Certified Access: Test-Time Compute on Conservative Physics-Structured AI Memory

Pratik Jawahar, Riccardo Maggioni, Jonas Petersen, Matteo Cozzi, Gian-Alessandro Lombardi, Marek Gajewski, Camilla Mazzoleni, Federico MartelliarXiv 2026 · arXiv

In the current era of "thinking" models, test-time compute (TTC)—the ability of a model to spend more time or energy on a difficult query—is becoming a cornerstone of AI performance. However, most TTC mechanisms like retries or early exits are treated as black boxes. We know they help, but we lack a structural understanding of what that extra compute "buys" us or how it impacts the model's internal stability.

A new paper, "Certified Access: Test-Time Compute on Conservative Physics-Structured AI Memory," presented at the NeurIPS 2026 Workshop on Long-Context Foundation Models, proposes a radical shift: treating AI memory as a physical system governed by the laws of dynamics to provide formal "certificates" for every extra compute step.

TL;DR

The authors introduce a framework for "certified access" in AI memory by structuring it as a Latent Dynamics Model. By using physics-inspired mechanisms like Lorentz squeezes and gated channels, they provide formal guarantees on energy conservation and information transport, enabling compute-efficient inference with mathematical stability bounds.

What Problem Does This Paper Address?

Modern deep learning models often allocate compute adaptively—for example, by taking more steps to "relax" into a solution. However, these mechanisms are typically ad-hoc additions. There is no structural account of how extra forward passes change the model's latent state or whether they might lead to instability (like the state "escaping" to infinity).

The authors address this by formalizing two types of retrieval failures:

  1. Kinematic Reach Failure: The target memory is simply too far away to reach within the model's "causal box" (a speed limit imposed by the architecture).
  2. Energetic Escape Failure: The target is nearby but trapped behind a "potential barrier" (an energy hill) that the current state doesn't have enough energy to cross.

Methods

The researchers build upon the Causal Learning Unit (CLU), a memory architecture where the latent state $(q, p)$ evolves like a physical particle under a learned Hamiltonian. To fix the failure modes mentioned above, they introduce two "certified" mechanisms:

  • Lorentz Squeeze ($S_\zeta^{(M)}$): Inspired by special relativity, this mechanism injects a bounded amount of energy to help the state "hop" over potential barriers. It acts as a hyperbolic rotation in phase space.
  • Gated Canonical Translation: This acts as a "wormhole" or non-local jump, allowing the state to bypass the causal speed limit and reach distant targets instantly.

Crucially, these aren't just heuristics. They come with a "certificate stack":

  • Jacobian Determinant ($\det J = 1$): Guarantees that the volume of information in phase space is preserved (no information is "crushed" or "blown up").
  • Bounded Energy Change: Quantifies exactly how much "fuel" is injected into the system.
  • Latch-Transport Law: Ensures that information stored in specific "protected" directions (Goldstone charges) is moved correctly rather than erased.

Datasets

The authors evaluated their framework using three distinct setups:

  1. Analytic Testbed: A custom-designed 2D and 4D symmetric double-well potential used to verify the physics certificates to machine precision.
  2. MQAR (Multi-Query Associative Recall): A standard benchmark for associative memory, used here to test "compute rationing"—how well the model can decide to stop early on easy queries.
  3. Lattice Routing Task: A task designed to evaluate the efficiency of non-local jumps across a grid-like memory structure.

Key Insights and Ideas

  • Physics as a Safety Rail: By using symplectic integrators (like velocity-Verlet), the model's "memory" behaves like a conservative physical system. This makes test-time compute predictable and mathematically verifiable.
  • Volume $\neq$ Stability: A major "aha" moment in the paper is that preserving phase-space volume ($\det J=1$) is necessary but not sufficient for stability. As shown in Figure 1, even with volume preservation, a state can escape to infinity if it jumps outside a "coercive" region of the potential. The authors show that a "certificate screen" is required to reject unsafe jumps.
Figure 1: The BIBO battery. Maximum excursion radius r∗= maxt ∥qt∥at horizon 2T, against the requested exit locus b. Exits inside the coercive component stay bounded for every arm. Beyond the coercive edge xb = 3.54 the unscreened arms grow as r∗∝T, while the screened gated channel holds at r∗= 0.09 because it refuses the jump rather than making it safe. The state-replacing map coincides exactly with the screen-ignored ablation, so the screen, not the jump, provides boundedness. Note b = 5.0: the energy change is exactly zero and the exit still escapes.
  • Transport vs. Erasure: The authors demonstrate that standard "state-replacing" maps (common in many AI architectures) effectively erase stored information by annihilating phase volume. Their gated channels, by contrast, transport the "Goldstone charge" perfectly, as seen in Figure 2(b).
Figure 2: Verification of the certificate stack on the designed testbed (Sec. 3.2). (a) Reach: landing rate against basin distance d on a symmetric double well (dim 2, 5 seeds, L = 2.5). (b) The certificate in force: outgoing against incoming Goldstone charge for 16 states drawn from the capture ball. Dashed line in (a): causal box L; grey band: crossover bracket. In (b), coset-tangent and across-coset are gated-channel jumps; the dashed line is Qout = Qin.

Key Results

  • Machine Precision Verification: On the analytic testbed, the certificates for volume and energy held to a relative error of $10^{-12}$ to $10^{-16}$.
  • Compute Rationing: In the MQAR task, the self-calibrating gate reduced the number of relaxation steps by up to $9.9\times$ for certain memory capacities while maintaining perfect accuracy.
  • Scaling Efficiency: The gated non-local routing maintained a flat FLOP count as distance increased, whereas traditional multi-hop diffusion methods saw costs grow linearly or quadratically.
  • Training Dynamics: Figure 3 shows that as the training budget increases, both storage fidelity and gated accuracy converge, though larger memory capacities require significantly more epochs to saturate.
Figure 3: Epoch-scaling frontier: CLU-EBM storage fidelity and gated accuracy against training budget for kv∈{32, 64, 96, 128}. Required epochs scale with capacity rather than saturating at a fixed ceiling, and kv32 over-trains between 2000 and 4000 epochs. Mean over 3 seeds (42–44).

Limitations and Future Work

The authors are candid about the trade-offs. Interestingly, they found that a "physics-free" MLP router actually outperformed their energy-gated version in raw accuracy and FLOP efficiency on the routing task. They argue that their contribution is the formalization and safety of the certificate stack, not necessarily setting new SOTA benchmarks.

Other noted limitations include:

  • Noise Sensitivity: The compute-rationing gate's accuracy degrades sharply when the input cues are noisy.
  • Local Selection: The Lorentz squeeze helps with "reach," but doesn't necessarily help the model choose between two nearby, equally valid attractors.
  • Future Directions: The team plans to explore "learned entrance-steering" for gated channels and testing these physics-structured memories on much larger, real-world latent potentials.

Summary updated

Abstract

Test-time compute mechanisms, such as retries, confidence-based escalation and non-local routing, are conventionally deployed as black-box additions to trained models. They typically add forward passes or halting heads, but give no structural account of what the extra compute buys or what it costs against the model's stability guarantees. We therefore propose a theoretical and empirical framework for test-time compute on a Latent Dynamics Model used as the memory. When an encoder's latent state $(q,p)$ is advanced by a damped symplectic (velocity-Verlet) step of a learned Hamiltonian, every mechanism that expands access has an exact phase-volume Jacobian, a bounded energy change and an exact latch-transport law. We split reachability failures into escape failures and reach failures. A Lorentz squeeze resolves escape failures by injecting bounded, re-absorbable energy. A gated canonical translation (a one-step channel across $C_T$) resolves reach failures at $\detJ=1$ exactly. We verify this certificate stack on a designed analytic testbed. Reaching just beyond $C_T$ with the squeeze requires energy that grows quadratically in the excess distance. The gated channel's energy cost equals the potential difference between its endpoints, which is zero on the matched channel measured here, and so does not grow with distance in this testbed. An untrained state-replacing map annihilates phase volume and erases stored charge, while the gated channel transports it. On a learned physics-structured memory, we show calibrated compute rationing and evaluate the scaling limits of non-local routing against a physics-free baseline. The main contribution is the formalization and verification of the certificate stack, not benchmark performance.
BibTeX
@misc{jawahar2026certified,
  author = {Pratik Jawahar and Riccardo Maggioni and Jonas Petersen and Matteo Cozzi and Gian-Alessandro Lombardi and Marek Gajewski and Camilla Mazzoleni and Federico Martelli},
  title = {{Certified Access: Test-Time Compute on Conservative Physics-Structured AI Memory}},
  year = {2026},
  eprint = {2610.39073},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.39073},
}
Figure