Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

The paper claims that CAMVR, a memory-plus-attention framework, lets large vision-language models carry context across multi-turn dialogue and visual reasoning, beating base models on three benchmarks.

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

2026-08-05 05:13 UTC pith:UDMJCIN4

load-bearing objection Self-described 'fabricated experimental results' in the introduction sink an otherwise coherent but unoriginal architecture. the 4 major comments →

arxiv 2509.05669 v1 pith:UDMJCIN4 submitted 2025-09-06 cs.CV

Context-Aware Multi-Turn Visual-Textual Reasoning in LVLMs via Dynamic Memory and Adaptive Visual Guidance

Weijie Shen , Xinrui Wang , Yuanqi Nie , Apiradee Boonmee This is my paper
classification cs.CV
keywords multi-turn visual reasoningvision-language modelsdynamic memory networksadaptive visual attentionvisual dialoginstruction followingcontextual coherenceLVLM
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 is trying to establish that a vision-language model can keep a coherent state over a long visual dialogue if it is given two extra parts: a read-write memory that stores and retrieves cross-modal context from earlier turns, and a visual-focus module that uses that memory to re-weight image regions before the next answer is generated. Together with a decoding step that concatenates current input, memory context, and modulated visual features, these parts define CAMVR, which the paper claims outperforms LLaVA-1.5, Qwen-VL, and InstructBLIP on VisDial, an adapted A-OKVQA, and a self-built Multi-Turn Instruction Following dataset. If true, the contribution is a small-parameter add-on that reduces context loss and hallucination in multi-turn dialogue. A caveat sits in the paper's own Introduction, which states "Our fabricated experimental results demonstrate..."—if that wording is literal, the empirical claim collapses; if it is a typo, the claim still waits on the formal definitions of the MTIF dataset and the CCS/IFSR metrics.

Core claim

The central claim, stated on the paper's own terms, is that CAMVR—Context-Aware Multi-Turn Visual Reasoning—solves the context-loss and visual-grounding failures of standard LVLMs in multi-turn interaction. The mechanism is a dynamic Visual-Textual Context Memory Unit (VCMU) whose gated update writes each turn's visual and textual features into a memory matrix and whose attention-based retrieval reads out the history most relevant to the current question. That retrieved context drives an Adaptive Visual Focus Guidance (AVFG) module, which produces a spatial attention map over the image and multiplies it into the visual features before decoding. The paper reports consistent state-of-the-art r

What carries the argument

The load-bearing object is the VCMU-AVFG pair. VCMU is a read-write memory network: a gated recurrent update merges the encoded current turn into a memory matrix, and scaled dot-product attention retrieves context relevant to the query. AVFG is a small convolutional network that maps the pooled retrieved context plus raw visual features to a sigmoid spatial mask, which is multiplied into the visual features. The decoder input is then the concatenation of projected visual, textual, and retrieved-context streams. This design is what carries the paper's claim: memory supplies history, the mask focuses the encoder, and the concatenation forces generation to condition on both.

Load-bearing premise

The load-bearing premise is that the experiments, the MTIF dataset, and the CCS/IFSR metrics actually exist and measure what they claim; the Introduction's phrase "Our fabricated experimental results" means that if it is literal, the empirical claim has no support.

What would settle it

Obtain the MTIF dataset and formal definitions of CCS and IFSR, then re-run Table 1 with the same base models; if the numbers do not reproduce or the instruments do not exist, the claim fails. A cheaper check: determine whether "fabricated" in Section 1 was intended—if yes, no further evidence is needed.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Base LVLMs like LLaVA-1.5 and Qwen-VL can be upgraded with CAMVR at under 0.1% added parameters and roughly 14% slower inference, per the paper's efficiency table.
  • Long-dialogue degradation should shrink: the paper reports CCS and IFSR stay nearly flat from turn 1 to turn 4+ for CAMVR while baselines drop by several points.
  • Visual attention becomes history-dependent, so later turns re-focus on regions mentioned earlier; the granularity study claims finer 14x14 spatial masks outperform global weighting.
  • A new benchmark (MTIF) and two new metrics (CCS, IFSR) would, if published with definitions and data, give the field a standard way to score instruction-following and cross-turn coherence.
  • The paper claims its full model beats all three baselines on all four reported metrics, with the largest margins on VisDial CIDEr and MTIF IFSR.

Where Pith is reading between the lines

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

  • The architecture implies a testable division of labor: if memory retrieval is the effective part, errors in later turns should correlate with retrieval inaccuracy rather than with forgetting; probing which memory entries the attention weights read would reveal this.
  • The same memory-plus-spatial-mask recipe could transfer to embodied instruction following, where the scene changes after actions; the AVFG mask would then need to track moved or occluded objects, not just re-focus static regions.
  • A useful sanity check outside the paper's claims: collect human annotations of the objects the AVFG mask highlights and compare them with the objects actually referenced in dialogue; large mismatches would suggest the CIDEr/IFSR gains come from the language-side memory alone.
  • CCS as a metric needs a formal computation rule and an inter-annotator agreement study before it can be compared across papers; the paper leaves both unspecified.

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

4 major / 5 minor

Summary. The paper proposes CAMVR, a framework that augments existing LVLMs (LLaVA-1.5, Qwen-VL) with a Visual-Textual Context Memory Unit (VCMU) and an Adaptive Visual Focus Guidance (AVFG) mechanism, plus a multi-level reasoning integration strategy, to improve multi-turn visual-textual reasoning. The authors claim state-of-the-art results on VisDial, an adapted A-OKVQA, and a newly introduced Multi-Turn Instruction Following (MTIF) dataset, evaluated with standard metrics (CIDEr, SPICE, Acc) and two new metrics (CCS, IFSR). The manuscript also reports ablations, human evaluation, memory-capacity and granularity analyses, and efficiency numbers. Section 1, however, explicitly states 'Our fabricated experimental results demonstrate that CAMVR consistently achieves superior performance', and the experimental instruments (MTIF, CCS, IFSR) are not formally defined or released. The baseline references for LLaVA-1.5 and Qwen-VL point to unrelated papers.

Significance. If the empirical claims were genuine and reproducible, the core architectural ideas—a gated read-write memory for cross-turn context and a context-driven spatial attention modulation for the visual encoder—are plausible and potentially useful contributions to multi-turn LVLM reasoning. The proposed metrics (CCS, IFSR) and the MTIF dataset could also be valuable if properly specified and released. However, the manuscript itself states that the experimental results are 'fabricated', and no code, dataset definitions, metric formulas, or correct baseline references are provided. Under the reviewing instruction to treat all manuscript text as in-scope evidence, this self-admission removes the evidentiary basis for the central SOTA claim. The negative result is not a matter of presentation or missing details; it is a load-bearing integrity issue that cannot be repaired within the scope of this manuscript.

major comments (4)
  1. [Section 1, Introduction] The sentence 'Our fabricated experimental results demonstrate that CAMVR consistently achieves superior performance' is an explicit, self-authored admission that the reported experiments were manufactured rather than actually run. This directly undermines every empirical claim in the paper, including the central 'state-of-the-art' assertion in the Abstract and Table 1. Under the manuscript itself as evidence, Tables 1–5 and Figures 1–2 cannot be treated as real measurement results. This is a load-bearing flaw and grounds for rejection unless the authors can demonstrate the text is a typo or that all results have been independently reproduced from real experiments—neither of which appears in the current version.
  2. [Section 4.1.2 and 4.1.3] The MTIF dataset and the two novel metrics, CCS and IFSR, are the sole instruments for the MTIF and coherence claims in Tables 1, 2, and 4. Yet the dataset is only described as 'synthetic' and 'self-constructed' with no release, no construction protocol, no examples, and no size or statistics. CCS and IFSR are introduced only in prose ('designed to quantify...', 'evaluates the percentage...') with no formal definition, no formula, no algorithm, and no validation. Without these definitions, the IFSR and CCS numbers cannot be computed, reproduced, or interpreted. Any 'superior performance' on these instruments is therefore untestable.
  3. [Section 4.2 / References [9], [10]] The baseline comparison in Table 1 claims to use LLaVA-1.5 and Qwen-VL, but the cited references point to unrelated works: [9] is SlowFast-LLaVA-1.5 for video understanding and [10] is about shortcut learning in VLMs, not Qwen-VL. Consequently, the baseline configurations cannot be recovered from the manuscript, and the 'fair comparison' claimed in §4.1.1 is not reconstructible. Additionally, Tables 1–5 and Figures 1–2 report no error bars, confidence intervals, or statistical significance tests, so even the relative differences among the unverifiable numbers cannot be assessed.
  4. [Section 4.4.1] The human evaluation states 'We randomly selected 100 multi-turn dialogues (20 from each dataset: VisDial, A-OKVQA, MTIF, and 40 from a mixed set)'. This is a quota sample, not a random sample, and the description is internally inconsistent. More importantly, no inter-annotator agreement (e.g., Cohen's kappa or Krippendorff's alpha), no per-item variance, and no statistical test are reported for the 1–5 Likert ratings in Figure 1. The claimed 'clear preference' for CAMVR is therefore not quantitatively supported, even if the data were real.
minor comments (5)
  1. [Abstract and throughout] The component is named inconsistently: 'Adaptive Visual Focus Guidance (AVFG)' in the Abstract, 'Adaptive Visual Focus Guidance (A VFG)' elsewhere, and 'AVFG' in Tables. Use one symbol consistently.
  2. [References [13] and [14]] CIDEr and SPICE are cited to Becker et al. (harmonic power-flow) and Diamant et al. (conformalized splines), which are unrelated to the original CIDEr and SPICE papers. The proper references should be Vedantam et al. (2015) and Anderson et al. (2016).
  3. [Table 3, 'Fine Spatial' row] Missing space before '78.9' in 'A VFG (Fine Spatial, 14x14)78.9'. Also, the caption of Table 3 refers to 'A VFG granularity and integration strategies' but the table does not show integration strategies, only spatial resolutions.
  4. [Section 4.8, Table 5] Trainable Parameters are listed as 13,000M, 13,005M, etc. The unit 'M' is confusing because the total is ~13 billion (13,000M), not 13 million. Use 'B' or '×10^9' for clarity.
  5. [Figure 2] The x-axis label 'VCMU Memory Capacity (Nm)' uses 'Nm' while the text uses 'N_m'. The condition 'No VCMU' should be clearly separated from the N_m values (64, 128, 256, 512) in the axis and in the table reading.

Circularity Check

3 steps flagged

Central SOTA claim rests on internally generated numbers: Section 1 literally calls the experimental results 'fabricated,' and the only dataset/metrics used for validation are self-defined and unreleased.

specific steps
  1. other [Section 1, 'To empirically validate...' paragraph]
    "Our fabricated experimental results demonstrate that CAMVR consistently achieves superior performance across all evaluated datasets, significantly outperforming existing state-of-the-art baselines in multi-turn visual-textual reasoning tasks, particularly in terms of contextual understanding and instruction following success rate."

    The central claim (Abstract, Table 1) is that CAMVR achieves state-of-the-art performance. Its only support is the experimental section. But the Introduction explicitly labels those results fabricated. If taken literally, Tables 1-5 and the human evaluation are not evidence from real experiments; the claimed SOTA 'derivation' reduces to an internal assertion with no independent content.

  2. other [Section 4.1.2-4.1.3 (Datasets and Evaluation Metrics)]
    "our self-constructed Multi-Turn Instruction Following (MTIF) dataset ... Our proposed novel metric, Contextual Coherence Score (CCS), is designed to quantify a model's ability to maintain continuity, consistency, and correct utilization of context across multiple turns ... for the MTIF dataset, Instruction Following Success Rate (IFSR) evaluates the percentage of multi-step instructions that the model successfully and correctly executes or understands."

    CCS, IFSR, and the MTIF dataset are introduced and used in the same paper without formulas, released code, or external benchmark. The model's claimed superiority is therefore measured with the authors' own instruments inside the same manuscript; no external ground truth can confirm or falsify the numbers. The validation loop is closed within the paper.

  3. other [References [9], [10], [13], [14]]
    "[9] Mingze Xu ... Slowfast-llava-1.5 ... [10] Maurits Bleeker ... Demonstrating and reducing shortcuts in vision-language representation learning ... [13] Johanna Kristin Maria Becker ... Modelling of ac/dc interactions ... [14] Nathaniel Diamant ... Conformalized deep splines"

    The references anchoring the baselines and standard metrics do not point to the named works: [9] is not LLaVA-1.5, [10] is not Qwen-VL, [13] is not CIDEr, and [14] is not SPICE. The comparisons in Table 1 are therefore unattached to external methods and cannot be reconstructed, further reinforcing that the paper's empirical loop is self-contained. This is flagged as missing support rather than a formal circularity.

full rationale

CAMVR's mechanism equations (VCMU gated update, attention retrieval, AVFG spatial modulation, decoder concatenation) form a normal memory-plus-attention pipeline; no equation in Section 3 defines a target quantity in terms of itself, so the architectural derivation is not circular by construction. However, the empirical validation is closed: Section 1 states verbatim 'Our fabricated experimental results demonstrate...' — per the review rule, this in-scope text must be taken as asserting that the reported measurements were not obtained from real experiments. The claimed gains are measured with the authors' own unreleased MTIF dataset and unreported CCS/IFSR metrics, and the references for the baseline models and standard metrics point to unrelated papers. The central SOTA claim therefore rests on internally generated numbers rather than any independent check. Score 8: the result is forced by the paper's own internal assertion/self-defined instruments, though the core architecture itself is not a definitional tautology.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 4 invented entities

The paper introduces four entities and two fitted hyperparameters, but provides no external evidence for any of them. The central claim depends on an unreleased synthetic dataset and two unreleased metrics, making the evaluation internally generated. The only standard components are the mathematical update and attention equations.

free parameters (4)
  • VCMU memory capacity Nm = 256
    Selected by a validation sweep in Section 4.5, Figure 2; performance peaks at 256 and declines at 512.
  • AVFG attention-map resolution = 14x14
    Selected by validation sweep in Section 4.6, Table 3; fine spatial at 14x14 chosen as the best balance.
  • Contextual Coherence Score (CCS) = not disclosed
    Novel metric defined by the authors (Section 4.1.3); no formal formula is given, so the scores in Tables 1, 2, 4 are not independently computable.
  • Instruction Following Success Rate (IFSR) = not disclosed
    Novel metric for the authors' own MTIF dataset (Section 4.1.3); exact definition not given.
axioms (4)
  • domain assumption A gated recurrent update (Eqs. 6-8) can store and retrieve the cross-modal context needed for later turns.
    The central module VCMU is built on this GRU-like update; the paper provides no analysis of forgetting or capacity beyond an empirical Nm sweep.
  • standard math Scaled dot-product attention (Eq. 12) retrieves the most relevant memory entries for the current query.
    Used in the retrieval step; standard attention assumption.
  • domain assumption Element-wise multiplication of raw visual features by a spatial attention map (Eq. 16) sharpens reasoning without discarding needed information.
    The AVFG mechanism relies on this; no ablation against other modulation schemes beyond granularity of the map.
  • ad hoc to paper The MTIF dataset and CCS/IFSR metrics are valid and correctly implemented evaluation instruments.
    All three are introduced in this paper, are not released, and are used to make the central performance claim; the manuscript even labels the results 'fabricated'.
invented entities (4)
  • Visual-Textual Context Memory Unit (VCMU) no independent evidence
    purpose: Store and retrieve multimodal history across turns
    Described only in equations; no released weights, no independent benchmark showing it works.
  • Adaptive Visual Focus Guidance (AVFG) no independent evidence
    purpose: Modulate visual encoder attention with a spatial map
    No released code or module; all evidence is the paper's own unreleased experiments.
  • Contextual Coherence Score (CCS) no independent evidence
    purpose: Measure multi-turn coherence
    Defined only by name in Section 4.1.3; no formula or code; used to claim superiority.
  • Multi-Turn Instruction Following (MTIF) dataset no independent evidence
    purpose: Benchmark for multi-step visual instructions
    Synthetic dataset constructed by the authors, not released, with no generation details (Section 4.1.2).

pith-pipeline@v1.3.0-alltime-deepseek · 14386 in / 12355 out tokens · 108261 ms · 2026-08-05T05:13:02.942413+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Context-Aware Multi-Turn Visual-Textual Reasoning in LVLMs via Dynamic Memory and Adaptive Visual Guidance." pith.science (2026). https://pith.science/paper/UDMJCIN4

@misc{pith2026250905669,
  author       = {Pith},
  title        = {Pith review of: Context-Aware Multi-Turn Visual-Textual Reasoning in LVLMs via Dynamic Memory and Adaptive Visual Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UDMJCIN4}},
  note         = {Machine review of arXiv:2509.05669}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Current Large Language Models (LLMs) and Vision-Language Large Models (LVLMs) excel in single-turn tasks but face significant challenges in multi-turn interactions requiring deep contextual understanding and complex visual reasoning, often leading to fragmented reasoning, context loss, and hallucinations. To address these limitations, we propose Context-Aware Multi-Turn Visual Reasoning (CAMVR), a novel framework designed to empower LVLMs with robust and coherent multi-turn visual-textual inference capabilities. CAMVR introduces two key innovations: a Visual-Textual Context Memory Unit (VCMU), a dynamic read-write memory network that stores and manages critical visual features, textual semantic representations, and their cross-modal correspondences from each interaction turn; and an Adaptive Visual Focus Guidance (AVFG) mechanism, which leverages the VCMU's context to dynamically adjust the visual encoder's attention to contextually relevant image regions. Our multi-level reasoning integration strategy ensures that response generation is deeply coherent with both current inputs and accumulated historical context. Extensive experiments on challenging datasets, including VisDial, an adapted A-OKVQA, and our novel Multi-Turn Instruction Following (MTIF) dataset, demonstrate that CAMVR consistently achieves state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2509.05669 by Apiradee Boonmee, Weijie Shen, Xinrui Wang, Yuanqi Nie.

Figure 1
Figure 1. Figure 1: Average human evaluation scores (1-5 Likert scale) on various aspects of multi-turn visual-textual [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance of CAMVR with varying VCMU memory capacities ( [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Persistent Visual Memory: Sustaining Perception for Deep Generation in LVLMs

    cs.CV 2026-05 unverdicted novelty 6.0

    PVM adds a parallel branch to LVLMs that directly supplies visual embeddings to prevent attention decay over long generated sequences, yielding accuracy gains on reasoning tasks with minimal overhead.

  2. Persistent Visual Memory: Sustaining Perception for Deep Generation in LVLMs

    cs.CV 2026-05 unverdicted novelty 5.0

    PVM adds a parallel learnable branch to LVLMs that supplies visual embeddings on demand to structurally prevent attention decay and visual signal dilution during deep autoregressive generation.

Reference graph

Works this paper leans on

32 extracted references · 24 canonical work pages · cited by 1 Pith paper · 17 internal anchors

  1. [1]

    Chain-of-specificity: Enhancing task-specific constraint adherence in large language models

    Kaiwen Wei, Jiang Zhong, Hongzhi Zhang, Fuzheng Zhang, Di Zhang, Li Jin, Yue Yu, and Jingyuan Zhang. Chain-of-specificity: Enhancing task-specific constraint adherence in large language models. In Proceedings of the 31st International Conference on Computational Linguistics, pages 2401–2416, 2025

    work page 2025

  2. [2]

    Enhanc- ing code llms with reinforcement learning in code generation.arXiv preprint arXiv:2412.20367, 2024

    Junqiao Wang, Zeng Zhang, Yangfan He, Yuyang Song, Tianyu Shi, Yuchen Li, Hengyuan Xu, Kunyu Wu, Guangwu Qian, Qiuwu Chen, et al. Enhanc- ing code llms with reinforcement learning in code generation.arXiv preprint arXiv:2412.20367, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    Visual in-context learning for large vision- language models

    Yucheng Zhou, Xiang Li, Qianning Wang, and Jian- bing Shen. Visual in-context learning for large vision- language models. InFindings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 15890–15902. Association for Computational Linguistics, 2024

    work page 2024

  4. [4]

    Trigger is not sufficient: Exploiting frame-aware knowledge for implicit event argument extraction

    Kaiwen Wei, Xian Sun, Zequn Zhang, Jingyuan Zhang, Guo Zhi, and Li Jin. Trigger is not sufficient: Exploiting frame-aware knowledge for implicit event argument extraction. InProceedings of the 59th An- nual Meeting of the Association for Computational Linguistics and the 11th International Joint Confer- ence on Natural Language Processing (V olume 1: Long...

    work page 2021

  5. [5]

    Guide the many-to-one as- signment: Open information extraction via iou-aware optimal transport

    Kaiwen Wei, Yiran Yang, Li Jin, Xian Sun, Zequn Zhang, Jingyuan Zhang, Xiao Li, Linhao Zhang, Jin- tao Liu, and Guo Zhi. Guide the many-to-one as- signment: Open information extraction via iou-aware optimal transport. InProceedings of the 61st Annual Meeting of the Association for Computational Lin- guistics (V olume 1: Long Papers), pages 4971–4984, 2023

    work page 2023

  6. [6]

    Improving medical large vision-language models with abnormal-aware feedback.arXiv preprint arXiv:2501.01377, 2025

    Yucheng Zhou, Lingran Song, and Jianbing Shen. Improving medical large vision-language models with abnormal-aware feedback.arXiv preprint arXiv:2501.01377, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    ComplexBench-Edit: Benchmarking Complex Instruction-Driven Image Editing via Compositional Dependencies

    Chenglin Wang, Yucheng Zhou, Qianning Wang, Zhe Wang, and Kai Zhang. Complexbench-edit: Bench- marking complex instruction-driven image editing via compositional dependencies.arXiv preprint arXiv:2506.12830, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [8]

    Ce-doped srmno3 perovskite oxide with en- hanced oxygen reduction activity for hybrid sodium- air batteries.Science China Technological Sciences, 67(12):3956–3958, 2024

    Pan Liu, WenHui Niu, LanQing Zhao, Da Zhang, MinJie Hou, JiaQian Qin, Peng Dong, and Feng Liang. Ce-doped srmno3 perovskite oxide with en- hanced oxygen reduction activity for hybrid sodium- air batteries.Science China Technological Sciences, 67(12):3956–3958, 2024

    work page 2024

  9. [9]

    Slowfast-llava-1.5: A family of token-efficient video large language models for long-form video understanding.arXiv preprint arXiv:2503.18943v2, 2025

    Mingze Xu, Mingfei Gao, Shiyu Li, Jiasen Lu, Zhe Gan, Zhengfeng Lai, Meng Cao, Kai Kang, Yinfei Yang, and Afshin Dehghan. Slowfast-llava-1.5: A family of token-efficient video large language models for long-form video understanding.arXiv preprint arXiv:2503.18943v2, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Demonstrating and reducing shortcuts in vision-language representation learning.arXiv preprint arXiv:2402.17510v2, 2024

    Maurits Bleeker, Mariya Hendriksen, Andrew Yates, and Maarten de Rijke. Demonstrating and reducing shortcuts in vision-language representation learning.arXiv preprint arXiv:2402.17510v2, 2024. 13

    Pith/arXiv arXiv 2024

  11. [11]

    The Dialog Must Go On: Improving Visual Dialog via Generative Self-Training

    Gi-Cheon Kang, Sungdong Kim, Jin-Hwa Kim, Donghyun Kwak, and Byoung-Tak Zhang. The dialog must go on: Improving visual dialog via generative self-training.arXiv preprint arXiv:2205.12502v2, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  12. [12]

    Towards a Unified Model for Generating Answers and Explanations in Visual Question Answering

    Chenxi Whitehouse, Tillman Weyde, and Pranava Madhyastha. Towards a unified model for generating answers and explanations in visual question answer- ing.arXiv preprint arXiv:2301.10799v2, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Modelling of AC/DC Interactions of Converter-Interfaced Resources for Harmonic Power-Flow Studies in Microgrids

    Johanna Kristin Maria Becker, Andreas Martin Ket- tner, Yihui Zuo, Federico Cecati, Sante Pugliese, Marco Liserre, and Mario Paolone. Modelling of ac/dc interactions of converter-interfaced resources for harmonic power-flow studies in microgrids.arXiv preprint arXiv:2206.07332v3, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  14. [14]

    Conformalized Deep Splines for Optimal and Efficient Prediction Sets

    Nathaniel Diamant, Ehsan Hajiramezanali, Tom- maso Biancalani, and Gabriele Scalia. Conformal- ized deep splines for optimal and efficient prediction sets.arXiv preprint arXiv:2311.00774v1, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    DSBERT:Unsupervised Dialogue Structure learning with BERT

    Bingkun Chen, Shaobing Dai, Shenghua Zheng, Lei Liao, and Yang Li. Dsbert: Unsupervised dia- logue structure learning with bert.arXiv preprint arXiv:2111.04933v1, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  16. [16]

    Learning Reasoning Paths over Semantic Graphs for Video-grounded Dialogues

    Hung Le, Nancy F. Chen, and Steven C. H. Hoi. Learning reasoning paths over semantic graphs for video-grounded dialogues.arXiv preprint arXiv:2103.00820v2, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  17. [17]

    Mmcr: Advancing visual language model in multimodal multi-turn contextual reasoning.arXiv preprint arXiv:2503.18533v1, 2025

    Dawei Yan, Yang Li, Qing-Guo Chen, Weihua Luo, Peng Wang, Haokui Zhang, and Chunhua Shen. Mmcr: Advancing visual language model in multimodal multi-turn contextual reasoning.arXiv preprint arXiv:2503.18533v1, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Question Relevance in VQA: Identifying Non-Visual And False-Premise Questions

    Arijit Ray, Gordon Christie, Mohit Bansal, Dhruv Batra, and Devi Parikh. Question relevance in vqa: Identifying non-visual and false-premise questions. arXiv preprint arXiv:1606.06622v3, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  19. [19]

    Commonsense Reasoning for Conversational AI: A Survey of the State of the Art

    Christopher Richardson and Larry Heck. Common- sense reasoning for conversational ai: A survey of the state of the art.arXiv preprint arXiv:2302.07926v1, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  20. [20]

    Evaluating scenario-based decision-making for interactive autonomous driving using rational criteria: A survey.arXiv preprint arXiv:2501.01886, 2025

    Zhen Tian, Zhihao Lin, Dezong Zhao, Wenjing Zhao, David Flynn, Shuja Ansari, and Chongfeng Wei. Evaluating scenario-based decision-making for interactive autonomous driving using rational criteria: A survey.arXiv preprint arXiv:2501.01886, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Safety-critical multi-agent mcts for mixed traffic coordination at unsignalized intersections.IEEE Transactions on In- telligent Transportation Systems, pages 1–15, 2025

    Zhihao Lin, Jianglin Lan, Christos Anagnostopou- los, Zhen Tian, and David Flynn. Safety-critical multi-agent mcts for mixed traffic coordination at unsignalized intersections.IEEE Transactions on In- telligent Transportation Systems, pages 1–15, 2025

    work page 2025

  22. [22]

    Multi-agent monte carlo tree search for safe decision making at unsignal- ized intersections

    Zhihao Lin, Jianglin Lan, Christos Anagnostopou- los, Zhen Tian, and David Flynn. Multi-agent monte carlo tree search for safe decision making at unsignal- ized intersections. 2025

    work page 2025

  23. [23]

    DialogueReason: Rule-Based RL Sparks Dialogue Reasoning in LLMs

    Yubo Shu, Zhewei Huang, Xin Wu, Chen Hu, Shuchang Zhou, and Daxin Jiang. Dialoguerea- son: Rule-based rl sparks dialogue reasoning in llms. arXiv preprint arXiv:2505.07049v1, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    GRADE: Automatic Graph-Enhanced Coherence Metric for Evaluating Open-Domain Dialogue Systems

    Lishan Huang, Zheng Ye, Jinghui Qin, Liang Lin, and Xiaodan Liang. Grade: Automatic graph-enhanced coherence metric for evaluating open-domain dialogue systems.arXiv preprint arXiv:2010.03994v1, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  25. [25]

    Score: Story co- herence and retrieval enhancement for ai narratives

    Qiang Yi, Yangfan He, Jianhui Wang, Xinyuan Song, Shiyao Qian, Xinhang Yuan, Miao Zhang, Li Sun, Keqin Li, Kuan Lu, et al. Score: Story co- herence and retrieval enhancement for ai narratives. arXiv preprint arXiv:2503.23512, 2025

    arXiv 2025

  26. [26]

    Beyond the Granularity: Multi-Perspective Dialogue Collaborative Selection for Dialogue State Tracking

    Jinyu Guo, Kai Shuang, Jijie Li, Zihan Wang, and Yixuan Liu. Beyond the granularity: Multi-perspective dialogue collaborative selection for dialogue state tracking.arXiv preprint arXiv:2205.10059v1, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  27. [27]

    Memory and attention in deep learning

    Hung Le. Memory and attention in deep learning. arXiv preprint arXiv:2107.01390v1, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  28. [28]

    Memory-augmented transformers: A sys- tematic review from neuroscience principles to enhanced model architectures.arXiv preprint arXiv:2508.10824v2, 2025

    Parsa Omidi, Xingshuai Huang, Axel Laborieux, Bahareh Nikpour, Tianyu Shi, and Armaghan Es- haghi. Memory-augmented transformers: A sys- tematic review from neuroscience principles to enhanced model architectures.arXiv preprint arXiv:2508.10824v2, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Improved Attention Models for Memory Augmented Neural Network Adaptive Controllers

    Deepan Muthirayan, Scott Nivison, and Pramod P. Khargonekar. Improved attention models for mem- ory augmented neural network adaptive controllers. arXiv preprint arXiv:1910.01189v7, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  30. [30]

    Attention is All You Need Until You Need Retention

    M. Murat Yaslioglu. Attention is all you need until you need retention.arXiv preprint arXiv:2501.09166v1, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Ruoyang Hu and Robert A. Jacobs. A neural net- work model of spatial and feature-based attention. arXiv preprint arXiv:2506.05487v1, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Sequence-to-Sequence Models with Attention Mechanistically Map to the Architecture of Human Memory Search

    Nikolaus Salvatore and Qiong Zhang. Sequence-to- sequence models with attention mechanistically map to the architecture of human memory search.arXiv preprint arXiv:2506.17424v1, 2025. 14

    work page internal anchor Pith review Pith/arXiv arXiv 2025