Contribution
Take sensitivity — the probability that the output changes under a random perturbation of one input token — which was previously studied for Boolean functions, extend it to real data modalities, and show it functions as a unified description of the transformer’s inductive bias.
Key results
The definition generalizes. Definition 4.1 computes sensitivity of a model
Φ over dataset D under a perturbation distribution P: Gaussian
N(0, σ²I) per patch for images, uniform over the vocabulary for synthetic
token tasks. The Boolean case is the anchor where theory is available.
The theoretical base is inherited and extended. On Boolean functions,
sensitivity relates to the degree of the representing multilinear polynomial
— Huang’s D(f) ≤ S(f)² — and to decision-tree size, so low sensitivity is
low complexity in several established senses. The paper proves transformers
show a low-sensitivity bias in the NTK regime by way of spectral bias, and
proves separately that low sensitivity implies better robustness.
The architecture separation. Transformers have lower sensitivity than MLPs, CNNs, ConvMixers and LSTMs, on both vision and language, compared at similar training accuracy — which is the control that makes the comparison mean something.
Three implications, in the paper’s own order. Sensitivity correlates with robustness and can be regularized for, improving robustness on CIFAR-10-C beyond the architecture’s default. It corresponds to flatter minima. And on modular addition it decreases while the training loss is flat, tracking the stages of grokking.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Transformers have lower sensitivity than other architectures | strong | multiple architectures, two modalities, matched training accuracy |
| C2 | Low sensitivity is a simplicity bias | strong in the Boolean case | established equivalences, plus an NTK-regime proof |
| C3 | Low sensitivity implies better robustness | proved in the Boolean setting, correlational elsewhere | theorem plus CIFAR-10-C |
| C4 | Sensitivity regularization improves robustness | moderate | an intervention, on one corruption benchmark |
| C5 | Sensitivity tracks flatness | moderate | correlational |
| C6 | Sensitivity is a grokking progress measure | suggestive | one synthetic task, modular addition |
Limitations
The theory is in the Boolean and NTK regimes; the experiments are not. The proofs live where analysis is possible and the interesting claims live on CIFAR and language. That gap is standard and should still be stated: nothing proves the extended sensitivity notion inherits the Boolean properties.
C6 rests on modular addition. The task where grokking is easiest to produce and least like anything deployed. A progress measure demonstrated there is a candidate, not an instrument.
The perturbation distribution is a choice. Sensitivity is defined relative
to P, and the numbers move with it. The paper fixes sensible choices per
modality; nothing shows the architecture ranking is invariant to them, which
is the shape of failure this record has been counting all week.
Architecture comparisons at matched training accuracy are not matched compute or matched parameters, so “transformers are less sensitive” carries whatever the training budget carries.
Bearing on the record
The grokking result has an existing home. SOTA-200 says to check whether an emergent capability is a metric artefact, and its sources include the reverse-engineering that recovered continuous progress beneath one network’s discontinuity. Sensitivity is a progress measure requiring no reverse engineering and defined for any architecture, which is a strictly more usable instrument — if it generalizes past modular addition. The record does not fold it into the practice on one synthetic task, and names it here so the next person checking an apparent jump knows a cheap candidate exists.
It is the record’s first architecture-comparable inductive-bias scalar. THEORY-041 describes conditioning drift, THEORY-035 describes sharpness rising to the step size’s tolerance, SOTA-012 relates sharpness to test error. All are properties of weights or of the landscape. This is a property of the learned function, which is what makes it comparable across architectures that share no parameterization.
The robustness ranking is not why this is filed. “Transformers are more robust than CNNs” is a leaderboard claim and DP-005 is about what adoption and comparison do not establish. The scalar is the contribution.
Open questions
- Does sensitivity track progress on a real plateau? Modular addition is the easy case. One language or vision run with an apparent capability jump, sensitivity plotted alongside, would turn C6 into something SOTA-200 could cite as an instrument.
- Is the architecture ranking invariant to the perturbation distribution?
One sweep over
Pwould settle whether this is a fact about models or a fact about the noise chosen to probe them. - Does sensitivity regularization cost accuracy? The robustness gain is reported; the trade, if any, is not prominent.