One of the most useful ways to read the recent interest in energy-based models, JEPA, transductive adaptation, and test-time training is not as a replacement of intelligence by one more fashionable technical recipe, but as evidence that a more adequate architecture must distinguish between at least two levels of competence. At the first level, a system must determine whether a local configuration is compatible, stable, or goal-aligned. At the second, more demanding level, it must determine what kind of compatibility matters in the present situation, what representation is relevant, what form of evaluation is legitimate, and which computational regime should be trusted for the current subproblem.
My claim is that the first level is where energy-based formulations and JEPA are especially illuminating, while the second is precisely where Meta-Rational Pragmatics begins. [EBM-2006] [LECUN-2022] [MRP-2026a]
Meaning as Compatibility: The Energy View
In the classical energy-based formulation, a model assigns a scalar energy to a configuration of variables, and inference consists in finding configurations with lower energy. LeCun and co-authors explicitly describe energy as a measure of compatibility: low energy corresponds to more compatible configurations, high energy to less compatible ones.
This is already important philosophically. It shifts the center of gravity away from exhaustive reconstruction and toward structured acceptability. The question is no longer only “What is the exact output?” but “Which candidate state best fits the observed constraints?” [EBM-2006]
JEPA and the Relocation of Prediction
JEPA sharpens this move. In LeCun’s 2022 programmatic paper, JEPA is presented as a non-generative architecture that predicts the representation of a target rather than the target itself, with the energy given by prediction error in representation space. The 2023 I-JEPA paper makes the same idea concrete: from a context block, the model predicts the representations of target blocks, and the design is intended to force the learned representation toward more semantic structure rather than pixel-level reconstruction.
This is the decisive point. JEPA is interesting not because it abandons prediction, but because it relocates prediction to a more abstract and potentially more relevant level. It does not try to say everything about the world. It tries to say what matters for the present dependency. [LECUN-2022] [I-JEPA-2023]
The MRP Generalization: Generalizing Energy Upward
From an MRP perspective, this is already close to the right intuition, but it is still not yet meta-rational enough. JEPA provides a machinery for local compatibility in a chosen representation space. MRP asks a prior question: how was that representation space chosen, under what policy of interpretation, for which bounded objective, and with which closure condition?
The official MRP formulation explicitly describes the framework as a missing layer between language and execution and defines MRP-VM as a runtime for governed interpretation, explicit routing between execution regimes, auditable frames, and bounded active theory frames. In that sense, MRP does not reject the energy view. It generalizes it upward. It treats energy minimization as one useful local discipline inside a larger architecture that must also decide what counts as a valid local discipline in the first place. [MRP-2026a] [MRP-2026b]
The Substantive Meta: Configuring the Regime
This is where the “meta” becomes substantive rather than rhetorical. In LeCun’s autonomous intelligence architecture, the configurator prepares the perception, world model, cost, and actor for the task at hand; the world model predicts future states; the cost computes an energy; and the actor searches for an action sequence that minimizes expected cost.
That architecture is already more structured than current end-to-end next-token continuation. It contains executive control, task-conditioned perception, simulated futures, and explicit optimization over predicted trajectories. But MRP extends the same logic one level higher. It treats the selection of frame, theory, solver, validator, and admissible evidence as first-class computational objects.
In other words, the actor in the JEPA-style picture minimizes energy within a configured regime, whereas the MRP layer asks whether the regime itself is appropriate, whether it should be revised, split, suspended, or replaced by another one. [LECUN-2022] [MRP-2026a] [MRP-2026b]
Plural Energy Landscapes
For that reason, the most natural MRP reinterpretation of energy is not a single universal scalar objective. It is a family of local compatibility functions attached to different regimes. In one subproblem, what matters may be predictive consistency in latent space. In another, exact symbolic satisfiability matters more than geometric closeness. In another, provenance, evidentiary admissibility, or regulatory closure matters more than raw predictive fit.
MRP therefore does not merely ask how to descend one energy landscape efficiently. It asks which landscape is relevant, when a landscape should be combined with another, and when low energy in one regime is still unacceptable because another regime has not been satisfied. This is fully aligned with the MRP claim that semantics is often a stabilized special case of constrained pragmatics rather than the universal starting point. [EBM-2006] [MRP-2026b]
Transduction and the Discipline of Situated Closure
The connection with transduction is equally strong. In the transductive picture associated with Vapnik, the problem is not first to learn a universally valid function in the abstract and only then to apply it; rather, given training data and the concrete test cases to be solved, one seeks the best classification of those cases.
This matters because MRP is best understood as a runtime for situated closure, not as a metaphysical promise of final semantic completeness. It is often better to form a bounded active theory frame that is good enough for the present problem than to pretend that every query must wait for a globally unified theory of everything. Transduction is therefore a useful analogue for MRP: solve the case at hand under disciplined constraints, instead of demanding universal closure before local competence becomes legitimate. [VAPNIK-2006] [MRP-2026a]
Test-Time Adaptation: From Adjustment to Reinterpretation
Test-time training and test-time adaptation push the same intuition further. The original test-time training paper proposes updating model parameters on the basis of the unlabeled test sample before prediction, so that the system can adapt to distribution shift instead of remaining frozen. More recent work such as TEA explicitly reframes this adaptation through an energy-based perspective, using the energy signal to align the model more effectively with the test distribution.
Technically, this is not yet MRP. But architecturally it points in the same direction. It recognizes that deployment is not the end of intelligence. Encountering the present case may legitimately modify the local decision process. [TTT-2020] [TEA-2024]
In MRP terms, this corresponds to the possibility that an active theory frame is not static. If reality resists the current interpretation, the system should not merely insist harder. It should be able to suspend, invalidate, and replan. A rise in incompatibility is not only a signal that the current action was poor. It may be a signal that the current framing of the problem was inadequate.
A non-meta system says: “my action sequence was wrong; adjust the policy.” A meta-rational system can also say: “my regime for interpreting the terrain was wrong; revise the frame, activate a different local theory, and only then re-optimize the policy.” This is precisely the step from local adaptation to governed reinterpretation. [TTT-2020] [TEA-2024] [MRP-2026b]
Beyond Orchestration: Disciplined Computability
This also clarifies why MRP is not just another ensemble or orchestration wrapper. The recent routing and orchestration literature shows that decomposition, routing, selective escalation, and validation can materially improve cost-quality tradeoffs. The MRP position, however, is slightly different.
It is not only that several models may be better than one. It is that many problems are not well-posed until the right regime of interpretation has been induced. The system should recursively decompose the problem until it reaches fragments for which a plausible and tractable mode of resolution exists. At that point, low-level mechanisms such as energy minimization, search, constraint solving, retrieval, or symbolic verification become appropriate. The gain is therefore not merely aggregate performance. It is disciplined computability. [MRP-2026c] [MRP-2026b]
Toward Human-Level Governance
Does this bring us closer to human-level intelligence than simple scaling of current language models? My answer is yes, but only in a qualified sense. JEPA, world models, energy-based evaluation, and test-time adaptation are closer to an architecture of situated agency than pure token continuation, because they introduce explicit internal state, simulation, task-conditioned abstraction, and local optimization over futures.
But they are still incomplete if they remain only object-level mechanisms. Human-level intelligence seems to require not only choosing good actions under a given framing, but also revising the framing itself, switching criteria, managing ambiguity, and deciding how much thought a problem deserves. That higher-order control problem is exactly where Meta-Rational Pragmatics becomes relevant.
On this reading, JEPA and energy-based methods are not competitors to MRP. They are candidate components inside it. [LECUN-2022] [MRP-2026a] [MRP-2026b]
Conclusion: The Road Beyond Scaling
Energy-based learning gives a language for local compatibility. JEPA gives a practical architecture for predicting relevant abstractions rather than raw detail. Transduction and test-time adaptation show that competence can be legitimately organized around the present case rather than only around universal prior training.
Meta-Rational Pragmatics takes the next step and asks how a system should govern the choice of frames, validators, routes, and interpreters under which those local mechanisms become meaningful. If that is right, then the road beyond today’s language models is not merely bigger models, nor merely better world models, but runtimes that can coordinate many partially adequate forms of computation under explicit meta-level control. [EBM-2006] [LECUN-2022] [TTT-2020] [TEA-2024] [MRP-2026a] [MRP-2026b]
References
- [EBM-2006] LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M. A., Huang, F. J. A Tutorial on Energy-Based Learning. 2006. Link: https://web.stanford.edu/class/cs379c/archive/2012/suggested_reading_list/documents/LeCunetal06.pdf
- [I-JEPA-2023] Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. 2023. Link: https://arxiv.org/abs/2301.08243
- [LECUN-2022] LeCun, Y. A Path Towards Autonomous Machine Intelligence. 2022. Link: https://openreview.net/pdf?id=BZ5a1r-kVsf
- [MRP-2026a] Alboaie, S. Meta-Rational Pragmatics. 2026. Link: https://agisystem2.com/MRP/meta-rational-pragmatics.html
- [MRP-2026b] Alboaie, S. Regime Selection and Tractable Computation as Regime Induction. 2026. Link: https://agisystem2.com/MRP/regime-selection-tractable-computation-regime-induction.html
- [MRP-2026c] Alboaie, S. Orchestration, Routing, and MRP-VM. 2026. Link: https://agisystem2.com/MRP/orchestration-routing-mrp-vm.html
- [TEA-2024] Yuan, Y., Xu, B., Hou, L., et al. TEA: Test-time Energy Adaptation. 2024. Link: https://openaccess.thecvf.com/content/CVPR2024/papers/Yuan_TEA_Test-time_Energy_Adaptation_CVPR_2024_paper.pdf
- [TTT-2020] Sun, Y., Wang, X., Liu, Z., Miller, J., Efros, A. A., Hardt, M. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts. 2020. Link: https://proceedings.mlr.press/v119/sun20b/sun20b.pdf
- [VAPNIK-2006] Vapnik, V. Empirical Inference Science. 2006. Link: https://www.cs.columbia.edu/~jebara/6998/NSF-2006.pdf