Pith. sign in

REVIEW 3 major objections 7 minor 38 references

Inference-time reliability control halves VLM hallucinations on museum queries

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 · glm-5.2

2026-07-08 07:47 UTC pith:ZVBGTI4Z

load-bearing objection Engineering paper with a real working system, but the GRPO-style selection reward overlaps with evaluation metrics, making the headline reliability gains partly circular. the 3 major comments →

arxiv 2607.06374 v1 pith:ZVBGTI4Z submitted 2026-07-07 cs.CV

VaseMuseum: Digital Intelligent Museum for Ancient Greek Pottery

Jiazi Wang , Nonghai Zhang , Qiushi Xie , Zeyu Zhang , Yufeng Chen , Yang Zhao , Ling Shao , Hao Tang This is my paper
classification cs.CV
keywords evidencedigitalmuseumvasemuseumancientanswerscontrolgreek
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.

This paper argues that the path to trustworthy AI in cultural-heritage settings lies not in building bigger vision-language models or curating larger domain-specific training sets, but in wrapping an off-the-shelf VLM with inference-time evidence controls. The authors present VaseMuseum, a system combining a virtual museum interface for ancient Greek pottery with a reasoning agent called VaseAgent. VaseAgent handles both 2D images and 3D artifact models, and when visual evidence is insufficient to answer a visitor's question, it retrieves external knowledge through an iterative search loop. The central claim is that two lightweight control layers—source-level filtering that removes unreliable or redundant search results before they enter the model's context, and response-level auditing that checks generated claims against the collected evidence pool—substantially improve answer trustworthiness without any gradient updates to the VLM. When extra compute is available, a training-free GRPO-style selector samples multiple answer trajectories and reranks them by reliability. Experiments show the full system cuts hallucination on knowledge-intensive queries from 44% to 22%, raises citation link validity from 29% to 60%, and improves neutrality under ambiguous evidence from 2.91 to 3.75 on a 0–5 scale, compared with a search-enabled VLM baseline.

Core claim

The paper's central finding is that inference-time reliability control—specifically, filtering retrieved evidence before generation and auditing draft answers against that evidence before output—yields larger reliability gains on knowledge-intensive cultural-heritage queries than unstructured retrieval alone. The mechanism carrying this argument is a two-stage evidence pipeline: source control applies validity filters, relevance-quality scoring, and diversity-aware selection to raw search hits, producing a compact evidence pool; response control then decomposes the draft answer into claim units, checks each claim against the pool for lexical support and cross-source conflict, and computes a

What carries the argument

The load-bearing mechanism is the response confidence score ψ (Eq. 4), a weighted combination of claim coverage, consistency, multi-source support, evidence-pool non-emptiness, and conflict penalties. When ψ falls below a threshold (0.52), the system switches to 'uncertain mode' and prepends an evidence-bounded preamble rather than returning an unsupported answer. A secondary mechanism is the GRPO-style selector (Eq. 5), which samples K independent controlled trajectories and picks the one with the highest aggregate reliability score R(·), measuring link validity, claim-evidence support, neutrality, and conflict awareness—all without updating the VLM backbone.

Load-bearing premise

The response confidence score ψ—a weighted sum of claim coverage, consistency, multi-source support, and conflict penalties with hand-tuned weights and a fixed threshold—accurately captures whether a generated answer is trustworthy. Both the response-control module and the GRPO-style selection mechanism depend on this score, so if it does not track real reliability, the system's reported improvements could reflect its own scoring artifacts rather than genuine gains.

What would settle it

If the response confidence score ψ were replaced by an independent trustworthiness measure (e.g., expert human ratings of answer accuracy and citation validity on held-out queries), and VaseAgent's improvements over the search-enabled baseline disappeared or reversed, the paper's central claim that inference-time reliability control improves trustworthiness would be undermined.

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

If this is right

  • Inference-time evidence filtering and answer auditing could be applied to any domain where VLMs must cite external sources—medical imaging, legal document analysis, or scientific figure explanation—without requiring domain-specific retraining.
  • The finding that uncontrolled search increases hallucination (from 54% to 79% on visual-only queries) while controlled search reduces it suggests that naive RAG is actively harmful in specialized domains, and that source-quality gates are a necessary complement to retrieval.
  • The trade-off between accuracy-oriented selection and reliability-oriented selection (Table 4) implies that museum-deployment systems may need to sacrifice some answer completeness to maintain citation integrity—a design tension relevant to any expert-assistant system.
  • The modularity of the source-control and response-control layers means they could be retrofitted onto existing museum chatbot deployments without changing the underlying VLM, offering a low-cost upgrade path for institutions with limited compute budgets.

Where Pith is reading between the lines

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

  • The response confidence score ψ uses hand-tuned weights (0.40, 0.30, 0.20, 0.10) and a threshold (0.52) introduced without sensitivity analysis. If ψ does not correlate with actual answer trustworthiness, the reported improvements could be artifacts of the system's own scoring function rather than genuine reliability gains.
  • The GRPO-style selection reward R(·) and the response-control reliability score share the same construction, creating a potential circularity: the system selects answers that score well on its own metric, then reports improvement on that same metric.
  • The evaluation set contains only 100 instances per task category, drawn from a single pottery collection linked to one external database (LIMC). Whether the reliability-control pipeline generalizes to other artifact types, languages, or knowledge sources with different coverage and conflict patterns remains untested.
  • The neutrality metric is human-rated on a 0–5 scale without reported inter-annotator agreement, leaving open whether the neutrality gains would replicate under independent evaluation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. This paper presents VaseMuseum, a multimodal agent framework for interactive digital museums of ancient Greek pottery. The system combines a virtual museum interface with VaseAgent, which performs vision-language reasoning, DeepResearch-style external knowledge retrieval, source-level evidence filtering, response-level answer calibration, and a training-free GRPO-style response selection mechanism. The authors evaluate the system on 300 instances across three query types (visual-only, visual-plus-knowledge, ambiguous) and report improvements in citation validity, hallucination reduction, and neutrality under ambiguity compared to search-enabled VLM baselines. The code and website are publicly available.

Significance. The paper addresses a practically important problem: reliable VLM assistance in cultural-heritage settings where evidence is incomplete, noisy, or contested. The modular, training-free design is a strength, as is the public release of code and an interactive website. The idea of wrapping retrieval-augmented generation with deterministic source and response control layers is reasonable and deployable. However, the central empirical claims are undermined by a structural circularity in the evaluation design (the GRPO selection reward shares components with the evaluation metrics) and by a substantial accuracy degradation that is not adequately analyzed. These issues are load-bearing for the paper's main contributions and require revision.

major comments (3)
  1. §3.4, Eq. (5) and Table 3: The GRPO-style selector's reward R(·) is described as aggregating 'link validity, claim–evidence support, neutrality under insufficient evidence, and conflict awareness' (§3.4). The evaluation metrics in Tables 2–3 include link validity and neutrality. This means the improvement from VaseAgent (w/o GRPO) to VaseAgent (Full) on link validity (44.07→60.00) and neutrality (3.04→3.75) is close to guaranteed by construction: the selector picks the candidate that maximizes a score containing these same dimensions, then the evaluation measures those same quantities. The paper needs to either (a) report evaluation metrics that are not components of R(·), or (b) provide an independent evaluation protocol (e.g., human expert ratings collected without knowledge of R(·) scores) to demonstrate that the improvements reflect genuine reliability gains rather than the selector-
  2. Table 2: VaseAgent (Full) degrades answer accuracy substantially compared to Qwen3-VL-8B (Direct): V-Only drops from 44% to 24%, V+K from 69% to 54%. The paper frames this as a reliability-accuracy trade-off but does not analyze whether this accuracy loss is acceptable for deployment. For a museum assistant, returning incorrect answers 76% of the time on visual-only questions is a serious concern. The paper should discuss this trade-off explicitly, report statistical significance given the small evaluation set (100 instances per split), and ideally provide intermediate operating points (e.g., different τ_conf values) to show whether better accuracy-reliability trade-offs exist.
  3. §3.3, Eq. (4) and §3.4: The response confidence score ψ uses hand-tuned weights (0.40, 0.30, 0.20, 0.10) and threshold (τ_conf = 0.52), and the GRPO selection reward R(·) also depends on these components. No sensitivity analysis, ablation over weight choices, or theoretical justification for these values is provided. Given that ψ directly controls whether the system enters 'uncertain mode' and that R(·) drives candidate selection, the robustness of the reported improvements to these parameter choices is unclear. A sensitivity analysis over at least the threshold τ_conf and the weight on c_cov would substantially strengthen the claims.
minor comments (7)
  1. §3.5: The evaluation set uses 100 instances per split. The paper should report inter-annotator agreement for the human-rated metrics (neutrality, groundedness) and confidence intervals for all reported numbers.
  2. Table 2: The 'G' (groundedness) metric is described in Table 1 but its computation is not detailed. Is it binary or continuous? How is it computed—automatically or by human raters?
  3. Figure 8b only shows hallucination rates for V-Only, not for V+K or Ambiguous. The V+K hallucination results (22% vs 44%) are central to the paper's claim and should be visualized for completeness.
  4. §3.3, Eq. (1): The notation 'h ∈ H_val ⇐⇒ sval(h) ≥ 0.40 and a(h) ≥ 0.50' is unusual. The double-arrow should be clarified or replaced with standard set-membership notation.
  5. The paper mentions 'Algorithm 7' and 'Algorithm 5' in the text but the figures are labeled 'Fig. 5' and 'Fig. 7'. The numbering inconsistency should be corrected.
  6. §4.1: The baselines use only Qwen3-VL-8B. Including at least one additional VLM backbone would strengthen the generality of the claims, which currently may be specific to one model family.
  7. The paper states 'K=4 controlled trajectories' but does not report variance across runs. Given stochastic sampling, reporting mean ± std over multiple seeds would be informative.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee identifies three major concerns: (1) structural circularity between the GRPO selection reward and evaluation metrics, (2) substantial accuracy degradation that is insufficiently analyzed, and (3) lack of sensitivity analysis for hand-tuned parameters. We agree that all three points are legitimate and require revision. Below we address each in turn.

read point-by-point responses

  1. Referee: GRPO selector reward shares components with evaluation metrics (link validity, neutrality), creating structural circularity. The improvement from w/o GRPO to Full on these metrics may be guaranteed by construction.

    Authors: The referee is correct that there is partial overlap between the reward components in R(·) and the evaluation metrics for link validity and neutrality. We acknowledge this as a genuine structural concern. However, we note three mitigating points. First, the reward R(·) aggregates multiple sub-scores (link validity, claim–evidence support, neutrality, conflict awareness) into a single composite, so the selector optimizes the aggregate, not any individual metric. The evaluation then measures these dimensions separately. The mapping from composite optimization to individual metric improvement is not mechanically guaranteed—e.g., a candidate could score high on conflict awareness but low on link validity, and the selector might still pick it if the composite is highest. Second, the improvement from Qwen+Search to VaseAgent (w/o GRPO) on link validity (28.89→44.07) and neutrality (2.91→3.04) is entirely independent of the GRPO selector, since the w/o GRPO variant does not use R(·) at all. This demonstrates that source and response controls alone produce genuine gains. Third, the accuracy degradation from Direct to Full (44→24 on V-Only) actually argues against circularity: if the selector were trivially optimizing evaluation metrics, we would not see accuracy drop. That said, we agree that the GRPO-specific improvement (44.07→60.00 on link validity, 3.04→3.75 on neutrality) is the part most vulnerable to the circularity critique, and the current manuscript does not adequately address this. We will revise the paper to: (a) add an explicit discussion of the overlap between R(·) components and evaluation metrics, acknowledging the limitation; (b) report the accuracy-oriented GRPO ablation (Table 4, which uses a different reward that does not include link validity or neutrality) revision: yes

  2. Referee: VaseAgent (Full) degrades answer accuracy substantially (V-Only 44→24%, V+K 69→54%). The paper does not analyze whether this trade-off is acceptable, does not report statistical significance, and does not explore intermediate operating points.

    Authors: The referee is correct that the accuracy degradation is substantial and insufficiently analyzed in the current manuscript. We will address this through several revisions. First, we will add an explicit discussion of the accuracy–reliability trade-off. The core design philosophy of VaseAgent is that in a museum setting, a confidently wrong answer with a fabricated citation is more harmful than a cautious answer that hedges or directs the user to authoritative sources. The 76% 'incorrect' rate on V-Only is partly inflated by the evaluation protocol: when VaseAgent enters uncertain mode and provides an evidence-bounded preamble rather than a direct answer, this is scored as incorrect under strict accuracy matching, even though such responses may be more appropriate for deployment. We will clarify this scoring behavior and also report a 'deployment-appropriateness' metric that credits cautious responses. Second, we will add statistical significance tests (bootstrap confidence intervals or McNemar's test) given the 100-instance splits. Third, we will add intermediate operating points by varying τ_conf (e.g., 0.40, 0.45, 0.52, 0.60) to show the accuracy–reliability frontier. We note that the accuracy-oriented GRPO variant in Table 4 already demonstrates that a different reward formulation yields V+K accuracy of 59% (vs. 54% for Full) at the cost of lower link validity (28.89 vs. 60.00), confirming that intermediate trade-off points exist and are controllable. We agree this should be discussed more prominently. revision: yes

  3. Referee: Hand-tuned weights (0.40, 0.30, 0.20, 0.10) and threshold (τ_conf = 0.52) in ψ lack sensitivity analysis, ablation, or theoretical justification. Robustness of improvements to these choices is unclear.

    Authors: The referee is correct that no sensitivity analysis is currently provided for the confidence score weights or the τ_conf threshold. We will add this analysis in the revision. Specifically, we will: (a) sweep τ_conf over at least four values (e.g., 0.40, 0.45, 0.52, 0.60) and report the effect on accuracy, hallucination rate, link validity, and neutrality; (b) perturb the weight on c_cov (currently 0.40) over a range (e.g., 0.25, 0.35, 0.40, 0.50) while proportionally adjusting the other weights, and report the same metrics; (c) add a brief discussion of why these particular values were chosen (they were selected on the held-out paraphrase set described in §3.5, not on the test set, though we will make this explicit). We acknowledge that we cannot provide a strong theoretical justification for the exact weight values—they are empirically motivated—and we will state this honestly. If the sensitivity analysis reveals that results are fragile to these choices, we will report that transparently. Our expectation, based on the ablation in Table 4 showing that removing entire reward components still yields reasonable (if degraded) performance, is that the system is moderately robust to weight perturbations, but we commit to verifying this empirically. revision: yes

Circularity Check

0 steps flagged

GRPO selector's reward R(·) is constructed from the same dimensions (link validity, neutrality, evidence support) used as evaluation metrics in Tables 2–3, making the reported improvements partially tautological rather than independent evidence of reliability gains.

full rationale

The paper's GRPO-style selector (Eq. 5) scores candidate trajectories using R(·), which aggregates link validity, claim–evidence support, neutrality, and conflict awareness (§3.4). The evaluation metrics in Tables 2–3 include link validity and neutrality—the exact same quantities. The improvement from VaseAgent (w/o GRPO) to VaseAgent (Full) on link validity (44.07→60.00) and neutrality (3.04→3.75) is thus partially guaranteed by construction: the selector picks the candidate that maximizes a score containing those metrics, then the evaluation measures those same metrics. However, hallucination rate (38→22%) is judged against external ground truth (exhibit observation, metadata, retrieved evidence) rather than directly against R(·), so not all reported improvements are circular. The partial overlap between the selection reward and the evaluation metrics constitutes a genuine but incomplete circularity: the link-validity and neutrality gains are largely artifacts of the selector optimizing its own reward components, while the hallucination reduction retains independent evidential force. The self-citations to VaseVQA/VaseVQA-3D ([1],[2]) are for dataset/benchmark construction, not load-bearing for the central reliability claim, and thus do not constitute circularity on their own. The hand-tuned weights in ψ (Eq. 4) and R(·) lack sensitivity analysis, but that is a correctness risk rather than circularity per se. The circularity is specifically in the reward-metric overlap for two of the four reported metrics. Score 6 reflects that one or more 'predictions' (link validity, neutrality improvements) reduce by construction while others (hallucination, accuracy) retain independent content, making the result partially but not fully circular.

Axiom & Free-Parameter Ledger

13 free parameters · 3 axioms · 1 invented entities

The paper introduces no new entities in the physics or mathematical sense. The 'GRPO-style' mechanism is a relabeling of existing techniques rather than a new entity.

free parameters (13)
  • sval weights = 0.55, 0.45
    Weights for accessibility proxy and text-sufficiency score in Eq. 1, set ad hoc without justification.
  • sval threshold = 0.40
    Validity threshold in Eq. 1, chosen without sensitivity analysis.
  • a(h) threshold = 0.50
    Minimum accessibility score in Eq. 1, ad hoc.
  • ϕ_pre weights = 0.35, 0.25, 0.20
    Weights for relevance, source-quality, and accessibility in Eq. 2, ad hoc.
  • τ_keep = 0.22
    Pre-diversity score threshold in Eq. 2, ad hoc.
  • N_src = 5
    Maximum number of selected hits, ad hoc.
  • λ_MMR = 0.72
    MMR relevance-diversity trade-off in Eq. 3, ad hoc.
  • β_dom coefficient = 0.04
    Domain diversity bonus coefficient in Eq. 3, ad hoc.
  • ψ weights = 0.40, 0.30, 0.20, 0.10
    Weights for coverage, consistency, multi-source, and evidence-non-empty in Eq. 4, ad hoc.
  • ψ conflict penalty = 0.35, 0.12
    Conflict penalty cap and per-conflict penalty in Eq. 4, ad hoc.
  • τ_conf = 0.52
    Confidence threshold for uncertain mode in Eq. 4, ad hoc.
  • τ_sup = 0.08
    Lexical overlap threshold for claim-evidence support, ad hoc.
  • K (group size) = 4
    Number of sampled trajectories for GRPO-style selection, chosen empirically (Fig. 10b) but on a small evaluation set.
axioms (3)
  • ad hoc to paper The response confidence score ψ (Eq. 4) is a valid proxy for answer trustworthiness
    The entire reliability-control framework depends on ψ accurately reflecting whether an answer is trustworthy. The weights and thresholds are hand-set without calibration against ground truth.
  • ad hoc to paper Lexical overlap above 0.08 (Jaccard) indicates claim-evidence support
    §3.3 Response Control uses τ_sup = 0.08 as the threshold for marking evidence as supporting a claim. This is a very low threshold and the choice is not justified.
  • domain assumption The VaseVQA-3D evaluation set is representative of real museum interactions
    §3.5 constructs the evaluation set from 518 LIMC-linked vases with 100 instances per task type. Whether this sample represents the distribution of real visitor queries is not validated.
invented entities (1)
  • None independent evidence
    purpose: N/A
    The paper introduces no new physical entities, particles, or forces. It is a system architecture paper.

pith-pipeline@v1.1.0-glm · 17198 in / 3908 out tokens · 228771 ms · 2026-07-08T07:47:57.064529+00:00 · methodology

0 comments
read the original abstract

Vision-language models (VLMs) have made interactive digital museums increasingly feasible by connecting 3D digitization with natural-language artifact exploration. However, in cultural heritage domains such as ancient Greek pottery, reliable VLM assistance is limited by two challenges. First, open-ended interpretation requires grounding fine-grained 2D/3D visual evidence in specialized curatorial knowledge, yet the retrieval process may introduce weak sources and unverifiable references. Second, when the available evidence is incomplete, noisy, or ambiguous, VLMs often produce confident but unsupported answers instead of calibrated uncertainty. To address these challenges, we propose VaseMuseum, a lightweight and modular multimodal agent framework for intelligent digital museums of ancient Greek pottery. VaseMuseum combines an interactive virtual museum with VaseAgent, which supports both 2D images and 3D artifacts through multimodal perception, 3D-aware reasoning, external knowledge retrieval, and inference-time reliability control. Specifically, VaseAgent retrieves evidence from authoritative web and museum knowledge sources, and source-level control selects diverse and verifiable evidence before generation. Meanwhile, response-level control checks generated claims against the evidence pool and encourages neutral, evidence-bounded answers when support is insufficient or conflicting. Moreover, a training-free GRPO-style selection mechanism favors responses with valid references and calibrated confidence without updating the VLM backbone. Experiments in a realistic digital museum simulation show that VaseMuseum improves citation validity, reduces hallucinations on knowledge-intensive queries, and produces more neutral answers under ambiguity compared with search-enabled VLM baselines.

Figures

Figures reproduced from arXiv: 2607.06374 by Hao Tang, Jiazi Wang, Ling Shao, Nonghai Zhang, Qiushi Xie, Yang Zhao, Yufeng Chen, Zeyu Zhang.

Figure 1
Figure 1. Figure 1: VaseVQA series development. historical or archaeological evidence from external sources. Retrieval may then bring low-quality pages, incomplete records, or unverifiable citations into the context. Second, cultural-heritage evidence is often partial, disputed, or underspecified. As a result, generic VLMs may produce fluent but unsupported explanations when expert cau￾tion is needed. Our previous VaseVQA ser… view at source ↗
Figure 2
Figure 2. Figure 2: System architecture of VaseMuseum. The virtual museum provides artifact observations and user interactions, while [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Virtual museum interaction workflow. After a user selects an exhibit and poses a question, VaseAgent analyzes [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Virtual museum interface. Users can inspect an [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pseudo-code of online VaseAgent inference. The agent alternates between VLM reasoning and tool execution; each search payload is filtered by source control before en￾tering the dialogue context, and the final draft is calibrated by response control. and a text-sufficiency score c(h) from the title and snippet. As defined in Eq. (1), a hit is retained only when sval(h) = 0.55 a(h) + 0.45 c(h), h ∈ Hval ⇐⇒ s… view at source ↗
Figure 6
Figure 6. Figure 6: VaseAgent reliability framework. Source control removes invalid, low-quality, and redundant search results [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pseudo-code of training-free GRPO-style response selection. When additional inference budget is available, multiple controlled trajectories are sampled with different seeds and reranked by reliability verifiers for link validity, evidence support, neutrality, and conflict awareness, with￾out updating the VLM. 3.5 Evaluation Data Construction We construct the evaluation set from the curated image collection… view at source ↗
Figure 8
Figure 8. Figure 8: Main comparison across baselines. The figure con [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Reliability metrics. VaseAgent improves external link [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ablation, scalability, and multi-dimensional evaluation of VaseAgent. [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Task-specific performance breakdown across V-Only, [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 12 internal anchors

  1. [1]

    VaseVQA-3D: Benchmarking 3D VLMs on ancient Greek pottery,

    N. Zhang, Z. Zhang, J. Wang, Y. Zhao, and H. Tang, “VaseVQA-3D: Benchmarking 3D VLMs on ancient Greek pottery,”arXiv preprint arXiv:2510.04479, 2025. 1, 2, 4

    work page arXiv 2025

  2. [2]

    VaseVQA: Multimodal agent and benchmark for ancient Greek pottery,

    J. Ge, T. Cheng, B. Wu, Z. Zhang, S. Huang, J. Bishop, G. Shepherd, M. Fang, L. Chen, and Y. Zhao, “VaseVQA: Multimodal agent and benchmark for ancient Greek pottery,” inFindings of the Association for Computational Linguistics: EACL 2026, 2026, pp. 1154–1167. 1, 3

    work page 2026

  3. [3]

    Colin Campbell, Sean Sands, Brent McFerran, and Alexis Mavrommatis

    Y. Cai, S. Cai, Y. Shi, Z. Xu, L. Chen, Y. Qin, X. Tan, G. Li, Z. Li, H. Linet al., “Training-free group relative policy optimization,” arXiv preprint arXiv:2510.08191, 2025. 2

    work page arXiv 2025

  4. [4]

    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

    C. Snell, J. Lee, K. Xu, and A. Kumar, “Scaling LLM test-time compute optimally can be more effective than scaling model parameters,”arXiv preprint arXiv:2408.03314, 2024. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Gtpo and grpo-s: Token and sequence-level reward shaping with policy entropy.arXiv preprint arXiv:2508.04349, 2025

    H. Tan, Z. Wang, J. Pan, J. Lin, H. Wang, Y. Wu, T. Chen, Z. Zheng, Z. Tang, and H. Yang, “GTPO and GRPO-S: Token and sequence-level reward shaping with policy entropy,”arXiv preprint arXiv:2508.04349, 2025. 2

    work page arXiv 2025

  6. [6]

    VQArt-Bench: A semantically rich VQA benchmark for art and cultural heritage,

    A. Alfarano, L. Venturoli, and D. N. Del Castillo, “VQArt-Bench: A semantically rich VQA benchmark for art and cultural heritage,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 396–406. 3

    work page 2025

  7. [7]

    A comparative study of vision language models for italian cultural heritage,

    C. Vitaloni, D. Shullani, and D. Baracchi, “A comparative study of vision language models for italian cultural heritage,”Heritage, vol. 8, no. 3, p. 95, 2025. 3

    work page 2025

  8. [8]

    Scalable 3D cap- tioning with pretrained models,

    T. Luo, C. Rockwell, H. Lee, and J. Johnson, “Scalable 3D cap- tioning with pretrained models,”Advances in Neural Information Processing Systems, vol. 36, pp. 75 307–75 337, 2023. 3

    work page 2023

  9. [9]

    View selection for 3D captioning via diffusion ranking,

    T. Luo, J. Johnson, and H. Lee, “View selection for 3D captioning via diffusion ranking,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 180–197. 3

    work page 2024

  10. [10]

    3D-R1: Enhancing Reasoning in 3D VLMs for Unified Scene Understanding

    T. Huang, Z. Zhang, and H. Tang, “3D-R1: Enhancing reasoning in 3D VLMs for unified scene understanding,”arXiv preprint arXiv:2507.23478, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  11. [11]

    LLaVA-3D: A simple yet effective pathway to empowering LMMs with 3D capabilities,

    C. Zhu, T. Wang, W. Zhang, J. Pang, and X. Liu, “LLaVA-3D: A simple yet effective pathway to empowering LMMs with 3D capabilities,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 4295–4305. 3

    work page 2025

  12. [12]

    3DRS: MLLMs need 3D- aware representation supervision for scene understanding,

    X. Huang, J. Wu, Q. Xie, and K. Han, “3DRS: MLLMs need 3D- aware representation supervision for scene understanding,”arXiv preprint arXiv:2506.01946, 2025. 3

    work page arXiv 2025

  13. [13]

    ShapeLLM: Universal 3D object understanding for embodied interaction,

    Z. Qi, R. Dong, S. Zhang, H. Geng, C. Han, Z. Ge, L. Yi, and K. Ma, “ShapeLLM: Universal 3D object understanding for embodied interaction,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 214–238. 3

    work page 2024

  14. [14]

    Prometheus: 3D-aware latent diffusion models for feed-forward text-to-3D scene generation,

    Y. Yang, J. Shao, X. Li, Y. Shen, A. Geiger, and Y. Liao, “Prometheus: 3D-aware latent diffusion models for feed-forward text-to-3D scene generation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 2857–2869. 3

    work page 2025

  15. [15]

    ReAct: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y. Cao, “ReAct: Synergizing reasoning and acting in language models,” inThe eleventh international conference on learning repre- sentations, 2022. 3

    work page 2022

  16. [16]

    A Survey of LLM-based Deep Search Agents: Paradigm, Optimization, Evaluation, and Challenges

    Y. Xi, J. Lin, Y. Xiao, Z. Zhou, R. Shan, T. Gao, J. Zhu, W. Liu, Y. Yu, and W. Zhang, “A survey of LLM-based deep search agents: Paradigm, optimization, evaluation, and challenges,”arXiv preprint arXiv:2508.05668, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Search-o1: Agentic search-enhanced large reasoning models,

    X. Li, G. Dong, J. Jin, Y. Zhang, Y. Zhou, Y. Zhu, P . Zhang, and Z. Dou, “Search-o1: Agentic search-enhanced large reasoning models,” inProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 2025, pp. 5420–5438. 3

    work page 2025

  18. [18]

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

    B. Jin, H. Zeng, Z. Yue, J. Yoon, S. Arik, D. Wang, H. Zamani, and J. Han, “Search-R1: Training LLMs to reason and lever- age search engines with reinforcement learning,”arXiv preprint arXiv:2503.09516, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    WebGPT: Browser-assisted question-answering with human feedback

    R. Nakano, J. Hilton, S. Balaji, J. Wu, L. Ouyang, C. Kim, C. Hesse, S. Jain, V . Kosaraju, W. Saunderset al., “WebGPT: Browser- assisted question-answering with human feedback,”arXiv preprint arXiv:2112.09332, 2021. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  20. [20]

    Deep Research: A Survey of Autonomous Research Agents

    W. Zhang, X. Li, Y. Zhang, P . Jia, Y. Wang, H. Guo, Y. Liu, and X. Zhao, “Deep research: A survey of autonomous research agents,”arXiv preprint arXiv:2508.12752, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Deep Research Agents: A Systematic Examination And Roadmap

    Y. Huang, Y. Chen, H. Zhang, K. Li, H. Zhou, M. Fang, L. Yang, X. Li, L. Shang, S. Xuet al., “Deep research agents: A systematic examination and roadmap,”arXiv preprint arXiv:2506.18096, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    MTA-Agent: An Open Recipe for Multimodal Deep Search Agents

    X. Peng, C. Qin, A. Yan, X. Yang, Z. Chen, R. Xu, and C.-S. Wu, “MTA-Agent: An open recipe for multimodal deep search agents,” arXiv preprint arXiv:2604.06376, 2026. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    MiroFlow: Towards high- performance and robust open-source agent framework for general deep research tasks,

    S. Su, S. Xing, X. Dong, M. Zhong, B. Wang, X. Zhu, Y. Chen, W. Wang, Y. Deng, P . Zhuet al., “MiroFlow: Towards high- performance and robust open-source agent framework for general deep research tasks,”arXiv preprint arXiv:2602.22808, 2026. 3

    work page arXiv 2026

  24. [24]

    JADE: Bridging the strategic-operational gap in dynamic agentic RAG,

    Y. Chen, E. Zhang, T. Hu, S. Wang, Z. Yang, M. Zhong, X. Wei, Y. Gao, Y. Wu, Y. Huet al., “JADE: Bridging the strategic-operational gap in dynamic agentic RAG,”arXiv preprint arXiv:2601.21916, 2026. 3

    work page arXiv 2026

  25. [25]

    WebWatcher: Breaking New Frontier of Vision-Language Deep Research Agent

    X. Geng, P . Xia, Z. Zhang, X. Wang, Q. Wang, R. Ding, C. Wang, J. Wu, Y. Zhao, K. Liet al., “WebWatcher: Breaking new fron- tier of vision-language deep research agent,”arXiv preprint arXiv:2508.05748, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    MedBrowseComp: Benchmarking Medical Deep Research and Computer Use

    S. Chen, P . Moreira, Y. Xiao, S. Schmidgall, J. Warner, H. Aerts, T. Hartvigsen, J. Gallifant, and D. S. Bitterman, “MedBrowseC- omp: Benchmarking medical deep research and computer use,” arXiv preprint arXiv:2505.14963, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    Agent-X: Evaluating Deep Multimodal Reasoning in Vision-Centric Agentic Tasks

    T. Ashraf, A. Saqib, H. Ghani, M. AlMahri, Y. Li, N. Ahsan, U. Nawaz, J. Lahoud, H. Cholakkal, M. Shahet al., “Agent-X: Evaluating deep multimodal reasoning in vision-centric agentic tasks,”arXiv preprint arXiv:2505.24876, 2025. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    Digital heritage and the rise of theory in museum computing,

    R. Parry, “Digital heritage and the rise of theory in museum computing,”Museum management and Curatorship, vol. 20, no. 4, pp. 333–348, 2005. 3

    work page 2005

  29. [29]

    The 3D vase museum: A new approach to context in a digital library,

    H.-y. Shiaw, R. J. Jacob, and G. R. Crane, “The 3D vase museum: A new approach to context in a digital library,” inProceedings of the 4th ACM/IEEE-CS joint conference on Digital libraries, 2004, pp. 125–134. 3

    work page 2004

  30. [30]

    A review of the de- velopment and application of generative technology in digital museums,

    J. Xu, L. Yan, R. Zhang, and M. Zhou, “A review of the de- velopment and application of generative technology in digital museums,”npj Heritage Science, vol. 13, no. 1, p. 589, 2025. 3

    work page 2025

  31. [31]

    Digital muse- ums and diverse cultural knowledges: Moving past the traditional catalog,

    R. Srinivasan, R. Boast, J. Furner, and K. M. Becvar, “Digital muse- ums and diverse cultural knowledges: Moving past the traditional catalog,”The information society, vol. 25, no. 4, pp. 265–278, 2009. 3

    work page 2009

  32. [32]

    What influences user contin- uous intention of digital museum: integrating task-technology fit (TTF) and unified theory of acceptance and usage of technology (UTAUT) models,

    F. Zheng, S. Wu, R. Liu, and Y. Bai, “What influences user contin- uous intention of digital museum: integrating task-technology fit (TTF) and unified theory of acceptance and usage of technology (UTAUT) models,”Heritage Science, vol. 12, no. 1, pp. 1–19, 2024. 3

    work page 2024

  33. [33]

    Museum websites and museum visitors: digital SUBMITTED TO IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 10 museum resources and their use,

    P . F. Marty, “Museum websites and museum visitors: digital SUBMITTED TO IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 10 museum resources and their use,”Museum management and cu- ratorship, vol. 23, no. 1, pp. 81–99, 2008. 3

    work page 2008

  34. [34]

    ScholarPeer: A Context-Aware Multi-Agent Framework for Automated Peer Review

    P . Goyal, M. Parmar, Y. Song, H. Palangi, T. Pfister, and J. Yoon, “ScholarPeer: A context-aware multi-agent framework for auto- mated peer review,”arXiv preprint arXiv:2601.22638, 2026. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  35. [35]

    OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments,

    T. Xie, D. Zhang, J. Chen, X. Li, S. Zhao, R. Cao, T. J. Hua, Z. Cheng, D. Shin, F. Leiet al., “OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments,”Advances in Neural Information Processing Systems, vol. 37, pp. 52 040–52 094,

    work page

  36. [36]

    OpenApps: Simulating environment variations to measure UI-agent reliabil- ity,

    K. Ullrich, J. Su, C. Shi, A. Subramonian, A. Bar, I. Evtimov, N. Tsilivis, R. Balestriero, J. Kempe, and M. Ibrahim, “OpenApps: Simulating environment variations to measure UI-agent reliabil- ity,”arXiv preprint arXiv:2511.20766, 2025. 4

    work page arXiv 2025

  37. [37]

    Lexicon iconographicum mythologiae classicae (limc),

    U. of Basel, “Lexicon iconographicum mythologiae classicae (limc),” Digital Cultural Heritage Archive, 2014, cC BY-NC 4.0 License. [Online]. Available: https://weblimc.org 6 Jiazi Wangis a second-year M.S. student at Bei- jing Jiaotong University, advised by Prof. Yufeng Chen. She received her B.S. degree from Bei- jing Jiaotong University. Her research i...

    work page 2014

  38. [38]

    Her interests include natural language processing and machine translation

    She joined Beijing Jiaotong University, Beijing, China, in 2014, as an Associate Pro- fessor. Her interests include natural language processing and machine translation. Yang Zhaois a Lecturer (Assistant Professor) at the Department of Computer Science and In- formation Technology, La Trobe University. Pre- viously, she was a Machine Learning Research Fell...

    work page 2014