Pith. sign in

REVIEW 3 major objections 5 minor 55 references

This paper establishes that once retrieved homolog evidence is available, a general LLM's enzyme classification accuracy becomes indistinguishable from a script that votes the EC numbers of the nearest neighbors, and that the apparent tie h

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 16:46 UTC pith:FTITY4BI

load-bearing objection A carefully built diagnostic benchmark with an unusually honest limitations section; the headline cancellation narrative is partly an artifact of an under-specified retrieval baseline, but the qualitative arbitration claim and the homology-availability law hold up well enough to warrant serious refereeing. the 3 major comments →

arxiv 2607.26397 v1 pith:FTITY4BI submitted 2026-07-29 cs.CL cs.LGq-bio.QM

Knowledge before Reasoning: EC-Reason-Bench, a Training-Free Diagnostic Benchmark for LLM Enzyme Classification

Linyu Li , Zhi Jin , Yichi Zhang , Dongming Jin , Yuanpeng He , Huanyao Zhang , Xuan Zhang , Gadeng Luosang
show 1 more author
Nyima Tashi
This is my paper
classification cs.CL cs.LGq-bio.QM
keywords enzyme classificationEC number predictionlarge language modelsretrieval-augmented reasoninghierarchical classificationdiagnostic benchmarkhomology availabilitymulti-label evaluation
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.

This paper tries to establish why general-purpose language models score almost zero when asked for a full four-level Enzyme Commission number, and how much of that failure can be reversed without retraining. It builds a diagnostic benchmark that separates four levers—output structure, external knowledge, reasoning structure, and reasoning robustness—and measures each with a training-free inference-time method against a shared zero-shot baseline. The central finding is that knowledge, not reasoning, is what the models lack: closed-book fourth-level accuracy is at best 0.16 for these models, while opening an offline cache of retrieved homologs lifts it to 0.70–0.76 and erases most differences between models and methods. Once evidence is available, the best LLM setting matches a script that simply votes the EC numbers of the nearest retrieved neighbors; that aggregate tie is a cancellation of a large gain on an adversarial split against a large loss on multi-label enzymes. The paper concludes that LLM reasoning over evidence acts as an arbiter of conflicting neighbors rather than a source of knowledge, and that no single-number leaderboard can expose this.

Core claim

The paper's core claim is that an LLM's apparent enzyme-classification ability is, in the open-book regime, almost entirely a function of which homologs can be retrieved, not of what the model knows or how it reasons. In aggregate the best of nine settings (0.728 full-set L4 accuracy, five-model mean) is indistinguishable from neighbor voting (0.723). But splitting the data shows the tie is a cancellation: the model converts about 19 more adversarial price items than voting (0.499 vs 0.372), even exceeding an oracle that always picks the best single channel, and loses about 19 promiscuous multi-label items (0.848 vs 0.938) by committing to one branch too early when evidence multiplies. The n

What carries the argument

The carrying mechanism is the four-lever diagnostic protocol, not a new model. A zero-shot free-generation baseline (B0) reproduces the reported near-zero failure; M1 walks the EC hierarchy level by level with valid candidates and an abstain option; M2 runs a budgeted agentic loop that decides which offline evidence channel to query and when to stop; M3 forces reasoning along the four biochemical axes of the EC number; M4 wraps any base reasoner with self-consistency and prefix voting. All methods share the same offline evidence cache, the same hierarchical scorer, and the same four test splits, and the LLM's contribution is isolated by comparing against a no-model neighbor-vote reference bu

Load-bearing premise

The load-bearing premise is that voting the EC numbers of the top sequence-homology neighbors, scored with the same any-label rule, is the right reference for what a non-reasoning retrieval baseline achieves; if a different or better-tuned voting rule—using additional evidence channels or smarter merging—erases the LLM's adversarial-split edge, the headline cancellation becomes an artifact of the chosen baseline rather than a property of LLM reasoning.

What would settle it

Run an LLM-free voting baseline that merges hits from all five evidence channels (or a learned top-1 channel selector) on the adversarial price split. If this baseline reaches or passes the 0.499 L4 accuracy achieved by the LLM agentic setting on that split, while keeping promiscuous accuracy near the 0.938 neighbor-vote level, then the 'arbiter, not knowledge' claim falls: the LLM's apparent gain would be recoverable by better retrieval without any reasoning. Alternatively, a paired bootstrap over per-sample differences between M2 and the neighbor vote on the 148 price items would show whethe

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

If this is right

  • If the paper is right, a single aggregate accuracy number for enzyme classification is not merely coarse but actively misleading; systems should be compared separately on homology-graded, adversarial, and multi-label splits.
  • General LLMs with no weight updates can be lifted from near-zero closed-book accuracy to specialist-level accuracy on evidence-rich and adversarial enzyme splits, provided retrieval is offline and inductive.
  • Closed-book format repair is not universally helpful: hierarchical cascades and chain-of-thought help only models that are willing to commit, and hurt abstention-prone models by converting silence into errors.
  • On homology-graded splits the choice of method matters little (about 0.05 spread), while on misleading or plural evidence splits it matters about three times more, so retrieval strategy should be chosen for the expected evidence regime.
  • The concrete practical handle for low-homology cases is a conflict-aware retriever or a lightweight router over evidence channels, not a stronger reasoner.

Where Pith is reading between the lines

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

  • I infer the same cancellation pattern may appear in other retrieval-augmented classification tasks: aggregate parity with a retrieval baseline can hide opposite-signed gains on hard subsets and losses on ambiguous or multi-label subsets, so diagnostic splits should become standard practice.
  • The paper's zero-net-contribution claim is anchored to one specific voting rule (top hits from the two sequence-homology channels, no merging). If a stronger LLM-free baseline—using structure or representation channels, or optimally merging hits—narrows or reverses the price-split gain, the 'arbiter, not knowledge' conclusion would need to be softened to 'arbiter relative to that reference.'
  • The 19-converted / 19-lost magnitudes rest on 148 and 209 samples and a single seed; the paper itself notes the intervals do not cover those per-split contrasts. A reader should treat the direction and cross-model sign consistency as the evidence, not the exact counts.

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 / 5 minor

Summary. EC-Reason-Bench is a training-free diagnostic protocol for explaining why general LLMs collapse on full EC-number prediction. The paper decomposes the failure into four levers (output structure, external knowledge, reasoning structure, robustness), implemented as one baseline (B0) and four inference-time methods (M1–M4) evaluated with five reasoning LLMs over 1,349 CARE test items in four splits. The headline findings are: (1) open-book access to a strictly inductive offline evidence cache raises L4 accuracy from ≤0.16 to 0.70–0.76; (2) whether closed-book cascade/CoT help or hurt depends on a model's abstention tendency; (3) the best LLM settings tie a 'plain neighbor voting' retrieval baseline in aggregate, but this tie masks a large gain on the adversarial 'price' split and an equally large loss on the multi-label 'promiscuous' split; and (4) open-book accuracy rises monotonically with sequence homology to the training database.

Significance. The paper is methodologically strong in several important respects: strictly inductive retrieval (training-only database, anonymized UniProt IDs), fully offline evidence cache, valid-EC rate guaranteed by construction for cascade methods, bootstrap confidence intervals, and an unusually candid appendix that discloses several limitations. The per-split arbitration story — LLM reasoning as an arbiter of conflicting evidence rather than a source of knowledge — is original and, if made robust, would be a valuable caution against relying on single-number leaderboards. The benchmark itself, with its four levers and matched controls, is a useful community asset. These strengths are real and should be credited. However, the quantitative form of the central claim depends on an underspecified and apparently underperforming retrieval baseline, and the main-text wording overstates the statistical support that the paper's own appendix qualifies. The qualitative direction of the arbitration finding may survive, but the exact 19-versus-19 cancellation and the aggregate 'tie' are not yet established as stated.

major comments (3)
  1. [§6 Finding 3; Tables 6(b) and 14] The 'plain neighbor voting' reference is underspecified and is not a competitive retrieval baseline. The 'BLAST plus HMMER' fusion scores below HMMER-only on the full set (0.723 vs 0.725), below HMMER-only on price (0.372 vs 0.378), and below both component channels on promiscuous (0.938 vs 0.957). No rule is given for how the top hits of the two channels are merged. Because the headline 19-versus-19 cancellation is computed against this reference, it is not robust to the choice of a reasonable retrieval baseline. Recomputing against HMMER-only changes the M2 deltas to +0.121 on price and −0.109 on promiscuous, and the M4-open deltas to +0.054 and −0.051, removing the exact symmetry. The paper must specify the fusion rule, justify it, and report the comparison against each channel separately and against a well-defined fusion of the two channels.
  2. [§6 Finding 3; §7 Conclusion; Appendix 'Statistical Significance'] The main text says the best LLM setting is 'indistinguishable' from neighbor voting and that the three settings 'tie overall,' but the appendix correctly states that overlapping bootstrap intervals do not establish equivalence, no equivalence margin was prespecified, and no paired between-model tests were run. The per-split contrasts of +0.127 and −0.090 rest on only 148 and 209 samples (about 19 items each), and the appendix says these contrasts are not covered by the reported intervals. The abstract and conclusion nonetheless state the cancellation without these caveats. Please either provide paired bootstrap intervals for the per-split deltas and a prespecified equivalence analysis, or rephrase the claims as directional and sign-consistent rather than as a tie and an exact cancellation.
  3. [§3 and §6 Finding 1; §7 Conclusion] The open-book condition supplies, by design, the EC numbers of homologous training neighbors in the evidence cache. The statement that 'external knowledge is decisive' is therefore close to a consequence of the protocol: the model is handed the answer in the retrieved label. The authors acknowledge this ('the retrieved evidence carries the EC numbers of homologous enzymes, so it puts the answer in context'), but the abstract and conclusion still present it as an empirical discovery. The 'law of homology availability' is similarly entangled, because the homology bins are computed from the same BLAST cache that provides the evidence. I recommend reframing Finding 1 as a sanity check of evidence use, or adding a control condition in which retrieved evidence contains sequences and homology scores but the EC labels are withheld, forcing the model to reason from similarity rather than read off
minor comments (5)
  1. [Fig. 2(C) vs Appendix B] Figure 2(C) lists 'pI' among the derived features, but Appendix B ('What a closed-book prompt contains') explicitly states that the descriptor block carries no isoelectric point. Please reconcile the figure with the text.
  2. [Appendix B, M3 subsection] The appendix discloses that M3 is the only method that receives the intrinsic-motif scan, so the M3−B0 contrast is not prompt-matched. Running B0 with the same motif list appended would cleanly isolate reasoning structure; as it stands, the direction of the confound is acknowledged but the clean control is missing.
  3. [Fig. 5] The annotation 'memorization' in the DeepSeek point label is speculative ('we suspect'). Please soften this to a hypothesis or add supporting evidence.
  4. [§6, experimental setup] All runs use a single seed, and the only repeated sampling is the k=3 self-consistency chains. This is disclosed in the appendix but should be stated in the main experimental section so that readers are not misled by the bootstrap intervals, which are sample-level only.
  5. [References] The LLaPA reference is listed as 'OpenReview; ICLR 2025 submission, withdrawn.' If the work was withdrawn, please cite the available version or explain why it is cited; if it was rejected/withdrawn, the citation format should be checked.

Circularity Check

1 steps flagged

Finding 4's 'law of homology availability' restates the open-book protocol (homology bins and fed evidence derive from the same retrieval cache); the headline arbitration result is empirical and not circular, though its fusion baseline is a robustness concern.

specific steps
  1. self definitional [Section 6, Finding 4 and Figure 6(b); Table 13 panel (d); abstract and conclusion]
    "Binning all 1349 items by the highest BLAST sequence identity between the test enzyme and the training database, open-book L4 climbs monotonically with homology... The mechanism is not subtle: the retrieved evidence carries the EC numbers of homologous enzymes, so it puts the answer in the context."

    Homology availability is measured from the same construction that defines the open-book input: Table 13(d) computes homology as 'the maximum local BLAST identity over the cached training hits,' and Appendix D defines open-book as 'the prompt may quote hits from a static evidence cache built offline from the training split alone.' The curve is pooled 'under B0', the baseline that pours cached EC numbers into the prompt, and Finding 3 shows B0·open (0.720) ≈ neighbor voting (0.723). So 'accuracy obeys a law of homology availability' largely restates the protocol: the model transcribes the answers the cache provides, and accuracy tracks whether those answers are correct. The paper concedes this ('the retrieved evidence carries the EC numbers of homologous enzymes, so it puts the answer in the

full rationale

The arbitration finding (Finding 3) is not circular: the 0.723 neighbor-vote vs 0.728 LLM aggregate and the 19-vs-19 item cancellation are measured comparisons between an independent fixed rule and black-box LLM outputs; no fitted parameter is renamed as a prediction and no equation collapses into its inputs. Its main fragility is different: the BLAST+HMMER fusion baseline underperforms HMMER-only on the full set (0.723 vs 0.725), on price (0.372 vs 0.378) and on promiscuous (0.938 vs 0.957) in Table 6(b), so the exact cancellation depends on an underperforming voting rule — a baseline-fairness/correctness risk, not a circularity, and the authors themselves disclose that the per-split gaps amount to only ~19 samples. Finding 1 (knowledge decisive) is a genuine ablation: open-book visibly raises accuracy, the closed/open intervals are disjoint, and the mechanism is transparently explained as answer-in-context. The self-citations (Li et al. 2025, 2026a, 2026b) appear once in the appendix for a general claim about knowledge organization and are not load-bearing. What partially reduces to the protocol is Finding 4: since the homology axis and the open-book evidence derive from the same offline cache and the curve is measured on the read-off-the-prompt baseline, the headline 'law' restates the construction. That partial circularity affects one of four headline findings, with the central arbitration claim retaining independent empirical content, hence score 4.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The benchmark's empirical claims do not rest on fitted constants; the free parameters listed are design choices that shape M2/M4 behavior and difficulty binning. The main domain assumptions are that CARE annotations and homology-based evidence are valid proxies for enzyme function, and that any-label scoring is acceptable on multi-label enzymes.

free parameters (5)
  • M4 vote threshold τ = 0.5
    Chosen by hand as the default for accepting a prefix-voting majority; directly controls how deep M4 commits (Section 4, M4).
  • M4 sampling count k = 3
    Fixed at 3 due to endpoint throughput; a larger k would change consensus depth, cost, and robustness (Section 4 and Appendix D).
  • M2 retrieval budget B = 4
    Global cap on agentic tool calls per enzyme; affects the accuracy versus retrieval-cost trade-off reported in Finding 3 (Section 4, M2).
  • Distractor mix ratio = ~70% siblings / 30% other subtrees
    Hand-set construction rule for cascade multiple-choice candidates; changes distractor hardness and hence the difficulty bins (Section 3 and Appendix B).
  • M3 intrinsic motif list = six hand-written regexes (GxGxxG, Walker-A, GxSxG, HExxH, D[TS]G, C[GP][PH]C)
    Hand-written heuristic motif scans shown only to M3; the authors state the M3 contrast is not prompt-matched and that this can only favor M3 (Appendix B).
axioms (5)
  • domain assumption CARE's EC list, splits, and labels are valid ground truth
    The 4963-leaf tree and the four splits are vendored with CARE; price annotations are assumed to be experimentally corrected, and this is not independently verified in the paper (Section 3 and Appendix B).
  • domain assumption Homology and structural similarity imply EC-function relevance
    Evidence channels (BLAST, HMMER, Foldseek, ESM-kNN, active-site motifs) are treated as carrying function-bearing evidence; the central open-book finding assumes these channels are meaningful proxies for enzyme function (Section 4, M2).
  • domain assumption LLM outputs can be reliably parsed and abstentions counted
    Unparseable or abstained levels are counted as wrong; the abstention-based Finding 2 depends on this parsing convention (Appendix D).
  • domain assumption Five proprietary endpoints are the models they claim to be
    The authors state endpoints were checked against model identity, but this cannot be independently verified from the paper text (Appendix D).
  • domain assumption Any-label match is acceptable for multi-label enzymes
    The promiscuous split is scored as correct if any one ground-truth label is predicted; the authors admit set-level precision/recall/F1 would be the stricter protocol and that this rule changes comparisons with published specialists (Appendix E).

pith-pipeline@v1.3.0-daily-deepseek · 35118 in / 9422 out tokens · 90756 ms · 2026-08-01T16:46:02.776730+00:00 · methodology

0 comments
read the original abstract

Enzyme function prediction is a hierarchical, knowledge-intensive form of protein function classification. Existing benchmarks expose an anomaly: general LLMs often get the coarse first level right, yet once asked for a complete EC number their accuracy at levels two through four drops to almost zero, while specialized models and tools stay usable. We propose EC-Reason-Bench, a training-free, diagnostic evaluation protocol built to answer two questions: why general LLMs score close to nothing on EC number prediction, and how much of that loss can be recovered without updating a single weight. We break enzyme classification ability into four orthogonal levers that can each be measured on their own: output structure, external knowledge, reasoning structure, and reasoning robustness. We test each lever with an inference-time method against a shared zero-shot baseline reproducing previously reported near-zero performance. Experiments with several strong reasoning LLMs yield four main findings. First, external knowledge is decisive and must precede reasoning: uniformly low closed-book performance rises sharply with open-book access, narrowing model gaps. Second, in closed-book settings, whether cascading and chain-of-thought help or hurt depends on a model's tendency to abstain. Third, once evidence is available the aggregate score of the best LLM setting is indistinguishable from simply voting the EC numbers of the nearest retrieved neighbors; that tie is an artifact of averaging, and it hides a large gain on adversarial evidence set against an equally large loss on multi-functional enzymes. Reasoning over evidence therefore acts as an arbiter of conflicting neighbors rather than as a source of knowledge, and no single-number leaderboard can see it. Fourth, accuracy obeys a law of homology availability.

Figures

Figures reproduced from arXiv: 2607.26397 by Dongming Jin, Gadeng Luosang, Huanyao Zhang, Linyu Li, Nyima Tashi, Xuan Zhang, Yichi Zhang, Yuanpeng He, Zhi Jin.

Figure 1
Figure 1. Figure 1: Motivation for EC-Reason-Bench. (a) The EC tree. (b) Level-wise accuracy under one-shot free generation: general LLMs collapse after the first level, while specialist and protein-aware methods decline smoothly. strong structural prior for free. The anomaly. As [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of EC-Reason-Bench. (A) Task and collapse. (B) Diagnosis of the near-zero failure into fixable formatting failures (A-type) and genuine knowledge failures (B-type). (C) Fully offline construction of the EC tree, the retrieval database and 1349 questions over four splits; domains there means motifs scanned from the sequence, not database annotation. (D) The zero-shot baseline B0 and the four orthog… view at source ↗
Figure 3
Figure 3. Figure 3: Level-wise accuracy for published baselines and our LLM settings. One panel per split; within each method the four bars run from L1 to L4, with L4 labeled on top. Grey: published baselines, in the order random, BLASTp (Altschul et al. 1990), Foldseek (Van Kempen et al. 2024), CLEAN (Yu et al. 2023), ESM-2 (Lin et al. 2023), ProtT5 (Elnaggar et al. 2021), S-PLM (Wang et al. 2025), GPT-4o-mini (Hurst et al. … view at source ↗
Figure 4
Figure 4. Figure 4: The specialist ceiling and the closed-to-open jump.(a) Best open-book LLM against published specialists, per split; promiscuous is not scored alike on the two sides. (b) Closed-book and open-book L4 per model. Bars are 95% bootstrap CIs. from <30%. The price gain is not a matter of picking the right channel, since it also clears our oracle best-channel retrieval bound of 0.426, which is granted per-sample … view at source ↗
Figure 7
Figure 7. Figure 7: M2 reliability diagrams. Self-reported confidence vs. accuracy per bin, one panel per model; the dashed line is perfect calibration and titles give ECE over the bins drawn, within 0.003 of the all-bin value tabulated in the appendix. echo of parametric memory rather than retrieval, since highly homologous enzymes are better studied and commoner in pretraining corpora. This curve also settles the mismatch f… view at source ↗
Figure 6
Figure 6. Figure 6: Where M2 searches, and what homology buys. (a) M2 tool calls per channel over the full set. (b) L4 against the highest train-test BLAST identity, all models pooled. is stable rather than random: M2 is the best open-book set￾ting on price for four of the five models and the worst on promiscuous for four of them. Selective retrieval and pre￾mature commitment are one mechanism seen from two sides, so the wrap… view at source ↗
Figure 8
Figure 8. Figure 8: Selective prediction with M2 risk-coverage curves. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: L4 accuracy for the open-book cascade across seven top-level EC classes. Rows are models and columns are classes; [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] 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

55 extracted references · 6 linked inside Pith

  1. [1]

    Journal of molecular biology , volume=

    Basic local alignment search tool , author=. Journal of molecular biology , volume=. 1990 , publisher=

    1990

  2. [2]

    Nucleic acids research , volume=

    HMMER web server: 2018 update , author=. Nucleic acids research , volume=. 2018 , publisher=

    2018

  3. [3]

    Nature biotechnology , volume=

    Fast and accurate protein structure search with Foldseek , author=. Nature biotechnology , volume=. 2024 , publisher=

    2024

  4. [4]

    Science , volume=

    Enzyme function prediction using contrastive learning , author=. Science , volume=. 2023 , publisher=

    2023

  5. [5]

    Science , volume=

    Evolutionary-scale prediction of atomic-level protein structure with a language model , author=. Science , volume=. 2023 , publisher=

    2023

  6. [6]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    ProtTrans: toward understanding the language of life through self-supervised learning , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2021 , publisher=

    2021

  7. [7]

    Advanced Science , volume=

    S-PLM: structure-aware protein language model via contrastive learning between sequence and structure , author=. Advanced Science , volume=. 2025 , publisher=

    2025

  8. [8]

    Proceedings of the National Academy of Sciences , volume=

    Deep learning enables high-quality and high-throughput prediction of enzyme commission numbers , author=. Proceedings of the National Academy of Sciences , volume=. 2019 , publisher=

    2019

  9. [9]

    elife , volume=

    ProteInfer, deep neural networks for protein functional inference , author=. elife , volume=. 2023 , publisher=

    2023

  10. [10]

    Nature Communications , volume=

    TopEC: prediction of Enzyme Commission classes by 3D graph neural networks and localized 3D protein descriptor , author=. Nature Communications , volume=. 2025 , publisher=

    2025

  11. [11]

    Briefings in Bioinformatics , volume=

    ProteinF3S: boosting enzyme function prediction by fusing protein sequence, structure, and surface , author=. Briefings in Bioinformatics , volume=. 2025 , publisher=

    2025

  12. [12]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Multimodal mixture-of-experts with retrieval augmentation for protein active site identification , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  13. [13]

    Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Improving Enzyme Prediction with Chemical Reaction Equations by Hypergraph-Enhanced Knowledge Graph Embeddings , author=. Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=

  14. [14]

    Nature Communications , volume=

    Trustworthy prediction of enzyme commission numbers using a hierarchical interpretable transformer , author=. Nature Communications , volume=. 2026 , publisher=

    2026

  15. [15]

    Advances in neural information processing systems , volume=

    Evaluating protein transfer learning with TAPE , author=. Advances in neural information processing systems , volume=

  16. [16]

    International conference on machine learning , pages=

    Protst: Multi-modality learning of protein sequences and biomedical texts , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  17. [17]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Prot2text: Multimodal protein’s function generation with gnns and transformers , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  18. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    ProtCLIP: Function-informed protein multi-modal learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  19. [19]

    International Conference on Learning Representations , volume=

    Saprot: Protein language modeling with structure-aware vocabulary , author=. International Conference on Learning Representations , volume=

  20. [20]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Retrieval-Augmented Language Model for Knowledge-aware Protein Encoding , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , editor =

    2025

  21. [21]

    arXiv preprint arXiv:2506.08954 , year=

    Protriever: End-to-end differentiable protein homology search for fitness prediction , author=. arXiv preprint arXiv:2506.08954 , year=

    Pith/arXiv arXiv

  22. [22]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    HeMeNet: Heterogeneous Multichannel Equivariant Network for Protein Multi-task Learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Care: a benchmark suite for the classification and retrieval of enzymes , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    Bioinformatics Advances , volume=

    EC-Bench: a benchmark for enzyme commission number prediction , author=. Bioinformatics Advances , volume=. 2026 , publisher=

    2026

  25. [25]

    Xie, Kun and Zhou, Peng and Zhang, Xingyi and Liu, Wei and Zhao, Peilin and Wang, Sibo and Jiang, Biaobin , booktitle=. Poinn. 2026 , url=

    2026

  26. [26]

    arXiv preprint arXiv:2402.13653 , year=

    PQA: Zero-shot Protein Question Answering for Free-form Scientific Enquiry with Large Language Models , author=. arXiv preprint arXiv:2402.13653 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  28. [28]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    STELLA: A multimodal LLM for protein functional annotation via unified sequence-structure encoding , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    2026

  29. [29]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    TIGER: Text-Informed Generalized Enzyme-Reaction Retrieval , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  30. [30]

    Jie Peng and Zijie Liu and Sukwon Yun and Yanyong Zhang and Tianlong Chen , year=

  31. [31]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Interleaved Tool-Call Reasoning for Protein Function Understanding , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  32. [32]

    arXiv preprint arXiv:2511.08024 , year=

    Knowledge-Augmented Long-CoT Generation for Complex Biomolecular Reasoning , author=. arXiv preprint arXiv:2511.08024 , year=

    arXiv

  33. [33]

    arXiv preprint arXiv:2205.10625 , year=

    Least-to-most prompting enables complex reasoning in large language models , author=. arXiv preprint arXiv:2205.10625 , year=

    Pith/arXiv arXiv

  34. [34]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  35. [35]

    arXiv preprint arXiv:2203.11171 , year=

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    Pith/arXiv arXiv

  36. [36]

    Advances in neural information processing systems , volume=

    Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in neural information processing systems , volume=

  37. [37]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    Pith/arXiv arXiv

  38. [38]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  39. [39]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Hierarchical text classification as sub-hierarchy sequence generation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  40. [40]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    A simple and effective framework for strict zero-shot hierarchical classification , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  41. [41]

    Advances in neural information processing systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

  42. [42]

    International conference on learning representations , volume=

    Self-rag: Learning to retrieve, generate, and critique through self-reflection , author=. International conference on learning representations , volume=

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    Self-discover: Large language models self-compose reasoning structures , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? , url =

    Yue, Yang and Chen, Zhiqi and Lu, Rui and Zhao, Andrew and Wang, Zhaokai and Yue, Yang and Song, Shiji and Huang, Gao , booktitle =. Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? , url =

  45. [45]

    Advances in Neural Information Processing Systems , volume=

    Grammar-aligned decoding , author=. Advances in Neural Information Processing Systems , volume=

  46. [46]

    International Conference on Machine Learning , pages=

    Convex calibrated surrogates for hierarchical classification , author=. International Conference on Machine Learning , pages=. 2015 , organization=

    2015

  47. [47]

    Nature Biotechnology , pages=

    Structural motif search across the protein universe with Folddisco , author=. Nature Biotechnology , pages=. 2026 , publisher=

    2026

  48. [48]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=

  49. [49]

    Nucleic Acids Research , volume=

    ExplorEnz: the primary source of the IUBMB enzyme list , author=. Nucleic Acids Research , volume=. 2009 , publisher=

    2009

  50. [50]

    Data Mining and Knowledge Discovery , volume=

    Evaluation measures for hierarchical classification: a unified view and novel approaches , author=. Data Mining and Knowledge Discovery , volume=. 2015 , publisher=

    2015

  51. [51]

    International conference on machine learning , pages=

    On calibration of modern neural networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  52. [52]

    International conference on machine learning , pages=

    Selectivenet: A deep neural network with an integrated reject option , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  53. [53]

    2026 , pages=

    Li, Linyu and Jin, Zhi and He, Yuanpeng and Jin, Dongming and Liu, Huanyu and Duan, Haoran and Zhang, Xuan and Zhang, Jin and Luosang, Gadeng and Tashi, Nyima , journal=. 2026 , pages=

    2026

  54. [54]

    Proceedings of the ACM Web Conference 2026 , pages=

    Learning to evolve: Bayesian-guided continual knowledge graph embedding , author=. Proceedings of the ACM Web Conference 2026 , pages=

    2026

  55. [55]

    Multi-View Riemannian Manifolds Fusion Enhancement for Knowledge Graph Completion , year=

    Li, Linyu and Jin, Zhi and Zhang, Xuan and Duan, Haoran and Wang, Jishu and Tao, Zhengwei and Zhao, Haiyan and Zhu, Xiaofeng , journal=. Multi-View Riemannian Manifolds Fusion Enhancement for Knowledge Graph Completion , year=