Pith. sign in

REVIEW 3 major objections 6 minor 19 references

Retrieval quality, not model complexity, drives drug response prediction

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Using an LLM to select biologically similar neighbor drugs and averaging their measured effects improves zero-shot prediction of molecular perturbation responses, particularly for unseen cell lines.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The unseen-cell-line gains likely come from LGR's test-time access to the held-out cell line's own response data, not from LLM-guided retrieval. the 3 major comments →

arxiv 2608.01734 v1 pith:JUNM3E7X submitted 2026-08-03 cs.LG

LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses

Betty Xiong , Jan-Christian Huetter , Gabriele Scalia , Tommaso Biancalani , Sepideh Maleki This is my paper
classification cs.LG
keywords molecular perturbation predictiontranscriptomic responseretrieve-and-aggregateLLM-guided retrievalzero-shot generalizationcell-line-specific retrievalsign accuracydrug response prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the main bottleneck in predicting how a drug perturbs gene expression in an unmeasured cell line is not the complexity of the prediction model but the quality of the retrieved drug neighbors. It proposes LLM-Guided Retrieval (LGR), which uses a large language model to rank, from a candidate pool of drugs already profiled in the target cell line, the compounds most likely to share the query drug's transcriptional program; a fixed mean of their observed effects forms the prediction. On the Tahoe-100M single-cell perturbation atlas, LGR consistently outperforms drug-mean, chemistry-based k-nearest-neighbor, and ChemCPA baselines, with the largest gains when generalizing to cell lines never seen in training. If correct, this reframes zero-shot perturbation prediction as a retrieval problem and suggests LLMs can serve as constrained biological priors rather than as end-to-end predictors.

Core claim

The paper's central claim is that a retrieve-and-aggregate estimator—selecting a small neighborhood of biologically related compounds and averaging their measured transcriptomic deltas—can match or beat learned end-to-end predictors for zero-shot molecular perturbation response, provided the neighborhood is chosen well. The selection is done by an LLM prompted to reason over mechanisms, pathways, and cell-line context, but constrained to a closed candidate pool measured in the target cell line. Across unseen-drug, unseen-cell-line, and open-world regimes on Tahoe-100M, LGR improves correlation, error, and especially sign accuracy of gene regulation over chemistry-based kNN and supervised bas

What carries the argument

The central object is the cell-line-restricted candidate pool P_c together with an LLM acting as a selector over it. For a query (cell line c, drug d), the LLM ranks the top-k candidates from P_c\{d} using mechanistic and pathway reasoning; outputs are parsed, validated against the pool, and the uniform mean of the retrieved drugs' observed control-subtracted expression deltas is the prediction. The mean aggregator is deliberately fixed and non-parametric, so any performance difference is attributed to neighborhood construction. This isolates retrieval quality as the variable of interest.

Load-bearing premise

The held-out cell lines and drugs are treated as unobserved by all methods, but the LLM's pretraining corpus may already contain published responses for exactly those cell lines and drugs, so the reported unseen-cell-line gains could reflect memorized literature rather than a generalizable retrieval mechanism.

What would settle it

Run LGR on held-out cell lines and drugs using an LLM whose training data provably excludes those specific perturbations (or retrain an open-weights model with those records withheld); if the correlation and sign-accuracy gains over mean baselines disappear, the effect is literature memorization, not retrieval. Alternatively, replace the LLM selector with a similarity score derived from an independent, non-overlapping perturbation atlas; if performance drops to chemistry-kNN levels, the LLM prior is the active ingredient.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Simple, transparent predictors paired with good retrieval can match or beat complex learned models in zero-shot settings.
  • Directional (sign) accuracy of gene regulation improves more than magnitude metrics, which matters for pathway and mechanism interpretation.
  • LLMs can act as constrained selectors that inject a biological prior without generating high-dimensional numerical outputs.
  • Cell-line-specific candidate pooling makes aggregation well-defined and enables clean closed-world evaluation.
  • Retrieval-based frameworks offer an interpretable alternative to end-to-end supervised perturbation models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If retrieval quality is the bottleneck, investing in richer or better-profiled candidate pools should improve prediction more than upgrading the aggregator—a direct test of the paper's thesis.
  • The reported unseen-cell-line gains may partly reflect literature memorization by the pretrained LLM; a truly closed evaluation would require an LLM trained without the held-out lines.
  • The sign-accuracy improvements suggest LGR could be used as a gene-set prioritization tool for pathway enrichment even when effect-size calibration is poor.
  • A hybrid that falls back to chemistry-based retrieval when LLM coverage is low may mitigate the reported instability without losing the biological prior.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes LLM-Guided Retrieval (LGR), a retrieve-and-aggregate method for predicting transcriptomic perturbation responses. For an unobserved (cell line, drug) pair, LGR uses a pretrained LLM to rank candidate drugs from a cell-line-specific candidate pool P_c, then predicts the response as the uniform mean of the observed expression deltas of the top-k retrieved candidates. The method is evaluated on the Tahoe-100M single-cell perturbation atlas under unseen-drug, unseen-cell-line, and open-world regimes, against drug mean, cell mean, PCA+RR, chemistry kNN, and ChemCPA baselines. The paper reports consistent improvements, with the largest gains in the unseen-cell-line regime, and attributes these gains to retrieval quality rather than predictor complexity.

Significance. If the empirical claims hold, the paper makes a useful contribution: it provides a transparent, parameter-free retrieve-and-aggregate baseline that can outperform more complex supervised predictors, and it demonstrates a concrete way to use LLMs as constrained selectors rather than as numerical predictors. The reproducibility measures — cached neighbor lists, deterministic parsing, fixed prompting, and a deliberately simple aggregator — are strengths. However, the central unseen-cell-line claim is currently undermined by a test-time information asymmetry between LGR and the baselines, and by the unaddressed possibility of LLM pretraining leakage. The paper's own Section 4.4.1 acknowledges the latter but treats it as a feature rather than as a confound. Because the main conclusions rest on the unseen-cell-line comparison, these issues must be resolved before the claims can be accepted.

major comments (3)
  1. [§4.1, Fig. 2, Eq. (2)-(4)] The unseen-cell-line evaluation gives LGR test-time access to the held-out cell line's own measured deltas, while the baselines are cell-agnostic. By Eq. (2) and Appendix B, the candidate pool P_c contains all drugs with measured responses in the target cell line, including the held-out cell line. LGR's prediction (Eq. 4) is the mean of those cell-line-specific deltas. In Fig. 2, drug mean, chemistry kNN, and PCA+RR are trained on other cell lines, and cell mean falls back to the global mean for unseen cell lines. Thus even a random selector over P_c would produce a cell-line-specific prediction and would likely outperform cell-agnostic baselines. The reported >0.15 Pearson improvement may reflect this privileged access to the target cell line's response distribution, not retrieval quality. The paper must add ablations with identical test-time access — e.g., random retrieval from P_c and
  2. [§4.1, §4.4.1] The closed-world evaluation restricts the candidate pool to drugs observed during training, but this restriction does not apply to the LLM's internal knowledge. The LLM is a pretrained model whose training corpus may contain published transcriptional responses for exactly the held-out cell lines and drugs. Section 4.4.1 explicitly attributes the unseen-cell-line gain to 'literature transfer' from the LLM's pretraining. This is a leakage concern: the reported gains may reflect memorization of test answers rather than a generalizable retrieval mechanism. The paper needs a control that blocks or quantifies this leakage, for example by comparing LGR against a non-LLM selector using only mechanism tags or other cell-line-agnostic features, or by evaluating on cell lines/drugs that postdate the LLM's training cutoff. Merely acknowledging the transfer does not establish that the closed-world co
  3. [Fig. 2-4] No error bars, confidence intervals, or significance tests are reported anywhere in the experimental section. The abstract and conclusions use language such as 'consistently improves' and 'strongest gains', but every figure shows only point estimates over what appears to be a single split. Since the differences between LGR and cell-mean/drug-mean are sometimes small on magnitude metrics (e.g., MAE in Fig. 3), the paper should provide variance estimates over bootstrap resamples of cell lines/drugs or over multiple data splits. Without this, it is impossible to assess whether the reported improvements are reliable.
minor comments (6)
  1. [§4.1] The description of the closed-world regime says the candidate pool 'available to all methods is restricted to drugs observed during training', but Appendix B defines P_c as all non-control molecules measured in cell line c, with no intersection with training drugs. For the unseen-cell-line task these are different sets. Please clarify exactly which candidate pool is used for each regime and for each baseline.
  2. [§3.1] The paper says it records k_used for each query to 'characterize selector coverage and its impact on variance', but no k_used statistics are reported. Since low k_used is mentioned as a failure mode in §4.5, please report the distribution of k_used in each regime.
  3. [§4.2] ChemCPA is absent from Fig. 2 because it cannot be evaluated on unseen cell lines. This is stated in the baseline section, but it would be helpful to repeat this in the figure caption so the reader does not wonder about the missing bar.
  4. [§4.4.1] 'The performance gain can be attributed to zero-shot learning via literature transfer' appears to be a fragment; the sentence reads as if an explanation was intended. Also, 'attributed' should be 'attributed' (typo in the provided text).
  5. [Fig. 2-4] The numeric labels under the bars are ambiguous (e.g., '0.39 0.23 0.37 0.20 0.59' in Fig. 2a appears to list bar values, but the alignment is unclear). Please format the figure labels so each value is directly above its bar.
  6. [Appendix A] The LLM is described only as 'GPT-5–class model'. For reproducibility, please specify the exact model name, version, and query date, as LLM behavior can vary across versions.

Circularity Check

1 steps flagged

Unseen-cell-line gains are partly forced by test-time access to the held-out cell line's own measured deltas via P_c, not purely by retrieval quality.

specific steps
  1. other [Section 3, Eq. (2)-(4); Section 4.1; Section 4.4.1]
    "Pc ={d ′ ∈ D: ∆c,d′ is measured and d ′ is non-control}. ... ˆ∆c,d = 1/|Nk(c,d)| Σ_{d′∈Nk(c,d)} ∆c,d′. ... The performance gain can be attributed to zero-shot learning via literature transfer, i.e. due to the LLM’s pre-training on literature about cell lines."

    In the unseen-cell-line split, the target cell line is held out from training, but Eq. (2) defines P_c as every drug with a measured delta in that same cell line. Eq. (4) then predicts the query delta as the mean of those measured cell-line-specific deltas. Thus LGR's prediction is an average of the held-out cell line's own observed responses; any nonempty neighbor set drawn from P_c provides cell-line-specific information that the cell-agnostic baselines (drug mean, chemistry kNN, PCA+RR) do not have. The reported >0.15 Pearson gain and doubled R^2 are therefore partly forced by test-time access to the target cell line's response distribution, independent of LLM retrieval quality. The attribution to 'literature transfer' in Sec 4.4.1 ignores this direct use of target-cell-line data. This

full rationale

The paper's method itself is not definitionally circular: LGR uses a fixed mean aggregator and a pretrained LLM, with no fitted parameters, and the prediction is not constructed to equal the target delta. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in by citation. The main circularity-adjacent issue is the unseen-cell-line evaluation protocol. Eq. (2) defines the candidate pool P_c as all drugs measured in the target cell line, and Eq. (4) aggregates those measured deltas. In the unseen-cell-line regime, this gives LGR access to the held-out cell line's own response distribution at test time, while the compared baselines are cell-agnostic. Consequently, the reported gains in that regime are partly forced by this test-time information, not solely by biologically informed retrieval. This is a significant confound affecting the paper's strongest claim. It is not, however, a full circularity: the method remains a legitimate retrieve-and-aggregate estimator, and the unseen-drug and open-world results retain independent content. A random or chemically blind retrieval-from-P_c baseline would be needed to isolate retrieval quality. Overall score 5 reflects that a central evaluation result is partially reducible by construction to test-time access to the target cell line's data, while the core algorithm is not circular.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim rests on the quality of the Tahoe-100M atlas and the assumption that LLM pretraining encodes enough pharmacology to select useful neighbors. No parameters are learned on test data; the only hand-set choices are k, the HVG count, and a hand-crafted mechanism-tag mapping.

free parameters (3)
  • k (number of neighbors) = 10
    Chosen by the authors for all experiments; not fitted to data, but a manual hyperparameter that affects results.
  • Top HVG count = 2000
    Metrics computed on top 2000 highly variable genes; choice affects all metrics.
  • Mechanism tag set and mapping to programs = internal pharmacological knowledge
    Hand-crafted category mapping used by the LLM prompt; not formally derived or released.
axioms (5)
  • domain assumption Tahoe-100M pseudobulk expression deltas accurately represent true perturbation effects
    Prediction targets and neighbor responses are derived from this atlas; any systematic error propagates to all methods.
  • domain assumption LLM pretraining encodes correct pharmacology and cell-line biology
    LGR's retrieval quality depends on the LLM's latent knowledge about drug mechanisms and cell-line contexts.
  • domain assumption Uniform mean of selected neighbor deltas is a suitable predictor for unseen perturbations
    This is the method's core inductive bias; no theoretical justification is provided.
  • domain assumption The candidate pool P_c (drugs profiled in the target cell line) contains informative neighbors
    If no biologically similar drug is measured in the target cell line, LGR cannot produce a meaningful prediction.
  • domain assumption Control-subtracted log-normalized deltas are comparable across cell lines
    Cross-cell-line aggregation in drug-mean baseline and LGR implicitly assumes comparability.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses." pith.science (2026). https://pith.science/paper/JUNM3E7X

@misc{pith2026260801734,
  author       = {Pith},
  title        = {Pith review of: LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUNM3E7X}},
  note         = {Machine review of arXiv:2608.01734}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Predicting transcriptomic responses to small-molecule perturbations across cell lines is central to drug discovery, but exhaustive profiling of drug-cell combinations is infeasible. We frame molecular perturbation prediction as retrieve-and-aggregate: approximate an unmeasured drug's response in a cell line by aggregating measured responses of a small set of biologically related compounds. We propose LLM-Guided Retrieval (LGR), where a large language model (LLM) ranks candidate neighbor drugs (restricted to those profiled in the target cell line); after which a fixed mean aggregator combines their observed expression deltas to form the prediction. We evaluate on the Tahoe-100M single-cell perturbation atlas under unseen-drug, unseen-cell-line, and open-world regimes. LGR consistently improves over drug mean, ChemCPA, and chemistry-based kNN baselines, with the strongest gains for unseen cell-line generalization, where it achieves higher correlation and lower error than mean baselines. Across settings, LGR improves directional (sign) accuracy of gene regulation, indicating better recovery of biologically meaningful perturbation effects even when magnitude-based metrics are similar. These results suggest that retrieval quality, rather than predictor complexity, is a key driver of zero-shot molecular perturbation prediction, and that LLMs can provide a useful biological prior when used as constrained retrieval modules.

Figures

Figures reproduced from arXiv: 2608.01734 by Betty Xiong, Gabriele Scalia, Jan-Christian Huetter, Sepideh Maleki, Tommaso Biancalani.

Figure 1
Figure 1. Figure 1: Overview of LLM-Guided Retrieval (LGR). Given a query drug and a target cell type, LGR uses a large language model (LLM) as a constrained retrieval module to identify pharmacologically similar drugs from a restricted candidate pool. The transcriptomic response of the query drug is then estimated by aggregating (mean pooling) the observed responses of the top-k retrieved drugs in the same cell type. 2019)),… view at source ↗
Figure 2
Figure 2. Figure 2: Unseen Cell-Line Generalization. Evaluation metrics are (a) cosine similarity, (b) MAE, (c) MSE, (d) sign accuracy, (e) Pearson correlation, and (f) R 2 of linear regression. 4.4 RESULTS 4.4.1 CLOSED-WORLD EVALUATION Unseen-cell-line generalization [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Unseen Drug Generalization. Evaluation metrics are (a) cosine similarity, (b) MAE, (c) MSE, (d) sign accuracy, (e) Pearson correlation, and (f) R 2 of linear regression. Drug mean kNN PCA Cell mean LGR 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 Cosine similarity (↑) 0.48 0.29 0.45 0.63 0.59 (a) Drug mean kNN PCA Cell mean LGR 57.5 60.0 62.5 65.0 67.5 70.0 72.5 75.0 MAE (↓) ×10 −3 0.0687 0.0736 0.0694 0.059 0.… view at source ↗
Figure 4
Figure 4. Figure 4: Open-World Evaluation. Evaluation metrics are (a) cosine similarity, (b) MAE, (c) MSE, (d) sign accuracy, (e) Pearson correlation, and (f) R 2 of linear regression. regime, where generalization across cellular context is required. In this setting, LGR outperforms mean-based baselines across correlation, error, and directional metrics, indicating that biologically informed neighborhood selection is particul… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages

  1. [1]

    Alexander Wolf and Fabian J

    Mohammad Lotfollahi and F. Alexander Wolf and Fabian J. Theis. scGen predicts single-cell perturbation responses. Nature Methods. 2019

    work page 2019

  2. [2]

    Predicting cellular responses to complex perturbations in high‐throughput screens

    Mohammad Lotfollahi and Anna Klimovskaia Susmelj and Carlo De Donno and Leon Hetzel and Yuge Ji and Ignacio L Ibarra and Sanjay R Srivatsan and Mohsen Naghipourfar and Riza M Daza and Beth Martin and Jay Shendure and Jose L McFaline‐Figueroa and Pierre Boyeau and F Alexander Wolf and Nafissa Yakubova and Stephan Günnemann and Cole Trapnell and David Lopez...

    work page 2023

  3. [3]

    Predicting cellular responses to novel drug perturbations at a single-cell resolution , year =

    Hetzel, Leon and B\". Predicting cellular responses to novel drug perturbations at a single-cell resolution , year =. Proceedings of the 36th International Conference on Neural Information Processing Systems , articleno =

    work page

  4. [4]

    Srivatsan and José L

    Sanjay R. Srivatsan and José L. McFaline-Figueroa and Vijay Ramani and Lauren Saunders and Junyue Cao and Jonathan Packer and Hannah A. Pliner and Dana L. Jackson and Riza M. Daza and Lena Christiansen and Fan Zhang and Frank Steemers and Jay Shendure and Cole Trapnell , title =. Science , volume =

    work page

  5. [5]

    and Thompson, John D

    Zhang, Jesse and Ubas, Airol A and de Borja, Richard and Svensson, Valentine and Thomas, Nicole and Thakar, Neha and Lai, Ian and Winters, Aidan and Khan, Umair and Jones, Matthew G. and Thompson, John D. and Tran, Vuong and Pangallo, Joseph and Papalexi, Efthymia and Sapre, Ajay and Nguyen, Hoai and Sanderson, Oliver and Nigos, Maria and Kaplan, Olivia a...

    work page 2025

  6. [6]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  7. [7]

    Roohani and Tony J

    Yusuf H. Roohani and Tony J. Hua and Po-Yuan Tung and Lexi R. Bounds and Feiqiao B. Yu and Alexander Dobin and Noam Teyssier and Abhinav Adduri and Alden Woodrow and Brian S. Plosky and Reshma Mehta and Benjamin Hsu and Jeremy Sullivan and Chiara Ricci-Tam and Nianzhen Li and Julia Kazaks and Luke A. Gilbert and Silvana Konermann and Patrick D. Hsu and Ha...

    work page 2025

  8. [8]

    Gene-embedding-based prediction and functional evaluation of perturbation expression responses with PRESAGE , year =

    Littman, Russell and Levine, Jacob and Maleki, Sepideh and Lee, Yongju and Ermakov, Vladimir and Qiu, Lin and Wu, Alexander and Huang, Kexin and Lopez, Romain and Scalia, Gabriele and Biancalani, Tommaso and Richmond, David and Regev, Aviv and H. Gene-embedding-based prediction and functional evaluation of perturbation expression responses with PRESAGE , year =

    work page

  9. [9]

    LangPert:

    Kaspar M. LangPert:. ICLR 2025 Workshop on Machine Learning for Genomics Explorations , year=

    work page 2025

  10. [10]

    Patterns

    Omar Kana and Rance Nault and David Filipovic and Daniel Marri and Tim Zacharewski and Sudin Bhattacharya , title =. Patterns. 2023

    work page 2023

  11. [11]

    and Meinl, Rico and Nourisa, Jalil and Tumiel, Jared and Tunjic, Tin M

    Sza ata, Artur and Benz, Andrew and Cannoodt, Robrecht and Cortes, Mauricio and Fong, Jason and Kuppasani, Sunil and Lieberman, Richard and Liu, Tianyu and Mas-Rosario, Javier A. and Meinl, Rico and Nourisa, Jalil and Tumiel, Jared and Tunjic, Tin M. and Wang, Mengbo and Weber, Noah and Zhao, Hongyu and Anchang, Benedict and Theis, Fabian J. and Luecken, ...

    work page

  12. [12]

    NeurIPS 2024 Workshop on AI for New Drug Modalities , year=

    PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis , author=. NeurIPS 2024 Workshop on AI for New Drug Modalities , year=

    work page 2024

  13. [13]

    2024 , eprint=

    GPT-4 Technical Report , author=. 2024 , eprint=

    2024

  14. [14]

    2024 , eprint=

    Gemma: Open Models Based on Gemini Research and Technology , author=. 2024 , eprint=

    2024

  15. [15]

    Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines

    Constantin Ahlmann-Eltze and Wolfgang Huber and Simon Anders. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods. 2025

    work page 2025

  16. [16]

    Nature , year=

    Large language models encode clinical knowledge , author=. Nature , year=

  17. [17]

    Briefings in Bioinformatics , volume =

    Luo, Renqian and Sun, Liai and Xia, Yingce and Qin, Tao and Zhang, Sheng and Poon, Hoifung and Liu, Tie-Yan , title =. Briefings in Bioinformatics , volume =. 2022 , month =

    work page 2022

  18. [18]

    Crawford and David Peck and Joshua W

    Justin Lamb and Emily D. Crawford and David Peck and Joshua W. Modell and Irene C. Blat and Matthew J. Wrobel and Jim Lerner and Jean-Philippe Brunet and Aravind Subramanian and Kenneth N. Ross and Michael Reich and Haley Hieronymus and Guo Wei and Scott A. Armstrong and Stephen J. Haggarty and Paul A. Clemons and Ru Wei and Steven A. Carr and Eric S. Lan...

    work page

  19. [19]

    Cell , volume =

    Aravind Subramanian and Rajiv Narayan and Steven M Corsello and David D Peck and Ted E Natoli and Xiaodong Lu and Joshua Gould and John F Davis and Andrew A Tubelli and Jacob K Asiedu and David L Lahr and Jodi E Hirschman and Zihan Liu and Melanie Donahue and Bina Julian and Mariya Khan and David Wadden and Ian C Smith and Daniel Lam and Arthur Liberzon a...

    work page

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.