Pith. sign in

REVIEW 5 major objections 6 minor 50 references

MTPChat: A Multimodal Time-Aware Persona Dataset for Conversational Agents

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read MTPChat introduces the first multimodal, time-aware persona dialogue benchmark, with two retrieval tasks showing that models struggle to use temporal order.

desk verdict A genuinely new multimodal time-aware persona dialogue resource with two sensible retrieval tasks, but the 'implicit temporal cues' claim is overstated and the missing data/code release is the main blocker. read the letter →

arxiv 2502.05887 v1 pith:V6AY7CTP submitted 2025-02-09 cs.CL cs.AI

Wanqi Yang , Yanda Li , Meng Fang , Ling Chen This is my paper · ORCID
classification cs.CLcs.AI
keywords MTPChatmultimodaldialoguetime-awarepersonamemorytemporalreasoningretrievaladaptivemoduleconversationalagents
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

To test whether conversational agents actually track time, the paper introduces MTPChat, a multimodal persona dialogue dataset that attaches explicit dates to both dialogue turns and the speaker's episodic memories, and constructs paired 'early' and 'later' versions of the same conversation. On top of this dataset, the paper defines two retrieval tasks: Temporal Next Response Prediction (TNRP), where a model must pick the response appropriate for the dialogue's date, and Temporal Grounding Memory Prediction (TGMP), where it must pick the memory that grounds a response or output 'No Memory' when the dialogue predates the memory. Experiments with strong text–image encoders show that both tasks are hard (best Recall@1 below 72%) and that removing time information significantly hurts TGMP, which the authors take as evidence that temporal awareness is a distinct, learnable capability. The paper also introduces an Adaptive Temporal Module that weights text and vision features, and reports consistent gains over fixed fusion baselines.

What carries the argument

The load-bearing object is the time-stamped triplet—(dialogue context, image, date) and (memory context, image, date)—together with the early/later conversation construction and the 'No Memory' class. The paper encodes the date as a plain string concatenated into the text, and its Adaptive Temporal Module (ATM) learns a sigmoid-gated blend of text and vision features per item. That module is what the paper credits for the improved integration of multimodal streams in the presence of date information.

What would settle it

Train a model on TGMP with all dialogue and memory text replaced by fixed placeholders, leaving only the yyyy/mm/dd dates; if this date-only model matches the full model's accuracy, the task is solvable by date comparison and does not measure temporal understanding.

Watch

Extended reading notes

Core claim

The central claim is that temporal order is a missing axis in persona-grounded dialogue, and that building a benchmark around it exposes a real limitation of current multimodal conversational models. MTPChat is constructed from MPChat by converting timestamps to yyyy/mm/dd dates and attaching them to dialogue and memory triplets (context, image, time); for each original 'later stage' conversation, the authors generate an 'early stage' counterpart with GPT-4, where the speaker lacks relevant memory and responds accordingly. TGMP operationalises grounding as a rule: if the dialogue date is earlier than a memory's date, the correct label is 'No Memory.' The paper's experiments show that CLIP+CLIP and SBERT+CLIP retrievers solve TNRP and TGMP only partially, that adding date strings to the input improves TGMP substantially, and that the proposed adaptive weighting of text and visual features outperforms attention, linear, and mean-pool fusion. The authors conclude that MTPChat provides the first multimodal, time-aware persona dialogue evaluation and that temporal reasoning in this setting remains an open problem.

Load-bearing premise

The TGMP task's rule that a dialogue predating a memory must be labeled 'No Memory' is the load-bearing premise, and because dates are written directly into the input text, a model could solve the task by comparing date strings instead of reasoning about content.

Editorial extensions

If this is right

  • MTPChat provides 18,973 conversations with dated dialogue and memory triplets, making it a reusable benchmark for temporal awareness in retrieval-based dialogue models.
  • The reported gap between temporally-aware and temporally-unaware inputs on TGMP (12.7% Recall@1) indicates that explicit time information is a key signal for grounding responses in persona memory.
  • The Adaptive Temporal Module improves TGMP Recall@1 to 71.82, outperforming attention, linear, and mean-pool fusion, suggesting that dynamically weighting modalities helps when text contains date tokens.
  • Because the paper includes both 'early' and 'later' responses for the same context, TNRP directly tests whether a model uses temporal cues to select a response rather than relying on lexical overlap.

Reading between the lines

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

  • Because date strings are concatenated into the text input, the reported gains from time information may partly reflect the model learning to parse and compare yyyy/mm/dd tokens; a date-only ablation (with content masked) would separate genuine temporal reasoning from format-based shortcutting.
  • The dataset recipe is portable: applying the same early/later construction to any timestamped persona dialogue corpus would yield a stress test for temporal grounding in other languages or domains, without requiring new annotations beyond the timestamps.
  • The GPT-4 generated early responses may introduce surface patterns (such as short 'I don't know' replies) that make TNRP easier than intended; human-written early responses or a human validation subset would measure how much of the task difficulty is due to time reasoning versus generation artifacts.
  • The current evaluation is retrieval-only; a generative variant that asks a model to produce a response conditioned on the dialogue date and memory set would test whether temporal awareness extends to generation, not just ranking.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper introduces MTPChat, a multimodal, time-aware persona dialogue dataset derived from MPChat by adding date annotations in yyyy/mm/dd format, generating "early-stage" conversations with GPT-4, and adding a "No Memory" category. It defines two retrieval tasks, Temporal Next Response Prediction (TNRP) and Temporal Grounding Memory Prediction (TGMP), and proposes an Adaptive Temporal Module (ATM) that fuses text and vision features through a learned sigmoid weighting. Experiments with SBERT+CLIP and CLIP+CLIP baselines report that MTPChat is challenging and that ATM improves TGMP performance while giving small gains on TNRP.

Significance. If the dataset and tasks are valid, MTPChat would be a useful resource for studying temporal order in persona-grounded multimodal dialogue, and the two retrieval tasks would provide a new testbed for temporal grounding. The proposed ATM is simple and model-agnostic. Credit is due for building on an existing multimodal persona dataset, for explicitly defining two temporally sensitive retrieval tasks, and for reporting multiple baselines. However, the absence of a data release, the lack of human validation of GPT-4-generated labels, and the explicit date strings in the inputs mean that the significance is currently conditional on substantial revisions.

major comments (5)
  1. [Abstract, Introduction, Section 3.1, Section 4.2] The paper repeatedly claims that MTPChat tests "implicit temporal cues" and avoids "relying on explicit timestamps" (Abstract; Introduction; Section 2). Yet Section 3.1 states that UTC strings are converted to "yyyy/mm/dd" and incorporated into both dialogues and memories, and Section 4.2 defines the TGMP label precisely by comparing the dialogue date with the grounding memory date. The temporal signal is therefore explicit in the input, and a model can solve the temporal component by comparing date tokens. This contradicts the central novelty claim and weakens the distinction from TimeIT. Please either reframe the claims as testing explicit temporal order, or add a stringent test such as masking or removing all date strings and showing that performance collapses, along with a date-comparison heuristic baseline, to demonstrate that content-based temporal reasoning is required.
  2. [Section 3.2, Appendix A] GPT-4-generated early-stage responses are treated as ground truth without any human validation, quality filtering, or inter-annotator agreement. The generation prompt in Appendix A instructs GPT-4 to produce a response indicating lack of familiarity when the provided memories are unrelated to the conversation, which is essentially the same temporal-assumption rule that defines the TGMP labels. This creates a risk of circular evaluation: the label-generation process encodes the very temporal assumption the tasks are intended to test. Please include human evaluation of a representative sample, report agreement statistics, and discuss how the label-generation assumptions affect task validity.
  3. [Section 6.2, Tables 2 and 3] The headline empirical claim that ATM "outperforms other feature integration methods" is only partially supported. On TNRP, the gains over the corresponding baselines are very small: for example, CLIP+CLIP+ATM versus CLIP+CLIP on (d, M) improves R@1 by 0.51 and MRR by 0.26, and for SBERT+CLIP the gains are 0.23 in both metrics. Larger gains appear on TGMP, but no error bars, confidence intervals, or significance tests are reported for any result. Please report multiple seeds with variance and significance tests, and temper the claim to indicate where the improvement is reliable.
  4. [Section 3.4, Section 10] The paper presents MTPChat as a dataset contribution, but no download URL, code release, or access instructions are provided. Section 10 discusses usage restrictions, yet a reader cannot obtain the data. For a dataset paper, this is a blocker for verification, reuse, and benchmarking. Please state the release plan, including a URL, license, and hosting details, or explicitly note if this version is anonymized and the link will be added in a camera-ready version.
  5. [Section 4.2, Section 3.4] The TGMP task may be shortcutted not only by date comparison but also by the 3:1 later-to-early conversation ratio and the deterministic pairing of early-stage dialogues with the "No Memory" category. A model that can identify an early-stage dialogue—for example, by response style or by cues in the GPT-4-generated text—could guess "No Memory" without performing temporal grounding. The paper should report TGMP performance separately for early-stage and later-stage dialogues, and analyze whether the "No Memory" decision is driven by content or by dialogue-stage features.
minor comments (6)
  1. [Section 2, References] The dataset referred to as "TimeIT" in Section 2 is cited as Ren et al. (2023), whose reference title is "Timechat". Please clarify whether the dataset name is TimeIT or TimeChat and use the name consistently.
  2. [Table 1, Section 2] The table header "TempLama" and the text "TEMPLAMA" should be unified to a single spelling.
  3. [Section 5] The Adaptive Temporal Module is described as an "adaptive temporal module", but the described mechanism is a per-instance sigmoid-weighted fusion of text and vision features; it does not explicitly model temporal order or temporal dependencies. Consider naming it "adaptive fusion module" or clarifying how it captures temporal structure beyond the date strings in the text encoder.
  4. [Figure 2] Figure 2 lacks clear axis labels and a description of what is plotted; please state whether the histograms show counts of conversations or memories and define the x-axis units.
  5. [Section 6.1, Table 5] The ablation "without time" should state explicitly that the date strings were removed from both dialogues and memories, and how the "No Memory" time was handled in this condition.
  6. [Appendix A, Table 6] The prompt template in Table 6 shows "Memories: [context]", which appears to be a placeholder for a single context rather than a list of memories; the template should be clarified.

Circularity Check

2 steps flagged · score 6.0 of 10

TGMP's temporal label is a date-comparison function over explicitly concatenated yyyy/mm/dd strings, so the 'implicit temporal cues' claim reduces by construction; GPT-4-generated labels encode the same assumption.

  1. self definitional [Section 3.1 'Time Annotations' and Section 4.2 'Temporal Grounding Memory Prediction']
    "We converted the UTC strings in MPChat dataset into date format “yyyy/mm/dd” and incorporated this feature into both the dialogue and memory components. ... when the time of the dialogue is later than the time of the grounding memory ... the model is capable of predicting the grounding memory. Conversely, if the time of the dialogue is earlier than that of the grounding memory ... the model must predict a “No Memory” outcome."

    The TGMP ground-truth label is defined by the ordering of the dialogue date t_d and memory date t_m, and both dates are explicitly part of the input: Section 3.1 puts yyyy/mm/dd strings into every dialogue and memory triplet, and the framework description says text and date information are concatenated for each entry. The temporal component of the 'prediction' is therefore not an inference from implicit cues but a comparison of two date tokens present in the input. Models that read the date strings can compute the 'No Memory' outcome from the date inequality alone, so the task's temporal signal reduces by construction to the explicit timestamps the paper claims to avoid.

  2. other [Section 3.2 'Time-Aware Conversations' and Appendix A]
    "we utilized GPT-4 ... to process a combination of inputs: the dialogue context, dialogue image, newly modified dialogue time, and speaker memories predating this new dialogue time. GPT-4 generated responses under the following guidelines: ..."

    The early-stage conversation labels that define the dataset's temporal dynamics are generated by GPT-4 using the same rule the proposed tasks test: memories predating the dialogue determine whether and how the speaker responds. These generated responses are then used as ground truth for TNRP, and early-stage conversations are aligned with the 'No Memory' category in TGMP. The benchmark therefore measures how well models reproduce GPT-4's instantiation of the authors' temporal-ordering assumption, not an independently verified human temporal-reasoning phenomenon. No human validation is reported for these labels, so the evaluation is tied to the same assumption that the dataset claims to discover rather than to an external standard.

full rationale

The paper's ATM framework comparison and retrieval experiments are ordinary empirical evaluations and are not circular; there is no load-bearing self-citation chain, and the framework's superiority claim is tested on the authors' own tasks just as in most dataset papers. The circularity is in the interpretative claim. TGMP's 'No Memory' label is a deterministic function of the yyyy/mm/dd strings that are explicitly concatenated into every dialogue and memory entry, so the temporal part of the task reduces to reading explicit date tokens, not to recognizing implicit temporal dynamics. The ablation in Table 5 (TGMP R@1 68.75 with time vs 60.99 without) shows content still matters, so dates do not fully trivialize the task, but the central novelty claim that MTPChat avoids explicit timestamps and tests 'implicit temporal cues' is contradicted by the dataset's own construction. The GPT-4 response-generation pipeline additionally feeds the same temporal-ordering assumption into the labels that later serve as ground truth, compounding the circularity. These issues concern the central contribution, so a score of 6 is appropriate: several task definitions reduce by construction to explicit date comparison and prompt-encoded assumptions, even though the framework evaluation itself is not circular.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

Central evaluation claims rest on hand-chosen hyperparameters, domain assumptions about timestamp fidelity and GPT-4 quality, and one constructed entity ('No Memory'). No independent external benchmark or human validation is provided.

free parameters (4)
  • Maximum number of memories m = 20
    Hand-selected hyperparameter limiting memory set size in training and evaluation (Section 6.1, Table 7).
  • Retrieval candidate count C = 100
    Hand-selected candidate pool size for TNRP and TGMP (Section 6.1, Table 7).
  • GPT-4 early response word cap = 40 words
    Prompt constraint used to generate early-stage responses (Section 3.2, Appendix A).
  • Later-to-early conversation ratio = 3:1
    Data construction ratio chosen by the authors, not derived from any principle (Section 3.4).
assumptions (5)
  • domain assumption UTC strings in MPChat, converted to yyyy/mm/dd, constitute valid temporal order for dialogues and memories.
    Section 3.1; timestamp noise or aggregation would make the temporal dynamics synthetic.
  • domain assumption GPT-4 generates realistic early-stage responses under the prompt guidelines.
    Section 3.2 and Appendix A; no human validation or quality check is reported.
  • domain assumption Original MPChat conversations and memories can serve unchanged as later-stage instances.
    Section 3.2; later-stage data inherits MPChat's collection biases and assumed temporal position.
  • domain assumption A dialogue earlier than a grounding memory should be paired with 'No Memory'.
    Section 4.2; defines TGMP labels and enables a date-comparison shortcut.
  • domain assumption The speaker's memory set is complete and correctly synchronized with dialogue time.
    Section 3.3; fabricated memories are avoided, but completeness of real Reddit histories is unverified.
invented entities (1)
  • 'No Memory' category with plain white image
    purpose: Acts as the correct grounding target for early-stage conversations with no relevant prior memory.
    Introduced in Section 3.3; a constructed category with no naturalistic counterpart, and its white image gives the vision encoder a constant input.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MTPChat: A Multimodal Time-Aware Persona Dataset for Conversational Agents." pith.science (2026). https://pith.science/paper/V6AY7CTP

@misc{pith2026250205887,
  author       = {Pith},
  title        = {Pith review of: MTPChat: A Multimodal Time-Aware Persona Dataset for Conversational Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6AY7CTP}},
  note         = {Machine review of arXiv:2502.05887}
}
read the original abstract

Understanding temporal dynamics is critical for conversational agents, enabling effective content analysis and informed decision-making. However, time-aware datasets, particularly for persona-grounded conversations, are still limited, which narrows their scope and diminishes their complexity. To address this gap, we introduce MTPChat, a multimodal, time-aware persona dialogue dataset that integrates linguistic, visual, and temporal elements within dialogue and persona memory. Leveraging MTPChat, we propose two time-sensitive tasks: Temporal Next Response Prediction (TNRP) and Temporal Grounding Memory Prediction (TGMP), both designed to assess a model's ability to understand implicit temporal cues and dynamic interactions. Additionally, we present an innovative framework featuring an adaptive temporal module to effectively integrate multimodal streams and capture temporal dependencies. Experimental results validate the challenges posed by MTPChat and demonstrate the effectiveness of our framework in multimodal time-sensitive scenarios.

Figures

Figures reproduced from arXiv: 2502.05887 by the authors.

Figure 1
Figure 1. An example of a multimodal, time-sensitive, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distribution of times across conversations and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the Temporal Next Response Prediction (TNRP) and Temporal Grounding Memory Prediction [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Architecture of our framework with Adaptive [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 12 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Jaewoo Ahn, Yeda Song, Sangdoo Yun, and Gunhee Kim. 2023. Mpchat: Towards multimodal persona-grounded conversation. arXiv preprint arXiv:2305.17388

    arXiv 2023

  5. [5]

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425--2433

    2015

  6. [6]

    Wenhu Chen, Xinyi Wang, and William Yang Wang. 2021. A dataset for answering time-sensitive questions. arXiv preprint arXiv:2108.06314

    arXiv 2021

  7. [7]

    Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jos \'e MF Moura, Devi Parikh, and Dhruv Batra. 2017. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326--335

    2017

  8. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    arXiv 2018

Show all 50 references
  1. [9]

    Bhuwan Dhingra, Jeremy R Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W Cohen. 2022. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257--273

    2022

  2. [10]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint...

    2020 arXiv

  3. [11]

    Jiazhan Feng, Qingfeng Sun, Can Xu, Pu Zhao, Yaming Yang, Chongyang Tao, Dongyan Zhao, and Qingwei Lin. 2022. Mmdialog: A large-scale multi-turn dialogue dataset towards multi-modal open-domain conversation. arXiv preprint arXiv:2211.05719

    2022 arXiv

  4. [12]

    Thibault F \'e vry, Livio Baldini Soares, Nicholas FitzGerald, Eunsol Choi, and Tom Kwiatkowski. 2020. Entities as experts: Sparse memory access with entity supervision. arXiv preprint arXiv:2004.07202

    2020 arXiv

  5. [13]

    Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282

    2020 arXiv

  6. [14]

    Yoonna Jang, Jungwoo Lim, Yuna Hur, Dongsuk Oh, Suhyune Son, Yeonsoo Lee, Donghoon Shin, Seungryong Kim, and Heuiseok Lim. 2022. Call for customized conversation: Customized conversation grounding persona and knowledge. In Proceedings of the AAAI Conference on Artificial Intel...

    2022

  7. [15]

    Zhen Jia, Abdalghani Abujabal, Rishiraj Saha Roy, Jannik Str \"o tgen, and Gerhard Weikum. 2018. Tequila: Temporal question answering over knowledge bases. In Proceedings of the 27th ACM international conference on information and knowledge management, pages 1807--1810

    2018

  8. [16]

    Zhen Jia, Soumajit Pramanik, Rishiraj Saha Roy, and Gerhard Weikum. 2021. Complex temporal question answering on knowledge graphs. In Proceedings of the 30th ACM international conference on information & knowledge management, pages 792--802

    2021

  9. [17]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  10. [18]

    Satwik Kottur, Jos \'e MF Moura, Devi Parikh, Dhruv Batra, and Marcus Rohrbach. 2019. Clevr-dialog: A diagnostic dataset for multi-round reasoning in visual dialog. arXiv preprint arXiv:1903.03166

    2019 arXiv

  11. [19]

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for C...

    2019

  12. [20]

    Deuksin Kwon, Sunwoo Lee, Ki Hyun Kim, Seojin Lee, Taeyoon Kim, and Eric Davis. 2023. What, when, and how to ground: Designing user persona-aware conversational agents for engaging dialogue. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguis...

    2023

  13. [21]

    Nyoungwoo Lee, Suwon Shin, Jaegul Choo, Ho-Jin Choi, and Sung-Hyun Myaeng. 2021. Constructing multi-modal dialogue dataset by replacing text with semantically relevant images. arXiv preprint arXiv:2107.08685

    2021 arXiv

  14. [22]

    Yanda Li, Chi Zhang, Gang Yu, Zhibin Wang, Bin Fu, Guosheng Lin, Chunhua Shen, Ling Chen, and Yunchao Wei. 2023. Stablellava: Enhanced visual instruction tuning with synthesized image-dialogue data. arXiv preprint arXiv:2308.10253

    2023 arXiv

  15. [23]

    Adam Liska, Tomas Kocisky, Elena Gribovskaya, Tayfun Terzi, Eren Sezener, Devang Agrawal, D’Autume Cyprien De Masson, Tim Scholtes, Manzil Zaheer, Susannah Young, et al. 2022. Streamingqa: A benchmark for adaptation to new knowledge over time in question answering models. In I...

    2022

  16. [24]

    Yuxian Meng, Shuhe Wang, Qinghong Han, Xiaofei Sun, Fei Wu, Rui Yan, and Jiwei Li. 2020. Openvidial: A large-scale, open-domain dialogue dataset with visual contexts. arXiv preprint arXiv:2012.15015

    2020 arXiv

  17. [25]

    Sewon Min, Julian Michael, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2020. Ambigqa: Answering ambiguous open-domain questions. arXiv preprint arXiv:2004.10645

    2020 arXiv

  18. [26]

    Nasrin Mostafazadeh, Chris Brockett, Bill Dolan, Michel Galley, Jianfeng Gao, Georgios P Spithourakis, and Lucy Vanderwende. 2017. Image-grounded conversations: Multimodal context for natural question and response generation. arXiv preprint arXiv:1701.08251

    2017 arXiv

  19. [27]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 3...

    2022

  20. [28]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  21. [29]

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training

    2018

  22. [30]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084

    2019 arXiv

  23. [31]

    Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. 2023. Timechat: A time-sensitive multimodal large language model for long video understanding. arXiv preprint arXiv:2312.02051

    2023 arXiv

  24. [32]

    Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? arXiv preprint arXiv:2002.08910

    2020 arXiv

  25. [33]

    Apoorv Saxena, Soumen Chakrabarti, and Partha Talukdar. 2021. Question answering over temporal knowledge graphs. arXiv preprint arXiv:2106.01515

    2021 arXiv

  26. [34]

    Paul Hongsuck Seo, Andreas Lehrmann, Bohyung Han, and Leonid Sigal. 2017. Visual reference resolution using attention memory for visual dialog. Advances in neural information processing systems, 30

    2017

  27. [35]

    Kurt Shuster, Samuel Humeau, Antoine Bordes, and Jason Weston. 2018. Image chat: Engaging grounded conversations. arXiv preprint arXiv:1811.00945

    2018 arXiv

  28. [36]

    Qingyu Tan, Hwee Tou Ng, and Lidong Bing. 2023. Towards benchmarking and improving the temporal reasoning capability of large language models. arXiv preprint arXiv:2306.08952

    2023 arXiv

  29. [37]

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

    2017

  30. [38]

    Denny Vrande c i \'c and Markus Kr \"o tzsch. 2014. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10):78--85

    2014

  31. [39]

    Shuhe Wang, Yuxian Meng, Xiaoya Li, Xiaofei Sun, Rongbin Ouyang, and Jiwei Li. 2021. Openvidial 2.0: A larger-scale, open-domain dialogue generation dataset with visual contexts. arXiv preprint arXiv:2109.12761

    2021 arXiv

  32. [40]

    Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652

    2021 arXiv

  33. [41]

    Yifan Wei, Yisong Su, Huanhuan Ma, Xiaoyan Yu, Fangyu Lei, Yuanzhe Zhang, Jun Zhao, and Kang Liu. 2023. Menatqa: A new dataset for testing the temporal comprehension and reasoning abilities of large language models. arXiv preprint arXiv:2310.05157

    2023 arXiv

  34. [42]

    Zhiyuan Wen, Jiannong Cao, Ruosong Yang, Shuaiqi Liu, and Jiaxing Shen. 2021. Automatically select emotion for response via personality-affected emotion transition. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 5010--5020

    2021

  35. [43]

    Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, and Shihang Wang. 2022. Long time no see! open-domain conversation with long-term persona memory. arXiv preprint arXiv:2203.05797

    2022 arXiv

  36. [44]

    Wanqi Yang, Yanda Li, Meng Fang, and Ling Chen. 2024 a . Enhancing temporal sensitivity and reasoning for time-sensitive question answering. arXiv preprint arXiv:2409.16909

    2024 arXiv

  37. [45]

    Wanqi Yang, Yunqiu Xu, Yanda Li, Kunze Wang, Binbin Huang, and Ling Chen. 2024 b . Continual learning for temporal-sensitive question answering. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1--9. IEEE

    2024

  38. [46]

    Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283--17297

    2020

  39. [47]

    Xiaoxue Zang, Lijuan Liu, Maria Wang, Yang Song, Hao Zhang, and Jindong Chen. 2021. Photochat: A human-human dialogue dataset with photo sharing behavior for joint image-text modeling. arXiv preprint arXiv:2108.01453

    2021 arXiv

  40. [48]

    Michael JQ Zhang and Eunsol Choi. 2021. Situatedqa: Incorporating extra-linguistic contexts into qa. arXiv preprint arXiv:2109.06157

    2021 arXiv

  41. [49]

    Yinhe Zheng, Guanyi Chen, Xin Liu, and Jian Sun. 2021. Mmchat: Multi-modal chat dataset on social media. arXiv preprint arXiv:2108.07154

    2021 arXiv

  42. [50]

    Peixiang Zhong, Chen Zhang, Hao Wang, Yong Liu, and Chunyan Miao. 2020. Towards persona-based empathetic conversational models. arXiv preprint arXiv:2004.12316

    2020 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.