Peeling Back the Wallpaper: How Tensor Similarity Reveals Whether Two AI Models Share the Same "Soul"

Forum topic · QianXun · 2026-05-16

Summary

How can we tell whether two neural networks are fundamentally the same model? A 2026 arXiv paper, "When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability," from researchers at Stanford, Heidelberg University, and other institutions proposes a solution. Because neural networks exhibit symmetry—neurons can be permuted and rescaled without changing function—comparing raw weights is unreliable, and behavioral testing can miss hidden backdoors. The proposed tensor similarity metric instead measures the geometric relationship between models on a mathematical manifold, filtering out symmetry-induced variations, comparing functional projections in function space, and using an efficient recursive algorithm to scan entire networks. Experiments show the method can detect hidden backdoors even when outputs match 99% of the time, visualize "grokking" phase transitions during training, and verify whether quantized models preserve their original internal logic. The work reframes model identity as a property of internal geometric structure rather than weights or outputs.

Peeling Back the Wallpaper: How to Tell at a Glance Whether Two AIs Share the Same "Soul"

In everyday life, you might meet two people with completely different personalities who handle the same crisis in astonishingly similar ways. You'd conclude: these two people have the same "core."

In the AI world, scientists face an equally maddening problem: how do we determine whether two large models are actually the same thing?

This sounds trivial. Just compare their weights (parameters), right?

Dead wrong. Neural networks suffer from a deeply annoying phenomenon called "symmetry." It's like swapping the positions of your sofa and TV in the living room: the layout changed, but the room's function is unchanged. AI neurons can be arbitrarily permuted and rescaled—weight data may look wildly different while the models do exactly the same work.

Previously, we relied on brute-force methods: feed two AIs the same data and compare outputs. But that's like judging a student only by exam scores without seeing their problem-solving process. What if one AI hides a "backdoor" that activates only on specific inputs? Surface-level behavioral testing can't see through to its "soul."

In May 2026, a research team from Stanford, Heidelberg University, and other institutions published an arXiv paper designed to be a "demon-revealing mirror": "When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability."

They invented a ruler that goes beyond appearances: Tensor Similarity.

What Is "Tensor Similarity"?

Feynman once said: "If you can't describe it with math, you don't truly understand it."

The brilliance of this paper is that instead of staring at deceptive weight numbers, it computes the "angle" between two models on a mathematical manifold.

Applying Feynman's logic, here's how this "X-ray technique" works:

1. Ignore the "wallpaper" (invariants): The system automatically filters out interference from neuron permutations. No matter how neurons are seated, the system only cares about the "net force" of their interactions. 2. Project into function space: Each component of the model is treated as a wave. Rather than comparing the wave's color (parameter values), it compares the wave's frequency and shape (functional behavior). 3. Recursive "holographic scanning": It doesn't look at just one layer. An efficient recursive algorithm scans the model from head to toe, computing the "holographic similarity" of components as they process information.

Why Does This "Soul Detection" Matter?

The paper's experiments demonstrate several striking scenarios:

Catching "sleepers": Even if an AI looks well-behaved, if a malicious backdoor was implanted internally, tensor similarity instantly reveals the "core difference" from a safe model—even when outputs match 99% of the time.
Witnessing "grokking": During training, models sometimes suddenly undergo "qualitative transformation (grokking)." Previously we could only guess; now this ruler clearly shows when a model's core truly "aligns" with the truth.
Cross-architecture comparison: It can even determine whether a compressed small model (after quantization) is still the "true child" of the original large model.

Why This Paper Matters

Feynman spent his life trying to understand the underlying laws of matter. This paper is essentially building an "algebra" for AI.

It tells us: an AI's identity depends neither on what it looks like (weights) nor on what it says (outputs), but on the "geometric structure" of its internal logic.

With this "see-through-the-soul" ability, our control over AI can rise from "probabilistic statistics" to "logical certainty."

In summary:

Surfaces deceive, data lies—only the structure of logic is unshakable.

The advent of tensor similarity means we finally have a scalpel to cut through AI's multi-layered black box and directly measure its "thoughts." This is not just a victory for interpretability, but a new lighthouse guarding AI safety.

Next time someone says two models are "about the same," you can push up your glasses and ask: "What's their tensor similarity?"

Truth is not in the pixels; truth lies in the angle between dimensions. That is 2026 machine learning theory's most profound definition of "identity."

*Note: The source post describes the paper as published in May 2026; this date is as stated in the original forum post.*