Pith. sign in

REVIEW 2 major objections 4 minor 42 references

A vision-language-model robot identifies most hazards in a virtual scene, and users report improved awareness when those hazards are annotated in VR.

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

2026-08-01 20:31 UTC pith:OMTM3LKM

load-bearing objection A useful integrated VR/ROS/VLM testbed whose detection results are acceptable but whose situational-awareness claim is overreached. the 2 major comments →

arxiv 2607.16582 v1 pith:OMTM3LKM submitted 2026-07-18 cs.RO cs.AIcs.CV

Autonomous VR-Based Risk Detection for Situational Awareness in Dangerous Settings

Mohammad Eskandari , Murali Krishna Varma Indukuri , Stephanie M. Lukin , Cynthia Matuszek This is my paper
classification cs.RO cs.AIcs.CV
keywords vision-language modelshuman-robot interactionvirtual realitysituational awarenesshazard detectionsafetysimulationuser study
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 argues that a robot equipped with a vision-language model (VLM) can autonomously identify the large majority of hazards in a simulated high-risk environment, and that presenting those hazards as annotated points of interest in virtual reality (VR) improves users' perceived situational awareness. The authors test this with a system in which a robot captures images while traversing a virtual makerspace, sends them to the VLM, and places interactive markers at detected hazard locations. The VLM identified 18 of 19 placed hazards, exceeding the paper's 80% detection threshold, and 26 of 27 study participants preferred the annotated scene over the unannotated one, rating it above the neutral midpoint on five measures. The paper's significance is in showing a concrete pipeline for VLM-based hazard communication in human-robot interaction, which, if it holds, could support navigation and decision-making in disaster response and industrial settings.

Core claim

The central claim is that a VLM-supported robot can detect at least 80% of hazards in a space and that VR annotations of those hazards make people feel more situationally aware. In the reported evaluation, the system placed 19 hazards in a post-earthquake makerspace; the VLM detected 18 (94.7% recall) with 4 false positives (21.1%). The user study, with 27 participants who first explored an unannotated baseline and then the annotated version, produced mean ratings of 4.16 to 4.60 on a 5-point scale across clarity, effectiveness, trust, usefulness, and future use, all significantly above the neutral midpoint, and 26 of 27 participants preferred the annotated environment. The authors interpret

What carries the argument

The core mechanism is a closed perception-to-display loop. A robot traverses a virtual scene, capturing still images at a fixed rate; each image is sent to a vision-language model with a structured prompt that asks for a JSON list of hazards, short safety messages, and whether to generate a point of interest. The system then places virtual markers—called Points of Interest (POIs)—in the environment at the corresponding locations. A POI is idle as a transparent red bouncing pin; when the user comes within one meter, it activates an information panel with the VLM's explanation and an optional audio playback generated separately. The load-bearing component is the VLM's contextual scene understa

Load-bearing premise

The claim that annotations improve situational awareness rests on the assumption that the high ratings participants gave to the annotated condition reflect the annotations themselves, rather than the fixed order of presentation or a novelty effect.

What would settle it

Run the same detection pipeline in a second simulated environment with a new set of ground-truth hazards, and run a user study with counterbalanced condition order and direct paired ratings of annotated vs. unannotated scenes; if the 80% detection threshold is not reached or the annotation preference fails to appear, the central claims are not supported.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the detection result generalizes, the same pipeline could let a drone or ground robot pre-scan a real disaster site and hand responders an annotated map before they enter, supporting proactive mitigation and preparedness.
  • Because the prompt deliberately avoids scenario-specific details, the system is meant to transfer across environments; a direct corollary is that the same code and prompt could label hazards in construction sites, industrial plants, or earthquake scenes without retraining.
  • The correlation data show that clarity of the hazard markers and usability of the interface were the strongest drivers of future-use intention, so improving marker legibility and panel ergonomics would likely be the most direct path to wider adoption.
  • The error analysis found that all false positives and the single false negative occurred when objects were densely packed, implying that accuracy will drop in cluttered scenes and that higher-resolution images or multi-view fusion may be needed there.

Where Pith is reading between the lines

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

  • A direct test of the situational-awareness claim would require a counterbalanced user study in which participants rate the annotated and unannotated scenes in random order; the current fixed-order design leaves the possibility that novelty or familiarity with the scene inflated the positive ratings.
  • The paper's energy estimate (~1.9 kWh for ~40 images) and its comparison to a lighter open-weight VLM together imply a concrete design trade-off for field deployment: on-device models would save power and protect privacy but likely lower hazard-detection accuracy, a trade-off the authors acknowledge but leave unquantified.
  • If these results transfer to the field, a sensible next step would be a confidence threshold or human-in-the-loop verification for low-confidence detections, since the paper's own failure cases include outright hallucinations that a user alone might not catch.
  • The reported 94.7% recall comes from a single scene with 19 hazards; replication in other environments, with different hazard types and densities, is needed before the 80% threshold can be treated as a stable property of the approach.

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

2 major / 4 minor

Summary. The paper presents a VR-based human-robot interaction framework in which a simulated robot explores a virtual makerspace and uses a VLM (GPT-4o) to identify hazards; these are presented to a human operator as points of interest in VR with visual and audio annotations. The authors report two hypotheses: H1 that the VLM-based robot detects at least 80% of hazards, and H2 that users feel improved situational awareness with annotated versus unannotated navigation. The detection experiment reports 18/19 hazards detected (94.7% recall) with 4 false positives; a 27-participant user study reports high Likert ratings for clarity, effectiveness, trust, usefulness, and future use, and 26/27 participants preferred the annotated condition.

Significance. If validated, the framework is a useful contribution to HRI for hazard communication, combining VLM perception with immersive VR. The paper's strengths are its concrete system description: the prompt is included, the model is identified, a comparison against an open-weight model is reported, failure cases are analyzed, and ethical/environmental considerations are discussed. The detection result is credible as a single-scene demonstration. However, the H2 claim rests on a design that cannot separate annotation effects from order and familiarity effects, so the central user-study claim is not currently supported.

major comments (2)
  1. [Section V (first paragraph, Table I)] The H2 conclusion is not supported by the reported data. The five Likert items were administered only for the annotated condition; Table I compares them to the scale midpoint (3), not to ratings of the unannotated baseline. All participants experienced both conditions in a fixed order (baseline first, then annotated), so the 26/27 preference for the annotated condition is confounded with learning/order effects. To support H2, the study needs counterbalanced ordering, the same rating items for both conditions, or an objective situational-awareness measure.
  2. [Section V (Results and Analysis)] The paper equates 'high ratings above midpoint' with 'improved situational awareness.' None of the five dimensions directly measures situational awareness (e.g., memory for hazards, SAGAT-style queries, or task performance). The construct validity of H2 is therefore weak. The authors should either use an established SA measure or reframe the claim as perceived usability/satisfaction rather than situational awareness.
minor comments (4)
  1. [Table I] The column header 'H1: median > 3' is confusing because H1 is already defined as the detection hypothesis in the paper. Rename this to 'Alternative: median > 3' or 'H0: median ≤ 3' to avoid ambiguity.
  2. [Fig. 3 and accompanying text] The text states that the Clarity–UI Usefulness correlation is 0.06, but the figure caption says 'relatively high relationships across all five metrics.' Clarify the interpretation and avoid overinterpreting correlations from N=27. Also explain how a near-zero correlation can coexist with both variables correlating strongly with Future Use.
  3. [Section IV.A] The ground-truth hazard inventory was established by the authors surveying the makerspace against OSHA standards. Provide more detail on how the 19 hazards were selected and validated (e.g., number of annotators, inter-rater reliability) to improve transparency and reproducibility.
  4. [Section V] The fixed presentation order is acknowledged in the procedure but not discussed as a limitation. Add a sentence in the Limitations section acknowledging that the fixed order may have inflated preference and ratings for the annotated condition.

Circularity Check

0 steps flagged

No circularity: the detection and user-study claims are empirical evaluations against an externally grounded hazard inventory and self-reports; no result reduces to its own inputs.

full rationale

The paper's central claims are empirical, not derivational. H1 is evaluated by comparing GPT-4o outputs to a 19-item hazard inventory that the authors established by surveying the makerspace against OSHA standards (29 CFR 1910); the reported 18/19 detection result is an external assessment of an unfine-tuned VLM, not a quantity fitted from or defined by the evaluation. H2 is assessed through Likert ratings and a preference question; although the fixed-order design and lack of a paired baseline weaken causal attribution, the ratings are independent self-reports rather than circular restatements of the hypothesis. The self-citations present in the paper (RIVR [22], 3DGS [4], Scarecrows [34]) are infrastructure, method, or taxonomy references with overlapping authors, but they are not load-bearing proofs of the detection or awareness claims. The limitations section explicitly notes the single-VLM scope and the absence of ablation, consistent with a contingent empirical result rather than a construction that forces the outcome. No equation, fitted parameter, or uniqueness argument reduces a prediction to its input, so there is no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central claims rest on an author-defined hazard inventory, a spatial mapping from image-level detections to 3D POIs, and self-reports as a proxy for situational awareness. No free physical parameters or invented entities appear; the system-level constants are interface choices and the a priori H1 threshold.

free parameters (3)
  • H1 success threshold = 80%
    A priori criterion for 'VLMs can identify most hazards'; chosen by the authors and determines whether H1 is judged supported.
  • Ground-truth hazard inventory = 19 hazards
    Reference list compiled by the authors against OSHA standards; all recall and precision values depend on this set's completeness and correctness.
  • POI activation radius = 1 meter
    Hand-set proximity threshold that triggers the annotation panel; it directly shapes the user experience measured in H2.
axioms (6)
  • domain assumption The 19 author-defined ground-truth hazards are correct and complete for the makerspace scene.
    Recall and precision are computed against this inventory, with no inter-rater reliability or independent audit reported (§IV-A, §V).
  • domain assumption Image-level VLM detection is sufficient to place POIs at or near actual hazard locations.
    POIs are placed at the location of images flagged as hazardous rather than at localized objects, so the 3D placement assumes the photo viewpoint is a good proxy for the hazard location (§IV-A).
  • domain assumption Self-reported Likert ratings and a single preference question measure situational awareness.
    No objective situational-awareness metric, task performance measure, or paired comparison against the baseline is used (§V).
  • ad hoc to paper Fixed presentation order does not bias the results.
    All participants saw the unannotated scene before the annotated scene; without counterbalancing, practice or novelty effects are confounded with the annotation condition (§V).
  • domain assumption GPT-4o outputs without fixed seeds are stable enough for evaluation.
    The paper states 'No fine-tuning was performed, and no seeds were specified,' and all queries were made in Oct-Nov 2024, so API behavior could vary (§IV-A).
  • domain assumption One VLM and one user-study scene are representative of general VLM/VR hazard communication.
    The Limitations section acknowledges reliance on a single VLM; H1 and the broader framing generalize only if this single-scene result transfers to other settings (§VII).

pith-pipeline@v1.3.0-alltime-deepseek · 10593 in / 15901 out tokens · 168999 ms · 2026-08-01T20:31:53.021245+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Autonomous VR-Based Risk Detection for Situational Awareness in Dangerous Settings." pith.science (2026). https://pith.science/paper/OMTM3LKM

@misc{pith2026260716582,
  author       = {Pith},
  title        = {Pith review of: Autonomous VR-Based Risk Detection for Situational Awareness in Dangerous Settings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OMTM3LKM}},
  note         = {Machine review of arXiv:2607.16582}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In high-risk environments such as disaster response, situational awareness depends not only on detecting hazards but also on communicating them clearly to human operators. Vision Language Models (VLMs) have shown strong potential for scene understanding in safety-critical settings, yet their value as part of human-facing robotic systems remains underexplored. We present a VR-based Human Robot Interaction framework for studying how VLM-assisted robots can support situational awareness in simulated hazardous environments. In our system, a robot explores a virtual scene and queries a VLM to identify potential hazards and annotate user-facing points of interest. These annotations are presented to a human operator through an immersive VR interface. This framework enables controlled evaluation of both robotic hazard identification and the communication of safety-critical information to users. Results from our study indicate that the annotated VR interface was preferred over the unannotated baseline and that participants reported high clarity, usefulness, and comfort when interacting with the system. These findings suggest that combining VLM-based robotic perception with immersive visualization is a promising approach for supporting situational awareness in hazardous settings.

Figures

Figures reproduced from arXiv: 2607.16582 by Cynthia Matuszek, Mohammad Eskandari, Murali Krishna Varma Indukuri, Stephanie M. Lukin.

Figure 1
Figure 1. Figure 1: An example of a hazardous environment used in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Our system pipeline for increasing situational awareness. The robot’s location and view data from the RIVR simulator [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The correlation matrix shows correlations between the [PITH_FULL_IMAGE:figures/full_fig_p005_3.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

42 extracted references · 1 canonical work pages

  1. [1]

    Explainable artificial intelligence: A survey,

    A. Adadi and M. Berrada, “Explainable artificial intelligence: A survey,”IEEE Access, vol. 6, pp. 52 138–52 160, 2018

    2018

  2. [2]

    Toward human-centered intelligent assistance system in man- ufacturing: Challenges and potentials for operator 5.0,

    C. Bechinie, S. Zafari, L. Kroeninger, J. Puthenkalam, and M. Tsche- ligi, “Toward human-centered intelligent assistance system in man- ufacturing: Challenges and potentials for operator 5.0,”Procedia Computer Science, vol. 232, pp. 1584–1596, 2024

    2024

  3. [3]

    The dark side of dataset scaling: Evaluating racial classification in multimodal models,

    A. Birhane, S. Dehdashtian, V . Prabhu, and V . Boddeti, “The dark side of dataset scaling: Evaluating racial classification in multimodal models,” inThe 2024 ACM Conference on Fairness, Accountability, and Transparency, 2024, pp. 1229–1244

    2024

  4. [4]

    3d gaussian splatting for human-robot interaction,

    S. Bowser and S. M. Lukin, “3d gaussian splatting for human-robot interaction,” inThe 1st InterAI Workshop: Interactive AI for Human- centered Robotics, 2024

    2024

  5. [5]

    Better safe than sorry? overreaction problem of vision language models in visual emergency recognition,

    D. Choi, S. Lee, and Y . Song, “Better safe than sorry? overreaction problem of vision language models in visual emergency recognition,” arXiv preprint arXiv:2505.15367, 2025

    arXiv 2025

  6. [6]

    On the use of simulation in robotics: Opportunities, challenges, and suggestions for moving forward,

    H. Choi, C. Crump, C. Duriez, A. Elmquist, G. Hager, D. Han, F. Hearl, J. Hodgins, A. Jain, F. Leve, C. Li, F. Meier, D. Negrut, L. Righetti, A. Rodriguez, J. Tan, and J. Trinkle, “On the use of simulation in robotics: Opportunities, challenges, and suggestions for moving forward,”PNAS, vol. 118, no. 1, p. e1907856118, 2021

    2021

  7. [8]

    Smart operators: How industry 4.0 is affecting the worker’s performance in manufacturing contexts,

    V . Di Pasquale, V . De Simone, S. Miranda, and S. Riemma, “Smart operators: How industry 4.0 is affecting the worker’s performance in manufacturing contexts,”Procedia Computer Science, vol. 180, 2021

    2021

  8. [9]

    Situation awareness: State of the art,

    M. R. Endsley and E. S. Connors, “Situation awareness: State of the art,” in2008 IEEE Power and Energy Society General Meeting - Conversion and Delivery of Electrical Energy in the 21st Century. Pittsburgh, PA, USA: IEEE, July 2008, pp. 1–4

    2008

  9. [10]

    Understanding human behaviour in industrial human–robot interac- tion by means of virtual reality,

    P. Fratczak, Y . M. Goh, P. Kinnell, A. Soltoggio, and L. Justham, “Understanding human behaviour in industrial human–robot interac- tion by means of virtual reality,” inProc. of the Halfway to the Future Symposium 2019. Nottingham, UK: ACM, 2019

    2019

  10. [11]

    Multi-sensory hmi for human- centric industrial digital twins: A 6g vision of future industry,

    B. Han and H. D. Schotten, “Multi-sensory hmi for human- centric industrial digital twins: A 6g vision of future industry,” in2022 IEEE Symposium on Computers and Communications (ISCC). IEEE, Jun. 2022, pp. 1–7. [Online]. Available: http://dx.doi.org/10.1109/ISCC55528.2022.9912932

    arXiv 2022

  11. [12]

    Hri challenges influencing low usage of robotic systems in disaster response and rescue operations,

    S. Hoque, F. F. Riya, and J. Sun, “Hri challenges influencing low usage of robotic systems in disaster response and rescue operations,” arXiv preprint arXiv:2401.15760, 2024

    Pith/arXiv arXiv 2024

  12. [13]

    The role of data and information quality during disaster response decision- making,

    V . Jayawardene, T. Huggins, R. Prasanna, and B. Fakhruddin, “The role of data and information quality during disaster response decision- making,”Progress in Disaster Science, vol. 12, p. 100202, 08 2021

    2021

  13. [14]

    Jentsch,Human-Robot Interactions in Future Military Operations

    F. Jentsch,Human-Robot Interactions in Future Military Operations. CRC Press, 2016

    2016

  14. [15]

    Copal: Corrective planning of robot actions with large language models,

    F. Joublin, A. Ceravola, P. Smirnov, F. Ocker, J. Deigmoeller, A. Belardinelli, C. Wang, S. Hasler, D. Tanneberg, and M. Gienger, “Copal: Corrective planning of robot actions with large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2310.07263

    Pith/arXiv arXiv 2023

  15. [16]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, T. Leimk ¨uhler, S. Bako, P. Kellnhofer, F. Rousselle, M. Nießner, and J. Nov´ak, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, 2023

    2023

  16. [17]

    Understanding large-language model (llm)-powered human-robot interaction,

    C. Y . Kim, C. P. Lee, and B. Mutlu, “Understanding large-language model (llm)-powered human-robot interaction,” inProc. of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, 2024, pp. 371–380

    2024

  17. [18]

    Vision-language models represent darker- skinned black individuals as more homogeneous than lighter-skinned black individuals,

    M. H. J. Lee and S. Jeon, “Vision-language models represent darker- skinned black individuals as more homogeneous than lighter-skinned black individuals,”arXiv preprint arXiv:2412.09668, 2024

    Pith/arXiv arXiv 2024

  18. [19]

    Power hungry processing: Watts driving the cost of ai deployment?

    S. Luccioni, Y . Jernite, and E. Strubell, “Power hungry processing: Watts driving the cost of ai deployment?” inThe 2024 ACM Confer- ence on Fairness, Accountability, and Transparency, 2024, pp. 85–99

    2024

  19. [20]

    Can llms keep a secret? testing privacy implications of language models via contextual integrity theory,

    N. Mireshghallah, H. Kim, X. Zhou, Y . Tsvetkov, M. Sap, R. Shokri, and Y . Choi, “Can llms keep a secret? testing privacy implications of language models via contextual integrity theory,” 2023

    2023

  20. [21]

    Information overload, stress, and emergency managerial thinking,

    S. Misra, P. Roberts, and M. Rhodes, “Information overload, stress, and emergency managerial thinking,”International Journal of Disaster Risk Reduction, vol. 46, p. 101615, 2020. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2212420920312644

    2020

  21. [22]

    A simulator for human-robot interaction in virtual reality,

    M. Murnane, P. Higgins, M. Saraf, F. Ferraro, C. Matuszek, and D. Engel, “A simulator for human-robot interaction in virtual reality,” in2021 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), 2021, pp. 470–471

    2021

  22. [23]

    Preliminary report: Norfolk southern railway train derailment, east palestine, ohio,

    National Transportation Safety Board, “Preliminary report: Norfolk southern railway train derailment, east palestine, ohio,” 2023

    2023

  23. [24]

    Virtual reality based space operations – a study of esa’s potential for vr based training and simulation,

    M. Olbrich, H. Graf, J. Keil, R. Gad, S. Bamfaste, and F. Nicolini, “Virtual reality based space operations – a study of esa’s potential for vr based training and simulation,” inVirtual, Augmented and Mixed Reality: Interaction, Navigation, Visualization, Embodiment, and Simulation, J. Y . C. Chen and G. Fragomeni, Eds. Cham: Springer, 2018, vol. 10909, p...

    2018

  24. [25]

    Hazard recognition under process safety management,

    OSHA, “Hazard recognition under process safety management,” 1992, 29 CFR 1910.119. [Online]. Available: https://www.osha.gov/process- safety-management/hazards

    1992

  25. [26]

    Risk analysis (assessment) using virtual reality technology—effects of subjective experience: An experimental study,

    P. Puschmann, T. Horlitz, V . Wittstock, and A. Sch ¨utz, “Risk analysis (assessment) using virtual reality technology—effects of subjective experience: An experimental study,” in26th CIRP Design Conference, 2016

    2016

  26. [27]

    MSTS: A Multimodal Safety Test Suite for Vision-Language Models,

    P. R ¨ottger, G. Attanasio, F. Friedrich, J. Goldzycher, A. Parrish, R. Bhardwaj, C. D. Bonaventura, R. Eng, G. E. K. Geagea, S. Goswami, J. Han, D. Hovy, S. Jeong, P. Jereti ˇc, F. M. Plaza-del- Arco, D. Rooein, P. Schramowski, A. Shaitarova, X. Shen, R. Willats, A. Zugarini, and B. Vidgen, “MSTS: A Multimodal Safety Test Suite for Vision-Language Models...

    2025

  27. [28]

    Modeling interaction in human–machine systems: A trust and trustworthiness approach,

    A. Sapienza, F. Cantucci, and R. Falcone, “Modeling interaction in human–machine systems: A trust and trustworthiness approach,” Automation, vol. 3, no. 2, pp. 242–257, 2022

    2022

  28. [29]

    Hear here: sonification as a design strategy for robot teleoperation using virtual reality,

    J. Simmons, A. Bown, P. Bremner, V . McIntosh, and T. Mitchell, “Hear here: sonification as a design strategy for robot teleoperation using virtual reality,” 2023

    2023

  29. [30]

    Virtual reality application for enhancing risk assessment skills,

    R. Slezak, N. Keren, and T. Finseth, “Virtual reality application for enhancing risk assessment skills,” inInterservice/Industry Training, Simulation, and Education Conference, 2018

    2018

  30. [31]

    Winoground: Probing vision and language models for visio-linguistic compositionality,

    T. Thrush, R. Jiang, M. Bartolo, A. Singh, A. Williams, D. Kiela, and C. Ross, “Winoground: Probing vision and language models for visio-linguistic compositionality,” 2022, version Number: 2. [Online]. Available: https://arxiv.org/abs/2204.03162

    Pith/arXiv arXiv 2022

  31. [32]

    Vr-based health and safety training in various high-risk engineering industries: A literature review,

    R. Toyoda, F. Russo-Abeg ˜ao, and J. Glassey, “Vr-based health and safety training in various high-risk engineering industries: A literature review,”International Journal of Educational Technology in Higher Education, 2022

    2022

  32. [33]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, M. Bosma, V . Y . Zhao, K. Guu, A. W. Yu, B. Lester, N. Du, A. M. Dai, and Q. V . Le, “Chain-of-thought prompting elicits reasoning in large language models,”arXiv preprint arXiv:2201.11903, 2022

    Pith/arXiv arXiv 2022

  33. [34]

    Scarecrows in oz: The use of large language models in hri,

    T. Williams, C. Matuszek, R. Mead, and N. Depalma, “Scarecrows in oz: The use of large language models in hri,”Transactions on Human-Robot Interaction, vol. 13, no. 1, Jan. 2024. [Online]. Available: https://doi.org/10.1145/3606261

    work page doi:10.1145/3606261 2024

  34. [35]

    Common inpainted objects in-n-out of context,

    T. Yang, T. Jordan, N. Liu, and J. Sun, “Common inpainted objects in-n-out of context,”arXiv:2506.00721, 2025. [Online]. Available: https://arxiv.org/abs/2506.00721

    Pith/arXiv arXiv 2025

  35. [36]

    Trustworthy anomaly detection: A survey,

    S. Yuan and X. Wu, “Trustworthy anomaly detection: A survey,” 2022

    2022

  36. [37]

    Large language models as zero- shot human models for human-robot interaction,

    B. Zhang and H. Soh, “Large language models as zero- shot human models for human-robot interaction,”arXiv preprint arXiv:2303.03548, 2023

    Pith/arXiv arXiv 2023

  37. [38]

    Large language models for human–robot interaction: A review,

    C. Zhang, J. Chen, J. Li, Y . Peng, and Z. Mao, “Large language models for human–robot interaction: A review,”Biomimetic Intelligence and Robotics, vol. 3, no. 4, p. 100131, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2667379723000451

    2023

  38. [39]

    Vision-language models for vision tasks: A survey,

    J. Zhang, J. Huang, S. Jin, and S. Lu, “Vision-language models for vision tasks: A survey,” 2023, arXiv:2304.00685

    Pith/arXiv arXiv 2023

  39. [40]

    Agent3d-zero: An agent for zero-shot 3d understanding,

    S. Zhang, D. Huang, J. Deng, S. Tang, W. Ouyang, T. He, and Y . Zhang, “Agent3d-zero: An agent for zero-shot 3d understanding,” Springer, pp. 186–202, 2024

    2024

  40. [41]

    Tinyllava: A framework of small-scale large multimodal models,

    B. Zhou, Y . Hu, X. Weng, J. Jia, J. Luo, X. Liu, J. Wu, and L. Huang, “Tinyllava: A framework of small-scale large multimodal models,” arXiv preprint arXiv:2402.14289, 2024

    Pith/arXiv arXiv 2024

  41. [42]

    Do llms understand visual anomalies? Uncovering llm’s capabilities in zero-shot anomaly detection,

    J. Zhu, S. Cai, F. Deng, B. C. Ooi, and J. Wu, “Do llms understand visual anomalies? Uncovering llm’s capabilities in zero-shot anomaly detection,” inProc. of the 32nd ACM International Conference on Multimedia, ser. Mm ’24. New York, NY , USA: Association for Computing Machinery, 2024, pp. 48–57

    2024

  42. [2025]

    Available: https://arxiv.org/abs/2501.12948

    [Online]. Available: https://arxiv.org/abs/2501.12948

    Pith/arXiv arXiv