acceptodds
Withdrawn from ICLR 2027

Input-relational Verification of Transformers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Transformers have been driving significant advances across a wide range of applications, yet their deployment in safety-critical domains require rigorous formal verification. Most existing verification techniques are designed to assess perturbation tolerance around individual inputs and are therefore insufficient to ensure model robustness in a global context. To address this limitation, we first propose a self-composition-based approach that enables end-to-end application of local verification techniques. However, due to the complexity of the composed model, this approach turns out to exacerbate the approximation error of local verifiers significantly. In this paper, we show that exploiting cross-execution relationships in intermediate layers can substantially improve verification precision. Specifically, we construct McCormick envelopes for the cross-execution differences of attention-layer and propagate these relational constraints to the output layer, yielding tighter output bounds. We implement this approach in a tool, , and develop two variants that prioritize efficiency and precision, respectively, to accommodate different verification requirements. We evaluate the approach on CCPP for regression and on Yelp and SST for classification. Across these benchmarks, the proposed methods produce tighter certified output bounds and certify more instances than the naive self-composition and the generic difference propagation, showing that our reasoning of the cross-execution differences is effective for achieving precise input-relational verification of Transformers.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Formal verification, Transformers, Hyperproperty

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2XZ2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies input-relational verification for Transformers. It considers pairs of nearby inputs and requires their outputs to remain close. It first introduces a self-composition baseline. It then proposes ReVerT, which explicitly propagates differences between two executions through the Transformer. The main technical step is to reformulate differences in attention operations so that cross-execution dependencies are preserved, and then apply McCormick relaxations. Two variants are proposed: ReVerT-F for faster bound propagation and ReVerT-P for tighter LP-based bounds. The method is evaluated on CCPP, Yelp, and SST. The reported results show tighter bounds than self-composition and naive difference propagation.

Critical strengths
  1. The paper identifies a useful precision issue in naive self-composition. Tracking cross-execution differences inside the network is more informative than abstracting two executions independently and only relating them at the output. The experiments show that this can substantially tighten certified bounds.
  2. The attention-specific reformulation is simple and practically useful. Rewriting the difference of bilinear attention terms before applying McCormick relaxation preserves dependencies that would otherwise be lost. The empirical results suggest that this improves verification precision.
Critical weaknesses
  1. The novelty relative to existing input-relational verification is not sufficiently established. The paper cites prior work on input-relational DNN verification, including Banerjee et al. (2024), but does not meaningfully compare ReVerT with these methods. The central idea of tracking relations across two executions is already the main principle of this literature. The paper mainly argues that no prior method targets Transformers. I do not think this is enough to establish a strong methodological contribution. The paper should explain what is fundamentally new beyond adapting existing relational verification ideas to Transformer attention.
  2. The problem formulation and evaluation are weaker than the paper’s framing suggests. Definition 1 is a fixed-scale pairwise output-sensitivity condition over a bounded domain. It is closely related to standard continuity or robustness notions, so I do not see a strong conceptual contribution in introducing it as “global robustness.” At the same time, the introduction gives broad safety-critical motivations, while the experiments use only CCPP, Yelp, and SST. The main certification-success experiment is based on only five CCPP tasks. This is not enough to support the broad claims made in the paper.
Questions for authors
  1. The paper cites prior work on input-relational verification of DNN, which is very close to the current topic. What is the precise technical difference between ReVerT and applying or extending those relational verification methods to Transformer attention? Which part of the proposed abstraction is fundamentally new?

  2. What is the conceptual significance of Definition 1? It appears to be a fixed-δ\delta, fixed-ϵ\epsilon pairwise output-sensitivity condition over a bounded domain, closely related to standard continuity notions. Why should this be viewed as a new “global robustness” formulation?

  3. Do the evaluated models use standard LayerNorm, or the no-variance LayerNorm described in the paper?

  4. Can the authors report bound comparisons on exactly the same set of tasks for all methods? The current averages may use different valid task subsets, which makes the comparison difficult to interpret.

Reviewer_Q33jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to generalize local robustness verification queries to a more global setting. This is realized through a parallel execution of local inputs of bounded distance, while bounding the difference between the intermediate states of the verification. The evaluation is done on typical datasets and transformer sizes considered in the verification community, with scalability to larger models used in practice still out of reach.

Critical strengths
  • The topic of the paper is well introduced and highly relevant for AI safety.
  • Generalizing verification queries to a more global context is insufficiently explored in most neural network verification works.
  • The paper gives sufficient background to follow along in the main section of the paper.
Critical weaknesses
  • The notation of the paper is somewhat vague. For example, it is often unclear if certain variables are vectors/matrices/sets. For example, I don't understand what you mean in l.197 with "T(x) ∈ yˆ′". Why is yˆ a set? You seem to use calligraphic letters for the domain D; maybe this could be adopted generally for sets.
  • I think the paper would greatly benefit form a running example.
  • The experimental setup is not clearly described. E.g., which perturbation radii ware used and how were those chosen?
  • It is also unclear how far the two inputs are apart, it only says "closely related" in l.053 with some abstract distance measure.
  • The approach does not compare its method properly against state of the art and only considers naive baselines (e.g. Tab. 1 omits the "main baseline" as described in Tab. 4). -> for these reasons, it's difficult to assess for me how well the approach performs.
  • The paper has some spelling mistakes, e.g., l.251 "witch" -> "which"
  • One could also explain in more detail how this approach differs from other input-relational approaches.
Questions for authors
  • How are the pair-wise inputs x,x′x, x' of a domain D\mathcal{D}, and perturbation radii δx\delta_x, and distances ϵy\epsilon_y chosen?
  • Would it also be possible to construct a perturbation hull containing both inputs in one query using traditional verifiers, i.e., use D\mathcal{D} directly to determine if e.g. a classification flips for any x∈Dx\in\mathcal{D}?
  • What kind of inputs can be verified using local verifiers and how does your approach extend it?
  • l.330: Could one optimize discarded constraints similar to CROWN?
Reviewer_z1vcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Neural network verification is limited applicability for transformers, but transformers are the current SOTA models. This paper extends global robustness verification with neuron-difference tracking to transformers. For this, it tracks and bounds state differences between two runs of a transformer. The evaluation shows that this produces tighter bounds than baseline approach.

Critical strengths

The paper tackles an important problem, as both verifying transformers and global robustness is essential for making verification practical for real-world applications. Although proofs are omitted (lines 341 to 343, Appendix H only provides a proof sketch), I can see the constructions are straightforward enough to be covered by existing work. Overall, I agree the approach provides valid bounds.

The motivation for why verifying transformers matters was easy to follow and written ok.

Critical weaknesses

I have several concerns:

  1. Contribution: The idea of using self-composition for verifying global robustness is widely established in literature [1,2,3]. The approach of propagating neuron differences was established by NeuriDIFF [2]. The core novelty I see in this paper is rearranging the dot products in Equation (8) to tighten bounds. What follows for propagating bounds through the dot product is straightforward extension of Shi et al. (2020). Overall, I consider this to provide insufficient novelty for ICLR.

  2. Missing references: Besides the three references listed above, the paper is missing citations in lines 34 to 40 (and I am indeed sceptical transformers have been applied in each of these domains). Lines 218 to 220 also lack references to the existing tools critizised in the sentence.

  3. Presentation: the background section is overly lengthy and repetitive. For example, Equation (4) duplicates line 128. Almost all equations are equations are unnecessarily restated 1:1 in the following text. Terms such as "Add&Norm" are not defined. The normalisation in Table 1 makes the bounds of ReVerT-P and ReVerT-F appear disproportionately tight.

  4. Transformer Architecture: The paper uses an uncommon variant of layer normalisation that omits the variance (line 124), instead of verifying the standard transformer architecture. As such, the results do not apply to typical real-world transformers.

  5. Experimental Setting: The provided code appears to test a different robustness definition than what is stated in the paper. See questions below.

[1]: Brandon Paulsen, Jingbo Wang, Chao Wang: ReluDiff: differential verification of deep neural networks. ICSE 2020: 714-726 [2]: Brandon Paulsen, Jingbo Wang, Jiawei Wang, Chao Wang: NEURODIFF: Scalable Differential Verification of Neural Networks using Fine-Grained Approximation. ASE 2020: 784-796 [3]: David Boetius, Stefan Leue: Verifying Global Neural Network Specifications using Hyperproperties. FoMLAS@CAV 2023: 71-82

Questions for authors

Q1. The provided code appears to verify perturbations of a single token instead of all tokens. How does this relate to your definition of robustness in Definition 1? Q2. What robustness definition was used in the experiments? Q3. What hardware were your runtime measurements taken on? What measures did you take to avoid contamination?

Formatting:

  • use citep or other command with brackets.
  • line 98: the growth of the needs for => the growing need for

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.