Computation and Language · arXiv:2610.29249 v1 · submitted 21 Oct 2026 · 92 references
MOSAIC: An Omni-Modal Benchmark for Long-Term Memory and Cross-Modal Reasoning in Conversational Agents
The paper says
“Omni-modal assistants must remember not only what users say, but also what they show and convey.”
more of the abstract
Existing long-term memory benchmarks are mostly text-and-image or report a single end-to-end score, so they cannot locate whether evidence is lost when it is represented, retrieved, delivered, or read. We introduce MOSAIC, a benchmark of 923 questions over persona-grounded histories of 11-24 sessions spanning text, image, video, and audio. MOSAIC constructs targeted spatial, temporal, and prosodic signals rather than decorative media, with the strongest causal validation for spatial information and selector-level validation for annotated prosody, and pairs them with a matched diagnostic protocol that freezes the question, reader, prompt, and budget while intervening on representation, retrieval, and delivery. Matched decomposition shows that representation loss is modality-dependent, residual reading dominates even under gold evidence, and abstention remains weak on these controls: strong answerable scores coexist with failures to refuse. MOSAIC thus turns omni-modal memory evaluation from a leaderboard into a diagnosis of write, retrieve, deliver, and read failures.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 22 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. MOSAIC is a well-executed omni-modal memory benchmark with a clever diagnostic protocol; the stage associations hold up, but video and audio strata are thin. This line is wrong
6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 92 references say.
6 minor objections
The paper says, at §3.2, Table 2
Pith says: minor objection 1 of 6
The counts for 'Audio Prosody' (450 facts, 399 items) overlap with other rows, making the per-stratum total >784. The paper notes this, but a clarifying sentence at the end of §3.2 would help readers avoid misinterpretation when scanning the table.
The paper says, at §5.1, first paragraph
Pith says: minor objection 2 of 6
The frozen reader is Gemini 3.5. The motivation (held out from dataset construction, MiMo-V2.5 used as probe) is clear, but it would be helpful to explicitly state whether the same reader is used across all seven matched conditions in Table 3; this is implied but could be stated directly.
show 4 more
The paper says, at Appendix K.4, Table 27
Pith says: minor objection 3 of 6
The 'N' column mixes unique QIDs and intervention cases. For example, 'Raw−ASR (prosody)' has N=450 (intervention cases) while 'Correct prosody−swapped' has N=399 (unique QIDs). The caption explains this in text, but a column header or footnote would improve clarity.
The paper says, at §6.2, Figure 5
Pith says: minor objection 4 of 6
The x-axis label 'Paired effect (95% CI)' is clear, but the unit (judge-score points) is stated only in the caption. Adding the unit to the axis label would make the figure more self-contained.
The paper says, at Appendix B.3
Pith says: minor objection 5 of 6
The prompt templates are described in prose. Including the full prompts (or pointers to the release) is sufficient, but a note that they are available in the code repository would help reproducibility.
The paper says, at Table 28
Pith says: minor objection 6 of 6
The table includes latency in seconds. The abbreviation 'Lat.' is not expanded in the caption or table notes; expanding it in the caption would help.
What would settle it
Replicate the full matched stage ladder (Table 3) using a different omni-modal reader (e.g., GPT-4o) on the 784 answerable QIDs under identical prompts and judges. If the rank ordering of stage losses and the qualitative finding that read dominates remain unchanged, the protocol's robustness is confirmed. If the read residual shrinks significantly or write costs shift, the stage attribution may be reader-dependent.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 22 Oct 2026
- Reader holds up
- Skeptic agrees
- Editor wrote the verdict above
- Referee accept: 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Here's my read on MOSAIC. It's a solid piece of benchmark work. The main thing you should know: they've built the first multi-session omni-modal (text+image+video+audio) memory benchmark, and the matched diagnostic protocol that separates write, retrieve, deliver, and read is genuinely new. It's not just another leaderboard; the stage-level analysis lets you pinpoint where a system loses evidence. The generation pipeline is careful—they do chain deletion tests, human audits, and transparently report the filtering lineage. The headline results (reading dominates residual error, retrieval is the next bottleneck, write-time textualization costs 9 points on spatial-heavy items, abstention is weak) are consistent across analyses.
Credit where due: the fixed-reader design and within-item modality interventions are well thought out. They go beyond most benchmarks by validating that their gaps are actually causal—swapped prosody tags shift scores, spatial descriptions recover some of the pixel gap. The limitations section is unusually honest: they call out stage labels as 'protocol associations,' video nulls are underpowered, audio prosody contrasts are inconclusive, refusal controls have form cues. That level of candor makes the paper trustworthy.
Soft spots. The main one is that video and audio strata are thin for the claims they want to make. The video perception cost (17 points) is striking but relies on 56–64 items per challenge; the audio nulls under gold delivery are inconclusive and the authors say so. The gold-versus-retrieved contrast is not a pure isolation of retrieval—distractors change. But they don't oversell it. The synthetic data is a limitation but not a fatal one for a diagnostic tool.
Who should read this: anyone working on multimodal conversational agents, memory systems, or RAG evaluation. It deserves peer review. I'd bring it to reading group and would cite it as a reference point for omni-modal memory evaluation.
Recommendation: the paper is ready for a serious venue. Minor expansion of video and audio strata would strengthen future work, but the current submission is sound enough to proceed.
The strongest objection
The paper's central claim—that MOSAIC's matched diagnostic protocol can separate omni-modal memory failures into WRITE, RETRIEVE, DELIVER, and READ stages—is well supported by the data and transparently qualified. The authors explicitly state that stage labels are 'protocol associations rather than fully isolated causal mechanisms,' acknowledge underpowered video strata, inconclusive audio contrasts, and residual question-form cues in refusals. The fixed-reader design and matched pair comparisons are internally consistent, and the main findings (read dominance, modality-dependent write costs, weak abstention) hold across multiple analyses. No additional load-bearing concern beyond those already recognized by the authors and the reader is apparent.
The reader's weakest_assumption highlights that stage drops are protocol associations rather than fully isolated causal mechanisms, and that video/audio strata are limited. The authors fully acknowledge both points (§8, Appendix F, Appendix G). My own reading finds no additional unresolved weakness: the matched protocol is carefully designed, the within-item modality interventions strengthen causal claims for spatial and prosody, and the use of a second reader (MiMo-V2.5) partially replicates the stage ladder. The concrete_test would further verify robustness to reader choice. The verdict should remain UNCHANGED as originally, because the central claim is well-evidenced and limitations are candidly disclosed.
The referee's report
If the results hold (and the evidence is strong), MOSAIC provides the first benchmark that separates omni-modal memory failures into actionable stages with a controlled diagnostic protocol. This moves evaluation from a single end-to-end leaderboard to a failure-mode diagnosis that can guide improvement in representation, retrieval, packing, and reading. The paper ships machine-checkable provenance, human audits, and fully specified run configurations. The matched-pair design and within-item modality interventions are methodologically rigorous. The benchmark's coverage of four modalities and cross-modal chaining fills a clear gap in prior work. The conclusions are appropriately qualified (protocol associations, underpowered strata, inconclusive contrasts). The release includes 10,765 multimodal artefacts and 139 refusal controls.
Pith's reading of the paper (a machine's summary, not the authors')
The paper introduces MOSAIC, a 923-question benchmark that exposes how conversational assistants lose information across text, image, video, and audio over multiple sessions. Rather than reporting a single end-to-end score, MOSAIC uses a matched diagnostic protocol to attribute failures to one of four stages: WRITE (representation compression), RETRIEVE (missing facts), DELIVER (dropped under media caps), and READ (integration error). The central finding is that reading dominates the residual error even when perfect evidence is provided, while visual perception, especially spatial configuration, suffers heavily from textualization. Audio and temporal signals show weaker or inconclusive sensitivity in contrastive tests, and all tested systems struggle to abstain reliably when evidence is absent.
What carries the argument. The matched four-stage protocol (WRITE, RETRIEVE, DELIVER, READ) with within-item modality interventions: each stage is made separately substitutable under a frozen reader, prompt, and budget, so that paired contrasts isolate which stage drops the score. The seven modality challenges (image: background, fine attribute, spatial; video: ordering, repetition, transition; audio: prosody) are construction-targeted to require native signals that captions or transcripts cannot preserve.
Core claim. Under a fixed reader, prompt, packer, and judge, the WRITE, RETRIEVE, DELIVER, and READ stages each contribute distinct, modality-dependent losses to omni-modal memory question answering. Spatial structure in images is the clearest example of a signal that no text surrogate adequately preserves, while video temporal contrasts remain underpowered and audio prosody shows selector-level sensitivity only when gold tags are swapped. The largest single associated loss (∼15 points) remains under gold symbolic facts, indicating that residual reading and grading error is a binding constraint.
The premise it rests on. Stage labels are protocol associations rather than fully isolated causal mechanisms because gold versus retrieved contexts differ in distractors and composition, and video strata (n=56–64) are underpowered to detect temporal compression effects.
If the paper is right (Pith's inferences, not the authors')
Representation compression (WRITE) costs 6–9 points overall, but for spatial-heavy items the loss can exceed 30 points, suggesting that current captioning and structured-text writers are inadequate for preserving relational visual information.
Retrieval costs ∼8 points in native pipelines; delivery under media caps costs another ∼4 points, but tight media caps (one per modality) collapse complete-chain delivery to 0.18, meaning that most multi-hop chains are broken before the reader sees them.
Reading is the largest residual loss even when gold symbolic facts are provided (∼15 points), so progress in answer integration and judge calibration is necessary alongside retrieval improvements.
Abstention is weak: no system in the evaluation both answers well and refuses reliably; strong answerable scores coexist with failures to decline unanswerable questions.
The structured writer recovers only a non-significant 1.4 points over generic captions/ASR, indicating that avoidable compression loss is small and that most textualization loss is inherent to captioning.
The 15.4-point residual under gold symbolic facts may include judge and output-format errors that are not true reading failures; separating those would refine the diagnostic power of the protocol.
Video perception costs are large (17.1 points) but measured on small strata (n=56–64); replicating with larger samples could determine whether temporal compression is truly lossy or whether the current reader is simply poor at video.
Audio prosody contrasts were near-null under gold delivery for this reader, but swapped gold tags showed a 4.6-point effect, suggesting that the reader can use symbolic prosody when provided; ASR losing prosody is the real bottleneck, not raw-audio perception.
The observation that no system occupies the upper-right region of the answer–abstention plot implies that calibration is currently dominated by a refusal heuristic (text-derived systems refuse too often, native systems too seldom), pointing toward meta-reasoning or uncertainty-aware retrieval as a required component.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Omni-modal assistants must remember not only what users say, but also what they show and convey. Existing long-term memory benchmarks are mostly text-and-image or report a single end-to-end score, so they cannot locate whether evidence is lost when it is represented, retrieved, delivered, or read. We introduce MOSAIC, a benchmark of 923 questions over persona-grounded histories of 11-24 sessions spanning text, image, video, and audio. MOSAIC constructs targeted spatial, temporal, and prosodic signals rather than decorative media, with the strongest causal validation for spatial information and selector-level validation for annotated prosody, and pairs them with a matched diagnostic protocol that freezes the question, reader, prompt, and budget while intervening on representation, retrieval, and delivery. Matched decomposition shows that representation loss is modality-dependent, residual reading dominates even under gold evidence, and abstention remains weak on these controls: strong answerable scores coexist with failures to refuse. MOSAIC thus turns omni-modal memory evaluation from a leaderboard into a diagnosis of write, retrieve, deliver, and read failures.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-22
The authors' main claim, as the classifier picked it out 0.63
“MOSAIC thus turns omni-modal memory evaluation from a leaderboard into a diagnosis of write, retrieve, deliver, and read failures.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.16 · states limitation 0.69 · human participants 0.28 · controlled comparison 0.57 · randomized assignment 0.03 · states quantitative result 0.91 · compares with prior results 0.30 · mentions code or data availability 0.05
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Omni-modal assistants must remember not only what users say, but also what they show and convey. | background 1.00 | 0.05 | 0.03 | asserted 1.00 |
| Existing long-term memory benchmarks are mostly text-and-image or report a single end-to-end score, so they cannot locate whether evidence is lost when it is represented, retrieved, delivered, or read. | gap 0.98 | 0.05 | 0.03 | asserted 0.99 |
| We introduce MOSAIC, a benchmark of 923 questions over persona-grounded histories of 11-24 sessions spanning text, image, video, and audio. | contribution 0.79 | 0.95 | 0.04 | asserted 1.00 |
| MOSAIC constructs targeted spatial, temporal, and prosodic signals rather than decorative media, with the strongest causal validation for spatial information and selector-level validation for annotated prosody, and pairs them with a matched diagnostic protocol that freezes the question, reader, prompt, and budget while intervening on representation, retrieval, and delivery. | method 0.96 | 0.96 | 0.07 | asserted 1.00 |
| Matched decomposition shows that representation loss is modality-dependent, residual reading dominates even under gold evidence, and abstention remains weak on these controls: strong answerable scores coexist with failures to refuse. | result 0.98 | 0.95 | 0.39 | asserted 1.00 |
| MOSAIC thus turns omni-modal memory evaluation from a leaderboard into a diagnosis of write, retrieve, deliver, and read failures. | interpretation 0.95 | 0.75 | 0.85 | asserted 1.00 |
What kind of paper the classifier reports this is: dataset or benchmark 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-22: 1 request, 5,066 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (12)
Showing 11 of the paper's 12 figures. Figures are the authors', reproduced from the paper.
References (92)
- [1] Zhang, Qizheng and Hu, Changran and Upasani, Shubhangi and Ma, Boyuan and Hong, Fenglu and Kamanuru, Vamsidhar and Rainton, Jay and Wu, Chen and Ji, Mengmeng and Li, Hanchen and Thakker, Urmish and Zou, James and Olukotun, Kunle , month = oct, year =. Agentic. doi:10.48550/arX... on PithPith's reviewarXivdoi:10.48550/arxiv.2510.04618
- [2] Rasmussen, Preston and Paliychuk, Pavlo and Beauvais, Travis and Ryan, Jack and Chalef, Daniel , month = jan, year =. Zep:. doi:10.48550/arXiv.2501.13956 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2501.13956
- [3] Wang, Yu and Chen, Xi , month = jul, year =. doi:10.48550/arXiv.2507.07957 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2507.07957
- [4] Anokhin, Petr and Semenov, Nikita and Sorokin, Artyom and Evseev, Dmitry and Kravchenko, Andrey and Burtsev, Mikhail and Burnaev, Evgeny , month = may, year =. doi:10.48550/arXiv.2407.04363 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2407.04363
- [5] Fang, Jizhan and Deng, Xinle and Xu, Haoming and Jiang, Ziyan and Tang, Yuqi and Xu, Ziwen and Deng, Shumin and Yao, Yunzhi and Wang, Mengru and Qiao, Shuofei and Chen, Huajun and Zhang, Ningyu , month = oct, year =. doi:10.48550/arXiv.2510.18866 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2510.18866
- [6] Yu, Hongli and Chen, Tinghong and Feng, Jiangtao and Chen, Jiangjie and Dai, Weinan and Yu, Qiying and Zhang, Ya-Qin and Ma, Wei-Ying and Liu, Jingjing and Wang, Mingxuan and Zhou, Hao , month = jul, year =. doi:10.48550/arXiv.2507.02259 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2507.02259
- [7] Chhikara, Prateek and Khant, Dev and Aryan, Saket and Singh, Taranjeet and Yadav, Deshraj , month = apr, year =. Mem0:. doi:10.48550/arXiv.2504.19413 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2504.19413
- [8] Wu, Di and Wang, Hongwei and Yu, Wenhao and Zhang, Yuwei and Chang, Kai-Wei and Yu, Dong , month = mar, year =. doi:10.48550/arXiv.2410.10813 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2410.10813
- [9] Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng , month = oct, year =. A-. doi:10.48550/arXiv.2502.12110 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2502.12110
- [10] Choubey, Prafulla Kumar and Su, Xin and Luo, Man and Peng, Xiangyu and Xiong, Caiming and Le, Tiep and Rosenman, Shachar and Lal, Vasudev and Mui, Phil and Ho, Ricky and Howard, Phillip and Wu, Chien-Sheng , month = oct, year =. Distill-. doi:10.48550/arXiv.2410.16597 , abstract = on Pithdoi:10.48550/arxiv.2410.16597
- [11] Luo, Haoran and E, Haihong and Chen, Guanting and Zheng, Yandan and Wu, Xiaobao and Guo, Yikai and Lin, Qika and Feng, Yu and Kuang, Zemin and Song, Meina and Zhu, Yifan and Tuan, Luu Anh , month = oct, year =. doi:10.48550/arXiv.2503.21322 , abstract = on Pithdoi:10.48550/arxiv.2503.21322
- [12] Salama, Rana and Cai, Jason and Yuan, Michelle and Currey, Anna and Sunkara, Monica and Zhang, Yi and Benajiba, Yassine , month = jul, year =. doi:10.48550/arXiv.2503.21760 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2503.21760
- [13] Zhang, Wentao and Kim, Woojeong and Deng, Yuntian , month = sep, year =. From. doi:10.48550/arXiv.2505.23765 , abstract = on Pithdoi:10.48550/arxiv.2505.23765
- [14] Hu, Yuyang and Liu, Shichun and Yue, Yanwei and Zhang, Guibin and Liu, Boyang and Zhu, Fangyi and Lin, Jiahang and Guo, Honglin and Dou, Shihan and Xi, Zhiheng and Jin, Senjie and Tan, Jiejun and Yin, Yanbin and Liu, Jiongnan and Zhang, Zeyu and Sun, Zhongxiang and Zhu, Yutao ... on PithPith's reviewarXivdoi:10.48550/arxiv.2512.13564
- [15] Long, Lin and He, Yichen and Ye, Wentao and Pan, Yiyuan and Lin, Yuan and Li, Hang and Zhao, Junbo and Li, Wei , month = oct, year =. Seeing,. doi:10.48550/arXiv.2508.09736 , abstract = on Pithdoi:10.48550/arxiv.2508.09736
- [16] Guo, Zirui and Ren, Xubin and Xu, Lingrui and Zhang, Jiahao and Huang, Chao , month = oct, year =. doi:10.48550/arXiv.2510.12323 , abstract = on Pithdoi:10.48550/arxiv.2510.12323
- [17] Wang, Piaohong and Tian, Motong and Li, Jiaxian and Liang, Yuan and Wang, Yuqing and Chen, Qianben and Wang, Tiannan and Lu, Zhicong and Ma, Jiawei and Jiang, Yuchen Eleanor and Zhou, Wangchunshu , month = dec, year =. O-. doi:10.48550/arXiv.2511.13593 , abstract = on Pithdoi:10.48550/arxiv.2511.13593
- [18] Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , month = feb, year =. doi:10.48550/arXiv.2310.08560 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2310.08560
- [19] Li, Zhiyu and Song, Shichao and Wang, Hanyu and Niu, Simin and Chen, Ding and Yang, Jiawei and Xi, Chenyang and Lai, Huayi and Zhao, Jihao and Wang, Yezhaohui and Ren, Junpeng and Lin, Zehao and Huo, Jiahao and Chen, Tianyi and Chen, Kai and Li, Kehang and Yin, Zhiqiang and Yu... on PithPith's reviewarXivdoi:10.48550/arxiv.2505.22101
- [20] Tan, Zhen and Yan, Jun and Hsu, I-Hung and Han, Rujun and Wang, Zifeng and Le, Long and Song, Yiwen and Chen, Yanfei and Palangi, Hamid and Lee, George and Iyer, Anand Rajan and Chen, Tianlong and Liu, Huan and Lee, Chen-Yu and Pfister, Tomas , editor =. In. Proceedings of the... on Pithdoi:10.18653/v1/2025.acl-long.413
show 72 more
- [21] Fang, Runnan and Cai, Shihao and Li, Baixuan and Wu, Jialong and Li, Guangyu and Yin, Wenbiao and Wang, Xinyu and Wang, Xiaobin and Su, Liangcai and Zhang, Zhen and Wu, Shibin and Tao, Zhengwei and Jiang, Yong and Xie, Pengjun and Huang, Fei and Zhou, Jingren , month = sep, ye... on Pithdoi:10.48550/arxiv.2509.13311
- [22] Wang, Zhaowei and Yu, Wenhao and Ren, Xiyu and Zhang, Jipeng and Zhao, Yu and Saxena, Rohit and Cheng, Liang and Wong, Ginny and See, Simon and Minervini, Pasquale and Song, Yangqiu and Steedman, Mark , month = oct, year =. doi:10.48550/arXiv.2505.10610 , abstract = on Pithdoi:10.48550/arxiv.2505.10610
- [23] Zhou, Chulun and Zhang, Chunkang and Yu, Guoxin and Meng, Fandong and Zhou, Jie and Lam, Wai and Yu, Mo , month = dec, year =. Improving. doi:10.48550/arXiv.2512.23959 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2512.23959
- [24] Sumers, Theodore R. and Yao, Shunyu and Narasimhan, Karthik and Griffiths, Thomas L. , month = mar, year =. Cognitive. doi:10.48550/arXiv.2309.02427 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2309.02427
- [25] Cheng, Xin and Zeng, Wangding and Dai, Damai and Chen, Qinyu and Wang, Bingxuan and Xie, Zhenda and Huang, Kezhao and Yu, Xingkai and Hao, Zhewen and Li, Yukun and Zhang, Han and Zhang, Huishuai and Zhao, Dongyan and Liang, Wenfeng , month = jan, year =. Conditional. doi:10.48... on PithPith's reviewarXivdoi:10.48550/arxiv.2601.07372
- [26] Behrouz, Ali and Zhong, Peilin and Mirrokni, Vahab , month = dec, year =. Titans:. doi:10.48550/arXiv.2501.00663 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2501.00663
- [27] Geva, Mor and Schuster, Roei and Berant, Jonathan and Levy, Omer , month = sep, year =. Transformer. doi:10.48550/arXiv.2012.14913 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2012.14913
- [28] Khandelwal, Urvashi and Levy, Omer and Jurafsky, Dan and Zettlemoyer, Luke and Lewis, Mike , month = feb, year =. Generalization through. doi:10.48550/arXiv.1911.00172 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.1911.00172
- [29] Wei, Rubin and Cao, Jiaqi and Wang, Jiarui and Kai, Jushi and Guo, Qipeng and Zhou, Bowen and Lin, Zhouhan , month = oct, year =. doi:10.48550/arXiv.2508.01832 , abstract = on Pithdoi:10.48550/arxiv.2508.01832
- [30] He, Zexue and Wang, Yu and Zhi, Churan and Hu, Yuanzhe and Chen, Tzu-Ping and Yin, Lang and Chen, Ze and Wu, Tong Arthur and Ouyang, Siru and Wang, Zihan and Pei, Jiaxin and McAuley, Julian and Choi, Yejin and Pentland, Alex , month = feb, year =. doi:10.48550/arXiv.2602.16313... on PithPith's reviewarXivdoi:10.48550/arxiv.2602.16313
- [31] Wu, Fangzhao and Qiao, Ying and Chen, Jiun-Hung and Wu, Chuhan and Qi, Tao and Lian, Jianxun and Liu, Danyang and Xie, Xing and Gao, Jianfeng and Wu, Winnie and Zhou, Ming , editor =. Proceedings of the 58th. 2020 , pages =. doi:10.18653/v1/2020.acl-main.331 , abstract = on Pithdoi:10.18653/v1/2020.acl-main.331
- [32] Hou, Yupeng and Li, Jiacheng and He, Zhankui and Yan, An and Chen, Xiusi and McAuley, Julian , month = mar, year =. Bridging. doi:10.48550/arXiv.2403.03952 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2403.03952
- [33] Jiang, Bowen and Yuan, Yuan and Shen, Maohao and Hao, Zhuoqun and Xu, Zhangchen and Chen, Zichen and Liu, Ziyi and Vijjini, Anvesh Rao and He, Jiashu and Yu, Hanchao and Poovendran, Radha and Wornell, Gregory and Ungar, Lyle and Roth, Dan and Chen, Sihao and Taylor, Camillo Jo... on Pithdoi:10.48550/arxiv.2512.06688
- [34] Meng, Yuxian and Wang, Shuhe and Han, Qinghong and Sun, Xiaofei and Wu, Fei and Yan, Rui and Li, Jiwei , month = may, year =. doi:10.48550/arXiv.2012.15015 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2012.15015
- [35] Zheng, Yinhe and Chen, Guanyi and Liu, Xin and Sun, Jian , month = may, year =. doi:10.48550/arXiv.2108.07154 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2108.07154
- [36] Feng, Jiazhan and Sun, Qingfeng and Xu, Can and Zhao, Pu and Yang, Yaming and Tao, Chongyang and Zhao, Dongyan and Lin, Qingwei , month = dec, year =. doi:10.48550/arXiv.2211.05719 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2211.05719
- [37] Jiayang, Cheng and Ru, Dongyu and Qiu, Lin and Li, Yiyang and Cao, Xuezhi and Song, Yangqiu and Cai, Xunliang , month = mar, year =. doi:10.48550/arXiv.2603.01966 , abstract = on Pithdoi:10.48550/arxiv.2603.01966
- [38] 2026 , note =. doi:10.48550/arXiv.2605.26667 , author = on PithPith's reviewarXivdoi:10.48550/arxiv.2605.26667
- [39] Ru, Dongyu and Qiu, Lin and Hu, Xiangkun and Zhang, Tianhang and Shi, Peng and Chang, Shuaichen and Cheng, Jiayang and Wang, Cunxiang and Sun, Shichao and Li, Huanyu and Zhang, Zizhao and Wang, Binjie and Jiang, Jiarong and He, Tong and Wang, Zhiguo and Liu, Pengfei and Zhang,... on Pith
- [40] Bei, Yuanchen and Wei, Tianxin and Ning, Xuying and Zhao, Yanjun and Liu, Zhining and Lin, Xiao and Zhu, Yada and Hamann, Hendrik and He, Jingrui and Tong, Hanghang , month = jan, year =. Mem-. doi:10.48550/arXiv.2601.03515 , abstract = on Pithdoi:10.48550/arxiv.2601.03515
- [41] Team, Gemini and Anil, Rohan and Borgeaud, Sebastian and Alayrac, Jean-Baptiste and Yu, Jiahui and Soricut, Radu and Schalkwyk, Johan and Dai, Andrew M. and Hauth, et al. , month = may, year =. Gemini:. doi:10.48550/arXiv.2312.11805 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2312.11805
- [42] OpenAI and Hurst, Aaron and Lerer, Adam and Goucher, Adam P. and Perelman, Adam and Ramesh, Aditya and Clark, Aidan and Ostrow, A. J. and Welihinda, et al. , month = oct, year =. doi:10.48550/arXiv.2410.21276 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2410.21276
- [43] Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei , month = feb, year =. Evaluating. doi:10.48550/arXiv.2402.17753 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2402.17753
- [44] Trivedi, Harsh and Balasubramanian, Niranjan and Khot, Tushar and Sabharwal, Ashish , month = may, year =. doi:10.48550/arXiv.2108.00573 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2108.00573
- [45] Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W. and Salakhutdinov, Ruslan and Manning, Christopher D. , month = sep, year =. doi:10.48550/arXiv.1809.09600 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.1809.09600
- [46] Meyer, Yev and Corneil, Dane , title =. 2025 , url = on Pith
- [47] Abootorabi, Mohammad Mahdi and Zobeiri, Amirhosein and Dehghani, Mahdi and Mohammadkhani, Mohammadali and Mohammadi, Bardia and Ghahroodi, Omid and Baghshah, Mahdieh Soleymani and Asgari, Ehsaneddin , editor =. Ask in. Findings of the. 2025 , pages =. doi:10.18653/v1/2025.find... on Pithdoi:10.18653/v1/2025.findings-acl.861
- [48] Li, Jiahao Nick and Zhang, Zhuohao (Jerry) and Ma, Jiaju , month = apr, year =. Proceedings of the 2025. doi:10.1145/3706598.3713448 , abstract = on Pithdoi:10.1145/3706598.3713448
- [49] Peng, Xiangyu and Qin, Can and Chen, Zeyuan and Xu, Ran and Xiong, Caiming and Wu, Chien-Sheng , month = jan, year =. doi:10.48550/arXiv.2510.03663 , abstract = on Pithdoi:10.48550/arxiv.2510.03663
- [51] Hönicke, Florian and Günther, Michael and Koukounas, Andreas and Akram, Mohammad Kalim and Martens, Scott and Sturua, Saba and Xiao, Han , month = may, year =. jina-embeddings-v5-omni:. doi:10.48550/arXiv.2605.08384 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2605.08384
- [52] Zhong, Wanjun and Guo, Lianghong and Gao, Qiqi and Ye, He and Wang, Yanlin , month = may, year =. doi:10.48550/arXiv.2305.10250 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2305.10250
- [53] Ou, Jiao and Lu, Junda and Liu, Che and Tang, Yihong and Zhang, Fuzheng and Zhang, Di and Gai, Kun , month = mar, year =. doi:10.48550/arXiv.2311.01677 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2311.01677
- [54] Liu, Ziyu and Chu, Tao and Zang, Yuhang and Wei, Xilin and Dong, Xiaoyi and Zhang, Pan and Liang, Zijian and Xiong, Yuanjun and Qiao, Yu and Lin, Dahua and Wang, Jiaqi , month = oct, year =. doi:10.48550/arXiv.2406.11833 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2406.11833
- [55] Xue, Haochen and Tang, Feilong and Hu, Ming and Liu, Yexin and Huang, Qidong and Li, Yulong and Liu, Chengzhi and Xu, Zhongxing and Zhang, Chong and Feng, Chun-Mei and Xie, Yutong and Razzak, Imran and Ge, Zongyuan and Su, Jionglong and He, Junjun and Qiao, Yu , month = mar, y... on PithPith's reviewarXivdoi:10.48550/arxiv.2502.11903
- [56] Mei, Jingbiao and Chen, Jinghong and Yang, Guangyu and Hou, Xinyu and Li, Margaret and Byrne, Bill , month = mar, year =. According to. doi:10.48550/arXiv.2603.01990 , abstract = on Pithdoi:10.48550/arxiv.2603.01990
- [57] Wang, Weiyun and Zhang, Shuibo and Ren, Yiming and Duan, Yuchen and Li, Tiantong and Liu, Shuo and Hu, Mengkang and Chen, Zhe and Zhang, Kaipeng and Lu, Lewei and Zhu, Xizhou and Luo, Ping and Qiao, Yu and Dai, Jifeng and Shao, Wenqi and Wang, Wenhai , month = oct, year =. Nee... on PithPith's reviewarXivdoi:10.48550/arxiv.2406.07230
- [58] Hu, Wenbo and Gu, Jia-Chen and Dou, Zi-Yi and Fayyaz, Mohsen and Lu, Pan and Chang, Kai-Wei and Peng, Nanyun , month = mar, year =. doi:10.48550/arXiv.2410.08182 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2410.08182
- [59] Hu, Yuanzhe and Wang, Yu and McAuley, Julian , month = mar, year =. Evaluating. doi:10.48550/arXiv.2507.05257 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2507.05257
- [60] Team, Qwen , month = apr, year =. Qwen3.5-. doi:10.48550/arXiv.2604.15804 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2604.15804
- [61] Jiang, Bowen and Hao, Zhuoqun and Cho, Young-Min and Li, Bryan and Yuan, Yuan and Chen, Sihao and Ungar, Lyle and Taylor, Camillo J. and Roth, Dan , year =. Know. doi:10.48550/ARXIV.2504.14225 , abstract = on Pithdoi:10.48550/arxiv.2504.14225
- [62] Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and Cheng, Zesen and Deng, Lianghao and Ding, Wei and Gao, Chang and Ge, Chunjiang and Ge, Wenbin and Guo, Zhifang and Huang, Qidong and Huang, Jie and Huang, Fei and Hui, Binyuan and Jiang, Shutong... on PithPith's reviewarXivdoi:10.48550/arxiv.2511.21631
- [63] Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya , file =. Robust. Proceedings of the 40th. 2023 , abstract = on Pith
- [64] Zhang, Saizheng and Dinan, Emily and Urbanek, Jack and Szlam, Arthur and Kiela, Douwe and Weston, Jason , editor =. Personalizing. Proceedings of the 56th. 2018 , pages =. doi:10.18653/v1/P18-1205 , abstract = on Pithdoi:10.18653/v1/p18-1205
- [65] Li, Dawei and Jiang, Bohan and Huang, Liangjie and Beigi, Alimohammad and Zhao, Chengshuai and Tan, Zhen and Bhattacharjee, Amrita and Jiang, Yuxuan and Chen, Canyu and Wu, Tianhao and Shu, Kai and Cheng, Lu and Liu, Huan , year =. From. doi:10.48550/ARXIV.2411.16594 , abstract = on Pithdoi:10.48550/arxiv.2411.16594
- [66] Talmor, Alon and Yoran, Ori and Catav, Amnon and Lahav, Dan and Wang, Yizhong and Asai, Akari and Ilharco, Gabriel and Hajishirzi, Hannaneh and Berant, Jonathan , month = apr, year =. doi:10.48550/arXiv.2104.06039 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2104.06039
- [67] Sung, Junyoung and Lyu, Seungwoo and Kim, Minjun and An, Sumin and Nagrani, Arsha and Seo, Paul Hongsuck , year =. doi:10.48550/ARXIV.2604.01634 , abstract = on Pithdoi:10.48550/arxiv.2604.01634
- [68] Chang, Yingshan and Narang, Mridu and Suzuki, Hisami and Cao, Guihong and Gao, Jianfeng and Bisk, Yonatan , month = mar, year =. doi:10.48550/arXiv.2109.00590 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2109.00590
- [69] Gutiérrez, Bernal Jiménez and Shu, Yiheng and Gu, Yu and Yasunaga, Michihiro and Su, Yu , month = jan, year =. doi:10.48550/arXiv.2405.14831 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2405.14831
- [70] Wang, Wenhui and Wei, Furu and Dong, Li and Bao, Hangbo and Yang, Nan and Zhou, Ming , month = apr, year =. doi:10.48550/arXiv.2002.10957 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2002.10957
- [71] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , month = mar, year =. doi:10.48550/arXiv.2210.03629 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2210.03629
- [72] Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and Zhang, Xin and Lin, Huan and Yang, Baosong and Xie, Pengjun and Yang, An and Liu, Dayiheng and Lin, Junyang and Huang, Fei and Zhou, Jingren , month = jun, year =. Qwen3. doi:10.48550/arXiv.2506.05176 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2506.05176
- [73] Kang, Jiazheng and Ji, Mingming and Zhao, Zhe and Bai, Ting , editor =. Memory. Proceedings of the 2025. 2025 , pages =. doi:10.18653/v1/2025.emnlp-main.1318 , abstract = on Pithdoi:10.18653/v1/2025.emnlp-main.1318
- [74] He, Junqing and Zhu, Liang and Wang, Rui and Wang, Xi and Haffari, Reza and Zhang, Jiaxing , month = oct, year =. doi:10.48550/arXiv.2409.15240 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2409.15240
- [75] Du, Bangde and Guo, Minghao and He, Songming and Ye, Ziyi and Zhu, Xi and Su, Weihang and Zhu, Shuqi and Zhou, Yujia and Zhang, Yongfeng and Ai, Qingyao and Liu, Yiqun , month = oct, year =. doi:10.48550/arXiv.2510.25536 , abstract = on Pithdoi:10.48550/arxiv.2510.25536
- [76] Wang, Hengyi and Shi, Haizhou and Tan, Shiwei and Qin, Weiyi and Wang, Wenyuan and Zhang, Tunyu and Nambi, Akshay and Ganu, Tanuja and Wang, Hao , month = feb, year =. Multimodal. doi:10.48550/arXiv.2406.11230 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2406.11230
- [77] Zhou, Yang and Zhao, Mingyu and Wang, Zhenting and Gu, Difei and Guo, Bangwei and Ye, Ruosong and Han, Ligong and Jin, Can and Metaxas, Dimitris N. , month = feb, year =. M. doi:10.48550/arXiv.2511.17729 , abstract = on Pithdoi:10.48550/arxiv.2511.17729
- [78] Kuratov, Yuri and Bulatov, Aydar and Anokhin, Petr and Rodkin, Ivan and Sorokin, Dmitry and Sorokin, Artyom and Burtsev, Mikhail , month = nov, year =. doi:10.48550/arXiv.2406.10149 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2406.10149
- [79] Hsieh, Cheng-Ping and Sun, Simeng and Kriman, Samuel and Acharya, Shantanu and Rekesh, Dima and Jia, Fei and Zhang, Yang and Ginsburg, Boris , month = aug, year =. doi:10.48550/arXiv.2404.06654 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2404.06654
- [80] Tie, Guiyao and Zhou, Xueyang and Gu, Tianhe and Zhang, Ruihang and Hu, Chaoran and Zhang, Sizhe and Sun, Mengqu and Zhang, Yan and Zhou, Pan and Sun, Lichao , month = jul, year =. doi:10.48550/arXiv.2505.16459 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2505.16459
- [81] Wang, Xiyao and Zhou, Yuhang and Liu, Xiaoyu and Lu, Hongjin and Xu, Yuancheng and He, Feihong and Yoon, Jaehong and Lu, Taixi and Bertasius, Gedas and Bansal, Mohit and Yao, Huaxiu and Huang, Furong , month = jan, year =. Mementos:. doi:10.48550/arXiv.2401.10529 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2401.10529
- [82] Chen, Howard and Pasunuru, Ramakanth and Weston, Jason and Celikyilmaz, Asli , month = oct, year =. Walking. doi:10.48550/arXiv.2310.05029 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2310.05029
- [83] Sarthi, Parth and Abdullah, Salman and Tuli, Aditi and Khanna, Shubh and Goldie, Anna and Manning, Christopher D. , month = jan, year =. doi:10.48550/arXiv.2401.18059 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2401.18059
- [84] Liu, Weijie and Tang, Zecheng and Li, Juntao and Chen, Kehai and Zhang, Min , month = aug, year =. doi:10.48550/arXiv.2408.16967 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2408.16967
- [85] Guo, Zirui and Xia, Lianghao and Yu, Yanhua and Ao, Tu and Huang, Chao , month = apr, year =. doi:10.48550/arXiv.2410.05779 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2410.05779
- [86] Paul, Swarna Kamal and Sharma, Shubhendu and Sareen, Nitin , month = may, year =. doi:10.48550/arXiv.2603.27910 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2603.27910
- [87] MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models , author=. 2026 , eprint= on Pith
- [88] Seeing, Listening, Remembering, and Reasoning: A Multimodal Agent with Long-Term Memory , author=. 2025 , eprint= on Pith
- [89] M2A: Multimodal Memory Agent with Dual-Layer Hybrid Memory for Long-Term Personalized Interactions , author=. 2026 , eprint= on Pith
- [90] Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning , author=. 2025 , eprint= on Pith
- [91] MuRAG: Multimodal Retrieval-Augmented Generator for Open Question Answering over Images and Text , author=. 2022 , eprint= on Pith
- [92] Enhancing Chat Language Models by Scaling High-quality Instructional Conversations , author=. arXiv preprint arXiv:2305.14233 , year= on PithPith's reviewarXiv
- [93] Chiang, Wei-Lin and Li, Zhuohan and Lin, Zi and Sheng, Ying and Wu, Zhanghao and Zhang, Hao and Zheng, Lianmin and Zhuang, Siyuan and Zhuang, Yonghao and Gonzalez, Joseph E. and Stoica, Ion and Xing, Eric P. , title=. 2023 , month= on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment