🤖 machine learning

LAWFUL: Law-Aligned Witness for Faithful Use of Latents

The paper introduces LAWFUL, a foundational framework that addresses key interpretability gaps to verify whether neural networks not only learn but also internally utilize formal physical laws, demonstrated by confirming the Mocap2Radar transformer's use of the Doppler frequency law despite its absence from the training data.

Original authors: Kevin Chen, Kenneth W. Parker, Anish Arora

Published 2026-08-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kevin Chen, Kenneth W. Parker, Anish Arora

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a magician pull a rabbit out of a hat. You know the trick works because the rabbit appears, but you don't know how it happened. Did the magician actually learn a secret spell, or did they just memorize a specific hand motion that happens to work for rabbits? This is the big question scientists are asking about artificial intelligence today. We have built computer brains (neural networks) that are incredibly good at predicting how the physical world works—like how a ball bounces or how sound waves travel. But are these computers actually "learning" the laws of physics, or are they just guessing really well based on patterns they've seen before?

To understand this, we need to look at two things. First, there's the idea of a "law," which is a strict rule the universe follows, like gravity always pulling things down. Second, there's "interpretability," which is the art of peeking inside the computer's brain to see what it's actually thinking. Usually, when we try to peek inside, we only see if the computer gets the right answer. But that's like checking if the rabbit is there without checking if the magician actually used magic. The big challenge is figuring out if the computer is using the real rule or just a lucky guess, especially when the rules involve things that can change smoothly and endlessly, like speed or time, rather than just simple "on/off" switches.

This is where a new framework called LAWFUL comes in. Think of LAWFUL as a super-strict detective for computer brains. Instead of just asking, "Did you get the right answer?" it asks, "Did you use the right logic to get that answer, even when we trick you with new situations?" The researchers built a system to test if a computer model truly understands a specific law of physics: the Doppler effect. You know this law from when an ambulance drives past you; the siren sounds high-pitched as it comes toward you and low-pitched as it moves away. The law says this pitch change depends exactly on how fast the ambulance is moving.

The team tested a smart computer model that was trained to predict this pitch change using video of a person moving (like a dancer with glowing dots on their joints) and radar data. Here's the tricky part: the computer was never told the speed of the person or the pitch of the sound. It had to figure it out all by itself from the raw video and radar signals. The researchers wanted to know: Did the computer actually learn the math formula for the Doppler effect, or did it just memorize the specific dance moves it saw during training?

Using their LAWFUL detective kit, they did something clever. They created "what-if" scenarios. They took the video of the dancer and mathematically sped up or slowed down the dancer's movement toward the radar, without changing the side-to-side motion. If the computer truly understood the law, its prediction of the pitch should change perfectly in sync with the speed change. If it was just guessing, the prediction would get messy.

The results were fascinating. The computer model did seem to understand the law. When they tested it on speeds it had never seen before, it still predicted the pitch correctly. But the real magic happened when they started turning off parts of the computer's brain to see which parts were doing the work. They found that the computer didn't need its whole brain to solve this. In fact, they identified a tiny "circuit" inside the model—just a few specific parts of its attention system—that was responsible for 91% of the correct physics. It was like finding that only three specific gears in a massive clock were actually making the hands move.

Even cooler, they discovered how those gears worked. The computer wasn't just memorizing numbers; it was using a specific pattern of "attention" to compare how far the dancer moved from one frame to the next. It was essentially doing a real-time calculation of speed, just like a human physicist would, but inside its own digital brain. The study suggests that the computer had indeed learned the formal law of physics, not just a pattern. However, the researchers are careful to note that this was a test on one specific law and one specific type of computer model. While it's a huge step forward in proving that AI can truly "understand" physics, it's a proof of concept for this specific case, not a guarantee that every AI understands every law of the universe. But for now, we have a witness that says: yes, sometimes the computer really is doing the math.

Technical Summary: LAWFUL – Law-Aligned Witness for Faithful Use of Latents

1. Problem Statement

When a neural network accurately predicts a physical system, it remains unclear whether the network has internalized the governing physical law as formal, structured knowledge or merely learned to mimic the input-output behavior within a narrow training distribution. Specifically, three distinct questions arise regarding models operating over continuous physical variables:

Representation: Has the network learned the law as structured knowledge?
Causal Usage: Is this representation causally utilized by the network's internal computation, or is it merely decodable by an external probe?
Domain Validity: Does the representation hold across the full domain of validity of the law, or only within the empirical training distribution?

Current mechanistic interpretability frameworks face four specific gaps when addressing these questions for laws over continuous variables:

Coverage-Aware Consistency: Existing causal abstraction methods quantify consistency only on narrowly sampled, discrete counterfactual pairs, lacking measures for continuous counterfactual families.
Domain-of-Validity Testing: No established procedure verifies if an identified "physics circuit" operates consistently across the law's full theoretical range rather than just the training distribution.
Invariant Verification: There is no method to verify the enforcement of physical invariants (forbidden behaviors) within the circuit, as existing tools target spurious correlations rather than mandated physical constraints.
Signal Flow Quantification: While attribution methods identify which edges matter, they do not quantify how continuous signals of derived physical quantities flow through the circuit (writing, transforming, transmitting, and reading).

2. Methodology: The LAWFUL Framework

The authors propose LAWFUL (Law-Aligned Witness for Faithful Use of Latents), a foundational framework that aligns a learned model with an explicit physics law to witness their agreement over continuous counterfactual states.

2.1. Core Components

The framework pairs a learned model $f: \mathcal{X} \to \mathcal{Y}_f$ with a physics model $g: \Omega \to \mathcal{Y}_g$ representing a known law. It utilizes three key mechanisms:

Bridges:
- Input Bridge ( $B$ ): Maps raw model inputs $\mathcal{X}$ to the physics space $\Omega$ (e.g., extracting radial velocity from motion capture data).
- Output Bridge ( $D_f, D_g$ ): Projects the outputs of both models into a shared observable space $\mathcal{Z}$ (e.g., Doppler centroid).
Perturbation Family ( $\mathcal{P}$ ): A set of physically meaningful operations $p_\alpha$ that generate counterfactual states in $\Omega$ . This allows testing the model's response to continuous variations (e.g., scaling radial velocity) rather than discrete swaps.
Consistency Score ( $M_P(f, g)$ ): A scalar functional $\Phi$ measuring the agreement between the bridged outputs of $f$ and $g$ across the perturbation family. A higher score indicates closer adherence to the physical law.

2.2. Identifying Physically Consistent Circuits

The framework defines a physically consistent circuit as a subgraph $C$ of the model's computational graph that recovers the consistency score under activation patching.

Sufficiency: Routing the perturbation only through $C$ (holding the rest of the network at base activations) should recover the model's physical consistency.
Necessity: Routing the perturbation through the complement ( $V \setminus C$ ) while holding $C$ at base activations should fail to recover the consistency.
Minimality: A circuit is $\tau$ -consistent if it satisfies deviation thresholds relative to the unpatched baseline, and it is minimal if no proper subset is also $\tau$ -consistent.

3. Experimental Instantiation: MoCap-to-Radar

The framework is instantiated on the MoCap2Radar transformer, a model trained to predict radar spectrograms from 3D motion-capture (MoCap) data.

The Law: The Doppler frequency law $f(t) = \frac{2v(t)}{\lambda}$ , where $f(t)$ is frequency, $v(t)$ is radial velocity, and $\lambda$ is wavelength.
The Challenge: Neither $f(t)$ nor $v(t)$ appears in the training data; the model must infer the relationship from raw 3D marker trajectories.
Perturbations: The input bridge extracts radial velocities. The perturbation family scales these velocities by a factor $\alpha$ (counterfactual scaling), effectively simulating different speeds while preserving tangential motion.
Validity Domain: The perturbation range is constrained by the radar's physical resolution limits ( $\alpha \in [-1.16, 1.16]$ ).

4. Key Results

4.1. Circuit Discovery

Baseline: The unpatched model achieves a high consistency score ( $M_0 = 0.988$ ), confirming it learns the law.
Circuit Identification: A greedy bottom-up search identified a minimal $\tau=0.9$ consistent circuit ( $C^\dagger$ ) comprising only 9 of the 18 total components (50% of the model).
Composition: The circuit is dominated by the Temporal Transformer block (7 out of 9 components), specifically including the MLP and attention heads $h_0, h_1, h_2, h_4, h_6, h_7$ . The Spatial Transformer contributes minimally (MLP and head $h_4$ ).
Performance: This reduced circuit recovers 91% of the model's original Doppler consistency.

4.2. Domain of Validity and Invariance

Radial Scaling: Within the radar-resolvable domain ( $|\alpha| \le 1.16$ ), the identified circuit $C^\dagger$ tracks the physics model $g$ closely, exhibiting a linear response to velocity scaling. Outside this domain, performance degrades, likely due to physical extrapolation limits.
Tangential Invariance: When tangential velocity (orthogonal to the radar line-of-sight) is scaled, the model and the circuit exhibit near-zero response, confirming adherence to the physical principle that tangential motion does not induce a Doppler shift.

4.3. Mechanistic Analysis

Emergence of Velocity: Linear probing reveals that radial velocity is not explicitly encoded in individual attention heads or the projection layer initially. Instead, it emerges as a linearly decodable representation only in the combined output of the temporal attention heads.
Mechanism of Action: By patching attention patterns (QK) versus value transformations (VAL), the authors found that attention patterns (QK) are the primary drivers of consistency. Replacing QK alone recovers consistency, whereas replacing VAL does not.
Interpretation: This suggests the model performs adaptive temporal differentiation. The attention patterns dynamically select frame pairs to compare, effectively computing velocity based on the target's speed (shorter gaps for fast motion, longer for slow), analogous to multi-scale signal processing.

5. Significance and Claims

The paper claims to provide the first principled framework for evaluating whether neural models embody physical laws over continuous variables.

Bridging Gaps: LAWFUL closes the gaps regarding coverage-aware causal consistency and domain-of-validity testing for continuous laws.
Faithful Use: It moves beyond "is the model accurate?" to "does the model use the law internally?" by identifying specific subgraphs responsible for physical consistency.
Foundation for Future Work: While the current instantiation focuses on a linear law (Doppler) and a specific transformer architecture, the framework lays the groundwork for addressing invariants and signal flow quantification in more complex physical AI systems.

The authors maintain modesty regarding limitations, noting that the framework currently relies on a pre-supplied grounding law, handles linear laws in this specific instantiation, and reduces complex spectrograms to a single centroid, potentially missing finer structural features.