acceptodds
Under review as a conference paper at ICLR 2027

Characterizing the Separability Gap in Prior-Data Fitted Networks

13 views

Abstract

Prior-data fitted networks (PFNs) are transformers pre-trained on simulated datasets to approximate the posterior predictive distribution (PPD) of a task prior. Evaluating this approximation outside conjugate settings has remained challenging: for the priors of current PFNs the posterior is intractable, and real-world datasets lack ground-truth PPDs. A second difficulty is conceptual, as the standard formalization of PFNs assumes a separable prior, under which the feature distribution and the labeling mechanism are a priori independent. The structural-causal-model priors of current PFNs are non-separable, since features and labels are outputs of one shared random mechanism. We show that this changes the minimizer of the training objective and can invalidate the fixed-query martingale tests used to diagnose Bayesian behavior in PFNs. We characterize the resulting separability gap, show that it affects interpretability and Bayesian optimization, and exploit it to turn PFNs into semi-supervised predictors that use additional unlabeled rows. We characterize the gap of TabICL architectures pre-trained on these priors, test which predictive standard pre-training recovers, and provide insights into how PFNs behave as Bayesian predictors.

Primary area:
foundation or frontier models, including LLMs
Keywords:
PFN, TFM, TabPFN, TabICL, tabular data

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_DvyVOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors work on PFNs, focusing on the distinction between query-updated posterior predictive distributions (PPDs) and context-updated PPDs, and on the problem that training targets the former while the community tends to assume the latter. They show that the two are identical under a separability assumption, which does not hold in any non-trivial real-world use cases.

Critical strengths
  • The authors clearly motivate and convincingly evaluate the difference between what PFN pretraining targets and what practitioners assume. This is relevant because it explores a clear discrepancy.
  • Clean experiments demonstrate the theoretical claims.

Overall, the theoretical and empirical evidence is sufficient in my opinion.

Critical weaknesses
  • While the experiments effectively demonstrate the problem, they lack evidence that it matters in practice, as they primarily rely on n=2n=2, while Proposition 2 shows that for larger contexts (at least asymptotically) the gap disappears.
Questions for authors
  • How large is the gap ∣π−π^∣|\pi - \hat \pi| under a realistic prior as a function of nn and dd?
  • Can you re-evaluate an existing PFN diagnostic with the corrected reference, e.g., the TabMGP martingale drift or the Hülskamp et al. amortization gap?
2 comments

We welcome the reviewer’s assertion of the sufficiency of theoretical and empirical evidence in our submission and appreciate the positive review.

In the following, we will address the reviewer’s points:

While the experiments effectively demonstrate the problem, they lack evidence that it matters in practice, as they primarily rely on n=2n=2

We would like to stress that only the experiments that need an exact reference use n=2n=2. This is due to the fact that these experiments require a finite task space so that we can exactly compute π\pi and πˉ\bar{\pi}. The practical evidence is our SemiTabICL model on the TALENT benchmark as well as UCI datasets, which provides direct evidence that non-separability matters in practice.

How large is the gap ∣π−π^∣|\pi - \hat \pi| under a realistic prior as a function of nn and dd?

We thank the reviewer for this question. We cannot directly assess the gap between the two predictives, as non-separable priors are in general not amenable to MCMC sampling. However, we extended our experiments with a more realistic prior based on a Bayesian neural network, as also used in the pretraining of the original TabPFN. For this, we mix two separable prior regimes to obtain a non-separable version together with a separable one, where the marginal feature and conditional label distributions are the same. The only difference between the two priors is thus their dependence structure.

We train a TabICL model on each prior and find that the difference, as expected, vanishes for large context sizes and only the PFN pretrained on the non-separable prior can use the query feature as evidence at small nn to identify which of the regimes it is tasked with. A figure and a detailed experimental description is available in the new Appendix C.4.

Our SemiTabICL experiment further shows that the difference matters in practice and enables substantial performance gains on well-established benchmarks, like the TALENT benchmark.

Can you re-evaluate an existing PFN diagnostic with the corrected reference, e.g., the TabMGP martingale drift [...]

We thank the reviewer for this interesting question. Our results imply that a PFN failing the martingale check is expected, because the exact Bayes predictor fails it too under a non-separable prior.

For a separable prior, we know that the approximation target of a PFN passes the martingale condition. To check whether this is also the case for a PFN trained on a separable prior, we conduct a martingale property check for the mixed prior setup discussed above, where we expect (more) violation for the PFN trained on the non-separable prior.

Results (Appendix C.5) show a small drift of the PFN pretrained on the separable prior, which is due to its approximation error. Compared against this baseline, the PFN pretrained on the non-separable prior shows substantially more drift. As expected, this drift stagnates as the sequential addition of features helps to identify the two different prior regimes in the mixture more and more.

Thanks for the explanations and not sure how I missed the n>2n> 2 in the experiment description (sorry for that).

I'm satisfied with the answer and will increase my score.

Reviewer_cxwgOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In supervised machine learning settings with feature and label sets XX and YY, we typically assume a data-generating structure of the form p(Y,X)=p(X)p(Y∣X)p(Y, X) = p(X) p(Y|X), where the features are sampled independently from the labels and therefore unlabelled datapoints provide no information about the labels of other unlabelled datapoints. However, the SCMs used to train large-scale PFNs for tabular data prediction result in a data-generating process that does not exhibit this property$\textemdash$the joint distribution over features and labels does not factorise in the above way, and so unlabelled points contain information about the labels of other unlabelled points.

The authors of this submission refer to this independence property as separability, and they investigate the pitfalls associated with erroneously assuming a separable data-generating process, as has invariably been done so far in the PFN/TFM community. In addition to these insights, they propose a new type of PFN that exploits the non-separability of real tabular data (via the TALENT benchmark) by conditioning on auxiliary unlabelled rows. It leads to impressive performance gains.

Critical strengths

S1 The paper is very well written.

S2 The paper tackles a problem with current PFN/TFM formulations that is very interesting, impactful, and likely to spark further research.

S3 The experiments in the paper are sufficient to support the claims, and they generally contribute meaningfully to the narrative of the paper. They are well-designed and make it easy to draw insightful conclusions.

Critical weaknesses

W1: Marginalisation Consistency. The authors pay no reference to the concept of marginalisation consistency, defined in PFNs/neural processes as π(y1∣x1,dn)≡∫π(y1,y2∣x1,x2,dn)q(x2)dx2dy2\pi(y_1|\mathbf{x}_1,d_n) \equiv \int \pi\bigl(\\{y_1,y_2\\}|\\{\mathbf{x}_1,\mathbf{x}_2\\},d_n\bigr) q(\mathbf{x}_2) \mathrm{d}\mathbf{x}_2 \mathrm{d}y_2for any valid density qq (including delta function distributions such that x2\mathbf{x}_2 is fixed arbitrarily). It is a different property to separability, but is nonetheless closely related. If a model exhibits marginalisation consistency, then it's predictions obtained through inductive and transductive protocols are the same. PFN architectures are designed so that marginalisation consistency is hard-coded as an inductive bias. In PFNs (and TNPs) this is typically achieved through masking so that each query token only attends to context tokens, and each context token only attends to other context tokens. Obeying marginalisation consistency means it is impossible for a PFN’s prediction at a given query to depend on the other queries.

Given how closely this is related to the concept of separability, the authors should add a line or two somewhere to delineate the two concepts and explain how they relate to each other. The final paragraph of section 3 should be updated to mention marginalisation consistency. The authors should also be more clear about the fact that they consider a single query in their analysis framework.

The marginalisation consistency of a PFN can be destroyed by removing the masking between query tokens. Of course, such a PFN would need to be mask-free during both training and testing in order to learn the mechanism by which the query tokens are updated by each other. I would appreciate a comment from the authors about mask-free PFNs. For example, if a practitioner knows their downstream data generating process is non-separable, should they use a SemiTabICL architecture or something like a mask-free PFN?

W2: Martingale Checks. In the paragraph starting on line 360, the authors point out that the martingale property should hold for a PFN trained under a separable prior, but not for one trained under a non-separable prior. Perhaps I have misunderstood, but the results in Figure 4 demonstrate that in practice neither PFN exhibits the martingale property$\textemdash$why? Also, equation 3 implies the martingale property should be tested by averaging predictions obtained from a length-1 autoregressive rollout, yet in Figure 4 the averaging seems to have been performed over an autoregressive rollout of length 1000$\textemdash$why? In its current state I struggle to fully grasp and therefore appreciate the value of this section, and would appreciate some further clarity from the authors.

Further Review Fields

Since there are no boxes for reviewers to enter minor correction suggestions or ask minor clarification questions, I use this box for that purpose as well. None of the below points should be interpreted as criticism; they are intended to help the authors in revising their manuscript for camera-ready or resubmission. The AC doesn't need to read the below.

Minor corrections/editorial suggestions

C1 In the abstract, the colon construction of “: for the priors” is unnecessarily stilted. My recommendation is to replace the string “: “ with “ because “, “ as “, or “ since “.

C2 In the abstract, the sentence starting with “A second difficulty is conceptual” actually would benefit from a (semi-)colon construction for better readability$\textemdash$my recommendation is to swap the string “, as ” with “; “.

C3 In the abstract, the comma in “separable prior, under which” should be removed.

C4 In the abstract, the final two sentences both being tricolons detracts from the readability and flow. It feels as though one of the sentences is a placeholder that the authors forgot to finalise. The same applies to the fact that “characterize” is used in both sentences$\textemdash$this leaves a sour taste in readers’ mouths and distracts from the actual content of the words. The final two sentences read as a list of six pursuits that the authors engage in without imparting the reader with any meaningful narrative or concrete takeaways. When re-writing these sentences, the authors should focus on the fact that readers of the abstract want to know: what research questions the authors pose, why these questions are interesting or important, what the findings are, and what the implications are.

C5 In the introduction, “A common data prior choice are structural causal models” should be altered to have the verb in the singular form “is” rather than the plural “are” (by analogy, when discussing lunchtime meal options, “a common choice are sandwiches” feels less correct than “a common choice is sandwiches”, though I would argue neither feel totally right). My recommendation would be to sneakily rewrite the sentence in a way such as “Structural causal models (SCMs) represent a common choice for the data prior”, where the question of verb plurality is pushed out of view.

C6 In machine learning, input-output pairs are usually referred to as “x-y pairs” and the mathematical notation typically follows this ordering convention (even though it is arbitrary). Is there any particular reason why the datapoints are consistently referred to in the opposite order in this submission? i.e., the authors introduce a context dataset as dn=yi,xi{ni=1}d_n = \\{y_i, \mathbf{x}_i\\}^n_\{i=1\}. This is highly pedantic of me, and I fully understand if the authors choose to ignore this very minor point, but to me it seems like an unnecessary (albeit very minor) hinderance on the paper’s readability.

C7 On line 41, the authors should avoid using the word “prior” here and should replace it with something like “antecedent”, “anterior”, “bigger”, “more fundamental”, “retrograde”, or so on. The use of a different meaning of the word “prior” in a paper about Bayesian priors could confuse non-native speakers (actually, it confused me, a native speaker, since I interpreted “prior question” to mean “question about the prior”).

C8 In the paragraph beginning on line 79, I believe I2I_2 should be emboldened throughout as I2\mathbf{I}_2 since it is a matrix.

C9 In the Model and notation paragraph of section 2, the sentence “Given a task p∼Πp\sim\Pi, let X1,…,Xn\mathbf{X}_1,…,\mathbf{X}_n denote the context features and X the query feature.” is unnecessary and should be removed (i.e. given the subsequent sentences).

C10 There is an inconsistency of notation between line 35 and line 118 where dnd_n is defined as yi,xiI=1n\\{y_i,\mathbf{x}_i\\}_{I=1}^n and (yi,xi){i=1}n(y_i,\mathbf{x}_i)_\{i=1\}^n. I assume the former is correct, but regardless, this inconsistency should be ironed-out one way or another. If braces are to be used to refer to the collection of points, then Dn=(Yi,Xi){i=1}nD_n = (Y_i,\mathbf{X}_i)_\{i=1\}^n, which is also on line 118, should be changed accordingly.

C11 On line 145, “like” should be swapped in favour of “such as” as its comparatively informal tone is not so appropriate in scientific writing.

C12 On line 234, “vanishes with the sample size” implies the gap decreases as the sample size decreases, which is wrong. I would change it to “vanishes with increasing sample size”.

C13 On line 421/422 (the first line of that paragraph), “a range of labeled training rows nn” should instead be something like “a range of the number of labeled training rows nn”. The clunkiness of this sentence can also be reduced by swapping “added as” with “of”, as has already been done in Figure 5’s caption.

C14 In Figure 5, it is my view that nn should be ordered from 16 on the lowest row to 96 on the highest row; the current layout feels weird. Also, a colourbar would really help the readability of this figure—it is a super figure with interesting results, but because the results are so different for the two TabICL variants, it initially seems like two separate colourmaps have been used and so the difference is less stark.

Minor Questions

Q1 In definition 1, what operation does ⊗\otimes represent?

Q2 Under a non-separable prior, when would the Martingale property hold under stochastic auxiliary features? In other words, for what qq does the following hold? EX{n+1}∼q(⋅), Y{n=1}∼π(⋅∣x{n+1},dn)[π(A∣x,D{n+1})]≡π(A∣x,dn)\mathbb{E}_{\mathbf{X}_\{n+1\}\sim q(\cdot),\ Y_\{n=1\}\sim \pi(\cdot|\mathbf{x}_\{n+1\},d_n)}[\pi(A|\mathbf{x},D_\{n+1\})] \equiv \pi(A|\mathbf{x},d_n) Would qq need to match the marginal distribution of x\mathbf{x} under the prior?

Questions for authors

No major questions.

3 comments

We thank the reviewer for the positive review and genuinely value the detailed assessment of our work as well as the great editorial suggestions that we have already incorporated in an updated version of this manuscript that we’ve uploaded to OpenReview.

We will address the reviewer’s points below:

[W1] The authors pay no reference to the concept of marginalisation consistency [...]

We thank the reviewer for the pointer to the concept of marginalization consistency, which is indeed closely related to our separability concept. We have incorporated all the suggested improvements into the updated manuscript version.

[W1] I would appreciate a comment from the authors about mask-free PFNs. For example, if a practitioner knows their downstream data generating process is non-separable, should they use a SemiTabICL architecture or something like a mask-free PFN?

Likewise, we appreciate the reviewer’s question about attention masks in PFNs that is also very to the point and are happy to comment on this. The two architectures differ in how they use the unlabeled information:

  • SemiTabICL keeps the unlabeled rows v\mathbf{v} separate from the also unlabeled queries. Its predictions are inductive whenever these unlabeled rows v\mathbf{v} and the test rows are disjoint. The number of unlabeled rows v\mathbf{v} can be chosen independently of the number of queries, within the range seen in pretraining. Also, SemiTabICL with a fixed set v\mathbf{v} gives per-query predictions that are invariant to the other test rows, so SemiTabICL keeps inductive and transductive settings explicitly distinct.
  • A mask-free PFN adds the unlabeled evidence to the test batch, which makes it transductive by construction. The prediction for a query depends on which other queries are in the batch.

So what should a practitioner choose? Separability is a property of the prior, not of a single data-generating process p0p_0. What matters is whether the feature distribuiton contains information about the conditional label distribution in the way the prior encodes, and both SemiTabICL as well as a mask-free PFN/TabICL can exploit unlabeled features if that is the case. If that is not the case, Prop. 3 implies that a non-separable prior can trade label fit for feature fit. Given this, we would recommend to use SemiTabICL over a mask-free PFN due to the explicit control over inductive and trasductive inference.

  • If the dependence encoded by the prior is plausible for the data and unlabeled rows other than the test rows are available, use SemiTabICL with those rows.
  • If the dependence is plausible but only the test rows are unlabeled, use SemiTabICL with the other unlabeled rows in the test batch, provided the evaluation protocol permits transductive inference (Section 3).

[...] the authors point out that the martingale property should hold for a PFN trained under a separable prior, but not for one trained under a non-separable prior. Perhaps I have misunderstood, but the results in Figure 4 demonstrate that in practice neither PFN exhibits the martingale property --- why?

We thank the reviewer for this question and agree that this section needs clarification, as there seems to be a misunderstanding of what Figure 4 actually shows. In the text, we have not explained sufficiently that the experiment involves no trained network. Both curves are exact predictives under the same non-separable ordered-cluster prior. πˉ\bar{\pi} is not the predictive of a separable prior. Under a separable prior πˉ=π\bar{\pi} = \pi and the drift is identically zero (Section 3), so that case is not plotted. Figure 4 thus shows that under non-separability the exact targets themselves violate Eq. (3), independent of any approximation error.

Also, equation 3 implies the martingale property should be tested by averaging predictions obtained from a length-1 autoregressive rollout, yet in Figure 4 the averaging seems to have been performed over an autoregressive rollout of length 1000 --- why?

We appreciate the reviewer’s question and are happy to provide clarification. We actually do not average over the steps of the rollout. At each step, we take the expectation over all future label sequences, computed exactly. Eq. (3) is the one-step condition, i.e., step k=1k=1 in the right panel and the remaining steps should show that the condition is violated throughout the rollout, not only at the first step. The left panel accumulates these one-step drifts.

We report the accumulation as well because martingale diagnostics (Falck et al., 2024; Fong et al., 2023; Ng et al., 2026) impute long sequences of rows, so their outcome depends on the accumulated drift rather than on a single step.

In the revision, we (i) have stated in the caption and text that both curves are exact predictives under the same non-separable prior, and (ii) have stated explicitly that Eq. (3) corresponds to k=1k=1 in the caption of Figure 4, that is now hopefully clearer.

In the following, we will reply to the reviewer’s minor questions:

In definition 1, what operation does ⊗\otimes represent?

The operator ⊗\otimes represents the product measure. We have added an explanation to the updated version of our manuscript.

Under a non-separable prior, when would the Martingale property hold under stochastic auxiliary features? In other words, for what qq does the following hold? EXn+1∼q(⋅),;Yn=1∼π(⋅∣Xn+1,dn)[π(A∣x,Dn+1)]=π(A∣x,dn)\mathbb{E}_{\mathbf{X}_{n+1} \sim q(\cdot),\\; Y_{n=1} \sim \pi(\cdot \mid \mathbf{X}_{n+1}, d_n)}\left[\pi(A \mid \mathbf{x}, D_{n+1})\right] = \pi(A \mid \mathbf{x}, d_n)

This is indeed a very interesting question. We have added a brief comment to our revised Appendix A.5, but answering in all generality is difficult.


Last, we want to quickly comment on the editorial suggestions that regard the abstract. First of all, we thank the reviewer for the great suggestions and the detailed reading of our work!

When reading C1 - C4, it became clear to us that the abstract is really not that great in the submitted version, so we took the liberty of rewriting it while taking into consideration the editorial suggestions. We hope that this new abstract version better conveys the key contents of the paper and are happy to do a further pass to improve it if the reviewer has objections against any change.

Once again, we would like to thank the reviewer for this great review of our work that has (already) led to substantial improvements of the manuscript and remain committed to answering any further questions during this discussion period.

I thank the authors for their thoughtful response and can confirm that they have addressed all of my concerns.

The new revision looks great, and I think the rewrite of the abstract makes for a huge improvement$\textemdash$the abstract was previously one of the few weaker areas in the paper. I am also satisfied with the new caption for Figure 4 as well as the comments regarding marginalisation consistency.

Due to ICLR's new scoring system, there is unfortunately no way for me to acknowledge the authors' improvements to the paper via a score increment. Regardless, I congratulate the authors on their hard work and hope to see their paper accepted. Well done!

Reviewer_vsKtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies how a PFN uses the features of a test point to update its beliefs about the task. It explains when this changes predictions and why some checks of Bayesian behavior can then be misleading. The authors test their claims on synthetic examples and introduce SemiTabICL, which uses unlabeled data and improves average accuracy by 0.63 percentage points across 150 datasets.

Critical strengths
  1. The paper shows that an exact Bayesian predictor can fail certain martingale checks. This is useful when deciding whether a failed check indicates a problem with the PFN.

  2. The synthetic experiments compare the networks with exact predictions, making the results easy to interpret. The real-data experiments also show a benefit from using unlabeled data.

Critical weaknesses
  1. At line 161, the authors claim that mixing priors that differ in both their feature distributions and labeling rules always produces a non-separable prior. However, Proposition 1 proves this only when both component priors are separable. Without this assumption, dependencies within the components can cancel, leaving a separable mixture. The broader claim therefore needs further justification or should be restricted to the assumptions of Proposition 1.

  2. It is still unclear how much the predictive gap matters in practice. The synthetic examples demonstrate that it exists, and SemiTabICL shows that additional unlabeled data can help. However, neither establishes how much this gap affects predictions or Bayesian optimization with existing production PFNs.

Questions for authors
  1. Proposition 1 assumes that both mixture components are separable, but line 161 makes the claim without this assumption. What supports this broader claim for the priors used by existing PFNs?

  2. Did you pretrain each model several times with different random seeds? If so, how much did SemiTabICL’s improvement on real data vary between runs?

  3. Could you measure the difference between the two predictive distributions using a more realistic SCM prior where both can be computed? How does the difference change with more labeled examples, and is it large enough to change prediction accuracy or uncertainty estimates?

2 comments

We thank the reviewer for the thoughtful review and appreciate the acknowledgment of the significance of our findings and the effectiveness of our experimental evaluation.

In the following, we address the reviewer’s questions:

Proposition 1 assumes that both mixture components are separable, but line 161 makes the claim without this assumption. What supports this broader claim for the priors used by existing PFNs?

We thank the reviewer for this sharp observation. Indeed, the sentence is not precise enough and we have clarified its meaning in the updated manuscript, where we restrict our statement to the assumptions of Proposition 1.

[W1] Without this assumption, dependencies within the components can cancel, leaving a separable mixture. The broader claim therefore needs further justification or should be restricted to the assumptions of Proposition 1.

We can construct non-separable regimes Πs\Pi_s, with marginals ΠX,s\Pi_{X,s} and ΠY∣X,s\Pi_{Y\mid X,s}, where s∈{0,1}s \in \{0,1\} indexes the mixture component. Let us define the shorthand notation Cs:=Πs−ΠX,s⊗ΠY∣X,s≠0C_s := \Pi_s - \Pi_{X,s}\otimes\Pi_{Y\mid X,s} \neq 0. Now, like the reviewer points out, we can construct a mixture of these two regimes, whose mixture is nevertheless separable. However, this requires the exact cancellation

(1−λ)C0+λC1=−λ(1−λ)(ΠX,1−ΠX,0)⊗(ΠY∣X,1−ΠY∣X,0).(1-\lambda)C_0 + \lambda C_1 = -\lambda(1-\lambda)(\Pi_{X,1}-\Pi_{X,0})\otimes(\Pi_{Y\mid X,1}-\Pi_{Y\mid X,0}).

The dependence within the regimes must exactly offset the dependence the mixture creates between them. The regimes used in current PFN priors are generated independently of one another, and we see no mechanism that would produce this cancellation.

Did you pretrain each model several times with different random seeds? If so, how much did SemiTabICL’s improvement on real data vary between runs?

The reviewer is correct in pointing this out. We have run the SemiTabICL experiment across 5 different seeds and report its results in the updated version of the manuscript. In short: The improvements we find are consistent across different seeds and the variability between seeds is small. We have also updated Fig. 5 accordingly and detailed results are provided in Appendix C.3.

Could you measure the difference between the two predictive distributions using a more realistic SCM prior where both can be computed? How does the difference change with more labeled examples, and is it large enough to change prediction accuracy or uncertainty estimates?

We thank the reviewer for this question. While we cannot directly assess the gap between the two predictives, as non-separable priors are in general not amenable to MCMC sampling, we have extended our experiments by an application of our Proposition 1, where we mix two separable prior regimes to obtain a non-separable version together with a separable one, where the marginal feature and conditional label distributions are the same. The only difference between the two priors is thus their dependence structure. We train a TabICL model on each prior and find that the difference, as expected, vanishes for large context sizes and only the PFN pretrained on the non-separable prior can use the query feature as evidence at small nn to identify which of the regimes it is tasked with. A figure and a detailed experimental description is available in the new Appendix C.4.

Our SemiTabICL experiment further shows that the difference matters in practice and enables substantial performance gains on well-established benchmarks, like the TALENT benchmark.

I thank the authors for their rebuttal. I think the paper is interesting, and my questions have been answered, so I am raising my soundness score from 2 to 3. I am keeping my overall score of 3, since the new mixture-prior experiment suggests the gap is small and mainly matters when there are few labeled examples.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.