASRA relative to Buchanan–Ma representation learning, and shows how ASRA Phase 3 decomposes directed exploration into novelty and usefulness under a step budget—turning episodic transitions into reusable exploration graphs rather than compressed latents.
Intelligence is framed not as optimization alone but as the search for increasingly useful representations of reality. Before learning can succeed, a system must discover appropriate state spaces, action semantics, evaluation criteria, and ontologies. This concept paper argues that prompts, benchmarks, world models, and ontologies are central substrates of intelligence, and introduces ASRA as a representation-first framework that infers semantic operators from observed transitions before constructing causal world models.
A step-by-step guide to how adaptive systems discover what actions mean from state transitions alone—without predefined action labels—using the ASRA transition-centric reasoning framework.
Scientific intelligence increasingly depends on systems that reason from interventions rather than merely fit observations. This review synthesizes conceptual foundations from model-based reinforcement learning, active inference, causal inference, information theory, and perturbation biology into a unified architecture-level view of adaptive scientific reasoning under uncertainty.
Adaptive Scientific Reasoning Architecture (ASRA) applied to decision biology: perturbation–response reasoning, world models, and intervention-centric scientific intelligence. Full text available as PDF (versions 1 and 2).
Phase 9 integrates Phases 1–8 into asra-v1.0-phase9: the final ARC Prize 2026 competition agent, unified architecture narrative, evaluation report framing, and Decision Biology extension section. The article presents integration theory and communication deliverables; a companion Kaggle notebook submits the v1.0 agent.
Phase 8 extends ASRA's transition-centric cognitive stack from grid worlds to perturbation–response biology: cell-state hashing, perturbation semantics, pathway hypotheses, and a biological transition graph on LINCS, OmniPath, scPerturb, and HCA context. The article presents theory and architecture; a companion Kaggle notebook deploys the asra-v0.9-phase8 bridge agent for ARC Prize 2026.
Phase 7 adds the Robustness & Generalization layer: failure analysis, Procgen/DMLab generalization benchmarks, memory mismatch and stuck detection, action waste reduction, and an evaluation dashboard—wrapping Phase 6 planning with self-monitoring. The article presents theory and architecture; a companion Kaggle notebook deploys RobustnessEngine guards for ARC Prize 2026.
Phase 6 adds the Planning & Strategy Invention layer: BFS/A* and MCTS-lite over observed transitions, a strategy library mapping goal templates to operator sequences, meta explore-exploit control, and plan repair—building on Phases 1–5. The article presents theory and architecture; a companion Kaggle notebook deploys PlanningEngine hints for ARC Prize 2026.
Phase 5 adds the Goal Inference and Hypothesis Engine: win-condition templates, progress detection, object roles, hypothesis ranking, and discriminating experiment planning—building on Phases 1–4. The article presents theory and architecture; a companion Kaggle notebook deploys GoalHypothesisEngine hints for ARC Prize 2026.
Phase 4 adds the Semantics and Causal Inference Engine: semantic signatures, causal transition models, hypothesis confirm/refute, counterfactual queries, and epistemic uncertainty over (state, action) pairs—building on Phases 1–3. The article presents theory and architecture; a companion Kaggle notebook deploys CausalSemanticsEngine hints for ARC Prize 2026.
Phase 3 extends ASRA with the Navigation and Memory Engine: exploration graphs, visitation memory, novelty versus usefulness scoring, compositional subgoals, strategy reuse, and transition replay—building on Phase 1 transitions and Phase 2 object-centric observation. The article presents theory and architecture; a companion Kaggle notebook deploys compact exploration hints for ARC Prize 2026.
Phase 1 established interactive intelligence from transition evidence alone; Phase 2 adds the Observation Engine—segmentation, object alignment, transform events, and rule candidates over ARC demonstration pairs, with compact object-scene hints feeding back into the interactive agent. This article presents the theory, architecture, and design principles for object-centric structure between Phase 1 logging and later memory, causality, and planning.
Phase 1 establishes the ASRA Experience Engine: transition logging, hash-stable state IDs, effect-based action semantics inference, uncertainty-directed exploration, dead-end taboo, Swarm orchestration, and competition-grade execution fidelity. The article presents full theory and architecture; a companion Kaggle notebook deploys asra-v0.1-phase1 for ARC Prize 2026.
Unified architecture reference for the Adaptive Scientific Reasoning Agent (ASRA): nine cognitive layers from experience logging through integration and submission, with read/write contracts, end-to-end data flow, storage schemas, deployment shapes, and links to Phase 1–9 SciLayer preprints.
Technical specification for the official ARC Prize 2026 Kaggle evaluation contract: validation vs scoring rerun, gateway sidecar integration, agent registration, submission.parquet schema, template-agent requirements, and shared ASRA notebook tooling. Documents the root cause of generic Kaggle Error failures and the gateway pattern that produced the first successful submission.
Most agent benchmarks assume documented tool semantics and static ground-truth answers. Real scientific inquiry requires agents to learn what interventions do from state transitions, then infer hidden mechanisms, design discriminating experiments, and predict held-out observables. ASDB unifies two complementary tracks: Action Semantics Discovery (inferring an action map φ̂(a) from unlabeled controls) and Scientific Discovery Evaluation (recovering hidden theory classes under an intervention budget). Both share one interaction loop but score different constructs. Linked A→B episodes, decoy falsification, tiered difficulty, and decomposable metrics aim at construct validity for adaptive scientific reasoning evaluation.
Empirical results for the ASRA Observation Engine on the Original ARC corpus (800 tasks): 100% rule-candidate coverage, ~98% cross-demo common-rule consistency, transform-event distributions, training vs evaluation complexity gradient, and branched-per-demo resolution for 17 exception tasks.
Comparative analysis of Buchanan, Pai, Wang, and Ma's representation-learning textbook (arXiv:2606.06624) and the Adaptive Scientific Reasoning Architecture (ASRA). Buchanan–Ma formalizes compressive memory and white-box deep representations; ASRA formalizes intervention loops, causal semantics, goal hypotheses, and experiment design. The programs are complementary: memory theory vs scientific reasoning under uncertainty.
We present Orbit Wars Phase 4, an open-source Kaggle agent unifying Nature Foundation Models (state–action–dynamics world modeling), Atlas-GS (2D Gaussian spatial value splat for target prioritization), and ASRA (hypothesis-driven simulation with five strategic theories tested per turn). The agent operates under a one-second decision budget in a continuous 2D RTS with orbiting planets, comets, and multi-agent free-for-all play.
We present Atlas-GS v1, the first end-to-end implementation of a 3D Gaussian world-modeling pipeline within the Nature Foundation Models (NFM) hierarchy. Atlas-GS ingests RGB-D observations, constructs a persistent Gaussian field, localizes against that map, persists world state across sessions, and logs state–action–state transitions. The system implements Phases 0–6 as a modular Python package with CLI tools and demo video generation—without requiring GPU hardware for v1 validation. We report empirical results on TUM RGB-D (4,018 Gaussians, 0.0102 m localization RMSE) and synthetic sequences.
We propose Nature Foundation Models (NFM), a research program for systems that learn representations, dynamics, causal structure, and mechanisms directly from interaction with the natural world. NFM organizes scientific intelligence as a hierarchy—NFM-Worlds, NFM-Robotics, Atlas, and Atlas-GS—with a shared state–action–dynamics abstraction and a seven-stage developmental pipeline from world representation to adaptive scientific reasoning. The central thesis is that scientific reasoning should emerge from increasingly sophisticated interactions with learned world models rather than from an independent symbolic module.
A practical technical essay on two recurring AI platform patterns—enterprise data infrastructure paired with maximum research flexibility (PyTorch-class) versus the same substrate with radical simplicity (Keras-class)—and how ASRA implements both lanes on the path toward scientific intelligence platforms.