Pith. sign in

REVIEW 3 major objections 5 minor 50 references

Planning Ahead with RSA: Efficient Signalling in Dynamic Environments by Projecting User Awareness across Future Timesteps

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that an assistive agent can keep human operators aware of fast-changing risks by planning entire sequences of alerts, simulating how each message will reshape a specific user's attention and beliefs over several future tim

desk verdict A clean RSA-based planning framework with a real novelty, but the headline claim rests on an unvalidated user model and a somewhat circular evaluation; worth serious review, not desk reject. read the letter →

arxiv 2510.23340 v1 pith:QCO6D6EY submitted 2025-10-27 cs.AI cs.CLcs.HC

Anwesha Das , John Duff , Jörg Hoffmann , Vera Demberg This is my paper
classification cs.AIcs.CLcs.HC
keywords RationalSpeechActframeworkPragmaticreasoningHuman-awareplanningAdaptivesignallingSituationalawarenessFinite-horizonBelieftrackingHuman-AIcollaboration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that assistive AI can communicate more effectively in dynamic, high-stakes settings if it treats communication as a planning problem: deciding what to say, when, and in what order, by projecting how each message will change a specific user's attention and beliefs several timesteps ahead. It extends the Rational Speech Act (RSA) framework, a Bayesian model of pragmatic communication, with temporal belief tracking, user-specific priors, and finite-horizon lookahead. On 800 simulated drone-monitoring trials, the full model outperforms greedy and user-agnostic variants, with the largest gains in the most difficult scenarios—many simultaneous risks, tightly clustered in time, and users who already partly anticipate some risks. The authors argue that multi-step planning and a realistic model of user awareness are jointly necessary for this effectiveness, and that the framework establishes a theoretical foundation for pragmatic, human-aware alerting in human-AI teams.

What carries the argument

The carrying mechanism is the pair of designer-specified functions AT and B (Eqs. 3–4): AT maps the RSA pragmatic listener's posterior over properties to an attention distribution, and B updates the user's belief over each property's value as a weighted blend of the true state and the previous belief, with the weight set by attention. Wrapped around this is a finite-horizon planner (Eq. 9) that enumerates utterance sequences, applies duration constraints so long messages block later alerts, and selects the sequence maximising cumulative reward over belief alignment on critical properties. The user-specific prior (Eq. 7) feeds the listener's prior with the previous belief state, so interpreta

What would settle it

A human-subject experiment in a comparable monitoring task: if operators' actual situational awareness (measured, for example, by accuracy on queries about critical states) is not higher after messages from the full model than after messages from the greedy or user-agnostic baselines—or if the model's predicted belief updating diverges systematically from measured beliefs—the central claim that effectiveness depends on combining planning with a realistic user model would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that an assistive agent can optimise situational awareness in a changing environment by planning entire sequences of alerts up to a fixed horizon, using a recursive Bayesian model of how each user interprets messages given their prior beliefs, and how that interpretation shifts their attention and updates their beliefs about the world. In the proposed d-RSA + Priors + Planning model, the agent evaluates every candidate utterance sequence by simulating, at each timestep, the user's belief state through an attention function and a belief-update rule, and scores the sequence by how well it aligns the user's beliefs with properties that are currently critical. On 800 simulat

Load-bearing premise

The belief-update rule B and attention function AT, which together define how the simulated user responds to messages, are assumed to be a realistic model of human awareness, but they are designer-specified and are not validated against human behaviour; the planner and the evaluation both rely on these same functions.

Editorial extensions

If this is right

  • Assistive agents in time-critical domains can reduce delays in raising awareness of risks the user does not know about, even at the cost of delaying mention of risks the user already anticipates.
  • Belief tracking is the main driver of prioritising low-awareness critical properties: the full model significantly shifts alert scheduling toward unconscious problems, whereas planning without user priors does not reliably do so.
  • The full model's messages are more diverse in specificity, using short or ambiguous alerts when the user's priors make them correctly interpretable—a concrete, observable signature of pragmatic behaviour.
  • The largest improvements over greedy baselines occur precisely in the hardest conditions: many simultaneous critical properties, tightly clustered onsets, and users with higher initial awareness.
  • The framework provides a template for extending RSA from single-turn, static reference games to multi-step, dynamic assistive communication, enabling joint optimisation of what, how, and when to communicate.

Reading between the lines

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

  • The paper's binary critical/non-critical model could be extended to graded urgency levels (e.g., gradual battery depletion vs. sudden rotor failure), and the same planning machinery would likely produce richer trade-offs between specificity, timing, and priority; this is a testable extension the authors list as future work.
  • Because the belief-update rule and attention function are designer-specified, the framework's predictions about optimal alert sequences become empirically testable only when those functions are fitted to human data (e.g., eye-tracking or response times); the qualitative finding that planning matters most with a realistic user model would probably survive such a replacement.
  • The explicit modelling of opportunity costs—where long messages occupy the attention channel—suggests a quantitative way to study alarm fatigue: the planner already trades off detail against timeliness, and could be adapted to cap alert frequency or measure the cost of over-alerting.
  • The anticipatory-alert behaviour observed in the full model (raising awareness of near-critical conditions before they become critical) implies that the model is effectively investing attention now to reduce the need for urgent alerts later; this could be quantified as a reduction in total alert burden across a horizon.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes an RSA-based adaptive signalling framework for dynamic human-AI settings. It extends a vanilla pragmatic speaker with linear belief tracking, user-specific priors, and finite-horizon planning over message sequences, with a lexicon that encodes message duration and a constraint that long utterances block subsequent alerts. In a simulated Drone World domain with 800 trials, the authors compare the full model (d-RSA + Priors + Planning) against greedy and/or user-agnostic ablations on cumulative reward and on timing/specificity metrics, and conclude that effectiveness depends crucially on combining multi-step planning with a realistic model of user awareness.

Significance. The formal framework is coherent, and the simulation design is detailed: 800 trials, varied critical-property counts and temporal dispersion, diverse simulated user beliefs, and exact breadth-first enumeration of utterance sequences. These are real strengths and make the internal optimization strategy transparent. If the central claim were externally validated, the framework would be a valuable extension of RSA to dynamic teaming. However, as presented, the headline evidence is internal: the full model optimizes the same reward function used for evaluation, and the 'realistic' user model is designer-specified rather than empirically validated. I therefore view the result as a promising simulation study whose stated conclusions currently outrun the evidence. This is fixable by reframing and by adding or targeting validation, so I do not recommend rejection.

major comments (3)
  1. [§6.1, Eq. 5/8/9] The evaluation metric is the objective. Eq. 9 selects the sequence maximizing cumulative reward R, and Section 6.1 reports exactly that R as the outcome. Since d-RSA + Priors + Planning exhaustively enumerates all legal sequences and the baselines are constrained variants of the same objective, the full model is guaranteed to dominate on total reward. The pairwise p-values in §6.1 therefore certify an optimization property, not empirical communicative effectiveness. To support the claim that the approach 'improves timely situational awareness', evaluate under a different user model than the planner's B/AT, or with human data; at a minimum, report the gap under a mismatched or held-out user model.
  2. [§4.3.1, Eq. 7] The user-specific prior P(p | b_{t-1}, StatCr) is the mechanism by which user awareness enters the pragmatic listener, but no defining formula is given. The text says it 'substitutes' uniform priors and the experiments simulate initial beliefs, yet the mapping from b_{t-1} and the critical set to a prior is unspecified. This is load-bearing: different definitions will change planner behavior and could change the ablation results. Please provide the exact formula, or state explicitly that this prior is a free design choice and analyze sensitivity to it.
  3. [§4.2.1, Eq. 4; §9.1.2] The abstract and conclusion describe B and AT as a 'realistic model of user awareness', but Eq. 4 is a linear interpolation that Section 9.1.2 says was chosen for simplicity, and AT directly maps the L1 posterior to attention. No human validation of either function is provided. Because B and AT are used both by the planner and in the reward that defines evaluation, the simulated results only describe behavior under the authors' assumptions. I recommend reframing the central claim as holding 'under the assumed B/AT model' and either adding a human validation or explicitly reserving realism for future work.
minor comments (5)
  1. [Eq. 4] The notation b_t(u_t | p_t) on the left-hand side appears to be a typo; the right-hand side defines a distribution over values v for each property p. Please clarify the indexing.
  2. [§6.2] The 'message-type entropy' and 'message specificity ratio' are used as quantitative outcomes, but their formulas are not given in the main text. Please define them precisely so the statistical comparisons are reproducible.
  3. [§8] The code repository URL is promised 'upon acceptance'; the attached supplementary files are not visible in the arXiv text. Since the model includes several unspecified choices, please make the code/data available or state clearly how to access them during review.
  4. [§9.2.1] There is a typo: 'trails' should be 'trials'. Also, the description of general-awareness users as 'high awareness' is defined only through the belief-generation distribution; it would help to state the numeric ranges used for the means and standard deviations.
  5. [§4.3.2, Eq. 8] The (X) placeholder is used to represent the blocking constraint, but it is not formally introduced as a lexicon item. Please define its role clearly and distinguish it from the 'silence' utterance '...'.

Circularity Check

1 steps flagged · score 6.0 of 10

Full model's advantage over baselines is built into its objective: Eq. 9 maximizes the same cumulative reward R used as the evaluation metric, with the unvalidated user model B/AT feeding both planning and scoring.

  1. self definitional [Sections 4.3.3/5.1.2/6.1 (Eqs. 8-9 vs. Eq. 6)]
    "𝑈∗ = arg max u1:H ∑_{t=1}^H R_t(u_t|s_t,Cr_t) (9) ... Pairwise comparisons show the d-RSA + Priors + Planning model outperforms all baselines (p s < 0.01), with increases associated with both planning and belief-tracking (p s < 0.01)."

    The evaluation metric is the cumulative reward R(u) (Eq. 8), which the full model is defined to maximize (Eq. 9). Since Eq. 9 searches all sequences (BFS in Section 5.1.2) and the baseline Eq. 6 chooses greedily at each timestep, the full model's total R is >= the baseline's total R on every trial by construction. Section 6.1 then reports this gap as evidence that 'd-RSA + Priors + Planning outperforms all baselines.' Additionally, R_t is computed from the model's own belief estimate B (Eq. 4) and attention AT (Eq. 3), so the same unvalidated user model is used both to plan and to score the outcome; the claim that this is a 'realistic model of user awareness' is assumed, not tested.

full rationale

The paper is a coherent simulation study and does not rely on self-citation chains or imported uniqueness theorems. The circularity is in the evaluation logic: the full model is defined as the argmax of cumulative reward R (Eq. 9), and the same R is the dependent measure in Section 6.1. Since the baseline models optimize only per-timestep or with uniform priors, the full model's total R is guaranteed to dominate them whenever the search is complete, as the paper's BFS enumeration ensures. The 'realistic model of user awareness' claim is also self-referential in that B and AT are designer-specified and are used both to plan and to simulate the user in evaluation; the paper concedes in Section 7.2 that perfect knowledge of user awareness is an idealization. This does not make the framework internally incoherent, but it means the headline result is a property of the authors' chosen objective and user model rather than evidence about real human awareness. Externally, the contribution stands as a proposed framework with an illustrative ablation, not as a validated prediction about human-agent communication. Score 6 reflects partial circularity: the main comparison reduces by construction, while the framework and qualitative analyses retain independent content.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new physical or ontological entities are introduced. The 'Silence' utterance and '(X)' lock are modeling devices within the existing state-utterance formalism, not entities with independent evidence requirements. The central free choices are the hand-crafted lexicon/durations and the unvalidated user belief model.

free parameters (3)
  • Rationality parameter α = 1 (assumed)
    Set to 1 for simplicity in the pragmatic speaker softmax (Technical Appendix §9.1.1); not fitted or varied.
  • Utterance duration vector k(u) = Not fully specified (e.g., specific message = 3, beep = 1)
    Hand-crafted lexicon durations encode the temporal opportunity cost central to planning; values are designer choices not derived from data.
  • Attention/reward variant choices = Direct AT; critical-only R
    Selected after qualitative review of 12 initial trials; not estimated from data.
assumptions (6)
  • domain assumption RSA recursive Bayesian inference is an adequate model of human pragmatic reasoning in this task
    The whole framework builds on RSA; no human data verify that operators update beliefs as L1.
  • ad hoc to paper The linear belief update B in Eq. 4 approximates real user belief revision
    Chosen for simplicity/interpretability; acknowledged as an idealization in §7.2; no empirical calibration.
  • ad hoc to paper Listener attention equals the pragmatic-listener posterior (AT direct mapping)
    Selected after qualitative review of 12 trials; binary alternative rejected; no independent evidence.
  • domain assumption Attentional zero-sum: a multi-timestep utterance blocks all other messages via '(X)'
    Encodes opportunity cost; plausible but not validated against human attention data.
  • domain assumption The reward R (critical-property belief alignment) is the correct measure of situational awareness
    Task-specific choice; alternative rewards were rejected for 'gaming'; no external definition.
  • ad hoc to paper Agent has perfect knowledge of the user's beliefs and future world states during planning
    Explicitly flagged in §7.2 as an idealization; necessary for BFS enumeration; likely false in practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Planning Ahead with RSA: Efficient Signalling in Dynamic Environments by Projecting User Awareness across Future Timesteps." pith.science (2026). https://pith.science/paper/QCO6D6EY

@misc{pith2026251023340,
  author       = {Pith},
  title        = {Pith review of: Planning Ahead with RSA: Efficient Signalling in Dynamic Environments by Projecting User Awareness across Future Timesteps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QCO6D6EY}},
  note         = {Machine review of arXiv:2510.23340}
}
read the original abstract

Adaptive agent design offers a way to improve human-AI collaboration on time-sensitive tasks in rapidly changing environments. In such cases, to ensure the human maintains an accurate understanding of critical task elements, an assistive agent must not only identify the highest priority information but also estimate how and when this information can be communicated most effectively, given that human attention represents a zero-sum cognitive resource where focus on one message diminishes awareness of other or upcoming information. We introduce a theoretical framework for adaptive signalling which meets these challenges by using principles of rational communication, formalised as Bayesian reference resolution using the Rational Speech Act (RSA) modelling framework, to plan a sequence of messages which optimise timely alignment between user belief and a dynamic environment. The agent adapts message specificity and timing to the particulars of a user and scenario based on projections of how prior-guided interpretation of messages will influence attention to the interface and subsequent belief update, across several timesteps out to a fixed horizon. In a comparison to baseline methods, we show that this effectiveness depends crucially on combining multi-step planning with a realistic model of user awareness. As the first application of RSA for communication in a dynamic environment, and for human-AI interaction in general, we establish theoretical foundations for pragmatic communication in human-agent teams, highlighting how insights from cognitive science can be capitalised to inform the design of assistive agents.

Figures

Figures reproduced from arXiv: 2510.23340 by the authors.

Figure 1
Figure 1. Impact of individualised alert sequences (x-axis) for [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Total reward (y-axis) for each model variant across scenario factors. (a) Reward as a function of the number of critical [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. (a) Message Specificity Ratio (y-axis) vs. Prior User Awareness of critical properties (x-axis). Specificity decreases [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 7 canonical work pages

  1. [1]

    Jessica S Ancker, Alison Edwards, Sarah Nosal, Diane Hauser, Elizabeth Mauer, Rainu Kaushal, and With the HITEC Investigators. 2017. Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system.BMC medical informatics and decision making17, 1 (2017), 36

    2017

  2. [2]

    Carolyn Jane Anderson. 2021. Tell Me Everything You Know: A Conversation Update System for the Rational Speech Acts Framework. InProceedings of the Society for Computation in Linguistics (SCiL) 2021. 244–253

    2021

  3. [3]

    Jacob Andreas and Dan Klein. 2016. Reasoning about pragmatics with neural listeners and speakers.arXiv preprint arXiv:1604.00562(2016)

    arXiv 2016

  4. [4]

    Immanuel Barshi. 2023. People are the Weak Link in the System. (2023). Invited Talk at Ben Gurion University

    2023

  5. [5]

    Mohan Baruwal Chhetri, Shahroz Tariq, Ronal Singh, Fatemeh Jalalvand, Cecile Paris, and Surya Nepal. 2024. Towards human-ai teaming to mitigate alert fatigue in security operations centres.ACM Transactions on Internet Technology24, 3 (2024), 1–22

    2024

  6. [6]

    Goodman, and Christopher Potts

    Reuben Cohn-Gordon, Noah D. Goodman, and Christopher Potts. 2017. Prag- matically Informative Image Captioning with Character-Level Inference. In Proceedings of NAACL-HLT 2018. 439–443. https://doi.org/10.18653/v1/N18-2070

    doi:10.18653/v1/n18-2070 2017

  7. [7]

    Goodman, and Christopher Potts

    Reuben Cohn-Gordon, Noah D. Goodman, and Christopher Potts. 2019. An Incremental Iterated Response Model of Pragmatics. InProceedings of the Society for Computation in Linguistics (SCiL) 2019. 81–90

    2019

  8. [8]

    Collins, Ilia Sucholutsky, Umang Bhatt, Kartik Chandra, Lionel Wong, Mina Lee, Cedegao E

    Katherine M. Collins, Ilia Sucholutsky, Umang Bhatt, Kartik Chandra, Lionel Wong, Mina Lee, Cedegao E. Zhang, Tan Zhi-Xuan, Mark Ho, Vikash Mansinghka, Adrian Weller, Joshua B. Tenenbaum, and Thomas L. Griffiths. 2024. Building Machines that Learn and Think with People.Nature Human Behavior8, 10 (2024), 1851–1863. https://doi.org/10.1038/s41562-024-01991-9

    doi:10.1038/s41562-024-01991-9 2024

Show all 50 references
  1. [9]

    Robert Dale and Ehud Reiter. 1995. Computational interpretations of the Gricean maxims in the generation of referring expressions.Cognitive science19, 2 (1995), 233–263

    1995

  2. [10]

    Anwesha Das, Zekun Wu, Iza Skrjanec, and Anna Maria Feit. 2024. Shifting Focus with HCEye: Exploring the Dynamics of Visual Highlighting and Cognitive Load on User Attention and Saliency Prediction.Proceedings of the ACM on Human- Computer Interaction8, ETRA (2024), 1–18

    2024

  3. [11]

    Judith Degen. 2023. The Rational Speech Act Framework.Annual Review of Linguistics9 (2023), 519–540. https://doi.org/10.1146/annurev-linguistics- 031220-010811

    2023 doi

  4. [12]

    Mica R Endsley. 2017. Toward a theory of situation awareness in dynamic systems. InSituational awareness. Routledge, 9–42

    2017

  5. [13]

    Frank and Noah D

    Michael C. Frank and Noah D. Goodman. 2012. Predicting Pragmatic Reasoning in Language Games.Science336 (2012), 998

    2012

  6. [14]

    Michael Franke. 2011. Quantity Implicatures, Exhaustive Interpretation, and Rational Conversation.Semantics & Pragmatics4 (2011), 1. https://doi.org/10. 3765/sp.4.1

    2011

  7. [15]

    Daniel Fried, Jacob Andreas, and Dan Klein. 2017. Unified pragmatic models for generating and following instructions.arXiv preprint arXiv:1711.04987(2017)

    2017 arXiv

  8. [16]

    Piotr J Gmytrasiewicz and Edmund H Durfee. 2001. Rational communication in multi-agent environments.Autonomous Agents and Multi-Agent Systems4, 3 (2001), 233–272

    2001

  9. [17]

    Goodman and Michael C

    Noah D. Goodman and Michael C. Frank. 2016. Pragmatic Language Interpreta- tion as Probabilistic Inference.Trends in Cognitive Sciences20, 11 (2016), 818–829. https://doi.org/10.1016/j.tics.2016.08.005

    2016 doi

  10. [18]

    Goodman and Andreas Stuhlmüller

    Noah D. Goodman and Andreas Stuhlmüller. 2013. Knowledge and Implicature: Modeling Language Understanding as Social Cognition.Topics in Cognitive Science5 (2013), 173–184. https://doi.org/10.1111/tops.12007

    2013 doi

  11. [19]

    Muaz Gülşen, Arslan, Sevban, and et al. 2025. The Effect of Alarm Fatigue on the Tendency to Make Medical Errors in Surgical Intensive Care Nurses: A Correlational Study Examining the Role of Moderating Factors. InHealthcare, Vol. 13. MDPI, 631

    2025

  12. [20]

    Hawkins, Hyowon Gweon, and Noah D

    Robert D. Hawkins, Hyowon Gweon, and Noah D. Goodman. 2021. The Division of Labor in Communication: Speakers Help Listeners Account for Asymmetries in Visual Perspective.Cognitive Science45 (2021), e12926. https://doi.org/10. 1111/cogs.12926

    2021

  13. [21]

    Prakash Jamakatel, Rebecca De Venezia, Christian Muise, and Jane Jean Kiam

  14. [22]

    Rhea Kapur, Robert Hawkins, D., and Elisa Kreiss. 2025. When More Words Say Less: Decoupling Length and Informativity in Image Description Evaluation. (2025). Unpublished manuscript, Stanford

    2025

  15. [23]

    Seth Karten, Siva Kailas, and Katia P Sycara. 2023. Emergent Compositional Concept Communication through Mutual Information in Multi-Agent Teams.. InAAMAS. 2391–2393

    2023

  16. [24]

    Thomas Keller and Malte Helmert. 2003. Trial-Based Heuristic Tree Search for Finite Horizon MDPs. InProceedings of the 13th International Conference on Automated Planning and Scheduling (ICAPS’03), Enrico Giunchiglia, Nicola Muscettola, and Dana Nau (Eds.)

    2003

  17. [25]

    Nafiz Hasan Khan and Carman Neustaedter

    Md. Nafiz Hasan Khan and Carman Neustaedter. 2019. An Exploratory Study of the Use of Drones for Assisting Firefighters During Emergency Situations. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. ACM, 1–14. https://doi.org/10.1145/3290605.3300502

    2019

  18. [26]

    Levente Kocsis and Csaba Szepesvári. 2006. Bandit Based Monte-Carlo Planning. InProceedings of the 17th European Conference on Machine Learning (ECML 2006), Johannes Fürnkranz, Tobias Scheffer, and Myra Spiliopoulou (Eds.), Vol. 4212. Springer-Verlag, 282–293

    2006

  19. [27]

    Jessy Lin, Nicholas Tomlin, Jacob Andreas, and Jason Eisner. 2024. Decision- Oriented Dialogue for Human-AI Collaboration.Transactions of the Association for Computational Linguistics12 (Aug. 2024), 892–911. https://doi.org/10.1162/ tacl_a_00679

    2024

  20. [28]

    1972.Bayesian statistics: A review

    Dennis Victor Lindley. 1972.Bayesian statistics: A review. SIAM

    1972

  21. [29]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved Baselines with Visual Instruction Tuning. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 26286–26296. https://doi.org/10.1109/ CVPR52733.2024.02484

    2024

  22. [30]

    Shipeng Liu, FNU Shrutika, Boshen Zhang, Zhehui Huang, and Feifei Qian. 2024. Effect of adaptive communication support on human-ai collaboration. AAAI 2025 Workshop on Advancing LLM-Based Multi-Agent Collaboration

    2024

  23. [31]

    Ziqiao Ma, Jing Ding, Xuejun Zhang, Dezhi Luo, Jiahe Ding, Sihan Xu, Yuchen Huang, Run Peng, and Joyce Chai. 2025. Vision-Language Models Are Not Pragmatically Competent in Referring Expression Generation. (2025). https: //doi.org/10.48550/arXiv.2504.16060 arXiv manuscript

    2025 doi arXiv

  24. [32]

    Daly, David Piorkowski, and John T

    Erik Miehling, Manish Nagireddy, Prasanna Sattigeri, Elizabeth M. Daly, David Piorkowski, and John T. Richards. 2024. Language Models in Dialogue: Con- versational Maxims for Human-AI Interactions. InFindings of the Association for Computational Linguistics: EMNLP 2024, Yaser ...

    2024 doi

  25. [33]

    Markus M Möbius, Muriel Niederle, Paul Niehaus, and Tanya S Rosenblat. 2014. Managing self-confidence.NBER Working paper17014 (2014)

    2014

  26. [34]

    John Mongan and Marc Kohli. 2020. Artificial intelligence and human life: five lessons for radiology from the 737 MAX disasters. , e190111 pages

    2020

  27. [35]

    Hawkins, Noah D

    Will Monroe, Robert D. Hawkins, Noah D. Goodman, and Christopher Potts

  28. [36]

    Will Monroe, Robert XD Hawkins, Noah D Goodman, and Christopher Potts

  29. [37]

    Marco Mussi, Alberto Maria Metelli, Marcello Restelli, Gianvito Losapio, Ricardo J Bessa, Daniel Boos, Clark Borst, Giulia Leto, Alberto Castagna, Ricardo Chavar- riaga, et al. 2025. Human-AI interaction in safety-critical network infrastructures. iScience28, 9 (2025)

    2025

  30. [38]

    Stephanie Rosenthal and Laura M Hiatt. 2020. Human-centered decision support for agenda scheduling. InProceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems. 1161–1168

    2020

  31. [39]

    Colors in context: A pragmatic neural model for grounded language understanding.Transactions of the Association for Computational Linguistics5 (2017), 325–338

    2017

  32. [40]

    Gregory Scontras, Michael Henry Tessler, and Michael Franke. 2021. A practical introduction to the Rational Speech Act modeling framework.arXiv preprint arXiv:2105.09867(2021)

    2021 arXiv

  33. [41]

    Scott Shappell, Cristy Detwiler, Kali Holcomb, Carla Hackworth, Albert Boquet, and Douglas Wiegmann. 2006. Human Error and Commercial Aviation Accidents: A Comprehensive, Fine-Grained Analysis Using HFACS. (2006)

    2006

  34. [42]

    Fardin Saad, Pradeep K Murukannaiah, and Munindar P Singh. 2025. Gricean Norms as a Basis for Effective Collaboration. InProceedings of the 24th Interna- tional Conference on Autonomous Agents and Multiagent Systems. 1812–1820

    2025

  35. [43]

    Zineng Tang, Lingjun Mao, and Alane Suhr. 2024. Grounding Language in Multi-Perspective Referential Communication. InProceedings of EMNLP 2024. 19727–19741. https://doi.org/10.18653/v1/2024.emnlp-main.1100

    2024 doi

  36. [44]

    Christopher D Wickens. 2008. Multiple resources and mental workload.Human factors50, 3 (2008), 449–455

    2008

  37. [45]

    Cyrus Sigari and Peter Biberthaler. 2021. Medical Drones: Disruptive Technology Makes the Future Happen.Unfallchirurg124, 12 (2021), 974–976. https://doi. org/10.1007/s00113-021-01095-3

    2021 doi

  38. [46]

    Tan Zhi-Xuan, Lance Ying, Vikash Mansinghka, and Joshua B Tenenbaum. 2024. Pragmatic Instruction Following and Goal Assistance via Cooperative Language- Guided Inverse Planning. InProceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems. 209...

    2024

  39. [48]

    Arianna Yuan, Will Monroe, Yue Bai, and Nate Kushman. 2018. Understanding the Rational Speech Act Model. InProceedings of the Annual Meeting of the Cognitive Science Society, Vol. 40

    2018

  40. [50]

    The choice of𝐴𝑇 directly influences the granularity of belief up- dates and reward computation

    showed this was overly restrictive – users lost all infor- mation about secondary properties even when utterances conveyed meaningful ambiguity, leading to overly conser- vative message selection. The choice of𝐴𝑇 directly influences the granularity of belief up- dates and rewa...

  41. [2017]

    https://doi.org/10.1162/tacl_a_00064

    Colors in Context: A Pragmatic Neural Model for Grounded Language Understanding.Transactions of the Association for Computational Linguistics5 (2017), 323–338. https://doi.org/10.1162/tacl_a_00064

    2017 doi

  42. [2024]

    International Joint Confernce on Artificial Intelligence(2024)

    A goal-directed dialogue system for assistance in safety-critical application. International Joint Confernce on Artificial Intelligence(2024)

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.