Pith. sign in

REVIEW 3 major objections 4 minor 39 references

The paper argues that image retrieval should be reframed as autonomous exploration over visual histories, because the clues that resolve a query and the images that answer it often live in different photos; its benchmark shows the best agen

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

2026-08-03 00:57 UTC pith:CDYQQGLM

load-bearing objection A genuinely new retrieval benchmark with a well-documented construction pipeline, but the necessity-of-agentic-reasoning claim is untested because the paper never runs a non-agentic multi-stage baseline. the 3 major comments →

arxiv 2602.10809 v2 pith:CDYQQGLM submitted 2026-02-11 cs.CV cs.IR

DeepImageSearch: Benchmarking Multimodal Agents for Context-Aware Image Retrieval in Visual Histories

Chenlong Deng , Mengjie Deng , Junjie Wu , Dun Zeng , Teng Wang , Qingsong Xie , Jiadeng Huang , Shengjie Ma
show 5 more authors
Changwang Zhang Zhaoxiang Wang Jun Wang Yutao Zhu Zhicheng Dou
This is my paper
classification cs.CV cs.IR
keywords agentic image retrievalvisual historiescontext-aware retrievalmultimodal agentsbenchmark constructionspatiotemporal reasoningmemory graphautonomous exploration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a meaningful class of image-retrieval queries cannot be answered by matching a query to each image on its own. In a personal photo history, the evidence that identifies the right event (say, a festival logo) and the target photo (the singer on stage) can appear in separate images, so a system must explore the corpus and chain evidence across photos. To test this, the authors introduce DISBench, a 122-query benchmark built from real photo albums, and a baseline agent framework with search, metadata, and visual-verification tools plus dual memory. State-of-the-art multimodal agents reach only 28.7% exact-match accuracy and 55.0 F1, while direct embedding retrieval is near chance, and error analysis points to breakdowns in multi-step planning and state tracking rather than perception. The paper also contributes a human-model pipeline that uses vision-language models to mine associations and humans to verify them, making context-dependent queries scalable.

Core claim

The central discovery is that corpus-level contextual reasoning—not better visual features—is the missing capability in current retrieval systems. The paper formalizes retrieval as estimating P(R|Q,C) over a chronological visual history, where an image's relevance depends on other images, and builds the first benchmark that requires this. On DISBench, the best agent scores 28.7 exact-match and 55.0 F1; direct multimodal-embedding retrieval stays near chance, and the gap is structural, not a scaling issue. Test-time scaling makes best-of-k jump to 60.8 F1, showing the reasoning paths exist in current models but are poorly prioritized. Error analysis attributes most failures to planning breakd

What carries the argument

The argument runs on a heterogeneous memory graph that captures how events, photos, visual clues, and people connect. Nodes represent photos, photosets (event boundaries), visual clues (distinctive entities), and persons; edges capture containment and, crucially, cross-event associations between the same clue reappearing in different photos. A vision-language-model pipeline mines candidate associations through retrieval plus verification, humans confirm them, and a balanced random-walk sampling strategy converts local subgraphs into natural-language queries that require multi-step reasoning. On the agent side, the ImageSeeker framework supplies fine-grained tools (semantic search, metadata f

Load-bearing premise

The benchmark's validity depends on treating a user's photoset groupings as faithful ground-truth event boundaries and on the VLM-mined association graph being complete; if a photoset merges multiple real events or a VLM misses a true reappearing clue, the annotated target set is incomplete and every reported score is deflated.

What would settle it

Give any agent the hidden photoset boundaries (the event labels that the benchmark deliberately hides) while keeping everything else identical: if exact-match accuracy does not rise substantially, the paper's claim that corpus exploration is the bottleneck would be undercut; if a strong independent-matching model given those same boundaries still cannot exceed near-chance recall, the paper's claim that independent matching is structurally insufficient would be confirmed.

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

If this is right

  • If the paper is right, retrieval benchmarks must stop scoring images independently; systems must be evaluated on their ability to explore and chain evidence across a corpus.
  • The reported gap (best exact match 28.7%) means current production photo-search products, which rely on independent matching, cannot answer a class of realistic memory queries without an agentic layer.
  • The dual-query structure implies progress will come from better long-horizon state tracking and association discovery, not just stronger embeddings or larger models.
  • Test-time scaling results suggest that sampling many exploration trajectories and selecting the best is a viable short-term route to large gains on this task.
  • The human-model pipeline offers a reusable recipe for building reasoning-heavy benchmarks where pure manual annotation is infeasible.

Where Pith is reading between the lines

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

  • A natural extension not explored in the paper is applying the association-graph-plus-verification pipeline to other sequential corpora—documents, messages, or sensor logs—to benchmark context-aware retrieval beyond photo albums.
  • The benchmark's text-only queries leave open whether richer query modalities (e.g., a reference photo of the logo) would reduce or merely relocate the difficulty; the paper's own design suggests converting references to text forces exploration rather than bypassing it.
  • The 6.1% retention rate implies many synthesized queries are trivially solvable or invalid; measuring the pipeline's yield as a quality signal for automatic query generation could be a useful diagnostic that the paper does not quantify beyond the rate.
  • Because EM is much lower than F1, models tend to over- or under-retrieve; a graded or rank-aware metric might reveal whether the bottleneck is recall (missing targets) or precision (false positives), and the paper's ablations suggest metadata tools mainly help precision.

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 / 4 minor

Summary. The paper proposes DeepImageSearch, a new paradigm that reformulates image retrieval as an autonomous, multi-step exploration task over a user's visual history, and introduces DISBench, a benchmark of 122 context-dependent queries built from YFCC100M. Queries are generated by a human-model collaborative pipeline: a VLM parses images into visual clues and persons, mines cross-photo associations, builds a memory graph, samples subgraphs, and synthesizes queries; human annotators then filter, refine, and exhaustively label targets. The paper also presents ImageSeeker, a modular agent framework with retrieval, metadata, and verification tools plus dual memory, and evaluates several closed- and open-source multimodal LLMs as agents. The best agent reaches EM 28.7 / F1 55.0, while single-pass embedding retrievers perform near chance; the authors conclude that agentic reasoning is necessary for this class of retrieval.

Significance. The paper addresses a real and underexplored gap: retrieval queries whose resolution depends on evidence distributed across multiple images rather than on the target's appearance in isolation. If the benchmark is valid, it could become a useful testbed for agentic and long-horizon visual reasoning. The dataset construction is unusually transparent: pipeline stages, retention rate (6.1%), annotator agreement (IoU 0.91), worked examples, and detailed appendices are provided. The ImageSeeker framework is a reasonable, modular baseline. However, the central claim that direct retrieval is structurally insufficient is weakened by a construction circularity, and the claim that agentic planning is necessary is not tested against a non-agentic multi-stage pipeline. The evidence supports the paper as a benchmark contribution, but not yet the stronger paradigmatic conclusion.

major comments (3)
  1. [Sec. 3.4 and Sec. 5.3] The benchmark's core property—that queries cannot be solved by direct semantic matching—is imposed by the annotation protocol rather than measured. The Quality Filtering stage explicitly retains only queries 'that cannot be solved through direct semantic matching' and discards 93.9% of candidates. Section 5.3 then reports that embedding models achieve near-chance performance (Table 2). This is circular: the baseline is guaranteed to fail by construction. To support the claim that this failure is intrinsic to the paradigm, the authors should either run direct retrieval on a set of queries not filtered by this criterion (e.g., a random sample of the 2,000 generated candidates), or report a quantitative measure of visual ambiguity for retained queries and show that a strong embedding model cannot separate targets from distractors even when the correct event context is provided as a prior.
  2. [Sec. 4, Sec. 5.4, Table 3] The abstract and conclusion state that agentic reasoning is necessary, but no non-agentic multi-stage pipeline is evaluated. A deterministic system—retrieve the anchor clue with an embedding model, use metadata filtering to isolate the event, then run appearance-based matching within that event—could plausibly solve many intra-event queries. Table 3 shows that removing metadata tools causes the largest F1 drop (5.7 points), which suggests the agent-vs-embedding gap in Tables 1–2 may largely reflect the availability of metadata filtering rather than agentic planning itself. The authors should add a scripted, non-LLM pipeline that uses the same tools (ImageSearch, FilterMetadata, GetMetadata, ViewPhotos) to determine whether the observed gap persists. If it does, the necessity claim is strengthened; if not, the conclusion should be substantially softened.
  3. [Secs. 3.3–3.4] The ground-truth target sets may be incomplete. Association edges in the memory graph are produced by top-5 retrieval plus VLM verification (Sec. 3.3, Appendix C.1), and human annotators search for targets using multimodal retrieval and metadata filters (Sec. 3.4). If the VLM misses an association, the top-5 truncation drops a relevant image, or a YFCC photoset mixes multiple real events, then valid target images will be absent from the ground truth, deflating EM/F1 for every system. The reported annotator IoU of 0.91 measures agreement on the candidate sets the annotators found, not recall against the true target set. The authors should perform a recall audit on a sample of queries—e.g., have an independent annotator exhaustively search with a different toolset, or inject known target images into the corpus and check whether they are recovered—before the absolute scores in Table 1 can b
minor comments (4)
  1. [Table 1] The GPT-5.2 row appears misformatted: the model name is concatenated with the numeric results ('GPT-5.2 10.5 38.0 ...'). Please fix. Also, the note says bold and underline indicate best and second-best, but underlining is not visible in the provided text.
  2. [Sec. 5.1] NDCG@k is reported for k=1,3,5,10; NDCG@1 is unusual because it equals precision at one for binary relevance. Consider reporting Precision@1 or clarifying the metric definition.
  3. [Sec. 3.2] The statement that YFCC100M photosets 'provide ground-truth event boundaries' is a strong assumption, since photosets are user-defined upload groupings and may contain multiple distinct events. A brief validation of this assumption (e.g., a manual audit of sampled photosets) would increase confidence in the construction.
  4. [Appendix C.1] The query synthesis/verification pipeline uses Qwen3-VL and Gemini-3-Pro, and the evaluated agent models include the same model families. A short discussion of possible contamination or bias from this overlap would be appropriate, even if the pipeline operates on a different subset than the eval.

Circularity Check

1 steps flagged

Direct-retrieval failure is built into DISBench's query-selection criterion; the absolute agent scores are independent evidence, but the 'necessity of agentic reasoning' claim partially reduces to a construction artifact.

specific steps
  1. self definitional [Sec. 3.4 (Human Verification and Refinement) -> Sec. 5.3 (Limitations of Direct Retrieval)]
    "Annotators then assess difficulty to ensure that queries cannot be solved through direct semantic matching. Specifically, we require that the corpus contains visually similar distractors that are indistinguishable from targets based on appearance alone, and that resolving this ambiguity requires contextual reasoning over events or temporal information. This criterion enforces multi-step exploration rather than single-shot retrieval. Applying these strict criteria, we retain 122 queries from 2,000 candidates, yielding a retention rate of 6.1%."

    The benchmark retains only queries that by definition 'cannot be solved through direct semantic matching' and 'require contextual reasoning.' Section 5.3 then reports that embedding-based direct retrieval performs near chance (Recall@3 10-14%, NDCG@5 13-17%) and calls this 'a fundamental ceiling of the paradigm rather than a deficiency in model capacity.' The poor direct-retrieval result is therefore guaranteed by the query-inclusion criterion rather than discovered empirically. The paper uses this forced failure to conclude that 'agentic reasoning' is necessary. What is not forced is the absolute agent difficulty (best EM 28.7, F1 55.0), which is a genuine finding about strong LLM agents on this curated set.

full rationale

The main circular step is the benchmark's self-definitional selection of hard queries. DISBench is constructed so that every retained query satisfies 'cannot be solved through direct semantic matching' and requires contextual reasoning (Sec. 3.4), and then the paper presents the failure of single-pass embedding retrieval as empirical evidence for the 'fundamental ceiling' of the independent-matching paradigm (Sec. 5.3). That specific prediction is forced by construction. However, the paper does not fit any numerical parameters and does not rely on load-bearing self-citations; the VLM-mined associations are human-verified, giving the target annotations independent content. The agent results (EM 28.7, F1 55.0 best overall) are not entailed by the selection criterion and provide real, non-circular evidence that current models find the curated queries difficult. The weakness lies in the inference from 'direct matching fails on these deliberately contextual queries' to 'agentic exploration is necessary' without testing non-agentic multi-stage pipelines that combine metadata filtering with appearance matching. That is an overreach but only partially circular; hence a score of 6 rather than 8.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

This is a benchmark paper, so there are no fitted physical parameters or invented entities. The central claim rests instead on hand-set construction thresholds and domain assumptions about event boundaries, VLM association completeness, human annotation completeness, and text-only query adequacy.

free parameters (5)
  • Association-retrieval top-k = 5
    Sec. C.1: top-5 candidates retrieved per clue within and outside the source photoset; hand-chosen, controls recall of cross-event associations.
  • Subgraph sampling edge limit = 40
    Sec. C.1-C.2: sampling terminates at 40 edges; hand-chosen to balance query complexity and construction cost.
  • Face detection/clustering thresholds = conf 0.68, cosine distance 0.65, min cluster 3
    Sec. C.1: hand-set thresholds determine which recurring people enter the memory graph.
  • User selection filters = >=2000 photos, >=90% metadata coverage, >=50 photosets, >1 year account history
    Sec. C.4: hand-set inclusion criteria; shape which visual histories are represented in the benchmark.
  • Agent evaluation hyperparameters = max turns 30, context 128K tokens, ImageSearch top_k 20, ViewPhotos cap 20
    Sec. D.4: hand-set; affect measured agent performance and could move the reported gaps.
axioms (4)
  • domain assumption YFCC100M photoset boundaries correspond to real-world events and give correct ground-truth event structure.
    Sec. 3.2: 'This structure provides ground-truth event boundaries for automated query construction'; if photosets split or merge events, target annotations inherit errors.
  • domain assumption VLM-generated visual-clue associations verified by a VLM (then humans) are correct and sufficiently complete to cover all valid reasoning paths.
    Sec. 3.3/C.3: association edges are mined with top-5 retrieval and verified by Qwen3-VL; missed associations mean queries may have unlabeled targets and annotators may not find them.
  • domain assumption Human exhaustive annotation fully identifies all qualifying images for each query.
    Sec. 3.4: annotators use multimodal search/filtering to find all targets; IoU 0.91 measures inter-annotator agreement on found sets, not completeness against the full corpus.
  • domain assumption Text-only queries converted from visual references preserve enough information for the task.
    Sec. 3.1: visual reference images are converted to textual descriptions; if conversion loses identity cues, the benchmark is harder than intended or solvable only by lucky matching.

pith-pipeline@v1.3.0-alltime-deepseek · 19421 in / 12894 out tokens · 124051 ms · 2026-08-03T00:57:35.388500+00:00 · methodology

0 comments
read the original abstract

Existing multimodal retrieval systems excel at semantic matching but implicitly assume that query-image relevance can be measured in isolation. This paradigm overlooks the rich dependencies inherent in realistic visual streams, where information is distributed across temporal sequences rather than confined to single snapshots. To bridge this gap, we introduce DeepImageSearch, a novel agentic paradigm that reformulates image retrieval as an autonomous exploration task. Models must plan and perform multi-step reasoning over raw visual histories to locate targets based on implicit contextual cues. We construct DISBench, a challenging benchmark built on interconnected visual data. To address the scalability challenge of creating context-dependent queries, we propose a human-model collaborative pipeline that employs vision-language models to mine latent spatiotemporal associations, effectively offloading intensive context discovery before human verification. Furthermore, we build a robust baseline using a modular agent framework equipped with fine-grained tools and a dual-memory system for long-horizon navigation. Extensive experiments demonstrate that DISBench poses significant challenges to state-of-the-art models, highlighting the necessity of incorporating agentic reasoning into next-generation retrieval systems.

Figures

Figures reproduced from arXiv: 2602.10809 by Changwang Zhang, Chenlong Deng, Dun Zeng, Jiadeng Huang, Junjie Wu, Jun Wang, Mengjie Deng, Qingsong Xie, Shengjie Ma, Teng Wang, Yutao Zhu, Zhaoxiang Wang, Zhicheng Dou.

Figure 1
Figure 1. Figure 1: Evolution of image retrieval paradigms. (a) Direct retrieval matches queries to images through visual semantic align￾ment. (b) Reasoning-intensive retrieval requires inference over external knowledge, but still evaluates each image independently. (c) DeepImageSearch demands corpus context awareness, where models must first locate target events within the visual history and then identify qualifying images t… view at source ↗
Figure 2
Figure 2. Figure 2: Two query types in DISBench. (a) Intra-Event queries locate a specific event and filter targets within it. (b) Inter-Event queries scan across events to verify recurring elements under temporal/spatial constraints. structured associations within the data. DeepImageSearch addresses this limitation by requiring corpus-level contex￾tual reasoning over visual histories. 2.2. Benchmarking Multimodal Agents Mult… view at source ↗
Figure 3
Figure 3. Figure 3: Semi-automated data construction pipeline. Starting from raw images, we first parse visual content to extract salient clues and person attributes, then mine latent associations across the corpus through retrieval and verification strategy. These elements are organized into a memory graph, from which we sample subgraphs via random walks to synthesize candidate queries for human verification. ically, we requ… view at source ↗
Figure 4
Figure 4. Figure 4: Dataset statistics of DISBench. (a) Query type distri￾bution shows a balanced split between intra-event and inter-event queries. (b) Target images span diverse themes including portraits, nature views, daily items, and scenic spots. 2025b), and Claude-Opus-4.5 (Anthropic, 2025a), as well as open-source models including Qwen3-VL-235B-A22B￾Thinking, Qwen3-VL-235B-A22B-Instruct, Qwen3-VL￾32B (Bai et al., 2025… view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of error categories across four representative models on DISBench [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Geographic distribution of geotagged photos in DISBench. Each point represents a photo location, covering all continents except Antarctica. To ensure data quality and diversity, we apply a two￾stage user selection pipeline. In the first stage, we filter YFCC100M users based on objective criteria. Each user must have at least 2,000 photos with a metadata coverage rate of at least 90%, a minimum of 50 photos… 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

39 extracted references · 12 linked inside Pith

  1. [1]

    Introducing claude opus 4.5, 2025 a

    Anthropic. Introducing claude opus 4.5, 2025 a . URL https://www.anthropic.com/news/claude-opus-4-5

    2025

  2. [2]

    Introducing claude sonnet 4.5, 2025 b

    Anthropic. Introducing claude sonnet 4.5, 2025 b . URL https://www.anthropic.com/news/claude-sonnet-4-5

    2025

  3. [3]

    Qwen3-vl technical report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.21631 2025

  4. [4]

    Seed1.6-embedding

    ByteDance Seed . Seed1.6-embedding. https://seed1-6-embedding.github.io/, June 2025. Model ID: doubao-embedding-vision-250615. Accessed: 2026-01-29

    2025

  5. [5]

    Moca: Modality-aware continual pre-training makes better bidirectional multimodal embeddings

    Chen, H., Liu, H., Luo, Y., Wang, L., Yang, N., Wei, F., and Dou, Z. Moca: Modality-aware continual pre-training makes better bidirectional multimodal embeddings. CoRR, abs/2506.23115, 2025 a . doi:10.48550/ARXIV.2506.23115

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.23115 2025

  6. [6]

    mme5: Improving multimodal multilingual embeddings via high-quality synthetic data

    Chen, H., Wang, L., Yang, N., Zhu, Y., Zhao, Z., Wei, F., and Dou, Z. mme5: Improving multimodal multilingual embeddings via high-quality synthetic data. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.), Findings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1, 2025 , volume ACL 2025 of Findi...

    2025

  7. [7]

    Generative thinking, corrective action: User-friendly composed image retrieval via automatic multi-agent collaboration

    Cheng, Z., Ma, Y., Lang, J., Zhang, K., Zhong, T., Wang, Y., and Zhou, F. Generative thinking, corrective action: User-friendly composed image retrieval via automatic multi-agent collaboration. In Antonie, L., Pei, J., Yu, X., Chierichetti, F., Lauw, H. W., Sun, Y., and Parthasarathy, S. (eds.), Proceedings of the 31st ACM SIGKDD Conference on Knowledge D...

    arXiv 2025

  8. [8]

    Embodiedeval: Evaluate multimodal llms as embodied agents

    Cheng, Z., Tu, Y., Li, R., Dai, S., Hu, J., Hu, S., Li, J., Shi, Y., Yu, T., Chen, W., Shi, L., and Sun, M. Embodiedeval: Evaluate multimodal llms as embodied agents. CoRR, abs/2501.11858, 2025 b . doi:10.48550/ARXIV.2501.11858

    Pith/arXiv arXiv doi:10.48550/arxiv.2501.11858 2025

  9. [9]

    N., Awasthi, A., Pan, X., Ahuja, C., Mishra, S

    Cui, X., Cheng, J., Chen, H., Shukla, S. N., Awasthi, A., Pan, X., Ahuja, C., Mishra, S. K., Yang, Y., Xiao, J., Guo, Q., Lim, S., Singh, A., and Fan, X. Think then embed: Generative context improves multimodal embedding. CoRR, abs/2510.05014, 2025. doi:10.48550/ARXIV.2510.05014

    doi:10.48550/arxiv.2510.05014 2025

  10. [10]

    Mind2web: Towards a generalist agent for the web

    Deng, X., Gu, Y., Zheng, B., Chen, S., Stevens, S., Wang, B., Sun, H., and Su, Y. Mind2web: Towards a generalist agent for the web. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.), Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orl...

    2023

  11. [11]

    Colpali: Efficient document retrieval with vision language models

    Faysse, M., Sibille, H., Wu, T., Omrani, B., Viaud, G., Hudelot, C., and Colombo, P. Colpali: Efficient document retrieval with vision language models. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net, 2025

    2025

  12. [12]

    A new era of intelligence with Gemini 3

    Google . A new era of intelligence with Gemini 3. https://blog.google/products-and-platforms/products/gemini/gemini-3/, November 2025. Accessed: 2026-01-29

    2025

  13. [13]

    J., Shu, Y., Song, C

    Gou, B., Huang, Z., Ning, Y., Gu, Y., Lin, M., Qi, W., Kopanev, A., Yu, B., Guti \' e rrez, B. J., Shu, Y., Song, C. H., Wu, J., Chen, S., Moussa, H. N., Zhang, T., Xie, J., Li, Y., Xue, T., Liao, Z., Zhang, K., Zheng, B., Cai, Z., Rozgic, V., Ziyadi, M., Sun, H., and Su, Y. Mind2web 2: Evaluating agentic search with agent-as-a-judge. CoRR, abs/2506.21506...

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.21506 2025

  14. [14]

    Hurst, A., Lerer, A., Goucher, A. P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., Madry, A., Baker - Whitcomb, A., Beutel, A., Borzunov, A., Carney, A., Chow, A., Kirillov, A., Nichol, A., Paino, A., Renzin, A., Passos, A. T., Kirillov, A., Christakis, A., Conneau, A., Kamali, A., Jabri, A., Moyer, A., Tam, A., ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.21276 2024

  15. [15]

    V., Sung, Y., Li, Z., and Duerig, T

    Jia, C., Yang, Y., Xia, Y., Chen, Y., Parekh, Z., Pham, H., Le, Q. V., Sung, Y., Li, Z., and Duerig, T. Scaling up visual and vision-language representation learning with noisy text supervision. In Meila, M. and Zhang, T. (eds.), Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event , volume 139 of...

    2021

  16. [16]

    Vlm2vec: Training vision-language models for massive multimodal embedding tasks

    Jiang, Z., Meng, R., Yang, X., Yavuz, S., Zhou, Y., and Chen, W. Vlm2vec: Training vision-language models for massive multimodal embedding tasks. In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net, 2025

    2025

  17. [17]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp.\ 1207--1216, Stanford, CA, 2000. Morgan Kaufmann

    2000

  18. [18]

    Li, J., Li, D., Xiong, C., and Hoi, S. C. H. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In Chaudhuri, K., Jegelka, S., Song, L., Szepesv \' a ri, C., Niu, G., and Sabato, S. (eds.), International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA , volume 162 ...

    2022

  19. [19]

    Qwen3-vl-embedding and qwen3-vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking

    Li, M., Zhang, Y., Long, D., Chen, K., Song, S., Bai, S., Yang, Z., Xie, P., Yang, A., Liu, D., et al. Qwen3-vl-embedding and qwen3-vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking. arXiv preprint arXiv:2601.04720, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Mm-browsecomp: A comprehensive benchmark for multimodal browsing agents

    Li, S., Bu, X., Wang, W., Liu, J., Dong, J., He, H., Lu, H., Zhang, H., Jing, C., Li, Z., Li, C., Tian, J., Zhang, C., Peng, T., He, Y., Gu, J., Zhang, Y., Yang, J., Zhang, G., Huang, W., Zhou, W., Zhang, Z., Ding, R., and Wen, S. Mm-browsecomp: A comprehensive benchmark for multimodal browsing agents. CoRR, abs/2508.13186, 2025. doi:10.48550/ARXIV.2508.13186

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.13186 2025

  21. [21]

    Mm-embed: Universal multimodal retrieval with multimodal LLMS

    Lin, S., Lee, C., Shoeybi, M., Lin, J., Catanzaro, B., and Ping, W. Mm-embed: Universal multimodal retrieval with multimodal LLMS . In The Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net, 2025

    2025

  22. [22]

    Vlm2vec-v2: Advancing multimodal embedding for videos, images, and visual documents

    Meng, R., Jiang, Z., Liu, Y., Su, M., Yang, X., Fu, Y., Qin, C., Chen, Z., Xu, R., Xiong, C., Zhou, Y., Chen, W., and Yavuz, S. Vlm2vec-v2: Advancing multimodal embedding for videos, images, and visual documents. CoRR, abs/2507.04590, 2025. doi:10.48550/ARXIV.2507.04590

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.04590 2025

  23. [23]

    Introducing GPT-5.2 : The most advanced frontier model for professional work and longrunning agents

    OpenAI . Introducing GPT-5.2 : The most advanced frontier model for professional work and longrunning agents. https://openai.com/index/introducing-gpt-5-2/, 2025. Accessed: 2026-01-29

    2025

  24. [24]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In Meila, M. and Zhang, T. (eds.), Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Vi...

    2021

  25. [25]

    Glm-4.5v and glm-4.1v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning, 2025

    Team, V., Hong, W., Yu, W., Gu, X., Wang, G., Gan, G., Tang, H., Cheng, J., Qi, J., Ji, J., Pan, L., Duan, S., Wang, W., Wang, Y., Cheng, Y., He, Z., Su, Z., Yang, Z., Pan, Z., Zeng, A., Wang, B., Chen, B., Shi, B., Pang, C., Zhang, C., Yin, D., Yang, F., Chen, G., Xu, J., Zhu, J., Chen, J., Chen, J., Chen, J., Lin, J., Wang, J., Chen, J., Lei, L., Gong, ...

    2025

  26. [26]

    A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., and Li, L

    Thomee, B., Shamma, D. A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., and Li, L. YFCC100M: the new data in multimedia research. Commun. ACM , 59 0 (2): 0 64--73, 2016. doi:10.1145/2812802

    doi:10.1145/2812802 2016

  27. [27]

    Multimodal reasoning agent for zero-shot composed image retrieval

    Tu, R., Sun, W., You, H., Wang, Y., Huang, J., Shen, L., and Tao, D. Multimodal reasoning agent for zero-shot composed image retrieval. CoRR, abs/2505.19952, 2025. doi:10.48550/ARXIV.2505.19952

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.19952 2025

  28. [28]

    Uniir: Training and benchmarking universal multimodal information retrievers

    Wei, C., Chen, Y., Chen, H., Hu, H., Zhang, G., Fu, J., Ritter, A., and Chen, W. Uniir: Training and benchmarking universal multimodal information retrievers. In Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., and Varol, G. (eds.), Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceed...

    doi:10.1007/978-3-031-73021-4 2024

  29. [29]

    J., Cheng, Z., Shin, D., Lei, F., Liu, Y., Xu, Y., Zhou, S., Savarese, S., Xiong, C., Zhong, V., and Yu, T

    Xie, T., Zhang, D., Chen, J., Li, X., Zhao, S., Cao, R., Hua, T. J., Cheng, Z., Shin, D., Lei, F., Liu, Y., Xu, Y., Zhou, S., Savarese, S., Xiong, C., Zhong, V., and Yu, T. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. In Globersons, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J. M., and Zhang,...

    2024

  30. [30]

    A survey on agentic multimodal large language models

    Yao, H., Zhang, R., Huang, J., Zhang, J., Wang, Y., Fang, B., Zhu, R., Jing, Y., Liu, S., Li, G., and Tao, D. A survey on agentic multimodal large language models. CoRR, abs/2510.10991, 2025. doi:10.48550/ARXIV.2510.10991

    doi:10.48550/arxiv.2510.10991 2025

  31. [31]

    A survey on multimodal large language models

    Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., and Chen, E. A survey on multimodal large language models. CoRR, abs/2306.13549, 2023. doi:10.48550/ARXIV.2306.13549

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.13549 2023

  32. [32]

    Sigmoid loss for language image pre-training

    Zhai, X., Mustafa, B., Kolesnikov, A., and Beyer, L. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 , pp.\ 11941--11952. IEEE , 2023. doi:10.1109/ICCV51070.2023.01100

    arXiv 2023

  33. [33]

    Magiclens: Self-supervised image retrieval with open-ended instructions

    Zhang, K., Luan, Y., Hu, H., Lee, K., Qiao, S., Chen, W., Su, Y., and Chang, M. Magiclens: Self-supervised image retrieval with open-ended instructions. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net, 2024 a

    2024

  34. [34]

    GME: improving universal multimodal retrieval by multimodal llms

    Zhang, X., Zhang, Y., Xie, W., Li, M., Dai, Z., Long, D., Xie, P., Zhang, M., Li, W., and Zhang, M. GME: improving universal multimodal retrieval by multimodal llms. CoRR, abs/2412.16855, 2024 b . doi:10.48550/ARXIV.2412.16855

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.16855 2024

  35. [35]

    J., Yan, R., Yao, Y., and Wang, L

    Zheng, X., Li, L., Yang, Z., Yu, P., Wang, A. J., Yan, R., Yao, Y., and Wang, L. V-mage: A game evaluation framework for assessing vision-centric capabilities in multimodal large language models. arXiv preprint arXiv:2504.06148, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    J., and Lian, D

    Zhou, J., Xiong, Y., Liu, Z., Liu, Z., Xiao, S., Wang, Y., Zhao, B., Zhang, C. J., and Lian, D. Megapairs: Massive data synthesis for universal multimodal retrieval. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T. (eds.), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2025, Vi...

    2025

  37. [37]

    F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Ou, T., Bisk, Y., Fried, D., Alon, U., and Neubig, G

    Zhou, S., Xu, F. F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Ou, T., Bisk, Y., Fried, D., Alon, U., and Neubig, G. Webarena: A realistic web environment for building autonomous agents. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024

    2024

  38. [38]

    Large language models for information retrieval: A survey

    Zhu, Y., Yuan, H., Wang, S., Liu, J., Liu, W., Deng, C., Dou, Z., and Wen, J. Large language models for information retrieval: A survey. CoRR, abs/2308.07107, 2023. doi:10.48550/ARXIV.2308.07107

    doi:10.48550/arxiv.2308.07107 2023

  39. [39]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...