acceptodds
Under review as a conference paper at ICLR 2027

Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder features

Abstract

Intrinsically disordered protein regions (IDRs) are ubiquitous across all kingdoms of life. These structurally heterogeneous regions play central roles in cellular processes such as transcriptional regulation, cellular signaling, and subcellular localization, yet their functional design remains challenging for existing generative methods. Structure-based design methods do not readily apply to IDRs, and existing protein language models are trained on full-length sequences, learning priors which are biased towards folded domains. Here, we present IDiom, an autoregressive protein language model trained on IDiom-DB, a dataset of 54 million predicted IDRs curated from the AlphaFold Database, enabling generation of diverse and biologically realistic IDRs. For explicit control over sequence features, we also introduce reinforcement learning with sparse autoencoder (SAE) features (RL-SAE), which rewards the model for generating sequences that activate a specified set of SAE features. Together, IDiom and RL-SAE enable the interpretable and composable design of IDRs that carry the sequence motifs and predicted functionality of natural IDRs. More broadly, RL-SAE could extend to other protein design settings where interpretable features provide useful design targets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein language models, intrinsically disordered regions, sparse autoencoders, reinforcement learning, protein design

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_1i8HOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces IDiom, an autoregressive protein language model trained on IDiom-DB, a collection of 54 million predicted intrinsically disordered regions curated from the AlphaFold Database. A fill-in-the-middle training format supports both unconditional IDR generation and generation conditioned on flanking protein sequences. The authors characterize the resulting sequences using sequence similarity, amino acid composition, patterning statistics, and several disorder predictors.The main methodological contribution is RL-SAE: the authors train a sparse autoencoder on IDiom representations, identify features enriched in function-associated sequence sets, and use reinforcement learning to reward the simultaneous activation of selected features. The reward also incorporates length and compositional-entropy penalties, with KL regularization to the base model. Importantly, reward features are evaluated using the frozen pretrained model and SAE. Experiments cover six compartment-association targets, transcriptional activation and repression domains, and combinations of chromosome-association and effector features. Comparisons include activation steering, supervised fine-tuning, and direct ProtGPS-reward optimization. The reported results support improved target-feature coverage and several improved computational readouts, while biological functionality of the generated sequences remains experimentally unvalidated.

Critical strengths
  1. The paper provides a concrete and effective mechanism for controlling multiple learned sequence features. RL-SAE turns a selected set of SAE features into an explicit sequence-level objective, rather than assuming that a sum of steering directions will successfully install all corresponding features. The reported average coverage of approximately 90% of the targeted top-30 features, compared with 24% for activation steering, is a substantial difference on the stated control objective. Evaluating rewards through a frozen pretrained model and SAE also keeps the feature definition fixed throughout optimization
  2. The feature-selection and composition experiments go beyond simply improving a single predictor score. The private-feature experiment provides an informative intervention: removing shared features reduces particular off-target predictions while largely preserving the intended target prediction. For example, the chromosome-targeted designs’ nucleolus score decreases from 0.85 to 0.27, while their chromosome score changes from 0.71 to 0.67. The chromosome-plus-effector experiments further show that the model can install approximately 53 of 60 requested features and organize incompatible acidic and basic features into different sequence regions without an explicit positional specification.
Critical weaknesses
  1. The described splitting and deduplication procedures do not clearly establish independent evaluation of the pretrained model. Appendix A describes duplicating the 54 million IDRs into prompted and unprompted examples, producing 108 million sequences, and then splitting these into training, validation, and test sets. Appendix B explicitly describes a random split of these examples. As written, this could place the prompted version of an IDR in training and its unprompted counterpart in validation or test. The paper does not state that these paired examples are assigned to the same partition.
  2. The evidence is stronger for surrogate-feature control than for reliable, function-specific biological design. The authors appropriately acknowledge that the compartment-associated sequences used for feature enrichment come from the same full-length proteins used to train ProtGPS. Thus, the main localization evaluation is not independent of the information used to define the design targets. DeepLoc is a useful additional readout, but the nuclear-versus-cytoplasmic evaluation cannot establish discrimination between the finer compartments central to the paper. Moreover, the compartment labels originate from full-length proteins, whereas the generated objects are isolated IDRs. The presented evidence does not establish that all extracted regions independently determine the corresponding localization.
Questions for authors
  1. Data Partitioning Are the prompt and non-prompt versions of each IDR assigned to the same partition? Are all IDRs of the same parent protein and their surrounding contexts also kept together? Please specify the actual partitioning unit and provide the number of overlapping focal IDRs and parent proteins in the training set, validation set, and test set.
  2. DisProt De-duplication Issue Does the de-duplication check in DisProt only target the core IDR in each training record, or does it also include all retained N-terminal and C-terminal contexts? In particular, could another retained IDR in the same parent protein potentially expose a benchmark region during pre-training? Review this possibility and determine which reported evaluation results will be affected and significantly influence my judgment if such overlaps exist.
  3. Function-specificity In addition to the output results of ProtGPS and coarse-grained DeepLoc, what evidence supports the fine specificity of the selected features on subcellular compartments? For example, can these features distinguish the IDRs of the target compartment from those of other human compartments with matching length and composition on the retained parental proteins? Conduct targeted analysis on such questions to help determine whether the selected features capture the compartment-specific information rather than broader sequence features.
2 comments

We thank Reviewer 1i8H for their helpful assessment of the manuscript and constructive comments. We have uploaded a new version of the manuscript.

Are the prompt and non-prompt versions of each IDR assigned to the same partition?

Yes, prompted and unprompted versions of each IDR are assigned to the same partition. We split IDR records into the different partitions before training, and we now clarify this in Appendix A.

Are all IDRs of the same parent protein and their surrounding contexts also kept together?

The IDRs of the same parent protein are not kept together, and data splitting was only done according to the focal IDR. We have screened for the number of focal IDRs in the validation and test sets which are found anywhere within the full sequence of any training record. We find that 147,915 of 270,775 validation IDRs (54%) and 147,682 of 270,777 test IDRs (54%) show up within a training record. The number of parents shared between training-validation is 139,863 and the number shared between training-test is 139,673.

We now remove from the validation and test sets those above records where the focal IDR is found anywhere within a training record. We then re-evaluate the model validation perplexity (Figure 1d), finding that for the 300M parameter model, unprompted IDR validation perplexity increases from 10.81 to 11.85, and prompted IDR perplexity increases from 9.58 to 10.82.

Does the de-duplication check in DisProt only target the core IDR in each training record, or does it also include all retained N-terminal and C-terminal contexts?

DisProt deduplication was also only applied to the focal IDR of each training record. We find that 913 of 1,663 DisProt evaluation focal IDRs (55%) are found anywhere within a training record, and we now remove those IDRs from the DisProt evaluation set. For the 750 remaining, we re-calculate model perplexities (Figure 1d), finding that the 300M unprompted DisProt perplexity increases from 6.72 to 8.48 and prompted DisProt perplexity increases from 5.28 to 7.34. We additionally update the other DisProt evaluations so they are only done for these 750, finding no major changes: all biophysical/sequence metrics maintain the same trends as previously. We thank the Reviewer for bringing up these points.

In addition to the output results of ProtGPS and coarse-grained DeepLoc, what evidence supports the fine specificity of the selected features on subcellular compartments?

We have additionally scanned the generated sequences for experimentally characterized functional sequence motifs, and we find that a high percentage of sequences generated after RL-SAE contain motifs that are expected within IDRs which associate to the target compartments. For these, we calculate motif occurrence rates for on- and off-target compartments, and we find that RL-SAE enriches the biologically expected motifs sequences for most compartments, with good specificity. We have updated Appendix P with this information.

For example, can these features distinguish the IDRs of the target compartment from those of other human compartments with matching length and composition on the retained parental proteins?

To test whether the top-30 SAE features can help distinguish IDRs of target compartments from those of other compartments, we train several linear probes and small MLPs to perform target compartment classification of the ProtGPS IDRs. We test 3 input variants:

(1) only amino acid composition and sequence length (as a 21 element vector of 20 amino acid frequencies and log(sequence length))

(2) only top-30 SAE feature activation vectors (as a 30 element vector of maximum feature activation magnitudes)

(3) amino acid composition, sequence length, and SAE feature activation vectors all together (as a 51 element vector)

We train a linear probe and small MLP for each case, using a 70/30 data split per compartment (2171 IDRs total for SAE feature enrichment and probe training, and 982 held out), and we ensure that there are no shared parent proteins between the two splits. The table reports held-out AUROC using length and amino acid composition alone or additionally using top-30 SAE features (MLP hidden dimension = 32):

CompartmentLinear: len + compLinear: len + comp + SAEMLP: len + compMLP: len + comp + SAE
Chromosome0.650.70 (+0.05)0.700.74 (+0.04)
Nucleolus0.710.75 (+0.04)0.740.80 (+0.05)
Stress granule0.650.63 (−0.02)0.690.69 (+0.00)
P-body0.520.56 (+0.03)0.500.51 (+0.02)
Nuclear speckle0.620.71 (+0.09)0.760.80 (+0.04)
Nuclear pore complex0.820.80 (−0.01)0.780.83 (+0.05)

The results demonstrate that SAE feature activations do provide modest improvements beyond using only length + composition, increasing AUROC values on average across all 6 compartments by +0.03 for the linear probe and +0.035 for the MLP.

Thank you for your reply. I believe the new analysis has enhanced the credibility of the paper and has addressed my main concerns within a reasonable range. Therefore, I have accordingly raised my score.

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

This paper studies the generation and controllable functional design of intrinsically disordered regions (IDRs). Existing protein generative models primarily focus on folded proteins and are difficult to apply directly to IDRs, which lack stable three-dimensional structures. Meanwhile, existing methods lack explicit control over sequence features associated with specific biological functions. The authors propose: (1) IDiom: An autoregressive protein language model trained on IDiom-DB, a dataset of 54M predicted IDRs curated from the AlphaFold Database using pLDDT, for IDR sequence generation. (2) IDiomSAE: A sparse autoencoder (SAE) trained on IDiom's hidden representations to extract biologically interpretable sequence features. (3) RL-SAE: A method that constructs rewards from SAE features and optimizes the generative model through reinforcement learning, enabling simultaneous activation and compositional design of multiple target features.

Experimental results show that IDiom generates IDRs with realistic amino acid composition, sequence patterning, and predicted disorder. RL-SAE significantly outperforms activation steering and SFT in target feature activation. It also achieves better predicted functional scores on tasks involving biomolecular compartment association and transcriptional regulation.

Critical strengths

S1. Extracting interpretable SAE features from hidden states and optimizing their activation through RL enables controllable IDR sequence design, demonstrating a degree of novelty.

S2. The experiments cover multiple biologically meaningful tasks and compare against baselines such as SFT and activation steering. The results are competitive and provide good interpretability.

Critical weaknesses

W1. Validation of biological functionality still relies entirely on computational metrics.

W2. The validity of IDR dataset construction and SAE feature selection is insufficiently validated. IDiom-DB primarily identifies predicted IDRs based on AlphaFold2 pLDDT, but low pLDDT does not always indicate intrinsic disorder. Systematic validation of IDiom-DB curation is lacking.

Questions for authors

Regarding IDiom-DB: Have the authors evaluated the consistency between pLDDT-based IDR segmentation and experimentally annotated IDRs?

1 comment

We thank Reviewer GQrV for their helpful assessment of the manuscript and useful comments. We have uploaded a new version of the manuscript.

Regarding IDiom-DB: Have the authors evaluated the consistency between pLDDT-based IDR segmentation and experimentally annotated IDRs?

We have now evaluated our AF2 pLDDT-based IDR predictions against the experimentally annotated IDRs from the DisProt set. For this, we use 1,378 parent proteins (the proteins in the DisProt set that have AFDB structures available), containing 1,592 DisProt-annotated IDRs. We find that our IDR prediction-extraction procedure recovers 72.0% of experimentally annotated residues overall. We additionally run another disorder predictor, AIUPred (Erdos, NAR, 2024), on these same sequences and find that it similarly recovers 72.3% of DisProt annotated disordered residues.

These calculations measure recovery of experimentally annotated disorder and therefore quantifies the false negatives. False positives cannot be determined here because residues that lack DisProt disorder annotations can't be assumed to be ordered. We have added this information into the manuscript's Appendix A.

We agree that low AF2 pLDDT does not always indicate intrinsic disorder, and we note that any choice of disorder predictor for the curation of a large-scale IDR dataset will have inaccuracies associated with it. For this work, we chose to use the AF2 pLDDT as our disorder predictor for the following reasons:

(1) AF2 pLDDT values are available at scale, via the AlphaFold Database (AFDB). This enables us to curate the 54M predicted IDRs without incurring the computational cost of running other disorder predictors from the literature on hundreds of millions of sequences.

(2) AF2 pLDDT-based disorder predictions are well characterized in the literature and have been shown to be a well-performing disorder predictor. Here, we point out several examples:

  • Tunyasuvunakool (Nature, 2021) evaluate the performance of AF2 pLDDT on the CAID DisProt-PDB benchmark, finding an AUROC of 0.897.
  • Piovesan (Protein Sci., 2022) also report the performance of AF2 pLDDT on the CAID DisProt-PDB benchmark, finding an AUROC of 0.904 and maximum F1=0.786.
  • Mehdiabadi (Proteins, 2026) evaluate AF2 disorder predictions on the CAID3 Disorder-PDB benchmark, finding an AUROC of 0.934. They calculate that the pLDDT threshold 68.2 maximizes the F1 score at 0.823.
  • Wilson (IJMS, 2022) compare AF2 pLDDT-based IDR prediction to other sequence-based disorder predictors. They find that on DisProt-PDB, the inverse pLDDT (1 - pLDDT/100) achieves an AUROC=0.905 and maximum F1=0.784. They show that pLDDT achieves the highest maximum F1 score vs. 11 other predictors, and that the pLDDT AUROC is among the highest.

(3) Our approach follows the procedure of Tesei (Nature, 2024), which is an established and reproducible procedure for the identification and extraction of AF2-predicted IDRs. Beyond the base procedure of Tesei, we additionally introduce the improvement of using SignalP (Teufel, Nat. Biotechnol., 2022) to filter out regions that are strongly predicted to be cleaved signal peptides.

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

The paper reports a decoder-only, autoregressive pLM trained on a dataset of presumed disordered sequences constructed based on AFDB’s (low) pLDDT scores. The model was trained with FIM, and allows conditioning of IDR sequence generation on flanking-termini, as well as unprompted sequence generation. RL-SAE approach is introduced in post-training to select generated IDR sequences that best activate a set of specified SAE features. By combining RL with SAE the model can support "composable design" based on largely interpretable features, and support specificity in designs by targeting "private" SAE features. The strategy is compared to activation steering, supervised fine-tuning, and an oracle RL baseline (RL-ProtGPS scores). The strategy and IDR designs are evaluated for localization to six biomolecular condensates and transcription activation and repression function.

Critical strengths
  1. IDR-only training The model is trained exclusively on disordered sequences (IDiom-lDR). This is a a big advantage of the work because it removes biases in generative priors that are typical of pLMs trained in the full sequence context (that includes structured domains).

  2. Design interpretability and "composability" Combining features activated under different RL-SAE conditions (as shown in Fig 4) is conceptually innovative. Using SAE to identify interpretable, sequence-level features bridges the advantages of pLM representations with interpretable IDR features, which have shown promise in functional inferences in more recent studies.

Critical weaknesses
  1. Clustering at 90% sequence identity for full-sequence context and length-biases Though understandable as a standard for pLM pretraining practice, this is heavily biased by conventions for structured proteins. It also suggests that sequence redundancy is never explicitly controlled in parts of the sequences that IDR(s) span (?). This is expected to be problematic for proteins dominated by highly similar structured domains, but potentially divergent IDRs, especially if these IDRs are on the shorter end - genuine IDR diversity would get removed from the training set on the basis of high similarity of structured domains. Using the 90% identity threshold also appears to be at odds with mmseq2 search parameters used subsequently for DisProt. There, IDR to IDR similarity is assessed and 50% identity threshold is used(?). This difference in approaches needs to be explained and reconciled (see also questions). Finally, because IDRs are difficult to align and similarity estimates on alignment bases are inadequate, authors should consider using e.g., SHARK to estimate sequence similarity (k-mers based), or using compositional and patterning-based evolutionary proxies (e.g. Forman-Kay, Moses PNAS 2026). Length restrictions bias is not discussed. IDRs in host proteins longer than 1,020 residues are completely discarded. This excludes IDRs in large and fully disordered proteins (many of which are associated with chromatin and biomolecular condensates). Similarly, fixing generated sequence lengths to 100 or 80 residues makes generation of long IDRs completely untested. Limitations section should acknowledge these biases. 2.Reward formulation and performance differences on different condensates The present formulation of reward does not acknowledge reality of multivalency, spacing and patterning of features in IDRs. Because a feature is counted in the reward as soon as it is present (at any residue), one would expect the reward to saturate as soon as many/all top-K features are encountered. This does not consider multivalency, strength, spacing or patterning of the features, all known to be important for condensate formation and localization. It would be important that the authors investigate if this might be the reason behind more poor performance on stress granules and P-bodies? As soon as the observed “functionality/localization” is less about a presence of any given motif and more about arrangement and patterning of features, the reward at present will likely prove inadequate.
  2. Robustness of the SAE feature set It is reported that the SAE layer is selected at fixed k = 32 and 16× expansion (Figure 7). 
Because the top-30 enriched feature sets (and the entire RL-SAE objective) are defined relative to one particular dictionary, it would be good to report how stable are the enrichment rankings and the resulting designs under different expansions (e.g. 8× or 32×) or different k values. In fact, Figures 8 and 10 shows variability across top-N (especially for ProtGPS scores). More evidence here would allow judging whether the selected SAE features cover robust sequence grammar or do they mainly come out a consequence of one particular dictionary’s factorization.
Questions for authors

Clarification requests i) Was clustering for sequence redundancy performed before IDR extraction (Appendix A) or after (Figure 1a)? The two imply different strategies. ii) What is the IDR-level redundancy of the final 54M set? E.g. what is the distribution of maximum IDR-vs-IDR identity within that final set? Given that 50% identity is already treated as leakage for the DisProt benchmark, how much of the training data fall above that threshold? iii) It is reported that the SAE layer is selected at fixed k=32 and 16× expansion (Figure 7). Have alternative dictionary sizes or k(s) been considered? Questions:

  1. How sensitive are the curated IDR boundaries to the annotation method? Could you report consistency between pLDDT-derived IDRs and a dedicated disorder predictor (e.g. AIUPred) on a sample of IDiom-DB?
  2. If alternative to alignment for measures of sequence similarity (e.g. SHARK score) were used, would clustering IDRs on those basis change the effective dataset size? And the diversity of the learned prior?
  3. How does the imposed 1 020-residue protein cap affect coverage of IDRs in large chromatin- and condensate-associated proteins?
  4. Would a different selection of k or dictionary size significantly impact IDiomSAE?
  5. What are the hypotheses for why RL-SAE strategy unperformed on stress granules, and especially P-bodies? Is this a feature-selection failure (top-30), or driven by the underlying condensate grammar, or a limitation of ProtGPS for those condensates?
  6. Can IDiom be conditioned on an existing IDR, without a dedicated RL run? Or how would this be done with RL-SAE strategy? If not possible, can this be stated explicitly as a limitation at present?
1 comment

We thank Reviewer MHfa for their helpful assessment of the manuscript and useful comments. We have uploaded a new version of the manuscript.

i) Was clustering for sequence redundancy performed before IDR extraction (Appendix A) or after (Figure 1a)?

Clustering is performed after IDR extraction, and we have clarified the wording in Appendix A and the main text.

ii) What is the IDR-level redundancy of the final 54M set?

To assess the IDR-level redundancy, we randomly sample 10,000 records from IDiom-DB and use MMseqs2 (80% coverage) to search the records' IDRs against IDRs in the rest of IDiom-DB:

Threshold % ID to IDiom-DB IDRSampled IDRsPercentage
>50% identity5,254 / 10,00052.54%
>70% identity3,971 / 10,00039.71%
>90% identity2,107 / 10,00021.07%

Next, for the 2,107 IDRs with >90% identity, we check the similarity of their flanking contexts to training records:

Threshold % ID for both C- and N- flanksMatching flank pairsPercentage
>50%1,330 / 2,10763.12%
>70%1,256 / 2,10759.61%
>90%767 / 2,10736.40%

We use the stricter 50% identity threshold for DisProt IDR removal to reduce the possibility that agreement between IDiom-generated and experimental DisProt IDRs was due to exposure during training.

Could you report consistency between pLDDT-derived IDRs and a dedicated disorder predictor (e.g. AIUPred) on a sample of IDiom-DB?

When we run AIUPred on 3,000 random parent proteins from IDiom-DB, we find that AIUPred classifies 64% of the pLDDT-based IDR residues as disordered. However, we additionally evaluate both the pLDDT-based predictions and AIUPred on the experimentally annotated DisProt IDRs, finding similar recall, with 72.0% and 72.3% of DisProt disordered residues correctly called disordered by pLDDT and AIUPred, respectively.

If alternative to alignment for measures of sequence similarity (e.g. SHARK score) were used, would clustering IDRs on those basis change the effective dataset size? And the diversity of the learned prior?

Using alternative measures of sequence similarity like the SHARK k-mer score for clustering is an innovative idea that could improve the data curation process by deduplicating IDRs with similar physicochemical properties but distinct sequences. We are unsure how such alternative clustering approaches would perform as such, but the exploration of such approaches is very well suited for future work on curating IDR datasets.

How does the imposed 1 020-residue protein cap affect coverage of IDRs in large chromatin- and condensate-associated proteins?

To assess how this cap affects the coverage of these large proteins, we check four sequence sets: human transcription factors, reviewed human UniProt proteins annotated for chromatin, DrLLPS proteins, and PhaSePro proteins. Across these sets, we find that the 1,020 cap excludes 8-21% of full proteins, 16-33% of pLDDT-predicted IDRs, and 24-46% of predicted IDR residues. We have included this information in Appendix A "Effect of the protein length cap" and updated the Limitations to bring up this point.

What are the hypotheses for why RL-SAE strategy unperformed on stress granules, and especially P-bodies?

We hypothesize that RL-SAE underperforms on P-bodies and stress granules because of feature selection variability. In our additional SAE experiments (below), we find that P-body and stress granule ProtGPS scores can vary substantially, depending on the SAE dictionary used. This suggests that the enriched selection of SAE features does not consistently capture the sequence features ProtGPS requires for high scores for those specific compartments.

Can IDiom be conditioned on an existing IDR, without a dedicated RL run?

IDiom could be used to propose mutant variants of an existing IDR by using Gibbs sampling (e.g. following method of Nayfach, Nat. Biotech, 2026). In this method, one would choose a residue position to mutate, calculate IDiom's log prob of the full IDR for each of the 20 possible amino acid substitutions at that position, then sample a new residue from the softmax of those log probs. This would require no RL and would enable the generation of mutant variants of existing IDRs.

iii) Have alternative dictionary sizes or k(s) been considered?

We have now tested four additional SAEs on layer-18 activations of IDiom. We have trained SAEs with 8x/32x expansions at k=32, and k=16/k=64 at 16x expansion (IDiomSAE is k=32, 16x). We then repeated the top-30 feature enrichment procedure for those SAEs and ran RL-SAE optimization. We find that nucleolus, NPC, and chromosome sequences consistently achieve high ProtGPS scores, while stress granule, P-body, and nuclear speckle scores can vary substantially, with several SAE dictionaries improving over the baseline SAE. We report these results in Appendix H ("Additional SAE experiments").

Comments on the paper

3 comments

Hi reviewers,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

I have read the authors rebuttal. The additional evaluation against experimentally annotated IDRs from DisProt, together with the supporting literature, adequately addresses my concern regarding the validity of the pLDDT-based dataset construction.

Overall, the rebuttal addresses my main concern about dataset construction, while some limitations remain. I maintain my original recommendation of Weak Accept (3).

We thank all Reviewers and the Area Chair for their time and effort in reviewing our manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.