Pith. sign in

REVIEW 4 major objections 7 minor 47 references

Driverless cars can treat passenger intent as a latent mental state, then turn it into planning objectives that keep rides aligned with human needs.

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 · grok-4.5

2026-07-11 15:27 UTC pith:LACKY36E

load-bearing objection Solid systems paper that packages ToM-style latent intent into a planner interface; the nuPlan numbers only prove feasibility, not passenger alignment. the 4 major comments →

arxiv 2607.04670 v1 pith:LACKY36E submitted 2026-07-06 cs.HC

Who Responds When the Driver Is Gone? A Framework for Human Intent Understanding

Xuewen Luo , Ding Fan , Ruiqi Chen , Ye Cao , Xiujin Liu , Bo Yu , Fengze Yang , Chenxi Liu This is my paper
classification cs.HC
keywords autonomous drivinghuman intentTheory of Mindhuman-in-the-looplatent human stateintent-conditioned planningin-cabin sensingdriverless mobility
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.

When there is no human driver, someone still has to notice when a passenger is nauseous, anxious with an infant, or otherwise needing a different ride. This paper argues that passenger intent is not just spoken commands; it is a latent cognitive state shaped by language, personal attributes, emotion and physical condition, behavior, and situation. The authors build a Holistic Intent Dataset that labels both explicit talk and those implicit cues, then train a Theory-of-Mind-style reasoner to infer a structured Latent Human State and convert it into a Human Intent Objective the planner can use. A hierarchical planner conditions both route choice and short-horizon trajectories on that objective. On a standard closed-loop driving benchmark the system stays competitive with strong planners while sharply improving structured intent inference over a pretrained language model. The point for a sympathetic reader is that fully driverless mobility needs an in-cabin mind-reader, not only better sensors on the road.

Core claim

Intent2Drive shows that modeling passenger intent as a structured latent cognitive state, inferred from both explicit language and implicit human cues and then translated into a planner-compatible Human Intent Objective, improves intent understanding and still supports competitive closed-loop driving when that objective conditions route and trajectory planning.

What carries the argument

The Human Intent Reasoner (HIR): a Theory-of-Mind-inspired pipeline that maps observable intent evidence to a Latent Human State (goal, urgency, condition, preference, risk tolerance) and then constructs a Human Intent Objective (trip objective plus driving objective) that the Hierarchical Intent-Conditioned Planner can use at route and trajectory levels.

Load-bearing premise

The load-bearing premise is that questionnaire, interview, and synthetic text descriptions of attributes, feelings, and behaviors are good enough stand-ins for the real latent states of passengers in live cars, so labels trained on that dataset transfer to real rides.

What would settle it

Run the same Latent Human State and Human Intent Objective pipeline on real in-cabin multimodal recordings (face, voice, posture, continuous passenger feedback) against independent human annotations of goals, urgency, and preferred driving style; if structured accuracy collapses relative to the paper's textual test set, or closed-loop rides no longer match passenger-rated comfort and intent, the central claim fails.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes Intent2Drive, a human-in-the-loop autonomous driving framework that treats passenger intent as a latent cognitive state rather than surface-level commands. It introduces the Holistic Intent Dataset (HID; 2,240 real+synthetic samples with hierarchical annotations), a Theory-of-Mind-inspired Human Intent Reasoner (HIR; LoRA-fine-tuned Qwen3-4B) that maps explicit language, implicit cues, and scenario context to a structured Latent Human State (LHS: goal, urgency, condition, preference, risk) and then to a planner-compatible Human Intent Objective (HIO: trip objective + driving objective), and a Hierarchical Intent-Conditioned Planner (HICP) that conditions route RL and diffusion trajectory generation on HIO. Experiments report large gains over pretrained Qwen3-4B on LHS attributes and HIO components (Tables 2–4), ablations crediting implicit cues and scenario, qualitative ToM case studies, and competitive nuPlan closed-loop scores (e.g., Val14 NR 93.02 / R 84.05) that the authors interpret as evidence that intent reasoning can be added without severe planning degradation.

Significance. If the central transfer claim holds—that structured LHS/HIO inference from multi-cue textual intent yields driving behavior better aligned with real passenger needs—the work would be a meaningful step from command-following AD toward passenger-responsive robotaxis, with a reusable annotation schema and an explicit cognitive intermediate representation. Strengths that should be credited even under a skeptical reading include: (i) a clearly articulated hierarchical pipeline (observable cues → LHS → HIO → hierarchical planning) that separates understanding from planning abstraction; (ii) substantial, multi-attribute gains of fine-tuned HIR over the pretrained backbone on structured labels (Tables 2–3) and a progressive ablation of implicit components (Table 4); (iii) competitive closed-loop feasibility on nuPlan after inserting the intent module (Table 1), showing the planner is not broken by conditioning; and (iv) an honest limitations discussion on missing multimodal in-cabin sensing and the simplicity of HIO. The contribution is primarily systems/framework and dataset design rather than a new planning algorithm or a validated cognitive model of intent.

major comments (4)
  1. The joint claim of “human-aligned planning” is only partially supported. Table 1 measures closed-loop feasibility (NR/R scores) relative to PDM, Diffusion Planner, LLM-ASSIST, etc.; it does not measure whether HIO-conditioned routes or trajectories better satisfy the passenger needs encoded in HIO. The abstract carefully says “preserving competitive closed-loop planning performance,” but the Conclusion states the system “achieves more human-aligned driving.” No preference study, human rating, or independent alignment metric (e.g., passenger preference over HIO-matched vs. mismatched plans, or success on intent-specific criteria such as comfort stops / hospital urgency) is reported. Without such a metric independent of HID’s own labels, the planning half of the central claim remains untested.
  2. Coupling between HID/HIR and the nuPlan evaluation is underspecified. HID is built from questionnaires, interviews, and GPT-4o samples; nuPlan provides external traffic scenarios without in-cabin passengers. The manuscript does not state how human-intent inputs (x_exp, x_imp, x_scn) or HIO z are obtained for Val14/Test14 closed-loop runs—default HIO, synthetic injection, oracle labels, or a fixed conservative objective. Equations (6)–(11) condition P on z, so the Table 1 numbers are uninterpretable as an intent-aware system until this interface is specified and ablated (with vs. without HIO; matched vs. mismatched HIO).
  3. LHS/HIO evaluation is supervised reproduction of an author-defined schema, not independent discovery of latent mental states. HIR is SFT’d on HID to predict the same LHS tuple m=(g,u,c,p,r) and HIO z=(o,d) that annotators (and GPT-4o under the same schema) produced (Eqs. 2–5; HID construction). Tables 2–3 therefore show that fine-tuning recovers the annotation protocol far better than the pretrained model—which is useful engineering—but do not validate that the five-factor LHS is cognitively complete or planning-sufficient, nor that questionnaire/interview text plus synthetic samples transfer to live in-cabin intent. The “ToM-inspired” framing should be scoped as structured intermediate supervision, and either external validation (held-out human raters of LHS/HIO quality; multimodal in-cabin data) or a clearer claim boundary is needed.
  4. Reactive closed-loop performance is materially below specialized planners (Table 1: Val14 R 84.05 vs. PDM-Closed 92.12 / LLM-ASSIST 92.20; Test14 R 81.35 vs. ~90+ for several baselines). The paper frames this as “competitive” and “no severe degradation,” which is defensible for a feasibility check, but if HIO-conditioned RL reward R_z and diffusion conditioning (Eqs. 8–10) are load-bearing for intent alignment, the reactive gap needs analysis: does intent conditioning trade safety/reactivity for comfort, or is HICP simply a weaker base planner? A controlled comparison of the same planner backbone with and without HIO under reactive agents would isolate the effect.
minor comments (7)
  1. Related Work, ToM paragraph: citation placeholder “(??)” remains unresolved.
  2. Eq. (1) and surrounding text: “multimodal intent evidence” is used in the training paragraph, but HID inputs as described are textual (talk, attributes, states, scenario). Clarify modality or reserve “multimodal” for the stated future extension.
  3. Table 1 formatting: “84.0587.20” appears to be a missing space/column break between Val14 R and Test14 NR for Intent2Drive.
  4. HID construction: report inter-annotator agreement (or agreement between real annotations and GPT-4o consistency checks) for LHS and HIO fields; “manual quality inspection” alone is hard to assess.
  5. HIO driving objective is treated as a discrete label (e.g., “Conservative”) in Tables 5–6 and evaluation, but the mapping from continuous factors (urgency, risk, condition) to that vocabulary is not fully specified in the main text.
  6. Figures 1–2 are referenced for HID pipeline and framework but, in the provided text, lack captions detailing field definitions; ensure self-contained figure legends in the camera-ready version.
  7. Minor wording: “always cannot be directly observed” → “cannot be directly observed”; “A V trust” spacing inconsistencies in case tables.

Circularity Check

1 steps flagged

HIR SFT reproduces author-defined LHS/HIO labels on HID (ordinary supervised learning); nuPlan closed-loop scores remain an independent external benchmark.

specific steps
  1. other [Methodology / Human Intent Reasoner / Training Strategy; HID construction; Tables 2–3]
    "HIR is trained via supervised fine-tuning (SFT) on the proposed HID using a Qwen3-4B backbone. Given the human intent information, the model is optimized to reason about the LHS … and to further generate the HIO. … our method significantly outperforms the pretrained Qwen3-4B model across all LHS attributes. … our method substantially improves HIO construction over the pretrained Qwen3-4B model."

    LHS = (g,u,c,p,r) and HIO = (o,d) are author-defined structured labels that already exist in every HID sample; SFT simply trains the LLM to emit those labels from the textual inputs. Reported “inference” and “construction” accuracies are therefore the training objective itself (plus a pretrained baseline), not an independent discovery or prediction of latent mental states. This is ordinary supervised evaluation on a self-constructed schema, not an equation-level tautology that forces the planning claims.

full rationale

The paper is an empirical systems/ML framework paper, not a first-principles derivation. HID supplies structured (Human Intent, LHS, HIO) annotations; HIR is SFT-trained to emit those same structured fields from textual inputs, so Tables 2–4 simply measure how well the fine-tuned model recovers the annotation schema versus a pretrained baseline. That is standard supervised learning, not a tautological reduction of a claimed scientific prediction to its fitted inputs. Route/trajectory planning is conditioned on the resulting HIO and evaluated on the external nuPlan closed-loop protocol (Table 1), which does not use HID labels; those numbers therefore supply independent evidence that injecting HIO does not destroy feasibility. Self-citations (Luo/Ding co-author works) appear only in related-work/intro and are not load-bearing uniqueness claims. No equation equates a derived quantity to a fitted parameter by construction, no uniqueness theorem is imported from the authors, and no ansatz is smuggled via self-citation. The residual mild circularity is only that “ToM reasoning” success is defined as matching the authors’ own latent-state schema; this does not force the planning results and warrants only a low score.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 4 invented entities

The central empirical claim rests on a hand-designed latent-state schema, a half-synthetic dataset, standard LLM/RL/diffusion machinery, and the domain premise that textual proxies plus ToM-style intermediate variables suffice for passenger-aligned planning. No free physical constants; free choices are architectural and annotation design decisions. Invented entities are framework constructs without independent external measurement handles beyond the authors’ labels.

free parameters (4)
  • LHS factor set (g,u,c,p,r)
    The five latent dimensions are chosen by the authors for ‘cognitive completeness’ and ‘planning sufficiency’; different factorizations would change all supervised targets and reported accuracies.
  • HIO discrete driving-objective vocabulary
    Trip objective and discrete styles such as ‘Conservative’ are author-defined abstractions that the planner reward and conditioning depend on; not derived from data.
  • HID real/synthetic split and GPT-4o generation policy
    Exactly 1,120 real + 1,120 synthetic samples and the conditioning/verification procedure are design choices that determine training distribution and reported gains.
  • LoRA / SFT and RL/diffusion training hyperparameters
    Rank, learning rates, reward weights R_z, diffusion noise schedule, etc., are not fully specified and affect the numbers in Tables 1–4.
axioms (5)
  • domain assumption Passenger intent is well-modeled as a latent Theory-of-Mind state that can be recovered from the listed explicit and implicit textual cues.
    Stated in Introduction and HIR section; underpins the entire hierarchical reasoning design.
  • ad hoc to paper The structured LHS tuple (goal, urgency, condition, preference, risk) is cognitively complete and planning-sufficient.
    Explicit design principles in Latent Human State Reasoning; not independently validated against alternative schemas.
  • ad hoc to paper GPT-4o synthetic samples that pass automatic consistency checks and manual inspection are distributionally useful for training HIR.
    HID construction section; half the training mass depends on this.
  • domain assumption Standard supervised fine-tuning, graph route RL, and conditional diffusion produce valid intent-conditioned plans when fed HIO.
    Methodology / HICP; relies on established ML practice rather than new theory.
  • domain assumption nuPlan closed-loop scores under non-reactive and reactive settings are an adequate proxy for planning feasibility after adding HIR.
    Experimental Setup and Table 1.
invented entities (4)
  • Latent Human State (LHS) no independent evidence
    purpose: Interpretable intermediate cognitive representation inferred by HIR before planning.
    Defined as m=(g,u,c,p,r); no external psychometric validation outside HID labels.
  • Human Intent Objective (HIO) no independent evidence
    purpose: Planner-compatible abstraction (trip objective + driving objective) derived from LHS.
    Defined as z=(o,d); interface invented for HICP conditioning.
  • Holistic Intent Dataset (HID) no independent evidence
    purpose: Provide structured supervision for explicit/implicit cues, LHS, and HIO.
    New 2,240-sample corpus; not previously published.
  • Human Intent Reasoner (HIR) / Hierarchical Intent-Conditioned Planner (HICP) no independent evidence
    purpose: End-to-end modules that implement ToM-style inference and intent-conditioned planning.
    System components of Intent2Drive; evaluated only inside this paper’s pipeline.

pith-pipeline@v1.1.0-grok45 · 17032 in / 3678 out tokens · 38863 ms · 2026-07-11T15:27:29.065964+00:00 · methodology

0 comments
read the original abstract

As autonomous vehicles progress toward fully driverless mobility, a critical question emerges: who understands and responds to passengers when the human driver is absent? Existing autonomous driving systems primarily optimize predefined navigation and control objectives from external scene observations, but they remain limited in perceiving and reasoning about in-cabin human intent. In this paper, we propose Intent2Drive, a unified framework for holistic human intent understanding and human-aligned planning. Instead of treating passenger intent as explicit commands alone, Intent2Drive models intent as a latent cognitive state shaped by language, personal attributes, emotional and physical conditions, behavioral signals, and situational context. To support this formulation, we construct a Holistic Intent Dataset (HID) that provides structured supervision over both explicit and implicit intent cues. Built upon HID, our Theory-of-Mind-inspired Human Intent Reasoner (HIR) infers a Latent Human State (LHS) and further translates it into a planner-compatible Human Intent Objective (HIO). We then introduce a Hierarchical Intent-Conditioned Planner (HICP) that incorporates HIO into route-level and trajectory-level planning, enabling driving behaviors to remain aligned with passenger needs across different planning horizons. Extensive experiments show that Intent2Drive improves structured human intent inference and HIO construction while preserving competitive closed-loop planning performance. These results demonstrate a promising step toward passenger-responsive autonomous driving systems that can reason about, interpret, and act upon human intent in driverless mobility.

Figures

Figures reproduced from arXiv: 2607.04670 by Bo Yu, Chenxi Liu, Ding Fan, Fengze Yang, Ruiqi Chen, Xiujin Liu, Xuewen Luo, Ye Cao.

Figure 1
Figure 1. Figure 1: Pipeline for HID Construction intent from observable cues; and (2) planning sufficiency, requiring the inferred representation to contain sufficient in￾formation for deriving planning objectives without directly encoding planning decisions. Accordingly, the LHS is de￾fined as m = (g, u, c, p, r), (2) where g denotes the user’s underlying goal, including both the trip purpose and destination, u denotes the … view at source ↗
Figure 2
Figure 2. Figure 2: Framework of Intent2Drive where G(·) denotes the HIO construction function imple￾mented by HIR. Compared with directly predicting plan￾ning objectives from observable human intent, the pro￾posed reasoning framework first infers the latent human state and subsequently derives the HIO, effectively sep￾arating human understanding from planning abstraction. Consequently, the HIO provides a compact, interpretab… 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

47 extracted references · 6 linked inside Pith

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2017 , eprint=

    Attention Is All You Need , author=. 2017 , eprint=

    2017

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    End-to-end autonomous driving: Challenges and frontiers , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2024 , publisher=

    2024

  13. [13]

    ACM Computing Surveys , volume=

    A survey of autonomous driving from a deep learning perspective , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  14. [14]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Drive as you speak: Enabling human-like interaction with large language models in autonomous vehicles , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  15. [15]

    arXiv preprint arXiv:2104.07246 , year=

    Human-in-the-loop deep reinforcement learning with application to autonomous driving , author=. arXiv preprint arXiv:2104.07246 , year=

    Pith/arXiv arXiv

  16. [16]

    What's Happening

    " What's Happening"-A Human-centered Multimodal Interpreter Explaining the Actions of Autonomous Vehicles , author=. Proceedings of the Winter Conference on Applications of Computer Vision , pages=

  17. [17]

    Engineering , volume=

    Toward human-in-the-loop AI: Enhancing deep reinforcement learning via real-time human guidance for autonomous driving , author=. Engineering , volume=. 2023 , publisher=

    2023

  18. [18]

    Computer-Aided Design and Applications , volume=

    User acceptance of autonomous vehicles: Review and perspectives on the role of the human-machine interfaces , author=. Computer-Aided Design and Applications , volume=

  19. [19]

    IEEE Transactions on Intelligent Transportation Systems , volume=

    Toward human-vehicle collaboration for automated vehicles: A review and perspective , author=. IEEE Transactions on Intelligent Transportation Systems , volume=. 2025 , publisher=

    2025

  20. [20]

    International Journal of Human--Computer Interaction , volume=

    Devising a high-level command language for the teleoperation of autonomous vehicles , author=. International Journal of Human--Computer Interaction , volume=. 2025 , publisher=

    2025

  21. [21]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Human-centric autonomous systems with llms for user command reasoning , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  22. [22]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Learning autonomous driving tasks via human feedbacks with large language models , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  23. [23]

    Active preference-based learning of reward functions , author=

  24. [24]

    arXiv preprint arXiv:2501.15564 , year=

    Diffusion-based planning for autonomous driving with flexible guidance , author=. arXiv preprint arXiv:2501.15564 , year=

    Pith/arXiv arXiv

  25. [25]

    Device , volume=

    Driving intention understanding for autonomous vehicles: Current approaches and beyond , author=. Device , volume=. 2024 , publisher=

    2024

  26. [26]

    arXiv preprint arXiv:2603.25740 , year=

    Drive My Way: Preference Alignment of Vision-Language-Action Model for Personalized Driving , author=. arXiv preprint arXiv:2603.25740 , year=

    arXiv

  27. [27]

    Autonomous Intelligent Systems , volume=

    ChatMPC: a language-driven model predictive control framework for adaptive and personalized autonomous driving , author=. Autonomous Intelligent Systems , volume=. 2025 , publisher=

    2025

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Lampilot: An open benchmark dataset for autonomous driving with language model programs , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [29]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Understanding natural language commands for robotic navigation and mobile manipulation , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  30. [30]

    2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI) , pages=

    Legibility and predictability of robot motion , author=. 2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI) , pages=. 2013 , organization=

    2013

  31. [31]

    Advances in neural information processing systems , volume=

    Cooperative inverse reinforcement learning , author=. Advances in neural information processing systems , volume=

  32. [32]

    Advances in neural information processing systems , volume=

    Inverse reward design , author=. Advances in neural information processing systems , volume=

  33. [33]

    arXiv preprint arXiv:2106.11810 , year=

    nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles , author=. arXiv preprint arXiv:2106.11810 , year=

    Pith/arXiv arXiv

  34. [34]

    Physical review E , volume=

    Congested traffic states in empirical observations and microscopic simulations , author=. Physical review E , volume=. 2000 , publisher=

    2000

  35. [35]

    Conference on Robot Learning , pages=

    Urban driver: Learning to drive from real-world demonstrations using policy gradients , author=. Conference on Robot Learning , pages=. 2022 , organization=

    2022

  36. [36]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Rethinking imitation-based planners for autonomous driving , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  37. [37]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  38. [38]

    arXiv preprint arXiv:2404.14327 , year=

    Pluto: Pushing the limit of imitation learning-based planning for autonomous driving , author=. arXiv preprint arXiv:2404.14327 , year=

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2510.11083 , year=

    Flow matching-based autonomous driving planning with advanced interactive behavior modeling , author=. arXiv preprint arXiv:2510.11083 , year=

    arXiv

  40. [40]

    arXiv preprint arXiv:2401.00125 , year=

    Llm-assist: Enhancing closed-loop planning with language-based reasoning , author=. arXiv preprint arXiv:2401.00125 , year=

    Pith/arXiv arXiv

  41. [41]

    Advances in neural information processing systems , volume=

    Deep reinforcement learning from human preferences , author=. Advances in neural information processing systems , volume=

  42. [42]

    Behavioral and Brain Sciences , volume=

    Does the chimpanzee have a theory of mind? , author=. Behavioral and Brain Sciences , volume=

  43. [43]

    Proceedings of the Annual Meeting of the Cognitive Science Society , year=

    Bayesian theory of mind: Modeling joint belief-desire attribution , author=. Proceedings of the Annual Meeting of the Cognitive Science Society , year=

  44. [44]

    The Springer Handbook of Robotics , year=

    Human-Robot Collaboration and Planning , author=. The Springer Handbook of Robotics , year=

  45. [45]

    arXiv preprint arXiv:2508.16947 , year=

    Drive As You Like: Strategy-Level Motion Planning Based on A Multi-Head Diffusion Model , author=. arXiv preprint arXiv:2508.16947 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2603.10441 , year=

    KnowDiffuser: A Knowledge-Guided Diffusion Planner with LLM Reasoning , author=. arXiv preprint arXiv:2603.10441 , year=

    arXiv

  47. [47]

    Extended abstracts of the CHI conference on human factors in computing systems , pages=

    Theory of mind in human-ai interaction , author=. Extended abstracts of the CHI conference on human factors in computing systems , pages=