Pith. sign in

REVIEW 3 major objections 6 minor 35 references

Logic-Guided Socially-aware Robot Navigation World Model

T0 review · 3 major / 6 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 NaviWM augments LLM navigation with a spatial-temporal world model and deductive chain-of-thought, lifting success to 0.8 in crowded simulations where plain LLM planners fail.

desk verdict NaviWM's world-model-plus-deductive-CoT combination is a genuine novelty with believable gains, but the 'always exists' guarantee is assumed, not proved, and the experiments need hardening before the claims can hold. read the letter →

arxiv 2510.23509 v3 pith:UENUODEK submitted 2025-10-27 cs.RO

Weizheng Wang , Obi Ike , Soyun Choi , Sungeun Hong , Aniket Bera , Byung-Cheol Min This is my paper
classification cs.RO
keywords socialrobotnavigationlargelanguagemodelsworldmodeldeductivereasoningchain-of-thoughtfirst-orderlogichierarchicalconstraintsspatial-temporalgraph
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 argues that large language models alone cannot reliably plan safe, socially compliant robot paths because they lack physical grounding and logical consistency. NaviWM addresses this by building a spatial-temporal world model of the scene and guiding the LLM through a deductive, step-by-step chain-of-thought, with social norms encoded as first-order logic constraints. The framework claims a hierarchy of guaranteed fallback behaviors: if ideal social spacing cannot be met, the robot still ensures collision avoidance and on-time arrival. In simulated crowded environments, NaviWM raises success rates to 0.8 and 0.7 with 5 and 10 humans, respectively, where most baseline LLMs score near zero, and cuts uncomfortable interactions substantially. The reader should care because it offers a concrete route to making LLM-based planners verifiable and safety-guaranteed in human spaces.

What carries the argument

The load-bearing mechanism is the hierarchical logical proposition: ∃a∈A ⇒ (Es∧Ed∧¬Ec∧Et) ∨ (Es∧¬Ec∧Et) ∨ (Ed∧¬Ec∧Et) ∨ (¬Ec∧Et), which orders social constraints by priority. A Gentzen-style natural-deduction proof tree guides the LLM step by step: each candidate action is verified against Level 1; on failure, the deduction re-applies with weaker constraints D2..D4, and the action satisfying only D4 (collision-free and on time) is always output as the ultimate fallback. The world model encodes the scene as a spatial-temporal graph whose vertices (robot/humans with attributes) and edges (relative distances, velocities, dynamics) ground the LLM's reasoning in measurable physical quantities.

What would settle it

Run the NaviWM planner in a simulated or real environment where a goal is deliberately placed such that every collision-free path to it takes longer than the time limit; if the planner either collides, times out, or outputs a non-executable action, the Level-4 fallback guarantee does not hold.

Watch

Extended reading notes

Core claim

NaviWM's central claim is that social navigation can be formalized as a hierarchical logical constraint-satisfaction problem, and that an LLM guided by a dedicated deductive chain-of-thought can solve it reliably. The system encodes activity-aware distancing, proxemic boundaries, collision avoidance, and a time limit as first-order logic predicates. Navigation is then reduced to finding an action satisfying the disjunction of four progressively relaxed constraint combinations, so that even when ideal social behavior is infeasible, a collision-free, on-time action is guaranteed to exist by construction. Combined with a spatial-temporal world model graph that feeds the LLM precise positions, v

Load-bearing premise

At every decision step there exists at least one robot action that is simultaneously collision-free and reaches the goal within the time limit; the paper asserts this rather than proving it.

Editorial extensions

If this is right

  • If NaviWM holds, LLM-based planners can offer formal social-safety guarantees rather than probabilistic behavior, since the hierarchical logic ensures a fallback action even when ideal social norms are infeasible.
  • The hierarchical fallback means social compliance degrades gracefully: the robot prefers activity-appropriate spacing, then general proxemics, then merely collision-free and timely motion, which is a principled way to handle crowded scenes.
  • Success in 10-human simulations suggests the approach scales to denser crowds where plain LLMs completely fail, pointing to a practical path for real-world social robots.
  • The ablation results indicate the world model, not just the logic prompt, drives the large reductions in uncomfortable interactions, so grounding in geometry and dynamics is essential for social compliance.

Reading between the lines

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

  • The formal guarantee depends on existence of a Level-4 action; in genuinely trap-like environments the time limit plus collision avoidance can be unsatisfiable, so the proof tree would have no valid leaf—this suggests a natural extension to verify or adaptively relax the time constraint.
  • The approach offers a template for other safety-critical LLM decision tasks: express requirements as a hierarchy of first-order logic constraints and force the LLM to walk a proof tree, rather than trusting free-form reasoning, which could generalize beyond navigation.
  • A testable extension: measure how the success/social-compliance trade-off shifts as the time limit shrinks; this would reveal when the Level-4 guarantee actually binds and how conservative the planner becomes in time-stressed scenarios.
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 / 6 minor

Summary. The paper proposes NaviWM, an LLM-based socially-aware robot navigation planner that couples a spatial-temporal world model with a deductive chain-of-thought prompt. It formalizes social and safety constraints in first-order logic as a four-level hierarchical disjunction, claiming that this structure guarantees the existence of a collision-free, timely action at every step. In an OpenAI-gym style simulation with 5 and 10 human agents, the authors report that NaviWM outperforms several LLM baselines and ablation variants in success rate, navigation efficiency, and social compliance metrics. The central advertised contributions are the world model, the deductive CoT framework, and the claimed 'formal guarantees' from combining logic with LLM reasoning.

Significance. The general idea of grounding LLM-based navigation with an explicit world model and logic-guided reasoning is timely and plausible, and the hierarchical constraint relaxation in Eq. (1) is a sensible way to manage trade-offs between social compliance and safety. If the empirical results were statistically solid, NaviWM would be a useful step toward more reliable and interpretable LLM-driven social navigation. The manuscript also makes its design choices fairly transparent, and the ablation set is helpful. However, the paper's headline claim of a 'provable guarantee' is not established: the existence of a Level-4 action is assumed rather than proved, and the verification step is an LLM self-check rather than an external formal verifier. The experimental evidence, as reported, also lacks the statistical grounding needed to support 'substantial performance improvements.'

major comments (3)
  1. [§III-A, Eq. (1)] The assertion that 'the disjunctive formulation ensures that a valid action always exists' is not proved. Equation (1) defines what a valid action is, but it does not show that such an action exists at every decision step. The proof tree in Fig. 3 assumes 'a∈A → Φ(a)' as step [2], effectively assuming the consequent. In the dense 10-human scenario with T_max=50s, there can be states where no collision-free, time-feasible trajectory exists (e.g., a moving human barrier). The paper reports SR=0.7 without analyzing whether failures come from this fundamental infeasibility rather than LLM/planning errors. The claimed 'provable guarantee' is unsupported; it should either be proved under explicit environmental assumptions or removed/reframed.
  2. [§V-A, §V-B, Table III] The empirical comparison lacks the information needed to support 'substantial performance improvements.' No number of runs, random seeds, error bars, or statistical tests are reported; many baseline SR values are 0 or 0.1, and with a single simulator and one table it is unclear whether the differences are meaningful. The baseline prompting protocol is also undocumented: it is not stated what prompt, examples, or world-model information the bare LLM baselines received, so the comparison may be confounded. At minimum, the authors should report multiple seeds, error bars, significance tests, and exact prompt templates for each condition.
  3. [§IV-C, Fig. 2 and Fig. 3] The 'verification step' is performed by the same LLM that generates the candidate action; Fig. 2 labels it 'self-validation.' This is not independent verification and does not constitute a formal proof. The Gentzen tree is a prompt-level guide, not a machine-checked derivation, and no external verifier or proof checker is used. Consequently, the claims of 'verifiable reasoning' and 'formal guarantees' in the Abstract and Introduction overstate what is demonstrated. The authors should implement a genuine verifier (e.g., checking the FOL constraints against the world model) or rephrase these claims as heuristic logic-guided reasoning.
minor comments (6)
  1. [§V-B and Table III] The metric name is introduced as 'Uncomfortable Frequency (UF)' but the table uses 'UI'; use one acronym consistently.
  2. [§VI-C] The text states that NaviWM achieved SR=0.8 in both the 5- and 10-human environments, but Table III reports SR=0.7 for the 10-human case. Please correct this inconsistency.
  3. [§VII] The conclusion says Abl-WM-GPT reduced GPT-4o's UI from 70 to 20, while Table III and §VI-B report 21. Update the number for consistency.
  4. [§III-A, Eq. (1)] The notation '∃a∈ A =⇒' is non-standard; write '∃a∈ A such that ...' and make the scope of the disjunction explicit. The current form is ambiguous.
  5. [Table III] The NaviWM row lacks spacing (e.g., '0.810.93'), making the table difficult to read. Ensure the typeset table clearly separates columns.
  6. [§V-A] The simulation environment [35] is not described. Please add details of the simulator, human motion model, episode termination conditions, and how the 50-second time limit is enforced.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'always exists' safety guarantee is Eq. (1)'s existential assumption restated as a theorem; empirical results remain non-circular.

  1. self definitional [Sec. III-A, Eq. (1)]
    "The disjunctive formulation ensures that a valid action always exists, as Level 4 requires only the most fundamental safety and task completion constraints."

    Eq. (1) is itself only the existential statement ∃a∈A (Level1∨...∨Level4), and 'a valid action' is defined as an action satisfying one of those levels. Claiming that the disjunction ensures existence is therefore restating the existential quantifier, not proving it. The paper never shows that an action satisfying even the weakest level (¬Ec∧Et) exists in every state; that existence is exactly the formal guarantee the paper claims to derive.

  2. other [Sec. IV-C, Fig. 3]
    "The verification proof tree is shown in Fig. 3, where Φ(a) ... denotes the objective event ... a∈ A →Φ(a) [2]∃a 0 ∈ A(∃E) [1,2] Φ(a0)"

    In the tree, a∈A→Φ(a) is an undischarged premise, not a derived theorem; from it the tree obtains Φ(a0). This is assuming the consequent: the existence of an action satisfying the objective (in particular the D4 fallback ¬Ec∧Et) is the premise, and the claimed 'guarantee' is merely that premise restated. The verification step therefore cannot establish existence; it only checks an action once one has already been assumed to exist.

full rationale

Most of the empirical pipeline is self-contained: NaviWM does not fit parameters to the reported SR/UI/HA/NP/NT metrics; those are measured against published LLM baselines and ablations, so the main performance claims are not circular. The self-citation to the Navistar-based OpenAI-gym environment [35] is a testbed citation and is not load-bearing for the derivation chain. However, the paper's headline formal contribution—the 'provable guarantee' that a valid action always exists—is circular. Eq. (1) merely states the existential condition, Sec. III-A then cites that disjunctive form as the reason existence is guaranteed, and Fig. 3's proof tree assumes a∈A→Φ(a) as a premise. The formal-safety contribution thus reduces to its input assumption by construction, even though the empirical SR=0.7 and social-compliance gains have independent experimental content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-set constraint parameters (time limit, activity distances, minimum distance) and on domain assumptions about perfect sensing, constant-velocity human motion, the existence of a legal action at every step, and the LLM's faithful execution of the logical procedure. No new physical entities are introduced.

free parameters (3)
  • T_max = 50 s
    Hand-chosen time limit for task completion ('within 50 seconds') in the Time-Constraint. It directly affects which actions satisfy E_t and the success-rate outcome.
  • Pref(H|a_H) = not stated in paper (example uses 1.5 m for watching phone)
    Per-activity preferred distances for Activity-Awareness. Values are not given; they determine Level 1 compliance and the HA metric. The paper references proxemic norms but does not list the table of values.
  • d_min = not stated
    Minimum universal distancing threshold in Distance-Awareness. Set by hand; affects Level 1/3 compliance.
assumptions (4)
  • domain assumption The environment provides ground-truth positions, velocities, and activities of all agents to the world model; no perception noise is modeled.
    Sec. IV-A constructs G_t from observation. Real robots would need perception, and errors could change decisions; the simulation bypasses this.
  • ad hoc to paper The Level-4 action (collision-free and on time) exists at every decision step.
    Sec. III-A asserts 'a valid action always exists' because Level 4 only requires ¬E_c ∧ E_t. No proof is given that such a path exists under T_max=50s in each episode.
  • domain assumption The LLM reliably executes the hand-crafted deductive CoT steps and its own verification step, producing actions that satisfy the declared constraints.
    The 'formal guarantees' depend on the LLM following the Gentzen-style deduction; the verification step in Fig. 3 is performed by the LLM, not a verified checker.
  • domain assumption Human motion is predicted by linear extrapolation of current velocities over a 0.5 s horizon.
    Fig. 2's sample query computes future positions as p + v·0.5; this assumes constant velocity and ignores acceleration or intent changes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Logic-Guided Socially-aware Robot Navigation World Model." pith.science (2026). https://pith.science/paper/UENUODEK

@misc{pith2026251023509,
  author       = {Pith},
  title        = {Pith review of: Logic-Guided Socially-aware Robot Navigation World Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UENUODEK}},
  note         = {Machine review of arXiv:2510.23509}
}
read the original abstract

Social robot navigation increasingly relies on large language models for reasoning, path planning, and enabling movement in dynamic human spaces. However, relying solely on LLMs for planning often leads to unpredictable and unsafe behaviors, especially in dynamic human spaces, due to limited physical grounding and weak logical consistency. In this work, we introduce NaviWM, a socially-aware robot Navigation World Model that augments LLM reasoning with a structured world model and a logic-driven chain-of-thought process. NaviWM consists of two main components: (1) a spatial-temporal world model that captures the positions, velocities, and activities of agents in the environment, and (2) a deductive reasoning module that guides LLMs through a multi-step, logic-based inference process. This integration enables the robot to generate navigation decisions that are both socially compliant and physically safe, under well-defined constraints such as personal space, collision avoidance, and timing. Unlike previous methods based on prompting or fine-tuning, NaviWM encodes social norms as first-order logic, enabling interpretable and verifiable reasoning. Experiments show that NaviWM improves success rates and reduces social violations, particularly in crowded environments. These results demonstrate the benefit of combining formal reasoning with LLMs for robust social navigation. Additional experimental details and demo videos for this work can be found at: https://sites.google.com/view/NaviWM.

Figures

Figures reproduced from arXiv: 2510.23509 by the authors.

Figure 1
Figure 1. The world model illustration of social robot navigation scenario: NaviWM constructs the world model from local observation to capture both agent vertex information and environmental semantic information with respect to spatial-temporal HRI features. TABLE II LOGICAL FORM EXAMPLES OF SOCIAL ROBOT NAVIGATION WORLD MODEL Function (obj.) Elements Example Templates Vertex (robot) (agent=r, location=p, velocity=v, target=… view at source ↗
Figure 2
Figure 2. The architecture of NaviWM: (1). NaviWM constructs a world model to represent environmental description from local observation; (2). The deductive CoT algorithm is encoded as prompt engineering for LLM reasoning guidance; (3). The inference chains are generated step-by-step with respect to logical procedure and self-validation step; (4). Final robot action is obtained in the final step of the deductive CoT. where th… view at source ↗
Figure 3
Figure 3. Verification proof tree in NaviWM illustrating the deductive reasoning process that derives the collision-free action [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 10 linked inside Pith

  1. [2]

    Tree of thoughts: Deliberate problem solving with large language models,

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y . Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  2. [1]

    TPO: Aligning large language models with multi-branch & multi-step preference trees,

    W. Liao, X. Chu, and Y . Wang, “TPO: Aligning large language models with multi-branch & multi-step preference trees,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=O0sQ9CPzai

    2025

  3. [3]

    Natural language embedded programs for hybrid language symbolic reasoning,

    T. Zhang, J. Ge, H. Luo, Y .-S. Chuang, M. Gao, Y . Gong, Y . Kim, X. Wu, H. Meng, and J. Glass, “Natural language embedded programs for hybrid language symbolic reasoning,” inFindings of the Association for Computational Linguistics: NAACL 2024, K. Duh, H. Gomez, and S. Bethard, Eds. Mexico City, Mexico: Association for Computational Linguistics, Jun. 20...

    2024

  4. [4]

    Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action,

    D. Shah, B. Osi ´nski, S. Levineet al., “Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action,” inConference on robot learning. PMLR, 2023, pp. 492–504

    2023

  5. [5]

    Multimodal llm guided exploration and active mapping using fisher information,

    W. Jiang, B. Lei, K. Ashton, and K. Daniilidis, “Multimodal llm guided exploration and active mapping using fisher information,”arXiv preprint arXiv:2410.17422, 2024

    arXiv 2024

  6. [6]

    Navgpt: Explicit reasoning in vision- and-language navigation with large language models,

    G. Zhou, Y . Hong, and Q. Wu, “Navgpt: Explicit reasoning in vision- and-language navigation with large language models,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 7, 2024, pp. 7641–7649

    2024

  7. [7]

    Social-llava: Enhancing robot navigation through human-language reasoning in social spaces,

    A. Payandeh, D. Song, M. Nazeri, J. Liang, P. Mukherjee, A. H. Raj, Y . Kong, D. Manocha, and X. Xiao, “Social-llava: Enhancing robot navigation through human-language reasoning in social spaces,”arXiv preprint arXiv:2501.09024, 2024

    arXiv 2024

  8. [8]

    L2r-nav: A large language model-enhanced framework for robotic navigation,

    X. Wu, Q. Li, C. Chen, X. Zhang, H. Zhao, and J. Niu, “L2r-nav: A large language model-enhanced framework for robotic navigation,” inInternational Conference on Knowledge Science, Engineering and Management. Springer, 2024, pp. 73–84

    2024

Show all 35 references
  1. [9]

    Language and sketching: An llm-driven interactive multimodal multitask robot navigation framework,

    W. Zu, W. Song, R. Chen, Z. Guo, F. Sun, Z. Tian, W. Pan, and J. Wang, “Language and sketching: An llm-driven interactive multimodal multitask robot navigation framework,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 1019–1025

    2024

  2. [10]

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

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, E. Chi, Q. Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” inAdvances in Neural Information Processing Systems, 2022

    2022

  3. [11]

    Large language models are zero-shot reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large language models are zero-shot reasoners,”Advances in neural information processing systems, vol. 35, pp. 22 199–22 213, 2022

    2022

  4. [12]

    Automatic chain of thought prompting in large language models,

    Z. Zhang, A. Zhang, M. Li, and A. Smola, “Automatic chain of thought prompting in large language models,”arXiv preprint arXiv:2210.03493, 2022

    2022 arXiv

  5. [13]

    Self-refine: Iter- ative refinement with self-feedback,

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yanget al., “Self-refine: Iter- ative refinement with self-feedback,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  6. [14]

    Knowledge-driven cot: Exploring faithful reasoning in llms for knowledge-intensive question answering,

    K. Wang, F. Duan, S. Wang, P. Li, Y . Xian, C. Yin, W. Rong, and Z. Xiong, “Knowledge-driven cot: Exploring faithful reasoning in llms for knowledge-intensive question answering,”arXiv preprint arXiv:2308.13259, 2023

    2023 arXiv

  7. [15]

    Graph of thoughts: Solving elaborate problems with large language models,

    M. Besta, N. Blach, A. Kubicek, R. Gerstenberger, M. Podstawski, L. Gianinazzi, J. Gajda, T. Lehmann, H. Niewiadomski, P. Nyczyket al., “Graph of thoughts: Solving elaborate problems with large language models,” inProceedings of the AAAI Conference on Artificial Intelligence, ...

    2024

  8. [16]

    Deductive verification of chain-of-thought reasoning,

    Z. Ling, Y . Fang, X. Li, Z. Huang, M. Lee, R. Memisevic, and H. Su, “Deductive verification of chain-of-thought reasoning,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  9. [17]

    Multimodal chain-of-thought reasoning in language models,

    Z. Zhang, A. Zhang, M. Li, H. Zhao, G. Karypis, and A. Smola, “Multimodal chain-of-thought reasoning in language models,”arXiv preprint arXiv:2302.00923, 2023

    2023 arXiv

  10. [18]

    Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies,

    M. Geva, D. Khashabi, E. Segal, T. Khot, D. Roth, and J. Berant, “Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies,”Transactions of the Association for Computational Linguistics, vol. 9, pp. 346–361, 2021

    2021

  11. [19]

    Skeleton-of- thought: Prompting llms for efficient parallel generation,

    X. Ning, Z. Lin, Z. Zhou, Z. Wang, H. Yang, and Y . Wang, “Skeleton-of- thought: Prompting llms for efficient parallel generation,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  12. [20]

    Analyzing chain-of-thought prompting in large language models via gradient-based feature attributions,

    S. Wu, E. M. Shen, C. Badrinath, J. Ma, and H. Lakkaraju, “Analyzing chain-of-thought prompting in large language models via gradient-based feature attributions,”arXiv preprint arXiv:2307.13339, 2023

    2023 arXiv

  13. [21]

    World models for math story problems,

    A. Opedal, N. Stoehr, A. Saparov, and M. Sachan, “World models for math story problems,” inFindings of the Association for Computational Linguistics: ACL 2023. Toronto, Canada: Association for Computational Linguistics, Jul. 2023, pp. 9088–9115. [Online]. Available: https://ac...

    2023

  14. [22]

    Towards general natural language understanding with probabilistic worldbuilding,

    A. Saparov and T. M. Mitchell, “Towards general natural language understanding with probabilistic worldbuilding,”Transactions of the Association for Computational Linguistics, vol. 10, pp. 325–342, 2022

    2022

  15. [23]

    Language models meet world models: Embodied experiences enhance language models,

    J. Xiang, T. Tao, Y . Gu, T. Shu, Z. Wang, Z. Yang, and Z. Hu, “Language models meet world models: Embodied experiences enhance language models,”Advances in neural information processing systems, vol. 36, 2024

    2024

  16. [24]

    Using large language model to solve and explain physics word problems approaching human level,

    J. Ding, Y . Cen, and X. Wei, “Using large language model to solve and explain physics word problems approaching human level,”arXiv preprint arXiv:2309.08182, 2023

    2023 arXiv

  17. [25]

    From task structures to world models: what do llms know?

    I. Yildirim and L. Paul, “From task structures to world models: what do llms know?”Trends in Cognitive Sciences, 2024

    2024

  18. [26]

    Reasoning with language model is planning with world model,

    S. Hao, Y . Gu, H. Ma, J. J. Hong, Z. Wang, D. Z. Wang, and Z. Hu, “Reasoning with language model is planning with world model,” inThe 2023 Conference on Empirical Methods in Natural Language Processing, 2023. [Online]. Available: https: //openreview.net/forum?id=VTWWvYtF1R

    2023

  19. [27]

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

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhouet al., “Chain-of-thought prompting elicits reasoning in large language models,”Advances in neural information processing systems, vol. 35, pp. 24 824–24 837, 2022

    2022

  20. [28]

    Solving olympiad geometry without human demonstrations,

    T. H. Trinh, Y . Wu, Q. V . Le, H. He, and T. Luong, “Solving olympiad geometry without human demonstrations,”Nature, vol. 625, no. 7995, pp. 476–482, 2024

    2024

  21. [29]

    Faithful logical reasoning via symbolic chain-of-thought,

    J. Xu, H. Fei, L. Pan, Q. Liu, M.-L. Lee, and W. Hsu, “Faithful logical reasoning via symbolic chain-of-thought,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), L.-W. Ku, A. Martins, and V . Srikumar, Eds. Ban...

    2024

  22. [30]

    MathGAP: Out-of-distribution evaluation on problems with arbitrarily complex proofs,

    A. Opedal, H. Shirakami, B. Schölkopf, A. Saparov, and M. Sachan, “MathGAP: Out-of-distribution evaluation on problems with arbitrarily complex proofs,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview. net/fo...

    2025

  23. [31]

    Untersuchungen über das logische schließen. i,

    G. Gentzen, “Untersuchungen über das logische schließen. i,”Mathema- tische zeitschrift, vol. 39, no. 1, pp. 176–210, 1935

    1935

  24. [32]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  25. [33]

    Deepseek-v3 technical report,

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruanet al., “Deepseek-v3 technical report,”arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  26. [34]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  27. [35]

    Navistar: Socially aware robot navigation with hybrid spatio-temporal graph transformer and preference learning,

    W. Wang, R. Wang, L. Mao, and B.-C. Min, “Navistar: Socially aware robot navigation with hybrid spatio-temporal graph transformer and preference learning,”2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), p. 11348–11355, Oct 2023

    2023

Pith tools

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