MIT's Self-Revising AI: A Category Theory Framework for Scientific Discovery

Forum topic · 小凯 · 2026-06-14

Summary

A forum post on zhichai.net discusses an MIT paper by Fiona Y. Wang and Markus J. Buehler (arXiv:2606.01444) that builds a mathematical foundation for AI-driven scientific discovery using category theory. The core insight distinguishes 'searching'—finding answers within a fixed vocabulary of concepts—from true 'discovery,' which changes the vocabulary itself by introducing new variables, operations, or validators. The framework represents knowledge states as copresheaves, transfers old evidence across regime shifts via left Kan extensions, and measures genuinely new content through 'discovery residuals.' An adversarial Builder/Breaker architecture with minimum description length (MDL) gating accepts revisions only when they improve compression efficiency. In protein mechanics, only 25 of 388 candidate laws (6.4%) were accepted, culminating in a 'modal-conditional compliance' law; in fiber network mechanics, an orientation-tensor anisotropic stiffness model won with ΔAIC = 123.87. Notably, non-monotonic R² across iterations is presented as evidence of healthy generalization rather than failure, since discovery involves deletion and compression, not just accumulation.

MIT researchers Fiona Y. Wang and Markus J. Buehler (Laboratory for Atomistic and Molecular Mechanics) propose a categorical framework for agentic AI scientific discovery in their paper *Self-Revising Discovery Systems for Science* (arXiv:2606.01444, 2026-05-31).

The Core Problem: Search ≠ Discovery

Existing AI discovery systems (ProtAgents, Sparks, SciAgents) operate within fixed representation frameworks. But real discovery—Newton introducing "force," Einstein introducing "spacetime curvature"—means changing the vocabulary itself: new effective variables, new admissible operations, new validators or tool types.

Categorical Semantics

A discovery regime is a tuple $b = (\mathcal{S}_b, \Gamma_b, V_b, L_b)$ : a schema category (the "vocabulary" of artifact types and operations), operation signatures, validators/gates, and a description-length functional (Occam's razor).

Typed artifact states: copresheaves $I_t: \mathcal{S}_b \to \mathbf{Set}$
Provenance graphs: the element category $\int_{\mathcal{S}_b} I_t$

Three operations are rigorously distinguished: retrieval (adding artifacts already expressible), search (new paths within a fixed schema), and discovery (changing the regime via a schema map $u: \mathcal{S}_b \to \mathcal{S}_{b'}$ ).

Kan Extensions and Discovery Residuals

Old evidence is transferred across a regime shift via the left Kan extension:

(\text{Lan}_u I_t)(A') = \text{colim}_{(A, f: u(A)\to A')\in(u\downarrow A')} I_t(A)

If a new type $$A'$$ receives no morphisms from the image of $$u$$ , the colimit is empty—the system must actively acquire new evidence. The discovery residual

\mathcal{R}(A') = I'_{t+1}(A') \setminus \text{im}(\bar{\rho}_{A'})

measures content not explainable by transfer—genuinely new discoveries.

Builder/Breaker with MDL Gating

A Breaker selects evidence exposing model failures; a Builder proposes symbolic DAG edits. Revisions are accepted only if MDL decreases:

L(M, D) = L_{\text{model}}(M) + L_{\text{data}}(D|M)

Protein mechanics results (388 proposals, 25 accepted = 6.4% acceptance rate):

\hat{B}^{(z)}_{pi} = \alpha + \beta \cdot \phi_{pi} \psi_{pi}

with $\alpha = -0.1332$ , $\beta = 0.2239$ , $\theta = 2.2678$ —a modal-conditional compliance law: local compliance expressed through slowest collective mode participation.

| Iteration | Transition type | k | R² | MDL gain | |-----------|-----------------|---|-----|----------| | 0 | Initial | 1 | 0.48 | - | | 1 | Regime split | 2 | 0.68 | +9.0 bits | | 2 | Ontological break | 2 | 0.54 | +37.3 bits | | 3 | Regime split | 1 | 0.41 | +54.3 bits |

The non-monotonic R² is not failure: each iteration evaluates on more heterogeneous, adversarial evidence sets. Monotonic R² would signal overfitting. Feature removal was the highest-yield operation (avg. $\Delta L_{\text{model}} = -27.4$ bits)—discovery includes deletion and compression.

Fiber Network Case

Comparing scalar fiber-count descriptors vs. orientation-tensor anisotropic stiffness proxies: $$M_1$$ wins with $\Delta\text{AIC} = 123.873782$ (nematic order $$S = 0.673115$$ , stiffness $$E = 119.4$$ kPa, linear fit $$R^2 = 0.999989$$ ).

Why It Matters

1. Discovery gets a mathematical definition—auditable and reproducible, not philosophical 2. Breaker demands structural evidence (counterexamples), not just true/false labels 3. 6.4% acceptance reflects the norm for genuine discovery; Occam's razor is a discovery mechanism, not decoration 4. Mechanics constrains AI: a rigorous mapping from state/loading/response/instability to artifact state/evidence pressure/model revision/mechanism failure

Limitations & Open Problems

Category-theoretic implementation is close to but not yet software-enforced; stochastic versions need Kleisli categories or probability monads
Kan-extension audits are post-hoc, not real-time
Five open questions: convergence of transferred artifact sequences, scaling laws for discovery (regime-enlargement rate rather than fixed-mechanism performance), verification tooling for agentic loops, learning foundation mechanism categories from corpora, and multicategorical/colored-operadic discovery

> "Discovery is not an accelerated form of search. Discovery is a meta-operation on search—it changes the search space itself."

Reference: Fiona Y. Wang and Markus J. Buehler. "Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence." arXiv:2606.01444, 2026.