Pith

open record

sign in
Browse

arxiv: 2606.12289 · v1 · pith:ZDVJNF2K · submitted 2026-06-10 · cs.LG · cs.AI· cs.NE

The Standard Interpretable Model: A general theory of interpretable machine learning to deductively design interpretable methods using Lagrangian mechanics

Pietro Barbiero , Giovanni De Felice , Mateo Espinosa Zarlenga , Francesco Giannini , Filippo Bonchi , Mateja Jamnik , Giuseppe Marra , Ruggero Noris This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 reserved 2026-06-27 10:49 UTCgrok-4.3pith:ZDVJNF2Krecord.jsonopen to challenge →

Figure 1
Figure 1. Figure 1: Standard Interpretable Model (SIM) yielding operational interpretability theories. ∗. Primary author. Contact: pietro.barbiero@ibm.com. †. Contributed to initial conceptualisation, technical discussions, writing process, and experiments. ‡. Contributed to technical discussions and writing… reproduced from arXiv: 2606.12289
classification cs.LG cs.AIcs.NE
keywords interpretable machine learningLagrangian mechanicsdeductive designStandard Interpretable Modelinterpretability symmetriesmachine learning theory
0
0 comments X

The pith

Lagrangian mechanics derives optimal interpretable models from user premises on interpretability

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents the Standard Interpretable Model as a theory that begins with premises defining interpretability for a specific user. These premises lead to symmetries and constraints that form a Lagrangian, with its minima representing the best interpretable models for that user. Models can be made interpretable either by tuning parameters in black-box systems or by building architectures that satisfy the constraints. This deductive approach addresses the lack of general theories in the field, aiming for consistent methods and evaluations. Readers would value it for providing a systematic foundation instead of relying on scattered techniques.

Core claim

The SIM summarises, in a set of premises, what interpretability is for a target user. From these premises, the SIM systematically derives interpretability symmetries and corresponding constraints, which shape the landscape of a Lagrangian whose minima correspond to optimal interpretable models. To reach the minima, one can either update the parameter values of an opaque model to make it more interpretable or compile constraints into an interpretable architecture.

What carries the argument

The Standard Interpretable Model (SIM) grounded in Lagrangian mechanics, which converts user interpretability premises into symmetries and constraints that define the optimization landscape.

If this is right

  • The SIM can identify limitations in existing interpretability methods such as traditional, concept-based, and mechanistic approaches.
  • It enables the design of new interpretable methods through a deductive process rather than ad-hoc development.
  • The theory informs the creation of core programming interfaces for interpretability tools.
  • It offers a structured basis for interpretability education and curricula.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Applying the SIM across different user groups could reveal how interpretability requirements vary systematically.
  • The framework might integrate with optimization techniques from physics to create hybrid interpretable-physics-informed models.
  • Testing the derived Lagrangians on benchmark datasets could quantify improvements in user-aligned interpretability.

Load-bearing premise

That user-defined premises about interpretability can be translated into symmetries and constraints within a Lagrangian mechanics formulation such that minimizing the resulting Lagrangian produces models that are verifiably optimal for the target user.

What would settle it

A user study showing that models obtained by minimizing the SIM Lagrangian do not better match the user's interpretability preferences than models from existing methods would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.12289 by Filippo Bonchi, Francesco Giannini, Giovanni De Felice, Giuseppe Marra, Mateja Jamnik, Mateo Espinosa Zarlenga, Pietro Barbiero, Ruggero Noris.

Figure 2
Figure 2. Figure 2: The Standard Interpretable Model characterises interpretable ML models through a Lagrangian L = T − V . The interpretability landscape V measures a model’s interpretability as a function of its parameters θ, where lower values of V cor￾respond to more interpretable and accurate models. The parameter dynamics T dictates how θ changes over time, determining how the landscape is explored. Applying the princip… view at source ↗
Figure 3
Figure 3. Figure 3: An example of a function f where ∇zf is not invariant to projections onto ∇zc. Symmetry II. Let Gc be the set of projections whose image is contained in the span of the concept gradients. A function f is interpretable with respect to the concept maps c if its local output variation is invariant under at least one such projection: ∃gc ∈ Gc such that gc.(∇zf) ⊤ = (∇zf) ⊤ (6) Example 2. Let z = (z1, z2, z3), … view at source ↗
Figure 4
Figure 4. Figure 4: Concept map architecturally implementing Constraint I: cw is a monotone trans￾formation of the partially ordered set {z(1), z(2), z(3)} defined by the map c [h] w . 18 [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Validation of Symmetry I. Top row: learned concept maps fitted to human￾assigned scores. Bottom row: the same predictions sorted by the human-induced ordering, where semantic preservation requires monotonicity. Right: MAE favours score fitting, while Constraint I reveals violations of the human preorder. Low prediction error alone does not imply preserved concept semantics. human-assigned scores while fail… view at source ↗
Figure 6
Figure 6. Figure 6: Validation of Symmetry II. Constraint II violation during training, measur￾ing misalignment between the prediction gradients ∇f and concept gradients ∇c. Optimising the constraint can reduce local gradient misalignment, while architec￾tural compilation satisfies the dependency by construction. 0.0 0.2 0.4 0.6 0.8 1.0 z1 0.00 0.25 0.50 0.75 1.00 z 2 DNN ∇f ∇c −25 −20 −15 −10 −5 0 5 10 15 20 0.0 0.2 0.4 0.6 … view at source ↗
Figure 7
Figure 7. Figure 7: Validation of Symmetry II. Top row: alignment between the predictor f and the concept map c. Coloured dots represent training samples. Dark curves show level sets of c, while colour gradients show level sets of f. If f depends only on c, these level sets should be aligned. Bottom row: relationship between predicted task values f(z) and concept value c(z). Local constraint optimisation can align f and c nea… view at source ↗
Figure 8
Figure 8. Figure 8: Validation of Symmetry III. Increasing the weight of Constraint III restricts the curvature of the learned concept formula, interpolating between flexible un￾constrained fitting and the architecturally compiled hypothesis space. The SIM makes reasoning complexity both tunable through optimisation and enforceable by design. 4.3 Applications to large-scale models Having validated the key symmetries of the SI… view at source ↗
Figure 9
Figure 9. Figure 9: Vision-language concept maps. Heatmaps show all pairwise comparisons of images with increasing red intensity. B/red (A/green) cells indicate that the model judges image B (A) as more red than image A (B). A semantics-preserving model should mark the upper triangle as B/red and the lower triangle as A/green. Pre-trained VLMs can produce inconsistent pairwise rankings, showing that label￾free concept annotat… view at source ↗
Figure 10
Figure 10. Figure 10: Label-free vision-language concept semantics can be perfectly fixed without training. Left: similarity matrix between test samples and prototyp￾ical examples ordered by red intensity. Right: pairwise ranking of test samples obtained by assigning each test sample to its nearest prototype. A semantics￾preserving ranking should mark the upper triangule as B/red and the lower triangle as A/green. Ordered prot… view at source ↗
Figure 11
Figure 11. Figure 11: In large-scale concept-based models, predictions depend on a small fraction of concepts. Blue: violation of Constraint II as the number of re￾tained supervised concepts increases. Predictions in Steerling depend almost exclusively on ∼ 500 concepts. Orange: the same analysis after masking all but the top-16 concept activations at test time. The concept dependence of Steerling predictions can be sparsified… view at source ↗
read the original abstract

As Artificial Intelligence models grow in complexity, interpretability has become an indispensable tool for understanding, debugging, and controlling their computations. However, interpretability lacks general theories to deductively design interpretable methods. This gap between theories and methods results in a fragmented literature and inconsistent evaluation protocols. To fill this gap, we introduce the Standard Interpretable Model (SIM), a general theory grounded in Lagrangian mechanics that enables the deductive design of interpretable methods. Specifically, the SIM summarises, in a set of premises, what interpretability is for a target user. From these premises, the SIM systematically derives interpretability symmetries and corresponding constraints, which shape the landscape of a Lagrangian whose minima correspond to optimal interpretable models. To reach the minima, one can either update the parameter values of an opaque model to make it more interpretable or compile constraints into an interpretable architecture. We empirically show that the SIM identifies and solves limitations of existing methods (including traditional, concept-based, and mechanistic interpretability), highlights underexplored research directions, and informs the design of core programming interfaces. Beyond being a research method, the deductive nature of the SIM offers pedagogical grounding for interpretability curricula and may shift the scientific community's perspective of a discipline that has long been fragmented.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces the Standard Interpretable Model (SIM), a general theory grounded in Lagrangian mechanics for deductively designing interpretable ML methods. It claims that a set of user-defined premises about interpretability can be systematically mapped to symmetries and constraints that define a Lagrangian L, whose minima (reached via parameter updates or architecture compilation) yield optimal interpretable models for the target user. The manuscript asserts that this framework identifies limitations of existing methods (traditional, concept-based, mechanistic), highlights new directions, and informs programming interfaces, while also offering pedagogical value.

Significance. If the claimed deductive mapping from arbitrary premises to symmetries, constraints, and verifiable Lagrangian minima were rigorously established with explicit general procedures and proofs, the SIM could provide a unifying framework that addresses fragmentation in interpretability research. The use of Lagrangian mechanics and Noether invariants is a potentially powerful formal tool if the translation is shown to be non-ad-hoc. However, the manuscript supplies no equations, derivations, or empirical verification, so its significance cannot be assessed beyond the level of an interesting but unsubstantiated proposal.

major comments (2)
  1. [Abstract] Abstract: The central claim that 'from these premises, the SIM systematically derives interpretability symmetries and corresponding constraints, which shape the landscape of a Lagrangian whose minima correspond to optimal interpretable models' is load-bearing but unsupported; no general procedure, example derivation, Euler-Lagrange equations, or proof is supplied showing that the minima enforce the original user premises without additional choices.
  2. [Abstract] Abstract: The assertion that 'we empirically show that the SIM identifies and solves limitations of existing methods' is presented without any data, tables, figures, or experimental setup, undermining the claim that the framework has been validated.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive report. We address the two major comments below. Our responses focus on clarifying the manuscript's scope while committing to targeted revisions for rigor.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'from these premises, the SIM systematically derives interpretability symmetries and corresponding constraints, which shape the landscape of a Lagrangian whose minima correspond to optimal interpretable models' is load-bearing but unsupported; no general procedure, example derivation, Euler-Lagrange equations, or proof is supplied showing that the minima enforce the original user premises without additional choices.

    Authors: The manuscript presents the SIM as a high-level deductive framework in which user premises are mapped to symmetries via Noether's theorem, with constraints then shaping the Lagrangian. Section 3 outlines the general procedure conceptually, but we agree that an explicit worked example with Euler-Lagrange equations and a verification that minima recover the premises is absent. We will add a self-contained derivation example (including the relevant equations) in the revised manuscript to make the mapping rigorous and non-ad-hoc. revision: yes

  2. Referee: [Abstract] Abstract: The assertion that 'we empirically show that the SIM identifies and solves limitations of existing methods' is presented without any data, tables, figures, or experimental setup, undermining the claim that the framework has been validated.

    Authors: The empirical component in the current manuscript consists of qualitative case analyses showing how the SIM framework exposes limitations in traditional, concept-based, and mechanistic interpretability approaches. No quantitative experiments, tables, or figures are included. We acknowledge that this falls short of a full empirical validation and will expand the relevant section with concrete illustrative examples (including at least one worked numerical case) to substantiate the claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation presented as independent mapping from premises

full rationale

The abstract and description outline a deductive process starting from user-defined premises about interpretability, deriving symmetries and constraints to form a Lagrangian. No quoted equations or steps in the provided text reduce the output (minima corresponding to optimal models) to the inputs by construction, self-citation, or fitted renaming. The framework claims to systematically derive from premises without evidence of the mapping being tautological or load-bearing on unverified self-citations. This is a normal non-finding for a high-level theoretical proposal whose concrete derivations would need to be inspected in the full equations.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that Lagrangian mechanics applies to interpretability and on the invented construct of the SIM itself; no free parameters are mentioned.

axioms (1)
  • domain assumption Lagrangian mechanics can be used to derive optimal interpretable models from user-defined premises about interpretability
    Invoked in the abstract as the mathematical foundation for the entire theory
invented entities (1)
  • Standard Interpretable Model (SIM) no independent evidence
    purpose: General theory enabling deductive design of interpretable methods
    Introduced in the abstract as the core contribution that summarises premises and derives constraints

pith-pipeline@v0.9.1-grok · 5795 in / 1343 out tokens · 32864 ms · 2026-06-27T10:49:21.553628+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 12 canonical work pages · 5 internal anchors

  1. [1]

    Barbiero, M

    36 P. Barbiero, M. E. Zarlenga, A. Termine, M. Jamnik, and G. Marra. Foundations of interpretable models.arXiv preprint arXiv:2508.00545, 2025. 11 F. Barez, T.-Y. Wu, I. Arcuschin, M. Lan, V. Wang, N. Siegel, N. Collignon, C. Neo, I. Lee, A. Paren, et al. Chain-of-thought is not explainability.Preprint, alphaXiv, page v1, 2025. 28 B. Bilodeau, N. Jaques, ...

    work page arXiv 2025

  2. [2]

    Sparse Autoencoders Find Highly Interpretable Features in Language Models

    34 H. Cunningham, A. Ewart, L. Riggs, R. Huben, and L. Sharkey. Sparse autoencoders find highly interpretable features in language models.arXiv preprint arXiv:2309.08600, 2023. 30 G. De Felice, A. C. Flores, F. De Santis, S. Santini, J. Schneider, P. Barbiero, and A. Ter- mine. Causally reliable concept bottleneck models.Advances in neural information pro...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    3 J. Feng, A. Kothari, L. Zier, C. Singh, and Y. S. Tan. Bayesian concept bottleneck models with llm priors.Advances in Neural Information Processing Systems, 38:93889–93920,

  4. [4]

    Fischer, M

    26, 36 M. Fischer, M. Balunovic, D. Drachsler-Cohen, T. Gehr, C. Zhang, and M. Vechev. Dl2: training and querying neural networks with logic. InInternational Conference on Machine Learning, pages 1931–1941. PMLR, 2019. 34 T. Freiesleben and G. K¨ onig. Dear xai community, we need to talk! fundamental miscon- ceptions in current xai research. InWorld confe...

    work page arXiv 1931

  5. [5]

    3, 25, 28, 33 D. Gunning. Explainable artificial intelligence (xai).Defense Advanced Research Projects Agency (DARPA), nd Web, 2(2), 2017. 35 D. Gunning, M. Stefik, J. Choi, T. Miller, S. Stumpf, and G.-Z. Yang. Xai—explainable artificial intelligence.Science robotics, 4(37):eaay7120, 2019. 35 S. Guo and B. Sch¨ olkopf. Physics of learning: A lagrangian p...

    work page arXiv 2017

  6. [6]

    11 M. R. Hestenes. Multiplier and gradient methods.Journal of optimization theory and applications, 4(5):303–320, 1969. 48 X. Hu, C. Rudin, and M. Seltzer. Optimal sparse decision trees.Advances in neural information processing systems, 32, 2019. 29 Y. W. Jie, R. Satapathy, R. Goh, and E. Cambria. How interpretable are reasoning ex- planations from prompt...

    work page internal anchor Pith review Pith/arXiv arXiv 1969

  7. [7]

    36 F. Klein. A comparative review of recent researches in geometry.Bulletin of the American Mathematical Society, 2(10):215–249, 1893. 3 M. J. Kochenderfer and T. A. Wheeler.Algorithms for optimization. Mit Press, 2019. 16 P. W. Koh, T. Nguyen, Y. S. Tang, S. Mussmann, E. Pierson, B. Kim, and P. Liang. Concept bottleneck models. InInternational Conference...

    2019

  8. [8]

    Lagrange

    19 J.-L. Lagrange. M´ ecanique analytique. 1788. 4 Y. LeCun. The mnist database of handwritten digits.http://yann. lecun. com/exdb/mnist/,

  9. [9]

    26 C. K. Lee, M. Samad, I. Hofer, M. Cannesson, and P. Baldi. Development and validation of an interpretable neural network for prediction of postoperative in-hospital mortality. NPJ digital medicine, 4(1):8, 2021. 3 S. Lie. ¨Uber die Integration durch bestimmte Integrale von einer Classe linearer partieller Differentialgleichungen. Cammermeyer, 1880. 13 ...

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  10. [10]

    Imposing Hard Constraints on Deep Networks: Promises and Limitations

    36 P. M´ arquez-Neila, M. Salzmann, and P. Fua. Imposing hard constraints on deep networks: Promises and limitations.arXiv preprint arXiv:1706.02025, 2017. 34 G. Marra, F. Giannini, M. Diligenti, and M. Gori. Lyrics: A general interface layer to integrate logic inference and deep learning. InJoint European Conference on Machine Learning and Knowledge Disc...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  11. [11]

    31 Plato.Timaeus. 3 E. Poeta, G. Ciravegna, E. Pastor, T. Cerquitelli, and E. Baralis. Concept-based explainable artificial intelligence: A survey.arXiv preprint arXiv:2312.12936, 2023. 19 43 The Standard Interpretable Model B. T. Polyak. Some methods of speeding up the convergence of iteration methods.Ussr computational mathematics and mathematical physi...

    work page arXiv 2023

  12. [12]

    Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations

    25 M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations.arXiv preprint arXiv:1711.10561, 2017. 34 M. Ranzato, C. Poultney, S. Chopra, and Y. Cun. Efficient learning of sparse representations with an energy-based model.Advances in neural information proce...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  13. [13]

    Schubert and P

    9, 34, 35 E. Schubert and P. J. Rousseeuw. Fast and eager k-medoids clustering: O (k) runtime improvement of the pam, clara, and clarans algorithms.Information Systems, 101:101804,

  14. [14]

    Selman and H

    19 B. Selman and H. Kautz. Knowledge compilation using horn approximations. InProceedings of the ninth National conference on Artificial intelligence-Volume 2, pages 904–909, 1991. 5, 18 B. Selman and H. Kautz. Knowledge compilation and theory approximation.Journal of the ACM (JACM), 43(2):193–224, 1996. 5, 18 R. R. Selvaraju, M. Cogswell, A. Das, R. Veda...

    work page arXiv 1991

  15. [15]

    7 A. Tarsky. The concept of truth in formalized languages.Logic, Semantics, Metamathe- matics, pages 152–278, 1956. 7 S. Tull, R. Lorenz, S. Clark, I. Khan, and B. Coecke. Towards compositional interpretability for xai.arXiv preprint arXiv:2406.17583, 2024. 3 A. M. Turner, L. Thiergart, G. Leech, D. Udell, J. J. Vazquez, U. Mini, and M. MacDiarmid. Steeri...

    work page arXiv 1956

  16. [16]

    Turpin, J

    12 M. Turpin, J. Michael, E. Perez, and S. Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting.Advances in Neural Information Processing Systems, 36:74952–74965, 2023. 28 M. Vandenhirtz, S. Laguna, R. Marcinkeviˇ cs, and J. E. Vogt. Stochastic concept bottleneck models.arXiv preprint arXiv:2406.1...

    work page arXiv 2023

  17. [17]

    In practice,Q c andQ f are obtained via singular value decomposition [Eckart and Young 1936], keeping singular vectors corresponding to singular values greater than a small thresholdε >0. 47 The Standard Interpretable Model Proof.Recall that (†)x∈S⇐ ⇒P Sx=x, whereP S is the orthogonal projector onto a subspaceS; (‡) for any matrixA,∥A∥ 2 F = tr(AA ⊤); and...

    1936