acceptodds
Under review as a conference paper at ICLR 2027

Presentation-Order-Gated Sensitivity Implants

Abstract

We construct a small, label-preserving modification of fine-tuning data whose effect depends on the batch order. Under a uniformly random order, the modified data train a model that behaves like one trained on clean data. Under the designated order, which presents the modified batches last, the same data train a model that predicts a prescribed class whenever a trigger pattern appears in its input. This order is not recorded by a fine-tuning run, and a checksum of the data does not cover it. In one pass of mini-batch gradient descent, later batches change the final weights more than earlier ones, and the base model determines how fast this effect decays with position. When the modified batches are presented last, they move the weights toward the implant. When a random order spreads them over the pass, the clean batches that follow them cancel their contribution. The modification of each batch is designed for the position at which that batch appears in training. Experiments on five tasks in images, audio and graphs confirm this behavior: the designated order implants the response, a random order of the same data does not, and the delivered data remain close to clean data at the base weights. The response requires the exact position of each modified batch, and it persists when the designated order is repeated over many epochs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
data poisoning, backdoor attack, clean-label, mini-batch gradient descent

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jH1YOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies label-preserving data modifications that induce a trigger response under a designated training order while remaining inactive under random orders. It proposes a trajectory-dependent construction and evaluates the effect across image, audio, and graph tasks.

Critical strengths

The order-dependent poisoning objective is interesting and highlights a training-process risk that dataset checksums alone cannot capture.

The experiments cover several modalities, and the appendices openly report unsuccessful settings and accuracy costs. These disclosures help clarify the phenomenon’s scope.

Critical weaknesses

First, the central approximation is insufficiently justified. Appendix A.1 drops $ (v^\top H_j y(t))$ on the grounds that it cancels across orders. However, the transverse trajectory depends on the order, so this cancellation does not generally follow. The omitted term could contribute at the same order as the curvature effect being analyzed.

Next, the normal behavior under random orders is not fully established. Table 1 reports random-order trigger sensitivity but omits its clean accuracy. Near-zero sensitivity alone does not establish equivalence to clean-data training. Moreover, the 64-epoch run has a residual ratio of 6.20 and random-order sensitivity of 0.057, so it does not demonstrate persistence together with the single-pass stealth properties.

Then, I find that Appendix B reports an unsuccessful public ResNet-20 setting before moving to custom CNNs. Accuracy costs are also substantial in some settings: on Speech Commands, the designated run is 19.9 points worse than clean-data fine-tuning.

Overall, I lean toward weak rejection. The conditional poisoning effect is interesting, but the theoretical explanation and the claim of normal random-order behavior need stronger support

Questions for authors

What assumptions justify cancelling (v⊤Hjy(t))(v^\top H_j y(t)) across orders? Can you bound its contribution or measure it against the retained curvature term along the actual trajectories?

Could you report random-order clean accuracy alongside matched clean-data runs, with uncertainty estimates? Is there a multi-epoch setting that simultaneously preserves task performance, low random-order sensitivity, and a residual ratio below one?

Does the effect hold on a standard pretrained architecture under fine-tuning settings that preserve clean accuracy? For the unsuccessful ResNet-20 case, is the proposed relaxation-length explanation supported by measurements?

Which split was used to select the working points from the 273 CIFAR-10 and 223 SVHN configurations? Were the selected configurations then evaluated on untouched data and fresh random orders?

6 comments
AuthorsOpenReview ↗

W1 and Q1: orthogonal term

The transverse term. You are right on both points. The transverse part y(t)y(t) of the displacement is determined by the order, so v⊤Hj y(t)v^\top H_j\, y(t) does not cancel between two orders. We have corrected the discussion in Appendix A.1.

We perform measurements along the actual trajectories as asked. The specific settings and results are as follows.

Setting.

  • Dataset: MNIST;
  • Net size 206,922206,922 (reported in paper);
  • Optimizer: plain SGD with η=0.05\eta = 0.05;
  • Batches: 448448 batches, 128128 samples in each batch;
  • Modification (kk):6464 batches,
  • Threshold distance: Δ=1.27\Delta = 1.27.

From w0w_0 we run the designated order and four random orders. At every step we take the batch presented at that step and compute one Hessian-vector product hj=Hjvh_j = H_j v, where HjH_j is the Hessian of that batch's loss at w0w_0; the trajectory quantities Z(t)Z(t) and y(t)y(t) are read from the actual weights w(t)w^{(t)}. Summing the exact one-step update along each trajectory gives

Z(m)=P+C+T+S,Z(m) = P + C + T + S ,

where P=−η∑tbσ(t)P = -\eta \sum_t b_{\sigma(t)} is the push, C=−η∑tZ(t) v⊤Hσ(t)vC = -\eta \sum_t Z(t)\, v^\top H_{\sigma(t)} v is the retained curvature term, T=−η∑ty(t)⊤hσ(t)T = -\eta \sum_t y(t)^\top h_{\sigma(t)} is the dropped transverse term, and SS is the remainder of second and higher order. PP depends only on the set of batches and is the same for all orders.

orderPPCCTTSSmeasured Z(m)Z(m)
designated2.618−1.480-1.480−1.541-1.5411.1760.773
random 12.618−1.928-1.928−1.399-1.3991.2360.527
random 22.618−1.845-1.845−1.426-1.4261.1560.503
random 32.618−2.050-2.050−1.276-1.2761.2270.520
random 42.618−1.831-1.831−1.411-1.4111.1480.524

CC and TT are the two halves of the same first-order term v⊤Hσ(t) u(t)v^\top H_{\sigma(t)}\, u(t), split only because the recursion tracks ZZ; together they pull ZZ back by 3.0213.021 in the designated order and by 3.2423.242 to 3.3273.327 in the four random orders.

Changes to the paper. All changes are marked in blue in the revised PDF.

  • Section 4, p. 5, the sentence between the statement of the two approximations and Eq. (7): we have corrected the statement on the orthogonal displacement.
  • Appendix A.1, p. 13, the sentence just after Eq. (22): we have corrected the discussion about the transverse term.
  • Appendix A.2, p. 13, "Size of the dropped terms": we have added the table above (as Table 2), and related discussion.
AuthorsOpenReview ↗

W2, part1 and Q2, part1: behavior under random orders

Thank you for pointing out this issue. Table 1 gave the base accuracy and two accuracy drops and left the random-order run out. The revised Table 1 (p. 5) reports the base accuracy and the three accuracies after fine-tuning, so the comparison you ask for can be read off directly.

AuthorsOpenReview ↗

W2 part2 and Q2, part2: overall: task performance, low random-order sensitivity, and a residual ratio

Yes, we frequently encounter runs that meet the criteria. For example, on SVHN, With the designated order repeated in each of four passes and Adam at a learning rate of 3×10−43 \times 10^{-4}, modifying 1616 batches, the three quantities hold together: the random-order sensitivity is +0.010+0.010, the designated sensitivity is +0.350+0.350, and the residual ratio is 0.810.81.

The 6464-epoch CIFAR-10 run you cite is reported for a different purpose, which we should have separated more clearly: it shows that the accuracy cost of the designated run falls as the number of passes grows, from 37.537.5 points at one pass to 7.07.0 points at 6464. You are right that it does not demonstrate the single-pass stealth properties at the same time. Its residual ratio is 6.206.20 because the construction of those runs does not bound the total gradient at w0w_0, and its random-order sensitivity is +0.057+0.057 because the modified data are seen 6464 times under any order. This row is about cost and not about stealth, and we point to the SVHN row above for a multi-epoch setting that meets all three conditions.

Reviewer_Q817OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a data poisoning method that controls the activation of a backdoor through the presentation order of mini-batches during fine-tuning. The authors show that the same modified training data can induce a backdoor under a designated order, while producing little backdoor response under a random order. To explain this phenomenon, they develop an approximate theory based on SGD dynamics. The proposed method is evaluated on five tasks involving image, audio, and graph classification.

Critical strengths

The paper shows a phenomenon in which the activation of a backdoor depends on the mini-batch presentation order, even when the same modified training data are used. The authors also propose an intuitive theoretical framework to explain the underlying mechanism.

Critical weaknesses
  1. The validity of the proposed theory is not sufficiently verified.

1-1. While the theory analyzes the parameter displacement Z, the experiments evaluate the sensitivity S. The relationship between these quantities remains unclear. In particular, the theoretical predictions for Z are not directly compared with measured values.

1-2. The derivation of Eq. (7) neglects the batch dependence of the Hessian and the contribution from parameter displacements orthogonal to the implant direction. However, the validity of these approximations is not sufficiently examined.

  1. The effectiveness of the proposed method under practical learning rates remains questionable.

2-1. Practical fine-tuning often requires small learning rates to maintain high accuracy. Learning rate scheduling and possible changes in curvature during training may further increase the relaxation length L relative to the epoch length m, potentially weakening the proposed effect. Indeed, in Appendix B, the authors report that fine-tuning with a small learning rate did not produce sufficient order dependence. However, the relationship between the learning rate, the effectiveness of the method, and classification accuracy has not been systematically investigated.

Questions for authors
  1. Could the authors directly compare the theoretically predicted Z with its measured values? Could they also provide a theoretical explanation of the relationship between Z and S, or at least investigate this relationship numerically?
  2. Could the authors numerically examine the differences in curvature between clean and modified batches, as well as the variation across batches? Could they also verify the validity of neglecting the contributions from parameter displacements orthogonal to the implant direction?
  3. Could the authors systematically investigate how the learning rate affects L/m, Z, the designated sensitivity, the random-order sensitivity, and classification accuracy? In particular, does the proposed method remain effective while maintaining high accuracy and sufficiently low random-order sensitivity? How does the learning-rate schedule affect the results?
  4. Could the authors examine how the curvature lambda changes during training? In particular, is the assumption of constant curvature in the multi-epoch analysis justified?
  5. (Minor question) Why is the residual ratio gamma for SVHN in Table 1 substantially larger than those for the other tasks? Could this be related to the relatively high classification accuracy of the base model?
6 comments
AuthorsOpenReview ↗

W1, Part 1 and Q1: how displacement Z and sensitivity S are linked

We perform the comparison. On CIFAR-10, we test 20 runs with different learning rates and block lengths. For each run we evaluate the closed form of Eq. (7) from the curvature and the push measured at w0w_0 and compare it with the displacement measured after the pass. The prediction is larger than the measured value in every run, so Eq. (7) is an upper bound on ZZ.

Nevertheless, the relation the construction relies on is valid in this empirical test: the sensitivity increases with the displacement. Sorting the twenty runs by Z/ΔZ/\Delta into four groups of five, the designated sensitivity rises through 0.0980.098, 0.1420.142, 0.2010.201 and 0.2660.266, while the random-order sensitivity stays between 0.0320.032 and 0.0380.038; the Spearman correlation between Z/ΔZ/\Delta and the designated sensitivity is 0.7740.774. We report both in the new Appendix A.6.

AuthorsOpenReview ↗

W1, part 2 and Q2: displacement variation and the orthogonal term

We examined the three quantities you asked for. The first and third are measured on MNIST: a network with 206,922206,922 parameters, plain SGD with η=0.05\eta = 0.05, one pass over 448448 batches of 128128 samples, 6464 of them modified, Δ=1.27\Delta = 1.27, mean curvature along vv over all batches λ=0.2065\lambda = 0.2065, relaxation length $L = 1/(\eta\lambda) = 97$ batches.

Curvature of clean and modified batches. At w0w_0 we computed v⊤Hjvv^\top H_j v for each of the 6464 modified batches before and after modification. The curvature along vv is 0.2630.263 (s.d. 0.1710.171) on the modified batches and 0.1870.187 (s.d. 0.1100.110) on the same batches unmodified. In the designated order the last 6464 steps are all modified batches, so their weights are products of (1−η v⊤Hjv)(1 - \eta\, v^\top H_j v) over modified batches. With 0.2630.263 in place of λ\lambda, the sum of these 6464 weights drops from 47.047.0 to 43.543.5, by 7.5%7.5\%. The measured displacements and the gap reported below include this effect, and the gap remains significant.

Variation across batches. On the CIFAR-10 base model (12801280 batches of 3232) the curvature along vv has mean λ=0.0472\lambda = 0.0472, standard deviation 0.06530.0653, minimum −0.0204-0.0204 and maximum 1.1421.142. Replacing each batch's curvature by a constant does not hold batch by batch. What Eq. (7) uses is the product of (1−η v⊤Hjv)(1 - \eta\, v^\top H_j v) over the batches presented after a given step. Its logarithm is

= -\eta \sum_j v^\top H_j v - \frac{\eta^2}{2} \sum_j (v^\top H_j v)^2 + O(\eta^3) .$$ At $\eta = 0.02$ the second-order term is $1.4 \times 10^{-3}$ of the first per batch. The first-order term over $n$ batches differs from $-\eta n \lambda$ by the fluctuation of a partial sum, at most about $\eta\, \mathrm{s.d.}\, \sqrt{n} = 0.02 \times 0.0653 \times \sqrt{1280} = 0.047$, so each weight is within about $5\%$ of $\rho^{\,n}$. In addition, the decomposition in the next paragraph uses each batch's own curvature and needs no constant. **Displacement orthogonal to $v$.** Reviewer jH1Y raised the same point. The justification in Appendix A.1, that the term cancels between the designated order and a random order, is wrong: the transverse part $y(t)$ of the displacement depends on the order. We measured the term along the actual trajectories of the designated order and four random orders. Summing the exact one-step update gives $Z(m) = P + C + T + S$, with the push $P = -\eta \sum_t b_{\sigma(t)}$, the curvature term $C = -\eta \sum_t Z(t)\, v^\top H_{\sigma(t)} v$, the transverse term $T = -\eta \sum_t y(t)^\top H_{\sigma(t)} v$, and the second-order remainder $S$. The five rows of the decomposition are Table 2 of the revised paper, in the new Appendix A.2, and we give the short form here. The ratio $|T|/|C|$ is $0.62$ to $1.04$, so the dropped term is not negligible within a single order. The push $P$ is identical across orders, so it contributes nothing to the difference between them. The gap is produced by the curvature term, $+0.434$, which is the mechanism of Section 4; the transverse term has the same sign in all five orders and reduces the gap by $0.163$, leaving a measured gap of $+0.255$, which is $24$ times the standard deviation $0.010$ of $Z(m)$ over the four random orders. The gap itself is measured, not predicted: the designated order moves the weights $0.773$ along $v$ against $0.518$ for the random orders, $1.49$ times as far. **Changes to the paper.** All changes are marked in blue in the revised PDF. - Section 4. We removed the reason "because the orders we compare share it" for dropping the orthogonal displacement and added: "The dropped term depends on the order; Appendix A.2 measures it on MNIST: the dropped term reduces the gap between the designated and a random order, but the gap remains significant." - Appendix A.1. We removed "it is shared by the two orders compared below and cancels in their difference" and added: "This term depends on the order through $y(t)$ and does not cancel between two orders; Section A.2 measures it along actual trajectories." - New Appendix A.2, "Size of the dropped terms". It gives the decomposition and the table above as Table 2, the conclusion that the gap is produced by the curvature term, as Eq. (7) describes, and remains significant, and the curvature of the modified batches before and after modification with its effect of $7.5\%$ on the designated displacement.
AuthorsOpenReview ↗

W2, part 1 and Q3: learning rate and "gentle" modification and sensitivity-accuracy-trade-off

We have performed the experiments you ask for and report them below. To interpret the results, it is useful to recall that the work is an existence claim, as the revised abstract now states: we report a potential weakness of a network that appears to have been trained normally

  • in a given setting, the presentation order of label-preserving data can decide whether a sensitivity is implanted. Many factors can affect the existence and the exposure of the order-gated sensitivity. Below we report the effect of the learning rate and its schedule.

Learning rate. Twenty runs on one base (CIFAR-10, 2,170,1782,170,178 parameters, clean accuracy 0.90240.9024, one pass over 12801280 batches of 3232 samples) cover five learning rates and two block lengths kk.

η\etaL/mL/mkkZ/ΔZ/\Deltarandomdesignatedacc. cleanacc. desig.
0.0020.0028.278.2732320.0900.0900.0330.0330.0900.0900.9030.9030.8110.811
0.0020.0028.278.2764640.1400.1400.0400.0400.1180.1180.9030.9030.8020.802
0.00630.00632.622.6232320.1930.1930.0340.0340.1300.1300.9030.9030.7750.775
0.00630.00632.622.6264640.2860.2860.0410.0410.1690.1690.9030.9030.7520.752
0.020.020.830.8332320.3530.3530.0320.0320.2200.2200.9010.9010.6820.682
0.020.020.830.8364640.4630.4630.0370.0370.2950.2950.9010.9010.6050.605
0.0630.0630.260.2632320.5970.5970.0370.0370.2710.2710.8900.8900.5890.589
0.0630.0630.260.2664640.7910.7910.0450.0450.1900.1900.8900.8900.6190.619
0.20.20.080.0832320.6830.6830.0250.0250.2260.2260.8720.8720.6920.692
0.20.20.080.0864641.2221.2220.0270.0270.4200.4200.8720.8720.5810.581

Z/ΔZ/\Delta is the measured displacement in units of the threshold distance, kk is the number of modified batches, and the two accuracy columns are the clean-data run and the designated run.

The row at η=0.02\eta = 0.02 is the one you ask for: the clean run is within 0.20.2 points of the base, the random-order sensitivity is 0.0320.032 to 0.0370.037, and the designated sensitivity is 0.2200.220 to 0.2950.295.

Schedule. Four schedules at the same working point, each with the designated order once and twelve random orders:

scheduleη\eta over last kkrandomdesignatedacc. cleanacc. randomacc. desig.
constant0.02000.0370.2410.9010.8910.628
cosine to 000.00000.0320.0160.9030.8950.904
cosine to 0.0020.0020.00200.0330.0620.9030.8940.781
step to 0.0020.002 at 90%0.00200.0320.0620.9020.8940.778

The two annealed schedules end at the same η\eta and give the same result, so what the schedule controls is the mean learning rate over the last kk batches.

Curvature. On the same base λ\lambda is 0.04840.0484 at w0w_0 and 0.05660.0566, 0.05000.0500, 0.04830.0483 at three points inside the pass. At the ends of four later passes it is 1.84031.8403, 1.39471.3947, 1.30011.3001, 1.20201.2020. Appendix A.1 now states that the constant-curvature step is a one-pass statement.

Lower modification. On CIFAR-10, modifying 0.31%0.31\% of the samples rather than 2.50%2.50\% gives a designated sensitivity of 0.1110.111 against 0.0170.017 under a random order. The designated run then keeps an accuracy of 0.7990.799 against 0.8060.806 for the clean-data run, and the residual ratio falls from 0.540.54 to 0.120.12 (Section 6).

Reviewer_Sp3cOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies data poisioning attack by manipulating the training batches order. It trains a model like clean data under a uniformly random batch order, and implants a response to a trigger pattern under a designated order that presents the modified batches last. A theoretical framework is developed. Experiments across five tasks confirm the behaviour that the designated order implants the response.

Critical strengths
  • This paper presents a different way to study clean-label poisioning. Unlike previous works that generally implant malicious behavior independent of data batch order, this work shows that a desinated order also makes a strong influence. The paper comes up with a theoretical framework that explains how the contribution of a batch to the final weights decays exponentially, hence supports the design of last batch.
Critical weaknesses
  • One biggest concern is about the strong assumption that may limited its practical application. The proposed attack assumes the exact order of batches are modifiable, thus modified batches can be presented last. However, a modern model usually uses distributed training, random augmentations and data shuffling in different epochs. This makes the attack method highly fragile.
Questions for authors
  1. Is there a practical application where the batch order can be manipulated? What about two training epochs (very common in VLM), will last batches in first epoch make any impacts on the attack?
  2. How does the method work with different optimizors, such as momentum or Sign-based methods?
5 comments
AuthorsOpenReview ↗

W1 (Part 1) and Q1: order assumption and multiple epochs

Reshuffling. Thanks for pointing out that reshuffling removes the effect. In revision, we clarify (Remark 1, after Proposition 1, p. 5) that the attack requires the modified batches to be presented in the final stretch of the whole run. Reshuffling after the first pass would reduce the push along vv (the direction of sensitivity) that the batches at the end of the first pass produced.

This phenomenon has been reported in Fig. 3 in the submission (See the discussion in Panel (a)). In the revision, we state the condition clearly and add a table (Appendix C.5, Table 9) that lists the reshuffled series of Fig. 3(a), together with a one-pass and a reused-order run on the same base.

A similar argument applies to the effect of the last batches in a non-final epoch -- batches in later epochs remove them from the final stage of training.

Where the order is fixed by the party that delivers the data. We report a weakness, and once the weakness exists the risk is real. How it would be exploited is outside the scope of this work and we have not attempted it. One possibility is that the party that modifies the data learns or arranges the random seed of the training party, and then knows which batches are presented last.

AuthorsOpenReview ↗

W1, Part 2: different training schemes

Random augmentation. Thanks for pointing out these factors.

We have performed the experiments you ask for and report them below. It is useful to recall that the work is an existence claim: we report a potential weakness of a network that appears to have been trained normally. Many factors can affect the existence and the exposure of the order-gated sensitivity. Reshuffling is the one that removes it, and we report it separately. Below we report random augmentation and distributed training.

Random augmentation. On the CIFAR-10 working point (base accuracy 0.90240.9024, one pass over 12801280 batches of 3232 samples, η=0.02\eta = 0.02, 6464 batches modified), with the designated order once and twelve random orders at each strength:

victim's looprandomdesignatedacc. cleanacc. randomacc. desig.
no augmentation0.0370.2410.9010.8910.628
random crop0.0260.1720.9010.8940.699
crop and flip0.0270.1810.9010.8940.746

The crop is a 44-pixel reflection pad with an independent random crop per sample; the flip is an independent random horizontal flip, which is the augmentation this base was pretrained with. In every row the designated order beats all twelve random orders, so the permutation pp is 1/13=0.0771/13 = 0.077.

Distributed training. With WW workers averaging their gradients the recursion keeps its form with the step index counted in global steps, so a batch carries the weight ρ\rho raised to the number of global steps presented after it. The modified block then has to fill the last global steps, its length is quantized in units of WW, and the relaxation length measured in batches becomes W/(ηλ)W/(\eta\lambda). The derivation is in Appendix A.5. We have not run a data-parallel experiment.

AuthorsOpenReview ↗

Q2: Optimizers

We performed the asked tests as follows.

Momentum. It keeps working with momentum, and the stopping point is as predicted. A heavy-ball buffer is a first-order low-pass filter with time constant τ=1/(1−β)\tau = 1/(1-\beta) batches, so a modified block of length kk gets the full position weight on each of its batches only once τ≲k\tau \lesssim k. On the same base (CIFAR-10, clean accuracy 0.90240.9024, one pass), with the learning rate of each coefficient scaled by 1−β1-\beta so that the asymptotic step is the same 0.020.02 in every row and the comparison is about the momentum and not about the step:

optimizerτ\taukkdesignated
plain gradient descent132+0.203+0.203
momentum 0.90.9108+0.092+0.092
momentum 0.90.91016+0.224+0.224
momentum 0.990.9910016+0.038+0.038
momentum 0.990.9910032+0.207+0.207

The random-order is +0.033+0.033 in the plain-descent row and the clean-data keeps an accuracy of 0.90130.9013. The transition sits where τ\tau and kk cross, as the filter argument says.

Adam. The effect is strongest at the middle learning rates: with k=32k = 32 at 10−410^{-4} the designated sensitivity is +0.355+0.355 with a clean-data accuracy of 0.89730.8973, and with k=64k = 64 at 10−310^{-3} it is +0.526+0.526 with 0.85770.8577. At 10−210^{-2} neither order responds, +0.003+0.003 for the designated order, because the clean-data run itself is down to an accuracy of 0.46010.4601 and the run has left the regime where a one-pass fine-tune means anything.

Sign-based optimizers. We have not run signSGD or Lion, and we do not claim the above extends to them. The nearest thing our data contain is Adam, which approaches a per-coordinate sign step when the second moment is small, and there the effect survives. A strict sign rule removes the magnitude of the gradient entirely, which is the quantity the construction uses, so we would expect the push per batch to saturate and the block to have to be longer; we have not measured this and we say so in the appendix rather than assert it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.