Pith. sign in

REVIEW 4 major objections 5 minor 58 references

Enhancing RLHF with Human Gaze Modeling

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that incorporating predicted human gaze into RLHF—either through gaze-augmented reward models or gaze-based token reward distribution—accelerates policy convergence by a factor of 1.5–2 while preserving final model…

desk verdict GazeDistrib is a fresh idea and the first to put gaze-informed rewards into an RLHF loop, but the 1.5-2x speedup claim rests on an undefined 'Steps to Conv.' metric, making the central result untestable as written. read the letter →

arxiv 2507.09016 v2 pith:GGN5JAIS submitted 2025-07-11 cs.LG

Karim Galliamov , Ivan Titov , Ilya Pershin This is my paper
classification cs.LG
keywords RLHFhumangazemodelingeyetrackingtotalreadingtimedenserewarddistributiontoken-levelpolicyoptimizationGRPO
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper aims to show that predicted human gaze can make RLHF cheaper. It reports that two gaze-based interventions—adding predicted gaze features to a reward model and using predicted reading time to spread a single sequence reward across tokens—reach the same validation score in roughly half the policy-optimization steps as standard sparse-reward RLHF, with final quality preserved or slightly improved. The speedup holds across PPO and GRPO runs on two model/dataset combinations. If the claim is right, alignment training becomes more affordable within a fixed compute budget, and human attention patterns gain a role in reward design that goes beyond preference ranking.

What carries the argument

The load-bearing mechanism is GazeDistrib's reward-redistribution rule, Eq. (1): for a sequence of $n$ tokens with predicted Total Reading Time values $t_1,\dots,t_n$ and a scalar reward $R$, each token receives $r_i = R \cdot e^{t_i}/\sum_{j=1}^n e^{t_j}$, a softmax over gaze values that gives more of the sequence reward to tokens readers dwell on longer. This couples token-level credit assignment to human attention rather than to the reward model's internal attention. The other component is a frozen gaze-prediction model of up to 125M parameters that supplies the TRT values; because it is far smaller than the 7B policy and reward models, its inference cost is negligible.

What would settle it

A control experiment that shuffles the predicted TRT scores across tokens within each sequence, keeping the reward scale and softmax temperature fixed, would settle the claim: if the 1.5–2x speedup persists under shuffled gaze weights, the specific gaze signal is not the cause of the acceleration.

Watch

Extended reading notes

Core claim

The central discovery is that gaze-informed RLHF converges faster than standard sparse-reward RLHF without sacrificing final quality. The paper demonstrates this in two ways: GazeRM, which integrates predicted gaze features into a reward model's first-layer token embeddings, and GazeDistrib, which redistributes a sequence-level reward $R$ across tokens according to predicted Total Reading Time via $r_i = R \cdot e^{t_i}/\sum_{j=1}^n e^{t_j}$. Across PPO runs on LLaMA-7B with HH-RLHF and Mistral-7B with OASST2, both approaches reach the same hold-out validation scores in about half the optimization steps; GRPO shows a smaller but consistent speedup. The paper reads this as evidence that fine-grained human attention signals address the sparse-reward problem in policy optimization, independent of the specific RL algorithm.

Load-bearing premise

GazeDistrib's benefit rests on the premise that predicted Total Reading Time marks the tokens that actually drive human preference judgments; if gaze dwell time does not track reward-relevant tokens, reallocating reward by gaze would not speed up learning.

Editorial extensions

If this is right

  • RLHF runs on the same preference data would need roughly half the PPO steps to reach the same hold-out validation score, directly lowering the compute cost of policy optimization.
  • GazeDistrib can be dropped into an existing RLHF stack without retraining the reward model; it requires only a small frozen gaze predictor, so adding it costs little at inference time.
  • The speedup transfers across two policy-optimization algorithms (PPO and GRPO) and two model/dataset pairs, suggesting the benefit is not an artifact of one architecture.
  • Gaze-informed reward models, previously shown to improve preference ranking, also accelerate full-pipeline training, so gains in reward quality carry over to training efficiency.
  • Since final performance is maintained or slightly improved, using gaze guidance is a near-free way to make larger alignment runs fit into a fixed compute budget.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A hybrid that combines GazeDistrib's gaze-based token weights with the reward model's internal attention weights might be more robust than either signal alone; the paper does not test this combination.
  • A direct test of the mechanism would replace predicted TRT with static weights biased toward content words (nouns, verbs, punctuation): if the speedup persists, the gaze model's PoS preference alone could explain the gain.
  • If gaze dwell time really marks reward-relevant tokens, the same signal could seed offline preference-learning methods with token-level importance, a direction the paper leaves untested.
  • Because the gaze predictor is trained on English reading corpora, the reward redistribution is only as portable as the gaze model; multilingual gaze predictors would be needed to transfer the method to other languages.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes two ways to inject predicted human gaze into RLHF: GazeRM, which augments a reward model with gaze features as in López-Cardona et al. (2025), and GazeDistrib, which redistributes a sequence-level reward to tokens proportionally to predicted Total Reading Time. Experiments with PPO and GRPO on HH-RLHF and OASST2 report that gaze-informed methods converge 1.5–2 times faster than sparse-reward baselines without compromising final hold-out reward scores. The authors also provide an analysis of gaze scores by part of speech and release code.

Significance. If the convergence-speed claim is correct and reproducible, the work is useful because RLHF is computationally expensive and gaze is an underused signal. The paper has several strengths: it evaluates with an external hold-out reward model, compares two online RL algorithms, and releases code. However, the central quantitative claim rests on an undefined metric, and the reported error bars are inconsistent with integer-valued training steps. The GazeDistrib mechanism also lacks validation of its key assumption. The result is therefore not yet testable as written, and the paper needs a major revision before the claim can be assessed.

major comments (4)
  1. [Section 5, Table 2] The paper's central claim is a 1.5–2x convergence speedup, but the only quantitative evidence, the 'Steps to Conv.' column in Table 2, is never defined. Section 4.3 defines the validation score but not what 'convergence' means: a fixed fraction of the final score, a threshold with patience, a plateau rule, or something else. It is also not stated whether the same criterion was applied to all methods, whether the criterion was chosen post hoc, or how the step count is computed across the three seeds. Moreover, the reported standard deviations are implausible for integer step counts: for example, 41.40±0.21 and 28.33±0.54 with three seeds cannot arise from integer-valued per-seed steps unless the metric is fractional or aggregated in an unspecified way. The definition of this metric and the aggregation procedure must be stated before the speedup claim can be evaluated.
  2. [Section 5 vs. Section 6] The conclusion states that both approaches 'accelerate convergence during policy optimization by a factor of 1.5-2 times', but Section 5 reports that the GRPO experiments show 'similar convergence speed improvements of approximately 1.3 times faster'. These numbers are inconsistent. If GRPO only achieves a 1.3x speedup, the blanket 1.5–2x claim in the abstract and conclusion overstates the result. Please reconcile the reported GRPO speedup with the headline claim, or qualify the conclusion accordingly.
  3. [Section 3.3, Eq. (1)] GazeDistrib's reward redistribution assumes that predicted Total Reading Time is a reliable proxy for which tokens determine human preference judgments, so that reallocating the sequence reward proportionally to exp(TRT) focuses learning on the right tokens. The paper motivates this with the part-of-speech table (Table 1), but it provides no evidence that the tokens with high predicted gaze are the ones that drive chosen/rejected preference labels. Without this link, the mechanism underlying GazeDistrib is unsupported. Please add an analysis or ablation that relates gaze predictions to preference-relevant tokens, or at least a sensitivity study showing that the convergence benefit is robust to the softmax temperature and to the choice of gaze metric.
  4. [Section 5, Table 2] The experiments confound model architecture with dataset: LLaMa is always trained on HH-RLHF and Mistral on OASST2. The text claims the results hold 'across different model architectures (LLaMa and Mistral) and datasets (HH-RLHF and OASST2)', but the current design cannot separate architecture effects from dataset effects. In addition, the final validation scores in Table 2 are not significantly different across methods: e.g., LLaMa Baseline 1.09±0.08 vs. GazeDistrib 1.22±0.12, and Mistral Baseline 1.21±0.15 vs. GazeRM 1.10±0.31. The 'maintaining or slightly improving final performance' claim is therefore not established by the reported numbers. Please either add a cross (e.g., LLaMa on OASST2 or Mistral on HH-RLHF) or soften the generalization claim.
minor comments (5)
  1. [Abstract and Section 3.3] There are typos: 'demonstate' in the abstract and 'levereging' in Section 3.3.
  2. [Figures 1–3] In the provided version, the learning curves contain only axis labels and captions without visible plotted data. Please ensure the final figures are legible and that the curves are described in the text quantitatively, since Table 2 does not define the steps metric.
  3. [Appendix A] The appendix is titled 'Implementation details and hparams tuning' but lists only datasets and model checkpoints. Please add the actual hyperparameters (learning rate, batch size, PPO/GRPO settings, number of steps, and reward-model training details) to support reproducibility.
  4. [References] The Mistral 7B reference appears to be a Master's thesis on LLM vulnerabilities, which is not the correct citation for the Mistral 7B model (Jiang et al., 2023). Please correct this and verify all other citations.
  5. [Section 5] The claim that the hold-out reward model 'achieves accuracy of 69.25 on HH-RLHF validation split' would be more informative with a confidence interval or at least a note on whether this is a point estimate from a single model; also clarify whether the same hold-out model is used for OASST2 and whether its accuracy on that split was checked.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the convergence claim is measured against an external hold-out reward model and no fitted parameter or self-citation forces the result.

full rationale

The paper's two proposals, GazeRM and GazeDistrib, rely on frozen gaze predictors trained on external eye-tracking corpora (Cop et al., 2016; Hollenstein et al., 2018) and on a reward model architecture reproduced from prior work by López-Cardona et al. (2025), none of which are authored by the present paper's contributors. The evaluation compares policies using a hold-out reward model (OpenAssistant's deberta-v3-large-v2) that was not used in any policy training run, and the reported convergence speedups and final validation scores are expressed relative to that external model. Equation (1) redistributes the sequence reward according to predicted Total Reading Time, but no parameter in that redistribution is fitted to the convergence measurements being reported, nor does the paper invoke any uniqueness theorem or self-authored result to make the gaze-based approach the only possible choice. The principal quantitative weakness is that the 'Steps to Conv.' metric in Table 2 is never defined and the learning curves are not readable in the provided text; this is an evaluation-specification and falsifiability concern, not circular reasoning. Because no load-bearing step reduces by construction to its own input, the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

Method has no invented entities; its new machinery is a reward-weighting formula with a fixed softmax temperature and an unspecified convergence criterion.

free parameters (2)
  • Implicit softmax temperature in Eq. (1) = 1 (fixed, not tuned)
    GazeDistrib weights rewards as r_i = R * exp(t_i) / sum_j exp(t_j). The temperature is implicitly 1; this controls how strongly rewards concentrate on high-TRT tokens and is a free modeling choice.
  • Convergence criterion for 'steps to convergence' = Not specified
    Table 2 reports that GazeDistrib reaches convergence in 22.33 steps (LLaMa/HH-RLHF) vs 41.40 for baseline, but the rule for declaring convergence is never stated. This hidden criterion drives the paper's principal efficiency claim.
assumptions (3)
  • domain assumption Predicted Total Reading Time (TRT) is a valid proxy for token importance in human preference judgments.
    GazeDistrib's token-level credit assignment assumes dwell time marks the tokens that drive preference; only the PoS correlation table supports this.
  • domain assumption The hold-out reward model (OpenAssistant deberta-v3-large) reflects true policy quality and is not hackable by gaze-informed policies.
    All cross-method comparisons use this single reward model; the paper gives no robustness check against reward hacking.
  • domain assumption Gaze models trained on sentence-reading corpora (GECO, ZuCo) transfer to dialogue text in HH-RLHF and OASST2.
    The paper filters to English but does not validate gaze prediction quality on the RLHF domains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing RLHF with Human Gaze Modeling." pith.science (2026). https://pith.science/paper/GGN5JAIS

@misc{pith2026250709016,
  author       = {Pith},
  title        = {Pith review of: Enhancing RLHF with Human Gaze Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GGN5JAIS}},
  note         = {Machine review of arXiv:2507.09016}
}
read the original abstract

Reinforcement Learning from Human Feedback (RLHF) aligns language models with human preferences but is computationally expensive. We explore two approaches that leverage human gaze modeling to enhance RLHF: (1) gaze-aware reward models and (2) gaze-based distribution of sparse rewards at token level. Our experiments demonstate that gaze-informed RLHF achieves faster convergence while maintaining or slightly improving performance, thus, reducing computational costs during policy optimization. These results show that human gaze provides a valuable and underused signal for policy optimization, pointing to a promising direction for improving RLHF efficiency.

Figures

Figures reproduced from arXiv: 2507.09016 by the authors.

Figure 1
Figure 1. Performance of LLaMa-7B-open-instruct models trained with and without gaze information on HH-RLHF efficient learning. The primary contributions of this work are: • integrating human gaze prediction models into RLHF pipelines through two alternative meth￾ods; • empirical evidence showing significant con￾vergence acceleration across two online-RL algorithms; • analysis of gaze-informed reward distribution effects on t… view at source ↗
Figure 2
Figure 2. Performance of Mistral-7B trained with and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Samples of human gaze modeling 7 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 29 canonical work pages

  1. [1]

    Alan Akbik, Duncan Blythe, and Roland Vollgraf. 2018. Contextual string embeddings for sequence labeling. In COLING 2018, 27th International Conference on Computational Linguistics , pages 1638--1649

    work page 2018

  2. [2]

    Yuntao Bai , Andy Jones , Kamal Ndousse , Amanda Askell , Anna Chen , Nova DasSarma , Dawn Drain , Stanislav Fort , Deep Ganguli , Tom Henighan , Nicholas Joseph , Saurav Kadavath , Jackson Kernion , Tom Conerly , Sheer El-Showk , Nelson Elhage , Zac Hatfield-Dodds , Danny Hernandez , Tristan Hume , and 12 others. 2022. https://doi.org/10.48550/arXiv.2204...

    arXiv doi:10.48550/arxiv.2204.05862 2022

  3. [3]

    Alex J Chan, Hao Sun, Samuel Holt, and Mihaela Van Der Schaar. 2024. Dense reward for free in reinforcement learning from human feedback. arXiv preprint arXiv:2402.00782

    arXiv 2024

  4. [4]

    Uschi Cop, Nicolas Dirix, Denis Drieghe, and Wouter Duyck. 2016. https://doi.org/10.3758/s13428-016-0734-0 Presenting geco: An eyetracking corpus of monolingual and bilingual sentence reading . Behavior Research Methods, 49

    doi:10.3758/s13428-016-0734-0 2016

  5. [5]

    Nora Hollenstein, Jonathan Rotsztejn, Marius Troendle, Andreas Pedroni, Ce Zhang, and Nicolas Langer. 2018. https://api.semanticscholar.org/CorpusID:54480751 Zuco, a simultaneous eeg and eye-tracking resource for natural sentence reading . Scientific Data, 5

    work page 2018

  6. [6]

    Fengqing Jiang. 2024. Identifying and mitigating vulnerabilities in llm-integrated applications. Master's thesis, University of Washington

    2024

  7. [7]

    Chris Kleinke. 1986. https://doi.org/10.1037/0033-2909.100.1.78 Gaze and eye contact. a research review . Psychological bulletin, 100:78--100

    doi:10.1037/0033-2909.100.1.78 1986

  8. [8]

    o pf, Yannic Kilcher, Dimitri Von R \

    Andreas K \"o pf, Yannic Kilcher, Dimitri Von R \"u tte, Sotiris Anagnostidis, Zhi Rui Tam, Keith Stevens, Abdullah Barhoum, Duc Nguyen, Oliver Stanley, Rich \'a rd Nagyfi, and 1 others. 2023. Openassistant conversations-democratizing large language model alignment. Advances in Neural Information Processing Systems, 36:47669--47681

    work page 2023

Show all 58 references
  1. [9]

    \'A ngela L \'o pez-Cardona, Carlos Segura, Alexandros Karatzoglou, Sergi Abadal, and Ioannis Arapakis. 2025. https://openreview.net/forum?id=uZgK0tcPqd Seeing eye to AI : Human alignment via gaze-based response rewards for large language models . In The Thirteenth Internation...

    2025

  2. [10]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    2017 arXiv

  3. [11]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  4. [12]

    Leandro von Werra, Younes Belkada, Lewis Tunstall, Edward Beeching, Tristan Thrush, Nathan Lambert, Shengyi Huang, Kashif Rasul, and Quentin Gallouédec. 2020. Trl: Transformer reinforcement learning. https://github.com/huggingface/trl

    2020

  5. [13]

    Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. 2019. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593

    2019 arXiv

  6. [14]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  7. [15]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  8. [16]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  9. [17]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems , 33:18661--18673, 2020

    2020

  10. [18]

    Codet5+: Open code large language models for code understanding and generation

    Yue Wang, Hung Le, Akhilesh Gotmare, Nghi Bui, Junnan Li, and Steven Hoi. Codet5+: Open code large language models for code understanding and generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages 1069--1088, 2023

    2023

  11. [19]

    Scaling down to scale up: A guide to parameter-efficient fine-tuning

    Vladislav Lialin, Vijeta Deshpande, and Anna Rumshisky. Scaling down to scale up: A guide to parameter-efficient fine-tuning. arXiv preprint arXiv:2303.15647 , 2023

    2023 arXiv

  12. [20]

    Low-rank adaptation of large language model rescoring for parameter-efficient speech recognition

    Yu Yu, Chao-Han Huck Yang, Jari Kolehmainen, Prashanth G Shivakumar, Yile Gu, Sungho Ryu Roger Ren, Qi Luo, Aditya Gourav, I-Fan Chen, Yi-Chieh Liu, et al. Low-rank adaptation of large language model rescoring for parameter-efficient speech recognition. In 2023 IEEE Automatic ...

    2023

  13. [21]

    Adaptive budget allocation for parameter-efficient fine-tuning

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512 , 2023

    2023 arXiv

  14. [22]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691 , 2021

    2021 arXiv

  15. [23]

    Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning

    Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems , 35:1950--1965, 2022

    1950

  16. [24]

    Codesearchnet challenge: Evaluating the state of semantic code search

    Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. Codesearchnet challenge: Evaluating the state of semantic code search. arXiv preprint arXiv:1909.09436 , 2019

    1909 arXiv

  17. [25]

    Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021

    2021

  18. [26]

    A neural framework for retrieval and summarization of source code

    Qingying Chen and Minghui Zhou. A neural framework for retrieval and summarization of source code. In Proceedings of the 33rd ACM/IEEE International Conference on Automated Software Engineering , pages 826--831, 2018

    2018

  19. [27]

    Survey of code search based on deep learning

    Yutao Xie, Jiayi Lin, Hande Dong, Lei Zhang, and Zhonghai Wu. Survey of code search based on deep learning. ACM Transactions on Software Engineering and Methodology , 33(2):1--42, 2023

    2023

  20. [28]

    Rosf: Leveraging information retrieval and supervised learning for recommending code snippets

    He Jiang, Liming Nie, Zeyi Sun, Zhilei Ren, Weiqiang Kong, Tao Zhang, and Xiapu Luo. Rosf: Leveraging information retrieval and supervised learning for recommending code snippets. IEEE Transactions on Services Computing , 12(1):34--46, 2016

    2016

  21. [29]

    Sniff: A search engine for java using free-form queries

    Shaunak Chatterjee, Sudeep Juvekar, and Koushik Sen. Sniff: A search engine for java using free-form queries. In Fundamental Approaches to Software Engineering: 12th International Conference, FASE 2009, Held as Part of the Joint European Conferences on Theory and Practice of S...

    2009

  22. [30]

    Nl-based query refinement and contextualized code search results: A user study

    Emily Hill, Manuel Roldan-Vega, Jerry Alan Fails, and Greg Mallet. Nl-based query refinement and contextualized code search results: A user study. In 2014 Software Evolution Week-IEEE Conference on Software Maintenance, Reengineering, and Reverse Engineering (CSMR-WCRE) , page...

    2014

  23. [31]

    Bimodal modelling of source code and natural language

    Miltos Allamanis, Daniel Tarlow, Andrew Gordon, and Yi Wei. Bimodal modelling of source code and natural language. In International conference on machine learning , pages 2123--2132. PMLR, 2015

    2015

  24. [32]

    Codebert: A pre-trained model for programming and natural languages

    Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, et al. Codebert: A pre-trained model for programming and natural languages. arXiv preprint arXiv:2002.08155 , 2020

    2002 arXiv

  25. [33]

    Graphcodebert: Pre-training code representations with data flow

    Daya Guo, Shuo Ren, Shuai Lu, Zhangyin Feng, Duyu Tang, Shujie Liu, Long Zhou, Nan Duan, Alexey Svyatkovskiy, Shengyu Fu, et al. Graphcodebert: Pre-training code representations with data flow. arXiv preprint arXiv:2009.08366 , 2020

    2009 arXiv

  26. [34]

    Comparison of graph embeddings for source code with text models based on cnn and codebert architectures

    Vitaly Anatolyevich ROMANOV and Vladimir Vladimirovich IVANOV. Comparison of graph embeddings for source code with text models based on cnn and codebert architectures. Proceedings of the Institute for System Programming of the RAS (Proceedings of ISP RAS) , 35(1):237--264, 2023

    2023

  27. [35]

    A novel neural source code representation based on abstract syntax tree

    Jian Zhang, Xu Wang, Hongyu Zhang, Hailong Sun, Kaixuan Wang, and Xudong Liu. A novel neural source code representation based on abstract syntax tree. In 2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE) , pages 783--794. IEEE, 2019

    2019

  28. [36]

    Retrieval-augmented generation for code summarization via hybrid gnn

    Shangqing Liu, Yu Chen, Xiaofei Xie, Jingkai Siow, and Yang Liu. Retrieval-augmented generation for code summarization via hybrid gnn. arXiv preprint arXiv:2006.05405 , 2020

    2006 arXiv

  29. [37]

    Multimodal representation for neural code search

    Jian Gu, Zimin Chen, and Martin Monperrus. Multimodal representation for neural code search. In 2021 IEEE International Conference on Software Maintenance and Evolution (ICSME) , pages 483--494. IEEE, 2021

    2021

  30. [38]

    Text and code embeddings by contrastive pre-training

    Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, et al. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005 , 2022

    2022 arXiv

  31. [39]

    Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment

    Lingling Xu, Haoran Xie, Si-Zhao Joe Qin, Xiaohui Tao, and Fu Lee Wang. Parameter-efficient fine-tuning methods for pretrained language models: A critical review and assessment. arXiv preprint arXiv:2312.12148 , 2023

    2023 arXiv

  32. [40]

    Exploring parameter-efficient fine-tuning techniques for code generation with large language models

    Martin Weyssow, Xin Zhou, Kisub Kim, David Lo, and Houari Sahraoui. Exploring parameter-efficient fine-tuning techniques for code generation with large language models. arXiv preprint arXiv:2308.10462 , 2023

    2023 arXiv

  33. [41]

    No more fine-tuning? an experimental evaluation of prompt tuning in code intelligence

    Chaozheng Wang, Yuanhang Yang, Cuiyun Gao, Yun Peng, Hongyu Zhang, and Michael R Lyu. No more fine-tuning? an experimental evaluation of prompt tuning in code intelligence. In Proceedings of the 30th ACM joint European software engineering conference and symposium on the found...

    2022

  34. [42]

    Codet5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation

    Yue Wang, Weishi Wang, Shafiq Joty, and Steven CH Hoi. Codet5: Identifier-aware unified pre-trained encoder-decoder models for code understanding and generation. arXiv preprint arXiv:2109.00859 , 2021

    2021 arXiv

  35. [43]

    Self-supervised learning: Generative or contrastive

    Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering , 35(1):857--876, 2021

    2021

  36. [44]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , p...

    2021

  37. [45]

    Contrastive multiview coding

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XI 16 , pages 776--794. Springer, 2020

    2020

  38. [46]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems , 33:5812--5823, 2020

    2020

  39. [47]

    Contrastive learning with cross-modal knowledge mining for multimodal human activity recognition

    Razvan Brinzea, Bulat Khaertdinov, and Stylianos Asteriadis. Contrastive learning with cross-modal knowledge mining for multimodal human activity recognition. In 2022 International Joint Conference on Neural Networks (IJCNN) , pages 01--08. IEEE, 2022

    2022

  40. [48]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning , pages 1597--1607. PMLR, 2020

    2020

  41. [49]

    Self-supervised contrastive bert fine-tuning for fusion-based reviewed-item retrieval

    Mohammad Mahdi Abdollah Pour, Parsa Farinneya, Armin Toroghi, Anton Korikov, Ali Pesaranghader, Touqir Sajed, Manasa Bharadwaj, Borislav Mavrin, and Scott Sanner. Self-supervised contrastive bert fine-tuning for fusion-based reviewed-item retrieval. In European Conference on I...

    2023

  42. [50]

    Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning

    Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing , 508:293--304, 2022

    2022

  43. [51]

    Xlcost: A benchmark dataset for cross-lingual code intelligence

    Ming Zhu, Aneesh Jain, Karthik Suresh, Roshan Ravindran, Sindhu Tipirneni, and Chandan K Reddy. Xlcost: A benchmark dataset for cross-lingual code intelligence. arXiv preprint arXiv:2206.08474 , 2022

    2022 arXiv

  44. [52]

    Staqc: A systematically mined question-code dataset from stack overflow

    Hu Yao and et al. Staqc: A systematically mined question-code dataset from stack overflow. In Proceedings of the World Wide Web (WWW'18) Conference , pages 135--144, 2018

    2018

  45. [53]

    Pytorrent: A python library corpus for large-scale language models

    Mehdi Bahrami, NC Shrikanth, Shade Ruangwan, Lei Liu, Yuji Mizobuchi, Masahiro Fukuyori, Wei-Peng Chen, Kazuki Munakata, and Tim Menzies. Pytorrent: A python library corpus for large-scale language models. arXiv preprint arXiv:2110.01710 , 2021

    arXiv

  46. [54]

    Search4code: Code search intent classification using weak supervision

    Nikitha Rao, Chetan Bansal, and Joe Guan. Search4code: Code search intent classification using weak supervision. In 2021 IEEE/ACM 18th International Conference on Mining Software Repositories (MSR) , pages 575--579. IEEE, 2021

    2021

  47. [55]

    Semantic code search for smart contracts

    Chaochen Shi, Yong Xiang, Jiangshan Yu, and Longxiang Gao. Semantic code search for smart contracts. arXiv preprint arXiv:2111.14139 , 2021

    2021 arXiv

  48. [56]

    Isadetect: Usable automated detection of cpu architecture and endianness for executable binary files and object code

    Sami Kairajärvi, Andrei Costin, and Timo Hämäläinen. Isadetect: Usable automated detection of cpu architecture and endianness for executable binary files and object code. In Proceedings of the Tenth ACM Conference on Data and Application Security and Privacy , CODASPY ’20. ACM...

    2020

  49. [57]

    Deepseek-coder: When the large language model meets programming--the rise of code intelligence

    Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming--the rise of code intelligence. arXiv preprint arXiv:2401.14196 , 2024

    2024 arXiv

  50. [58]

    Rouge 2.0: Updated and improved measures for evaluation of summarization tasks

    Kavita Ganesan. Rouge 2.0: Updated and improved measures for evaluation of summarization tasks. arXiv preprint arXiv:1803.01937 , 2018

    2018 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.