Peeling Off 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 if two neural networks are fundamentally the same? Comparing weights fails because of permutation and scaling symmetries, and behavioral tests can miss hidden backdoors. A 2026 arXiv paper from researchers at Stanford, Heidelberg University, and other institutions, titled "When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability," introduces Tensor Similarity as a principled metric. The method computes angles between model components on a mathematical manifold, filters out symmetries (invariants), compares functional projections in function space, and uses an efficient recursive algorithm to scan entire models. Experiments show it can detect implanted backdoors even when outputs match 99% of the time, visualize grokking during training, and verify whether quantized models retain their original internal logic. The work establishes a kind of algebra for AI identity, grounding sameness in the geometric structure of internal computation rather than weights or outputs.

Peeling Off 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 nonetheless handle the same crisis in strikingly similar ways. At that moment you'd sigh: these two have the same "core."

In the AI world, scientists face an equally hair-pulling puzzle: how do we determine whether two large models are actually the same thing?

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

Dead wrong. Neural networks suffer from a deeply annoying phenomenon called "symmetry." It's like swapping your living room's sofa and TV — the layout has changed, but the "living room's" function is completely unchanged. A network's neurons can be arbitrarily reordered and rescaled. The weight data may look wildly different, yet the models may be doing exactly the same job.

Previously, we could only use a "brute-force" approach: feed both AIs the same data and see whether their outputs look alike. But that's like judging a student only by exam scores without seeing how they solve problems. What if one of the AIs hides a "backdoor" that only triggers on specific inputs? Such surface-level testing (behavioral similarity) simply cannot see through to the "soul."

In May 2026, a research team from Stanford, Heidelberg University, and other institutions released an exposé-style arXiv paper: "When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability."

They invented a measuring stick that goes beyond appearances: Tensor Similarity.

What Is "Tensor Similarity"?

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

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

Following 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 your neurons are seated, the system only cares about the "net force" of their interactions. 2. Find projections in "function space": it treats each component of the model as a wave. Instead of comparing the wave's color (parameter values), it compares the wave's frequency and shape (function). 3. Recursive "holographic scanning": it doesn't just look at one layer; a highly efficient recursive algorithm scans the model from head to toe, computing a "holographic similarity" between components as they process information.

Why Does This "Soul Detection" Matter?

The paper demonstrates several striking scenarios through experiments:

Catching "sleeper agents": even if an AI looks well-behaved, if a malicious backdoor has been implanted internally, tensor similarity can instantly reveal its "core difference" from a safe model — even when their outputs match 99% of the time.
Witnessing "grokking": during training, models sometimes undergo sudden "qualitative transformations (epiphanies)." We used to have to guess; now, with this ruler, we can clearly see exactly when a model's core begins to truly "align" with the truth.
Cross-architecture comparison: it can even tell you whether a slimmed-down model (after quantization) is still genuinely the "offspring" of the original large model.

Why Is This Paper Important?

Feynman spent his life trying to understand the underlying laws of matter. This paper, in effect, builds an "algebra" for AI.

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

Once we possess this "see-through-the-soul-at-a-glance" ability, our control over AI will advance from "probability and statistics" to "logical certainty."

To sum up:

Appearances deceive, data lies, but the structure of logic is unshakable.

The arrival of tensor similarity means we finally have a scalpel to cut through the AI black box's endless layers and directly measure its "thoughts." This is not only a victory for interpretability, but a new lighthouse guarding AI safety.

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

Truth isn't in the pixels; truth lies in the angle between dimensions. That is the deepest definition of "identity" that 2026's machine learning theory has given us.