acceptodds
Withdrawn from ICLR 2027

Attributing HOW, Not Just WHICH: Counterfactual Response Trajectories for Diffusion Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion models have achieved remarkable success in image generation, yet tracing their outputs to individual training examples remains challenging. Existing attribution methods often compress factor-specific effects into scalar responses, making distinct internal changes indistinguishable. This is particularly limiting for diffusion models, where semantic factors emerge through evolving representation dynamics during denoising. We therefore reformulate diffusion data attribution as attributing factor-induced internal response trajectories. In this paper, we propose a novel Concept Attribution method through Dynamic Trajectories(CADT). We argue that attribution should therefore ask not only which examples matter, but also how their influence unfolds during generation. Specifically, we construct matched counterfactual pairs at identical noisy states to isolate factor-specific representation displacements, and model their directional and magnitude evolution across denoising as dynamic attribution signatures. For each training example and generated query, CADT extracts stage-wise feature vectors and integrates them along the denoising process to form a trajectory descriptor. Applying the same construction across the training set yields a bank of factor-specific trajectory descriptors. The covariance statistics of this bank are then used to construct . CADT uses this covariance-aware positive-semidefinite kernel to calibrate the query and training representations, and compares the calibrated query trajectory with each training trajectory to produce the final training-sample attribution scores. Experiments on multiple public datasets show consistent improvements over existing diffusion attribution baselines across hierarchical, compositional, and style attribution.

Primary area:
interpretability and explainable AI
Keywords:
diffusion models, interpretability, Data Attribution, Internal Representations

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_vR52OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies factor-specific training-data attribution for diffusion models. It argues that existing methods summarize influence as one scalar per training example, which mixes semantic factors and hides when and how influence arises during denoising. It therefore proposes to attribute factor-induced internal response trajectories instead. The method, CADT (Concept Attribution through Dynamic Trajectories), ranks training images by their relevance to one factor (shape, colour, class, style or object) of a generated query. Each image, whether a training image or the query, is forward-noised to a set of timesteps. At each noisy state, the frozen denoiser is run twice, once with the target condition y_c and once with a matched reference y_ref(c). The difference of the intermediate activations at selected layers is called the counterfactual response (Eq. 6).

The experiments cover three settings: a synthetic shape/colour benchmark with a held-out compositional combination; CIFAR-100, with hierarchical retrieval from a model conditioned only on animal/plant and with 20% label flipping; and ArtBench-2/5 with an SD-v1.4 LoRA, for pure-style and object–style queries. All main results are reported as P@K against factor labels, in comparison with TRAK, D-TRAK, DAS and Journey-TRAK, and CADT obtains the highest precision in most settings.

Critical strengths

A gradient-free formulation of factor-specific attribution. CADT isolates a factor by contrasting the denoiser’s activations under a target condition and a matched reference condition at the same noisy state, and it needs only forward passes. Existing concept-level attribution, notably Concept-TRAK, requires a per-example parameter gradient for every training image. If validated, a forward-only alternative could make factor-level attribution cheaper and simpler to deploy. This is the main element that distinguishes CADT from prior work, and it is why I see potential in the direction. It does not change my recommendation, however. The paper does not explain why this activation-based score should reflect training influence. It reports no runtime or memory comparison, and the descriptor bank must be rebuilt for every target/reference pair. It also provides no counterfactual evidence that the retrieved examples are actually used by the model.

Critical weaknesses

Weakness 1: The claimed research gap does not exist, and the method does not answer the question it raises. The paper frames CADT as factor-specific, trajectory-level attribution, but both halves of this framing are already covered by prior work. Concept-TRAK (ICLR 2026) addresses the same task with the same mechanism: it contrasts the target condition with a condition from which the concept has been removed. It also introduced the synthetic benchmark used in §5.2. Yet the paper mentions Concept-TRAK only in passing, does not explain how CADT differs from it, and reports no comparison with it. This holds even on Concept-TRAK’s own benchmark, although the authors have an implementation (App. J). On the temporal side, Journey-TRAK already attributes individual denoising steps, and later methods (D-TRAK, DAS, K-FAC/EK-FAC influence functions, Concept-TRAK) aggregate per-timestep terms by design; the paper’s own Eq. A31 decomposes Concept-TRAK scores over timesteps. CADT also does not deliver the promised “how”. For a given factor, it still outputs one scalar per training example, summed over all layers and timesteps (Eq. 14), and every experiment measures “which”. Its trajectory is obtained by forward-noising clean images, not from the generation process. This is consequential because the paper’s contribution rests on this framing. Without it, what remains is a change of signal (activations instead of gradients) and a whitened kernel, and neither is justified or supported by the ablations: in Table A1, plain cosine outperforms K_λ.

Weakness 2: There is no evidence that the score measures training influence. Gradient-based attribution is grounded in a counterfactual, namely how a utility changes when an example is removed and the model is retrained, and the paper itself states this definition in §3.2. CADT replaces gradients with a similarity between forward activations, but it gives no argument linking this score to that counterfactual. The leave-one-descriptor-out result concerns an auxiliary regression in descriptor space, and App. D concedes that the signal is not defined through training-induced parameter updates. The experiments do not close this gap. All main results are P@K against factor labels, which measures category retrieval and would be maximized by simply looking up images with the same label, yet no similarity baseline is reported. LDS is dismissed in App. E.1, although its utility can be made factor-specific and the authors already use the official ArtBench subsets to tune DAS (l. 867–868). The deletion experiments in App. I either evaluate no ranking (I.1) or compare only against random deletion (I.2), without same-style or other-method controls. The headline comparison is also structurally biased: in Table 1, a baseline that produces a single ranking cannot exceed 50% on average, and Concept-TRAK is absent. This is consequential because the paper’s central claim is attribution. Without either a theoretical link or causal validation, the results show only that CADT retrieves images sharing the query’s factor, not that the model relies on them.

Questions for authors

Clarification requests

  1. For a training example x_i, which condition pair does CADT evaluate in Eq. 6: the query’s y_c / y_ref(c), or a pair derived from x_i’s own label? Does CADT use training labels anywhere? This determines how §5.4 should be read. If training descriptors do not depend on training labels, then recovering flipped animals reflects content similarity rather than label-driven influence.
  2. In Tables 1 and 5, is each baseline scored with a single ranking per query against both factors? If so, a baseline’s average P@K in Table 1 cannot exceed 50%, and D-TRAK and DAS sit at exactly 50.00. Please state this explicitly and explain why you consider the comparison fair.
  3. Table 3 cannot be interpreted as written. For every method, the P@1 values of the two source groups sum to more than 100% (135% for CADT), and all values are multiples of 1/185. How are the queries divided between the two columns, and which ranking and candidate pool does each column use?
  4. Please specify the descriptor construction for each experiment. Are queries forward-noised from the final generated image? How many probes S are used, and is one noise draw shared across timesteps within a probe (otherwise the velocity in Eq. A3 differences independently noised states)? Which layers, timesteps and reference values y_ref(c) are used, and what are the block weights w_b? §4.4 says the weights are given in App. A, but they are not there.
  5. In Table 6, which component accounts for the jump from “+ Multi-seed stability” (35.24) to “Full CADT” (50.59)? Since plain cosine outperforms Kλ in Table A1, why is Kλ the default, and how does cosine perform on Tables 1, 2, 4 and 5?
  6. Appendix J conditions the CIFAR-100 model on fine classes, whereas §5.3 states that the model is conditioned only on animal/plant. Which model is used in App. J?
  7. DAS and Journey-TRAK are near chance on CIFAR-100 (Table 2). On the synthetic benchmark, D-TRAK’s factor bias is the opposite of what Concept-TRAK reports for the same setup. Which timesteps, projection dimensions and damping did you use for each baseline? Did you verify the implementations, for example by reproducing published LDS values?
  8. For the object–style experiment, what is the base rate of cat or dog content in the ArtBench training sets, and how accurate is the Qwen3-VL judge on a labelled subset? With 11 and 30 queries, an object P@1 of 9.09% on ArtBench2 corresponds to a single query.

Requests for additional evidence (these could change my assessment)

  1. Can you provide an argument for why the activation-based score K_λ measures training influence? For example, derive a link to a model-level counterfactual via representer points, a gradient factorization, or the closed-form empirical denoiser. The leave-one-descriptor-out result concerns an auxiliary regression in descriptor space, and App. D concedes that the signal is not defined through training-induced parameter updates. If no such link exists, would you reframe CADT as factor-specific similarity retrieval?

  2. Please add counterfactual validation. You already use the official ArtBench 50% subsets to tune DAS (l. 867–868), so you can report CADT’s LDS on them, both with the standard utility and with a factor-level utility such as a style classifier’s probability on fixed-seed regenerations. In addition, please run leave-top-K-out retraining for every method, with random same-label deletion as a control, and measure the drop in target-factor presence. For App. I.2, please also report the deletion ratio, the number of runs, the probe set, and the quantity plotted in Fig. A7.

  3. Please report Concept-TRAK in Tables 1–5. It is listed as a baseline in §5.1, the synthetic benchmark of §5.2 is its own, and you already have an implementation (App. J). Please also explain how CADT differs from it methodologically. In addition, please include baselines that receive the same factor information, such as TRAK or D-TRAK with a contrastive utility or with Concept-TRAK’s set-level protocol. Please also include feature-similarity baselines: U-Net features without the contrast, CLIP/DINO, and CSD for style.

  4. Can CADT output and validate stage-wise information, i.e., when or in which layer a training example acts? One test: check whether per-stage contributions predict the timestep profile of the change in denoising error after deleting that example, as in Fig. A6. If it cannot, would you present CADT as factor-specific “which” attribution and drop the “how” claim?

  5. Please ablate the core components. First, the contrast itself: H(x_t, y_c) alone vs. H(y_c) − H(y_ref(c)). Second, a trajectory vs. a single timestep with spatial resolution held fixed. Third, the choice of y_ref(c). Please also report runtime and memory relative to TRAK, D-TRAK and Concept-TRAK, given that the descriptor bank must be rebuilt for every target/reference pair.

7 comments
AuthorsOpenReview ↗

Response to Reviewer vR52 (1/6): W1, research gap and positioning

Thank you for the exceptionally detailed review. We carried out the work closely following your requests R1–R5; questions shared by several reviewers are answered in the General Response (GR) 2–6.

Factor-specific attribution and Concept-TRAK. Concept-TRAK and CADT both isolate the target factor through a condition contrast. They differ in where the contrast is applied and in the signal that is compared: Concept-TRAK writes the contrast into a scalar utility and compares projected parameter gradients, whereas CADT builds the condition contrast on the internal activations of the frozen denoiser and compares how the response is distributed over layers and evolves over noise levels. The quantitative comparison in all settings of Tables 1–5 and the implementation are in GR 2/6. One further clarification: the benchmark of Sec. 5.2 was built by us following the design of Concept-TRAK, whose original data are not public, so the numbers in the two papers cannot be compared directly.

Temporal side. Journey-TRAK computes attributions for individual sampling steps, and D-TRAK, DAS and Concept-TRAK aggregate per-timestep scalar terms. CADT keeps an ordered, layer-resolved sequence of condition responses together with their change between adjacent noise levels; a simple sum cannot retain these structures. In Table 6, adding the temporal direction to response energy raises mid-level P@5 from 13.73 to 31.57. The new single-timestep control shows that the full trajectory is better than the best single timestep on fine-class retrieval (P@5 21.68 vs. 19.89); the best timestep also depends on the target factor, e.g. t=900t=900 for shape, t=600t=600 for color and t=100t=100 on CIFAR-100 (see R5 in 6/6).

The "how" claim (also R4). By "how" we mean how the internal response to the target factor is distributed over network layers and how it evolves along the denoising timesteps; CADT does provide this structural information. The attribution score of Eq. 14 is a sum of block interactions indexed by layer and timestep, so every attribution comes with a layer–timestep decomposition. The paper analyzes this structure: Figs. 4–5 and Table A2 show that shape and color depend on different network layers (d1 and d0) and share a late interval (t=300→100t=300\to100); App. J shows that two training examples with nearly identical Concept-TRAK scores have different layer–time response patterns; Fig. A6 shows that the effect of deleting training data also varies over timesteps. The new ablations further show that the stage structure matters for attribution: restricting the descriptor to the late interval raises mid-level P@5 from 35.24 to 50.59 (Table 6; component-wise decomposition in our response to ubCe, 4(b)), and the most informative stage differs across factors. The per-example deletion-and-retraining profile test suggested in R4 requires larger-scale experiments and could not be completed in this round; we regard it as a complementary validation of the existing layer–timestep analysis (see 5/6).

Trajectory from forward-noising, not from generation. This is a deliberate design. The response trajectory is defined as the ordered responses of the trained denoiser over a sequence of noise levels; queries and training examples are forward-noised in the same way, so that every stage is compared at a matched noise level. A training image has no corresponding generation path, so the sampling path of a query could not be aligned with the training examples. The settings are given in Q4 (3/6).

Signal and kernel. The conclusion of Table A1 is confirmed on label flip and on the synthetic model (cosine better) and reversed on ArtBench style (style head of ArtBench5 object–style: KλK_\lambda 90.00 vs. cosine 46.67). No scoring function dominates; results of all three on every main task are in GR 3/6, and we agree that the descriptor matters more than the kernel. Whether activations are a justified signal is addressed in GR 4/6 and GR 5/6.

AuthorsOpenReview ↗

Response to Reviewer vR52 (2/6): W2, evidence for training influence

Link to the counterfactual. The counterfactual definition of training influence requires model-level evidence. Following R1, we derived the corresponding exact model-level link under the closed-form empirical denoiser (GR 4/6): the first-order term of the leave-one-out change of the condition-contrast energy decomposes into a kernel weight that depends on the observed training label and on the log-SNR, the inner product of the condition responses of the query and of the training example, and bounded remainders; the full energy change also contains a second-order term. This decomposition gives CADT's response-matching signal a theoretical basis in the idealized model, while training influence in actual networks is tested by the deletion-and-retraining experiments (GR 5/6).

P@K is maximized by label lookup; no similarity baseline. This holds when labels are reliable, and it is why we added CLIP, DINOv2, CSD and text-retrieval baselines (GR 3/6). In label flip, looking up the corrupted training label does not help: flipped animals and genuine plants carry the same observed label. CLIP text retrieval scores 0.00% on flipped sources, CLIP image similarity 8.11%, Concept-TRAK 12.43%, CADT 37.84%. How this result should be read, given that CADT uses training labels, is discussed under Q1 (3/6).

LDS. We agree that both standard and factor-level LDS are valuable counterfactual tests. The utility of standard LDS is the whole-image denoising loss, which does not answer the same question as CADT's factor-specific score; factor-level LDS responds in a threshold-like rather than linear way when the target factor is common (consistent with the argument of Concept-TRAK), and it requires retraining many subset models, beyond our compute budget for this round (GR 6/6). We therefore added the leave-top-K-out retraining that you also suggested (GR 5/6).

Deletion experiments. We ran the test you propose in R2: leave-top-K-out retraining with random same-label deletion as control (GR 5/6). CADT-guided deletion gives the largest increase in target-class denoising error, is the only deletion whose 95% interval excludes 0, and is significantly above deletions guided by CLIP and Concept-TRAK; only 8–24% of the examples it selects belong to the target fine class, yet the effect is comparable to deleting 200 same-class images. The missing details of App. I.2 are given in GR 5/6, and its random group is a same-style control; I.1 indeed evaluates no ranking, which the text already states.

Structural bias of Table 1. This is explained under Q2 (3/6); Concept-TRAK results in the setting of Table 1 are in GR 2/6.

In sum. Around W2, we added three kinds of direct evidence: a model-level theoretical link (GR 4/6), feature-similarity baselines (GR 3/6), and deletion-and-retraining experiments with a same-label random control (GR 5/6).

AuthorsOpenReview ↗

Response to Reviewer vR52 (3/6): Clarifications Q1–Q4

Q1. Which condition pair for a training example? Are training labels used? A pair derived from the example's own observed label; CADT does use training labels. CIFAR-100: ycy_c is the observed (possibly flipped) class and yrefy_{\rm ref} the other class. Synthetic: the example's own shape/color with only the target factor flipped. ArtBench pure style: its own caption "a {style} painting" vs. the empty prompt. The only exception is the object head of object–style, where all training images use the query's object word vs. the empty prompt.

Sec. 5.4 should be read as follows: flipped animals and genuine plants have the same observed label (plant), so the label cannot separate them; among identically labeled examples, CADT separates them by the model's condition-contrast response (flipped sources 37.84 / 38.27 / 37.51 / 34.42, chance ≈ 5.9%; genuine plants 97.30 / 95.46 / 95.51 / 94.53). We tested your alternative directly with a label-free variant in which every training example gets the query's condition pair:

CADT (own label)Label-free
Hierarchy P@5 (mid / fine)50.59 / 21.6839.62 / 16.00
Label flip, average P@167.5745.41
– flipped animal P@137.845.95
– genuine plant P@197.3084.86
Synthetic OOD P@1 (shape / color)91.20 / 90.8089.30 / 23.50

Without labels, the share of flipped sources in the top-10 of flipped-animal queries falls from 37% to 6% and that of ordinary animals rises from 2% to 60%: pure response similarity retrieves "animal content", not "animals learned as plants". This matches the derivation in GR 4/6, where the observed label enters only through the gate ϱm\varrho_m: the label decides in which condition an example took part in training, and the response separates examples within that condition. So you are right that recovering flipped animals is not label-free; but it is not label lookup either.

Q2. Single ranking and the 50% bound in Table 1. Yes. Each whole-image baseline scores both factors with the same single ranking, and since the query's factor combination is held out, a training example shares at most one factor with it; shape and color precision sum to at most 100%. Concept-TRAK results in the setting of Table 1 are in GR 2/6, and the contrastive TRAK/D-TRAK are in our response to GQFn, 1 (cont.).

Q3. Table 3. The table contains 370 queries, all generated under the plant condition and split by a frozen classifier trained on clean labels into 185 visually-animal queries (relevant set: the 1,574 animals relabeled as plant) and 185 visually-plant queries (relevant set: the 6,298 genuine plants). Each column is computed only on its own 185 queries, hence multiples of 1/185; the two columns are independent results on two disjoint query groups, not a split of one ranking, so they need not sum to 100%. Each column uses the method's ranking for that query (the whole-image ranking for whole-image methods, the plant-vs-animal factor ranking for CADT and Concept-TRAK) over all 26,500 training candidates.

Q4. Descriptor construction.

SyntheticCIFAR-100ArtBench
Query statesforward-noised final generated imagesameVAE latent of the generated image, forward-noised
Probes SS1 (training), 5 (query)samesame
Noise within a probeone draw shared across timestepssamesame
Layers (resnets.0)10: down0–3, mid0–1, up0–36: down0–2, mid, up1–27: down0–2, mid, up1–3
Timesteps13-point grid 900→100, refined at low noise; 200→100 selected on the 500 selection queries900→100, step 100; last third900→100, step 100; last third
Referenceflip only the target factorthe other classstyle: empty prompt; object–style: object prompt (style head), "in the style of {style}" (object head)
Block weights wbw_b0.5 energy family, 0.5 velocity family, uniform over layerssamesame

Since one noise draw is shared across timesteps within a probe, the velocity of Eq. A3 differences responses under the same noise. The block weights were missing from App. A; this was our omission, and they are given in the table above.

Reviewer_GQFnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on a limitation of conventional training data attribution methods, which aggregate the contribution of each training sample into a single scalar value. To address this, the authors propose CADT, a method that estimates attribution separately for different semantic factors. Specifically, the method prepares counterfactual conditions in which only the target concept is modified, and extracts responses specific to the target concept by computing the difference in the model's internal representations at the same noise state. Furthermore, it captures how the magnitude and direction of these responses change along the denoising process of the diffusion model, and uses these as features for attribution estimation. However, simply comparing these features is not sufficient to distinguish variations common across the entire training set from those that are informative for attribution. To solve this, the authors introduce a kernel that accounts for the covariance structure of the training data features, and by emphasizing discriminative feature directions, they achieve more accurate attribution estimation.

Critical strengths

The underlying motivation, namely that attribution based on a single scalar value is insufficient and that computing attribution scores corresponding to each semantic factor is needed, addresses a topic of significant practical importance. In addition, since this method does not require retraining, I understand that it can be executed with a practical computational cost, similar to methods such as D-TRAK.

Critical weaknesses

While I appreciate the problem this paper addresses, I do not feel able to recommend acceptance at this time for the following reasons. In particular, I regard the insufficient comparison with Concept-TRAK as an issue that cannot be overlooked. Adding comparison results against Concept-TRAK during the rebuttal period would certainly help to resolve this concern. However, this comparison is essential for validating the effectiveness of the proposed method, and it should have been included in the initial submission. Even if the comparison results are added during the rebuttal period, the problem remains that a fundamental experiment supporting the paper's core claim was missing from the initial version. For this reason, I am currently inclined against acceptance.

  1. Comparison with key prior work is insufficient.
  • Concept-TRAK is the prior work that addresses the problem setting closest to this paper, so a comparison with it is essential. However, although it is listed as a baseline in Section 5.1 and its scores are computed on CIFAR-100 in Appendix J, no results are shown in Tables 1 through 5, and no explanation is given for this omission.
  • In the main tables, comparisons are made with methods such as D-TRAK and TRAK. However, these are primarily attribution methods targeting the entire image, and they are not designed for concept-specific attribution as CADT is. Therefore, demonstrating superiority over these methods alone does not sufficiently validate the effectiveness of the proposed method. A comparison with at least Concept-TRAK is necessary, yet only a partial comparison is provided in the appendix.
  1. Comparison with simple similarity-based baselines is insufficient.
  • This paper mainly uses P@K for evaluation, but there is no comparison with similarity based on image features such as CLIP. Since even a simple similar-image retrieval might rank training samples containing the target concept highly, it is necessary to clarify how much better CADT actually performs.
Questions for authors

Nothing in particular. Please explain why no comparison with Concept-TRAK was conducted.

3 comments
AuthorsOpenReview ↗

Response to Reviewer GQFn

Thank you for the review and for recognizing the practical value of the problem and the computational advantage of not requiring retraining. Questions shared by several reviewers are answered in the General Response (GR) 2–6; your concerns are answered point by point below.

1. Why Concept-TRAK was not compared, and the comparison itself. The comparison with Concept-TRAK should indeed appear in the main tables; we have now completed it in all settings of Tables 1–5. Why it was missing from the submission, our implementation aligned with the official repository, and the full results are in GR 2/6. In short, CADT is better on hierarchical retrieval (P@5 50.59/21.68 vs. 44.97/15.62), on label flip (flipped-source P@1 37.84 vs. 12.43) and on ArtBench5 object–style (style head P@1 90.00 vs. 80.00, object head 30.00 vs. 0.00), and both saturate on pure style; on the synthetic data, Concept-TRAK is better on color and worse on shape. In deletion and retraining, the paired difference in the increase of target-class denoising error (CADT minus Concept-TRAK) is +0.34 [+0.13, +0.55]×10⁻⁴ (GR 5/6).

1 (cont.). TRAK and D-TRAK are whole-image methods. We agree that outperforming them alone does not validate factor-specific attribution. Therefore, in addition to Concept-TRAK, we gave TRAK and D-TRAK the same factor information through a contrastive utility, in a literal and a stop-gradient form: each keeps its training gradients, and its query utility ff is replaced by a contrast between the target and the reference condition at the same noisy state, either literally, f(yc)−f(yref)f(y_c)-f(y_{\rm ref}), or with a stop-gradient on the reference branch, ∥ϵθ(xt,yc)−sg[ϵθ(xt,yref)]∥2\Vert\epsilon_\theta(x_t,y_c)-{\rm sg}[\epsilon_\theta(x_t,y_{\rm ref})]\Vert^2 (whose gradient has the direction of the Concept-TRAK utility gradient).

MethodSynthetic OOD P@1 (shape / color)Pure style P@1 (AB2 / AB5)AB2 obj.–style P@1 (style / object)AB5 obj.–style P@1 (style / object)
TRAK, literal contrast29.10 / 65.1096.65 / 90.3381.82 / 9.0963.33 / 3.33
D-TRAK, literal contrast100.00 / 100.00100.00 / 100.00100.00 / 0.00100.00 / 0.00
TRAK, stop-gradient53.00 / 85.8099.48 / 94.50100.00 / 0.0073.33 / 6.67
D-TRAK, stop-gradient21.40 / 75.9098.85 / 93.58100.00 / 18.1883.33 / 30.00
CADT91.20 / 90.80100.00 / 100.00100.00 / 9.0990.00 / 30.00
  • CIFAR-100: on hierarchical retrieval, CADT's mid/fine P@5 is 50.59/21.68, above D-TRAK (45.30/18.54); CADT is the best method from K=5 on, and its margin grows with K; at K=1 D-TRAK is slightly higher on mid-level retrieval (50.27 vs. 49.73). With the contrastive utility, D-TRAK's fine-class P@5 drops to 11.89: because the model is conditioned only on animal/plant, this contrast emphasizes kingdom-level differences rather than the hierarchy below them. On label flip, the contrastive D-TRAK has average P@1/5/10/50 67.30/66.27/65.81/65.45 and recovers 77.30% of flipped sources, but only 57.30% of genuine plants.
  • When the target factor is itself a conditioning input (synthetic factors, ArtBench style): the literal contrastive D-TRAK reaches 100/100 on the synthetic factors and saturates ArtBench style — the contrast reads the factor label of a training example through the gradient of the condition embedding (or caption). With a stop-gradient on the reference branch (the direction of the Concept-TRAK utility gradient), it drops to 21.40 / 75.90 on the synthetic data; on factors that are not condition inputs (CIFAR mid-level/fine classes, ArtBench objects), the literal form has no such advantage.

2. Simple similarity baselines. We added CLIP ViT-L/14 and DINOv2 ViT-L/14 image similarity, CSD for style, and CLIP text retrieval (full results in GR 3/6). Simple similar-image retrieval does not reach CADT on the CIFAR-100 hierarchy (best feature baseline 33.62/14.00 vs. 50.59/21.68), on label flip (flipped-source P@1 8.11 vs. 37.84) or on ArtBench5 style (CSD 97.35 vs. 100). On the pixel-level synthetic factors and on the ArtBench object head (ArtBench5 DINOv2 P@1 60.00 vs. 30.00), however, it is as good or better; there the objects are concepts the pretrained model already knows, and only 2.84% / 4.72% of the training images contain them. We will report both sides in the paper.

Thank you for your response. I have two questions.

  • I would like to confirm one point regarding your implementation of Concept-TRAK. The results of Concept-TRAK on your synthetic benchmark, particularly for shape, seem to differ considerably from the results reported in the original paper. I understand that a direct comparison is not possible because the benchmark construction (and architecture) and the metric differ. You also mention that Concept-TRAK is very sensitive to the regularization parameter. What are the grounds for considering your implementation to be faithful to the original method? For example, it would help me assess the validity of the comparison if you could clarify whether you verified the implementation by reproducing results reported in the original paper, and how the performance varies with the regularization parameter.
  • While you added removal-and-retraining experiments, the evaluation does not seem to include a counterfactual test such as LDS, which directly measures how well attribution scores predict counterfactual changes. I understand that designing such an experiment for this type of task may be difficult. On the other hand, in the added CIFAR-100 removal experiment, CADT does not seem to show a significant difference from random deletion. I am interested in this area, and this is a question from curiosity: what kind of experiment do you think would be appropriate for validating factor-specific attribution? Or do you consider the current experiments sufficient?

Thank you for the follow-up and for the careful reading.

Q1. Faithfulness of the Concept-TRAK implementation

What we aligned. The public Concept-TRAK repository covers only the pipeline of the AbC qualitative experiments, and we implement Concept-TRAK following its code: the DPS training objective at the 10 states of a 50-step DDIM inversion, per-timestep gradient normalization, the concept utility of Eq. (4), the default random-mask projection, and the damping of App. E.4, λ = λ_r · 0.1 · mean eig(F) with λ_r ∈ [10⁻⁴, 10⁴]. For class-conditional models, the train-side inversion uses the guidance of 2 from App. E.2; since our models have no null condition, the unconditional branch uses a random condition.

What differs on the synthetic benchmark. The original experiment uses a LightningDiT, 64×64 images, 16 OOD queries and P@10, and, per App. E.4, sweeps the damping from 10⁻⁴ to 10⁴ times its recommended value and reports the best result. We use a U-Net, 32×32 images and 1,000 OOD queries, and select λ_r on 500 queries disjoint from the test queries.

Sensitivity to λr\lambda_r (OOD / ID P@10, 1,000 main queries each; * = value selected on the disjoint selection split):

λr\lambda_r1e-41e-31e-21e-11101e21e31e4
Shape, OOD1.910.324.230.837.2*46.741.914.40.8
Shape, ID99.793.680.473.677.7*77.377.892.995.9
Color, OOD25.248.758.460.053.157.484.5*100.0100.0
Color, ID92.180.284.593.293.782.380.4*55.758.7

The color head comes close to the original result (0.90), whereas the shape head does not. The shape gap is not caused by the selection rule: even the best λr\lambda_r chosen on the test queries themselves gives only 46.7. For λr≥103\lambda_r\geq10^{3}, the color head returns red training images for almost every query (OOD 100%, while ID falls to 56–59%), which is why we select λr\lambda_r on the selection split using both OOD and ID queries.

Reproduction. We have not yet reproduced the numbers reported in the original paper. Its simple color/shape benchmark is not public, and the experiments on its other datasets take a long time to run, so a reproduction was not feasible at this stage. We plan to release the code of our method and of the baseline experiments so that the comparison can be verified.

Q2. Counterfactual validation of factor-specific attribution

CIFAR-100. You are right that CADT is not significantly different from same-class random deletion (paired difference +0.10 [−0.14, +0.33]×10⁻⁴). Two aspects of this comparison are worth noting. (i) Same-class random deletion removes 200 of the ≈450 images of the target class; this is a class-level intervention, and its effect on that class does not depend on any attribution method. Only 8–24% of the 200 images deleted by CADT belong to the target class, yet CADT reaches a comparable effect (+0.24 [+0.01, +0.48]), and its effect exceeds that of CLIP-guided deletion (paired difference +0.23 [+0.01, +0.46]) and Concept-TRAK-guided deletion (+0.34 [+0.13, +0.55]); against D-TRAK the interval includes 0. (ii) In line with Concept-TRAK's argument about LDS, a common concept is preserved as long as enough supporting images remain, so deleting a few hundred images changes the utility little and the intervals of all methods are wide. The test with a same-style random control and 8 training seeds (App. I.2) has higher power: CADT minus random is 1.05 [0.88, 1.23]×10⁻⁴, positive for both queries.

What experiments we think would validate factor specificity. As another reviewer also suggested, LDS with a factor-specific utility is the natural extension: e.g., the contrast energy ∥ϵθ(xt,yc)−ϵθ(xt,yref(c))∥2\|\epsilon_\theta(x_t,y_c)-\epsilon_\theta(x_t,y_{\mathrm{ref}(c)})\|^2 at the query's noisy states, or a factor classifier's probability on fixed-seed regenerations. Two further points seem important to us. First, the test should measure specificity, not only accuracy: deletions guided by the style head should change the style utility more than deletions guided by the object head, and vice versa (a 2×2 "attribution head × utility" matrix). Second, because of the threshold behavior described above, the subsets should control how much support of the factor remains (e.g., deletion curves over several K, or a planted rare factor with known support), rather than being uniform 50% subsets. Under our protocol, the LDS part requires 64 subset models per seed (≈76 / 190 GPU hours per seed on ArtBench2 / ArtBench5), which was beyond the budget of this round; we will pursue this experiment in future work. We believe the current experiments provide, to a certain extent, evidence for the effectiveness of our method in factor-specific attribution, as well as a validation of its effect on the model after deletion and retraining.

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

This paper introduces CADT, a method for identifying training examples associated with a particular factor in a diffusion model's output. It compares target and reference conditions at the same noisy state, records how the internal responses change during denoising, and matches the resulting descriptors using a covariance-aware kernel. Experiments on synthetic data, CIFAR-100, and ArtBench show improvements in retrieving examples with the requested shape, class, or style. The paper also includes a reconstruction interpretation and deletion-and-retraining experiments.

Critical strengths
  1. Identifying which training examples support an object's identity or its style is a useful question. The proposed condition contrast gives a concrete way to examine these factors separately.

  2. The experiments cover several settings, including hierarchical classes and corrupted labels. The gain from adding temporal direction to response energy is substantial and supports this part of the method.

  3. Appendix I goes beyond retrieval by examining changes after data deletion and retraining. The comparison of examples with similar Concept-TRAK scores also illustrates the information that a final scalar score leaves out.

Critical weaknesses
  1. My main concern is whether CADT identifies training influence or mainly retrieves semantically related images. The exact removal result concerns the descriptor bank, and the authors acknowledge that it does not describe diffusion-model retraining. Appendix I helps, but its attribution-guided deletion experiment compares against balanced random deletion. Would CADT still have an advantage over random deletion restricted to the same factor or class? Comparisons with other attribution methods on query-specific deletion effects would also help answer this question.

  2. Table A1 raises a question about the kernel. With the same descriptors, ordinary cosine improves macro P@5 from 66.86 to 73.41, and flipped-animal P@5 from 38.27 to 52.54. Why is the covariance-aware kernel preferred? I would like to see this comparison on the other main tasks, since the current result suggests that the descriptor contributes more than the proposed scoring function.

  3. Section 5.1 lists Concept-TRAK as a baseline, but its retrieval results are absent from Tables 1–5. Appendix J examines pairs with similar Concept-TRAK scores, which does not establish better attribution accuracy. Given how closely related the two methods are, a direct quantitative comparison under the same factor definitions and query selection seems necessary. A feature-retrieval baseline would be useful as well.

  4. Some experimental details need clarification. How sensitive is CADT to the reference condition? What exactly is added between the penultimate and full-CADT rows in Table 6, where a large part of the improvement occurs? The ArtBench object–style results also use only 11 and 30 accepted queries. More queries and uncertainty estimates would make these results more convincing.

Questions for authors

see Critical Weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer ubCe (1/2)

Thank you for the constructive review and for recognizing the value of the problem, the experimental coverage and the role of temporal direction. Questions shared by several reviewers are answered in the General Response (GR) 2–6; your four concerns are answered below.

1. Training influence or semantic retrieval? Same-class control and other methods. We ran both experiments you asked for (GR 5/6): leave-top-200-out retraining on CIFAR-100 for CADT, Concept-TRAK, D-TRAK and CLIP similarity, with random deletion within the same class as control. CADT-guided deletion raises the target-class denoising error by +0.24 [+0.01, +0.48]×10⁻⁴, the largest of all deletions and the only one whose 95% interval excludes 0; the paired differences against CLIP, Concept-TRAK and D-TRAK are +0.23 [+0.01, +0.46], +0.34 [+0.13, +0.55] and +0.17 [−0.04, +0.39]×10⁻⁴. As for the same-class random control you specifically asked for: it deletes 200 target-class images directly, whereas only 8–24% of the 200 images selected by CADT belong to the target fine class, and the two produce comparable changes in denoising error (+0.24 vs. +0.15; paired difference +0.10 [−0.14, +0.33]×10⁻⁴); restricting CADT to ranking within the class gives an effect close to same-class random deletion. CADT's effect therefore cannot be explained by "picking target-class examples"; it is tied to the examples it selects outside the class. For comparison, CLIP similarity also selects mostly out-of-class examples (2–18% in the target class), yet its deletion effect is close to zero (+0.01). For the ArtBench experiment of App. I.2, the random-deletion group is drawn from the same same-style candidate pool as CADT, so the experiment already contains a same-factor random control; the complete setup is in GR 5/6.

Two further pieces of evidence bear on the "semantic retrieval" question. (i) Generic similarity (CLIP, DINOv2, CSD) is far below CADT on CIFAR-100 hierarchical retrieval and label flip, though it is better on the ArtBench object head (GR 3/6). (ii) Under the idealized empirical denoiser, the first-order term of the model-level leave-one-out energy change can be decomposed exactly, and its inner product of condition responses is theoretically linked to CADT's stage-wise matching signal (GR 4/6); model-level changes in actual networks are supported by the deletion-and-retraining experiments above.

2. Why the covariance-aware kernel rather than cosine? Because no scoring function dominates, and KλK_\lambda is the one we analyze theoretically; results of all three on every main task are in GR 3/6. Your reading of Table A1 is confirmed on label flip and on the synthetic model, where cosine is better (label flip average P@1 74.59 vs. 67.57). On ArtBench style the opposite holds by a wide margin (style head of ArtBench5 object–style, P@1: KλK_\lambda 90.00, SλS_\lambda 96.67, cosine 46.67; enlarged query set: 94.87 / 97.44 / 70.94); on the CIFAR-100 hierarchy the three are close (50.59/21.68, 50.22/22.22, 49.19/20.81). We agree with you that the descriptor contributes more than the scoring function, and will state this in the paper.

3. Concept-TRAK and feature-retrieval baselines. We implemented Concept-TRAK following its official repository and compared it with CADT in all settings of Tables 1–5, with the same factor definitions and queries (why the comparison was missing from the submission, the implementation and the full comparison are in GR 2/6); CLIP, DINOv2, CSD and CLIP text retrieval are in GR 3/6. In short, CADT is better than Concept-TRAK on CIFAR-100 hierarchical retrieval (P@5 50.59/21.68 vs. 44.97/15.62), on label flip (flipped-source P@1 37.84 vs. 12.43) and on ArtBench object–style, and both saturate on pure style. We agree that this comparison belongs in the main tables. We also added gradient baselines that receive the same factor information (TRAK / D-TRAK with a contrastive utility; results in our response to GQFn, 1 (cont.)).

AuthorsOpenReview ↗

Response to Reviewer ubCe (2/2)

4(a). Sensitivity to the reference condition. CADT's factor specificity comes from a reference that changes only the target factor and keeps the other factors. We ablated the contrast itself (H(xt,yc)H(x_t,y_c) alone) and the choice of yrefy_{\rm ref}:

VariantSynthetic OOD P@1 (shape / color)CIFAR-100 hierarchy P@5 (mid / fine)CIFAR-100 flipped-source P@1
Default: change only the target factor91.20 / 90.8050.59 / 21.6837.84
No reference49.00 / 50.1047.30 / 18.2730.81
Null (all-zero) reference6.00 / 94.00––
Both factors flipped0.80 / 13.30––

Without a reference, the two factor heads on the synthetic data degenerate to nearly the same ranking, and both CIFAR-100 tasks drop; with a null reference or with both factors flipped, factor specificity is lost (ID precision stays above 89%). The CIFAR-100 model is a two-class conditional model, in which the all-zero condition is equivalent to animal and there is only one factor, so only the default and no reference can be compared.

4(b). What is added between the penultimate and the full row of Table 6? This step also changes the temporal support, which the original table did not show: rows 1–4 use all timestep intervals, whereas the full method uses only the last third of the trajectory (t=300→100t=300\to100 on CIFAR-100). This was a presentation error on our side. The table below adds the components one at a time on the late support:

Variant (late support, KλK_\lambda)Mid-level P@5Fine-class P@5
Velocity only46.1116.59
+ Energy50.1621.24
+ Multi-seed stability50.2721.62
+ Second-order reliability (= Full CADT)50.5921.68
Energy only, all timesteps (Table 6, row 1)13.732.65

The gain from 35.24 to 50.59 therefore comes mainly from the late temporal support together with the energy term. Single-timestep variant: on hierarchical retrieval the best single timestep reaches 49.78/19.89 and the full trajectory 50.59/21.68, so the trajectory adds a gain on fine-class retrieval; at the same time, the best timestep varies with the target factor (t=900t=900 for shape and t=600t=600 for color on the synthetic data).

4(c). Only 11 and 30 object–style queries; uncertainty. You are right that the intervals on the original query sets are wide (ArtBench2 object-head P@1 9.09%, i.e. 1 of 11 queries). The small number came from generating few candidates, not from strict filtering. We enlarged the query sets with the original pipeline to 115 / 117 valid queries and report bootstrap intervals; full results on both datasets, object base rates and the judge audit are in GR 6/6. The style results are confirmed with narrow intervals; on the object head, CADT is clearly above Concept-TRAK but below CLIP/DINOv2 retrieval, because cats and dogs are concepts the pretrained model already knows, and only 2.84% / 4.72% of the training images contain them.

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

Related papers

Discussion (0)

Sign in to comment.