Pith. sign in

REVIEW 5 major objections 5 minor 42 references

MELLON - Multimodal Enhanced LLM for Online Navigation

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read MELLON, a multimodal web navigation agent, raises step-wise action accuracy on WebShop from 5.11% to 14.37% after one epoch of training.

desk verdict A course project with a real architecture and honest negative results, but the headline claim about task completion accuracy is not supported by the reported step-wise metric on a curated test set. read the letter →

arxiv 2608.09121 v1 pith:QZFVBYPX submitted 2026-08-10 cs.AI

Ruiyu Li , Haoyang Cai , Zhitong Guo , Tong Hu This is my paper · ORCID
classification cs.AI
keywords MELLONWebShopmultimodalnavigationvisual-textalignmentfrozenLLMparameter-efficienttrainingReActpromptingimitationlearning
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 sets out to show that web navigation agents benefit from seeing product images as well as reading page text, and that this can be achieved without retraining the language model. Its main proposal, MELLON, feeds aligned visual embeddings from a ViT and Q-Former through a single trainable projection layer into a frozen CodeLlama, trained with a cross-entropy action-prediction loss. On the WebShop benchmark the authors report step-wise action accuracy climbing from 5.11% to 14.37% after one training epoch, an average gain of 9.26 percentage points. Two alternative proposals, VQAgent and Multimodal Ranker, perform worse than their baselines, which the paper interprets as evidence about when multimodal input helps. The core claim is that parameter-efficient visual alignment can give a frozen LLM meaningful web-navigation ability.

What carries the argument

The load-bearing mechanism is the projection layer: a single trainable linear mapping placed between the Q-Former's visual output and the frozen LLM's embedding space. It converts visual information into prompt tokens inserted at an <ImageHere> slot in a shortened ReAct prompt, allowing the LLM to reason over text and image together. Because only the projection layer and Q-Former are trained, the whole adaptation fits in one epoch on a 24 GB GPU.

What would settle it

Train MELLON for one epoch as described and evaluate it on the standard 500-instance WebShop test set, reporting success rate and reward score. If the score and success rate do not exceed the ReAct baselines of 66.6 and 40.0 percent, the claimed 9.26-point improvement does not carry over to the benchmark's standard metric.

Watch

Extended reading notes

Core claim

The paper's central claim is that a frozen large language model can become a multimodal web navigation agent through a small trainable bridge. MELLON uses an EVA ViT to encode product images, a Q-Former to compress them, and a projection layer to turn the result into soft prompt tokens that CodeLlama can read alongside the WebShop accessibility tree and a ReAct-style reasoning prompt. After one epoch of training the projection layer and Q-Former with a cross-entropy generative loss, step-wise action accuracy on the paper's 338-instance test set rises from 5.11% to 14.37%, which the abstract reports as a 9.26% increase in task completion accuracy. The authors also find that a ViT-based encoder with a trainable Q-Former matches the ResNet baseline on score and success rate, while a frozen Q-Former does not improve over ResNet. On their own terms, the result is that multimodal alignment improves an LLM's action selection on WebShop without unfreezing the LLM.

Load-bearing premise

The result depends on treating step-wise action accuracy on a 338-instance test set, which the authors curated because it was hard for baselines, as task completion accuracy; if that metric and test set do not represent the standard WebShop benchmark, the headline improvement may not hold.

Editorial extensions

If this is right

  • MELLON can be adapted to new frozen LLMs without retraining the model, so stronger base models should translate directly into better web agents.
  • The one-epoch training budget makes the approach practical for environments where collecting large trajectory datasets is expensive.
  • The authors' error analysis suggests frozen CodeLlama's empty or invalid outputs are the main remaining bottleneck, so better prompting or light LLM tuning is the next lever.
  • The architecture is defined on WebShop but is not benchmark-specific; the same projection-layer recipe could be applied to other multimodal web environments.

Reading between the lines

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

  • Editorial inference: the reported gain is measured on step-wise action accuracy, not success rate or reward score; on the standard 500-instance WebShop test the improvement could be smaller or absent, and the paper does not provide that number.
  • Editorial inference: because the test set was curated to be challenging for baselines, the 9.26-point gain may overstate average-case benefit if hard cases are over-represented.
  • Editorial inference: the failure of VQAgent and Multimodal Ranker suggests that in WebShop, text usually carries more information than images; a testable implication is that MELLON's gains concentrate on attributes that appear only in images, such as color or material.
  • Editorial inference: a direct comparison on standard WebShop success and reward metrics against ReAct and the WebShop IL baseline would settle whether the step-wise gain translates into better task completion.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes three multimodal enhancements for the WebShop web-navigation benchmark: MELLON, which projects aligned ViT/Q-Former visual embeddings into a frozen CodeLlama model with redesigned ReAct prompts; VQAgent, which frames WebShop as a multiple-choice VQA task using BEiT3; and Multimodal Ranker, which augments action selection with BERT and CLIP similarity scores. The authors report that one epoch of MELLON training improves action accuracy from 5.11% to 14.37% on a 338-instance test set, while VQAgent and Multimodal Ranker both underperform the IL baseline. The abstract interprets the MELLON gain as a 9.26% improvement in 'task completion accuracy,' but Section 4.5 defines the reported metric as step-wise action accuracy, and no task-level success rate or reward score is reported for MELLON on the standard 500-instance WebShop test set.

Significance. If the step-wise accuracy gain were accompanied by task-level improvements on the standard WebShop test set, the parameter-efficient recipe of freezing the LLM and ViT while training only the projection layer and Q-Former would be a useful contribution, and the paper is candid in reporting that VQAgent and Multimodal Ranker fail to beat the baseline. However, the central positive claim is currently supported only by a non-public, nonstandard test set and a metric that does not measure task completion. The paper's only task-level evidence involving the visual architecture shows no improvement over the ResNet baseline, so the significance as stated is not established.

major comments (5)
  1. [Abstract; §4.5; §5.1.1] The abstract and Section 5.1.1 claim a 9.26% improvement in 'task completion accuracy,' but Section 4.5 explicitly defines the MELLON metric as step-wise accuracy, i.e., the fraction of steps where the predicted action matches the human action. No success rate or WebShop reward score for MELLON is reported on the standard 500-instance test set. The 9.26% gain is therefore a step-wise accuracy increase on the paper's own test set, not evidence of improved task completion; the conclusion in Section 6 itself reverts to 'step-wise accuracy,' confirming the discrepancy.
  2. [§4.3] The MELLON test set consists of 338 instances 'curated from the midterm report' and described as 'particularly challenging for baseline models,' but the curation procedure and the set itself are not described, justified, or released. Reporting results only on this nonstandard set makes the 9.26% gain non-comparable to existing WebShop results such as ReAct, IL, and IL+RL, and selecting instances specifically hard for baselines can inflate the apparent improvement. The paper should evaluate MELLON on the standard 500-instance test set or provide a clear argument that the curated set is representative.
  3. [§5.1.1, Table 3] The only task-level results connected to MELLON's visual architecture show that ViT+Trainable QFormer achieves a score of 39.92, identical to the ResNet baseline, and a success rate of 9.6% versus 9.8%. This is direct evidence that the improved step-wise action accuracy does not translate into improved task completion, and it contradicts the abstract's claim. The authors should report full MELLON success rate and score on the standard test set and reconcile these numbers with the claimed task-completion improvement.
  4. [§3.1.2] This section states that 'we integrated the WebShop completion score into the loss function as a contributing factor,' but the loss function defined immediately afterward is a pure cross-entropy loss over action tokens with no completion-score term. Either the score integration should be described precisely, for example as a weighted auxiliary term, or the claim should be removed; as written, the training objective is not actually specified.
  5. [§2.2 versus §5.1.2] Section 2.2 reports that ReAct achieves a score of 66.6 and a success rate of 40.0 on WebShop, but Table 4 lists ReAct (CodeLlama-7b) with a score of 24.98. These numbers are inconsistent, and the discrepancy is not explained. Because ReAct is a key comparison baseline, the paper must clarify the configuration and evaluation conditions used in each table.
minor comments (5)
  1. [Throughout] There are several typos and formatting errors, including 'backbond' in §3.1.1, 'batsh size' in §4.4, 'we doesn't train' in §5.1.1, and 'MELON' instead of 'MELLON' in Table 7.
  2. [§4.3] The MELLON training set of 8,237 instances is not related to the standard WebShop training split of 10,587 instances; please state explicitly how the training instances were filtered or subsampled.
  3. [§5.1.1, Table 2] The phrase '9.26% increase' is ambiguous between an absolute percentage-point gain and a relative gain; from Table 2 it is an absolute increase from 5.11% to 14.37%, so the wording should be clarified.
  4. [§3.1.2] The transition from the initially trained projection layer to the 'dual-focused' training strategy is described only qualitatively; a precise description of the data collection, loss weighting, and training schedule would be needed for reproducibility.
  5. [§5.1.3 and Appendix] Several textual references such as 'example 7', '11', '12', and '13' should be replaced with consistent figure or appendix numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MELLON's reported 9.26% gain is an empirical held-out measurement, not a derivation from its own inputs.

full rationale

The paper's central MELLON result is obtained by training a projection layer and Q-Former on human WebShop trajectories using a cross-entropy generation loss (Sec. 3.1.2) and comparing step-wise action accuracy before and after one epoch of training on a separate 338-instance test set (Secs. 4.3, 4.5, 5.1.1). This is an empirical measurement, not a derivation from the conclusion. There is no fitted parameter that is later renamed as a prediction: the 9.26% figure is simply 14.37% minus 5.11% in Table 2. The paper does not rely on load-bearing self-citations; all cited architectural components (ViT, Q-Former, BLIP-2, ReAct, WebShop) are external prior work, and no uniqueness theorem from the authors is invoked. VQAgent and Multimodal Ranker are reported as negative results and are evaluated with standard WebShop reward/success metrics, so their conclusions are not circular either. The genuine concerns in this paper are metric-validity and comparability issues: calling step-wise action accuracy 'task completion accuracy' (Sec. 5.1.1 vs Sec. 4.5) and using a test set 'curated from the midterm report' that is 'particularly challenging for baseline models' (Sec. 4.3). These are validity and comparability concerns, not circularity, because they do not make the measured outcome equivalent to the training signal or to a self-cited premise. Under the stated circularity definitions, the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a standard loss function, the appropriateness of WebShop as a benchmark, the equation of step-wise accuracy with task completion, and a nonstandard test set chosen to be difficult for baselines. No new physical or mathematical entities are introduced.

free parameters (3)
  • Projection layer weights = Learned on WebShop trajectories (not reported)
    The single projection layer maps Q-Former outputs into the LLM embedding space; it is trained on human action sequences and updated with a cross-entropy loss (Section 3.1.2).
  • Q-Former fine-tuned weights = Fine-tuned on WebShop data (details not specified)
    The Q-Former is trained along with the projection layer. Its weights are updated on WebShop trajectories, which directly influences the model's visual representations.
  • Multimodal Ranker learnable balance parameter = Trained with MSE loss (not reported)
    A learnable scalar balances BERT score and CLIP score in the ranker (Section 3.3). It is fit to WebShop reward labels.
assumptions (4)
  • standard math Cross-entropy loss is an appropriate objective for action generation
    The paper uses a standard cross-entropy loss for training MELLON's action outputs (Section 3.1.2). This is a common and accepted assumption in supervised learning.
  • domain assumption WebShop is a valid and representative benchmark for multimodal web navigation
    The paper assumes WebShop's product pages, with text and images, are a good stand-in for real e-commerce navigation. This is stated in Section 4.1 without independent justification.
  • ad hoc to paper Step-wise action accuracy is a meaningful proxy for task completion
    Section 4.5 defines accuracy as the fraction of steps where the agent's action matches the human action. The abstract and conclusion reinterpret this as "task completion accuracy." Whether step-wise agreement predicts successful purchases is an adopted assumption, and the paper provides no correlation evidence.
  • ad hoc to paper The curated 338-instance test set is representative enough for evaluation
    Section 4.3 states the test set is "specifically curated from the midterm report" and contains instances "particularly challenging for baseline models." This selection is an ad hoc choice that biases the reported improvement and is not the standard WebShop test set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MELLON - Multimodal Enhanced LLM for Online Navigation." pith.science (2026). https://pith.science/paper/QZFVBYPX

@misc{pith2026260809121,
  author       = {Pith},
  title        = {Pith review of: MELLON - Multimodal Enhanced LLM for Online Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QZFVBYPX}},
  note         = {Machine review of arXiv:2608.09121}
}
read the original abstract

Web navigation agents are capable of addressing various types of tasks on different websites. Current baselines on web navigation are either unimodal or lack strong reasoning abilities given multimodal inputs. Focusing on the WebShop benchmark, a real-world website simulation, we explore the alignment of text and images, as well as multimodal reasoning and planning abilities, to enhance the performance of web navigation agents. We propose three innovative multimodal enhancements: Multimodal Enhanced LLM for Online Navigation (MELLON), VQAgent, and Multimodal Ranker. MELLON demonstrates a significant improvement in task completion accuracy, with a 9.26% increase after just one epoch of training. Our findings suggest the necessity of further exploration into multimodal approaches, with a focus on more extensive training and alignment strategies to enhance the effectiveness of web navigation agents.

Figures

Figures reproduced from arXiv: 2608.09121 by the authors.

Figure 1
Figure 1. Example Trajectory for WebShop 1 github.com/ZhitongGuo/11-777-MMML-Project arXiv:2608.09121v1 [cs.AI] 10 Aug 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. MELLON Architecture 3.1.1. BETTER ENCODERS For the model architecture, we propose to use pre-trained visual encoders to generate image encoding of web screen￾shots, inspired by (Huang et al., 2023). More specifically, we can build a ViT backbond (Fang et al., 2022) coupled with a pre-trained Q-Former (Li et al., 2023). Unlike tradi￾tional convolutional neural networks, ViT relies on a trans￾former architecture. It d… view at source ↗
Figure 4
Figure 4. VQAgent Architecture Given a dataset D′ = {(o, A(o), a∗}M′ i=1 of M′ = 9558 samples from the training human trajectories, we train our model using imitation learning, which is a learning tech￾nique that tries to mimic human action in a task (Ho & Ermon, 2016). We encode observation o, which include the image and detail of the product we are trying to decide whether to buy, as well as the instruction, into embeddings… view at source ↗
Figures from the paper (9 more)
Figure 3
Figure 3. Figure 3: BEiT3 Model Architecture (Wang et al., 2022) For this purpose, we selected the state-of-the-art BEiT3 model (Wang et al., 2022), known for its versatility in lan￾guage, vision, and vision-language tasks, and its capability in both unimodal and multimodal settings. BEiT…
Figure 5
Figure 5. Figure 5: Multimodal Ranker Architecture 4. Experimental Methodology 4.1. WebShop In this paper, we are using WebShop as our benchmark to evaluate our agent. WebShop is a simulated e-commerce website environment that contains 1.18 million real-world products and 12, 087 crowd-so…
Figure 6
Figure 6. Figure 6: Imitation Learning Architecture Proposed by WebShop (Yao et al., 2023a) The model for choice-based imitation forecasts a probability distribution across all potential click actions, A(o), for a given observation, o. The model aims to increase the like￾lihood of the but…
Figure 8
Figure 8. Figure 8: Model Accuracy Comparison: Baseline vs VQAgent From the plot, it is not hard to see that at the first few epochs, VQAgent and baseline perform similarly in terms of accu￾racy. However, accuracy for VQAgent drops afterwards, while the baseline model has been learning an…
Figure 7
Figure 7. Figure 7: Baseline Trajectory (left) vs MELLON Trajectory (right) 5.2. VQAgent 5.2.1. RESULTS Model Score Success Rate (%) Baseline 39.92 9.8 VQAgent 31.26 5.7 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 10
Figure 10. Figure 10: Comparative analysis of WebShop Reward against BERT and CLIP scores. into the ranking algorithm might not effectively meet the goal of high task scores. The relationship between CLIP similarity and WebShop reward highlights potential issues in CLIP-based ranking for t…
Figure 9
Figure 9. Figure 9: WebShop Reward vs BERT + CLIP Scores The plots revealed a broad dispersion and lack of significant correlation. Higher WebShop scores did not consistently align with higher combined similarity scores, suggesting that BERT and CLIP amalgamation does not reliably predict…
Figure 13
Figure 13. Figure 13: Some images that provide context irrelevant to the task description, or did not provide more context compared to the textual descriptions [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]
Figure 12
Figure 12. Figure 12: Example Product Image from WebShop Title: Pointed Toe Pumps Shoes, Stiletto Plaid Leather High Heels, Sexy Elegant Party Wedding Work Court Shoes,Wine red,40 Description: Detailed parameters:Upper material: microfiber check pattern Inside material: leather Shoe sole m…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 8 canonical work pages

  1. [1]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  2. [2]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  3. [3]

    M. J. Kearns , title =

  4. [4]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  5. [5]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  6. [6]

    2022 , eprint=

    PaLM: Scaling Language Modeling with Pathways , author=. 2022 , eprint=

    2022

  7. [7]

    Suppressed for Anonymity , author=

  8. [8]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

Show all 42 references
  1. [9]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  2. [10]

    2023 , journal=

    Language Is Not All You Need: Aligning Perception with Language Models , author=. 2023 , journal=

    2023

  3. [11]

    Junnan Li and Dongxu Li and Silvio Savarese and Steven Hoi , year=

  4. [12]

    arXiv preprint arXiv:2305.16355 , year=

    PandaGPT: One Model To Instruction-Follow Them All , author=. arXiv preprint arXiv:2305.16355 , year=

    arXiv

  5. [13]

    arXiv preprint arXiv:2305.11854 , year=

    Multimodal Web Navigation with Instruction-Finetuned Foundation Models , author=. arXiv preprint arXiv:2305.11854 , year=

    arXiv

  6. [14]

    arXiv:2306.06070 , year=

    Mind2Web: Towards a Generalist Agent for the Web , author=. arXiv:2306.06070 , year=

    arXiv

  7. [15]

    arXiv:2307.12856 , year=

    A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis , author=. arXiv:2307.12856 , year=

    arXiv

  8. [16]

    arXiv:2207.01206 , year=

    WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents , author=. arXiv:2207.01206 , year=

    arXiv

  9. [17]

    arXiv:2307.13854 , year=

    WebArena: A Realistic Web Environment for Building Autonomous Agents , author=. arXiv:2307.13854 , year=

    arXiv

  10. [18]

    arXiv:2301.13280 , year=

    WebUI: A Dataset for Enhancing Visual UI Understanding with Web Semantics , author=. arXiv:2301.13280 , year=

    arXiv

  11. [19]

    arXiv:1808.09132 , year=

    Mapping natural language commands to web elements , author=. arXiv:1808.09132 , year=

    arXiv

  12. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    WIERT: Web Information Extraction via Render Tree , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2023 , month=. doi:10.1609/aaai.v37i11.26546 , abstractNote=

    2023 doi

  13. [21]

    CoVA: Context-aware Visual Attention for Webpage Information Extraction , journal =

    Anurendra Kumar and Keval Morabia and Jingjin Wang and Kevin Chen. CoVA: Context-aware Visual Attention for Webpage Information Extraction , journal =. 2021 , url =. 2110.12320 , timestamp =

    2021 arXiv

  14. [22]

    Muslea, Ion and Minton, Steve and Knoblock, Craig , year =

  15. [23]

    Advances in neural information processing systems , volume=

    Neural discrete representation learning , author=. Advances in neural information processing systems , volume=

  16. [24]

    arXiv preprint arXiv:2206.08916 , year=

    Unified-IO: A Unified Model for Vision, Language, and Multi-Modal Tasks , author=. arXiv preprint arXiv:2206.08916 , year=

    arXiv

  17. [25]

    arXiv preprint arXiv:2304.10592 , year=

    MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models , author=. arXiv preprint arXiv:2304.10592 , year=

    arXiv

  18. [26]

    2022 , eprint=

    WebGPT: Browser-assisted question-answering with human feedback , author=. 2022 , eprint=

    2022

  19. [27]

    2023 , eprint=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. 2023 , eprint=

    2023

  20. [28]

    2022 , eprint=

    EVA: Exploring the Limits of Masked Visual Representation Learning at Scale , author=. 2022 , eprint=

    2022

  21. [29]

    arXiv preprint arXiv:2304.11477 , year=

    Llm+ p: Empowering large language models with optimal planning proficiency , author=. arXiv preprint arXiv:2304.11477 , year=

    arXiv

  22. [30]

    arXiv preprint arXiv:2307.02485 , year=

    Building Cooperative Embodied Agents Modularly with Large Language Models , author=. arXiv preprint arXiv:2307.02485 , year=

    arXiv

  23. [31]

    doi:10.5281/zenodo.5143773 , url =

    Ilharco, Gabriel and Wortsman, Mitchell and Wightman, Ross and Gordon, Cade and Carlini, Nicholas and Taori, Rohan and Dave, Achal and Shankar, Vaishaal and Namkoong, Hongseok and Miller, John and Hajishirzi, Hannaneh and Farhadi, Ali and Schmidt, Ludwig , title =. doi:10.5281...

    doi

  24. [32]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Reproducible scaling laws for contrastive language-image learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  25. [33]

    ICML , year=

    Learning Transferable Visual Models From Natural Language Supervision , author=. ICML , year=

  26. [34]

    2022 , url=

    Christoph Schuhmann and Romain Beaumont and Richard Vencu and Cade W Gordon and Ross Wightman and Mehdi Cherti and Theo Coombes and Aarush Katta and Clayton Mullis and Mitchell Wortsman and Patrick Schramowski and Srivatsa R Kundurthy and Katherine Crowson and Ludwig Schmidt a...

    2022

  27. [35]

    2023 , eprint=

    Code Llama: Open Foundation Models for Code , author=. 2023 , eprint=

    2023

  28. [36]

    2023 , eprint=

    AgentBench: Evaluating LLMs as Agents , author=. 2023 , eprint=

    2023

  29. [37]

    2017 , eprint =

    Reinforced Mnemonic Reader for Machine Reading Comprehension , author =. 2017 , eprint =

    2017

  30. [38]

    2023 , eprint=

    Vision Transformer with Quadrangle Attention , author=. 2023 , eprint=

    2023

  31. [39]

    2016 , eprint=

    VQA: Visual Question Answering , author=. 2016 , eprint=

    2016

  32. [40]

    2022 , eprint=

    Image as a Foreign Language: BEiT Pretraining for All Vision and Vision-Language Tasks , author=. 2022 , eprint=

    2022

  33. [41]

    2016 , eprint=

    Generative Adversarial Imitation Learning , author=. 2016 , eprint=

    2016

  34. [42]

    arXiv preprint arXiv:2304.15010 , year=

    Llama-adapter v2: Parameter-efficient visual instruction model , author=. arXiv preprint arXiv:2304.15010 , year=

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.