Pith. sign in

REVIEW 5 major objections 6 minor 29 references

Can LLMs Learn to Map the World from Local Descriptions?

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Continual pre-training on local distance-and-azimuth sentences lets a small language model reconstruct the global layout: held-out pair distances reach 0.11% error and azimuths 0.79%, with absolute coordinates readable from hidden states.

desk verdict Solid synthetic-grid evidence that a small LLM can build a global spatial map from local relational text, but missing connectivity checks and release details keep it from being fully verified. read the letter →

arxiv 2505.20874 v1 pith:QYNGLJ5M submitted 2025-05-27 cs.CL

Sirui Xia , Aili Chen , Xintao Wang , Tinghui Zhu , Yikai Zhang , Jiangjie Chen , Yanghua Xiao This is my paper
classification cs.CL
keywords spatialcognitioncontinualpre-traininglatentrepresentationsnavigationworldmodelssyntheticenvironmentdistanceandazimuthpredictionshortestpathplanning
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 asks whether a language model can assemble one coherent map from many small local descriptions, the way a person builds a mental map from sentences like 'the library is 100 meters southeast of the park.' It argues yes, in a synthetic 100×100 grid: after continual pre-training on templated pairwise distance-and-azimuth sentences and route descriptions, a 0.5B-parameter language model predicts relationships between unseen pairs of places, with distance error around 0.11% and azimuth error around 0.79% at an 80:20 split. Its hidden states encode absolute coordinates, and trajectory training lets it plan shortest paths between place pairs whose direct path was never seen, reaching 83.63% exact accuracy in the bridged-exposure setting. If this is right, global spatial cognition can emerge from fragmented relational language without coordinates or vision, at least in a clean simulated environment. The paper's own limitations section adds that the training sacrifices general language ability and that the internal mechanism is not fully analyzed.

What carries the argument

Two-stage continual pre-training is the engine. In stage one, the model reads templated sentences of the form 'the distance from p_i to p_j is X km, and the azimuth is Y degrees' for many pairs of points of interest on a synthetic 100×100 grid; in stage two it reads multi-step route descriptions generated from shortest-path computations on the grid's weighted road graph. This data stream is what lets a small language model internalize a global layout. The work the machinery does is shown by two probes: a multilayer-perceptron regressor that recovers absolute (x,y) coordinates from the model's last hidden states (below 1 km mean absolute error), and a compositional probe that predicts held-out distance and azimuth from the concatenated hidden states of two points. The machinery makes the paper's claim testable: if the hidden states did not carry coordinate information, the probe would fail.

What would settle it

Train the same model on two groups of places with no relational sentence connecting the groups, then ask it for the distance and azimuth between a place in one group and a place in the other: if accuracy stays high, the connectivity premise is unnecessary, and if it collapses, the reported global-map inference depends on the training graph being connected.

Watch

Extended reading notes

Core claim

The paper's central claim is that a language model, trained only on local relative descriptions and local trajectory descriptions, constructs a globally coherent spatial map and encodes it in its internal representations, all without coordinates, images, or a complete map. Concretely, after continual pre-training on pairwise distance-and-azimuth sentences, the model predicts relations for held-out point pairs with about 0.11% distance error and 0.79% azimuth error at an 80:20 split; probes read absolute coordinates from its hidden states with sub-kilometre error; and the latent layout is distance- and angle-consistent with the true geography. After trajectory training, the model plans shortest paths between held-out pairs with 83.63% exact match accuracy in the bridged-exposure setting, even though those specific paths were never shown. The claim is explicitly qualified: the environment is a noise-free synthetic grid, the model is small, and navigation is not robust to perturbations, with recovery depending on training-data frequency of turning points.

Load-bearing premise

The load-bearing premise is that the relational training pairs form one connected graph that pins down every point's location relative to the rest, so that if a group of places were never linked in, no model could infer where it sits.

Editorial extensions

If this is right

  • A language model can act as a spatial world model built purely from natural-language observations of local geometry.
  • Pairwise distance-and-azimuth sentences are enough to recover absolute layouts, so map induction can be posed as an ordinary language-modelling objective.
  • Multi-step route descriptions teach road topology, not only place positions, enabling shortest-path plans between pairs never seen connected.
  • Absolute coordinate information is readable from internal representations, which supports compositional spatial reasoning tasks such as judging whether a point lies inside a region.
  • Navigation is brittle under perturbations, so the induced road model is fragmented rather than a continuous global map.

Reading between the lines

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

  • The near-perfect metrics likely depend on the training pair graph being connected and richly overcomplete; the paper does not report connectivity or pair-count statistics, so this precondition should be checked before transferring the method to real data.
  • Real human descriptions add noise, missing links, and inconsistent reference frames; a natural stress test is to add azimuth noise or randomly delete 50% of the relational sentences and measure how fast distance and azimuth errors grow.
  • A minimal-data experiment, using just enough pairwise links to connect the whole map instead of the full dense set, would separate genuine global inference from interpolation over a dense relational graph.
  • The model's failure to recover after road-name perturbations suggests it may rely on frequent turning-point patterns rather than a general pathfinding computation; renaming roads or reweighting edges after training would test this directly.
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

5 major / 6 minor

Summary. This paper asks whether an LLM, continually pre-trained on synthetic relational descriptions (distances and azimuths between points of interest) and local shortest-path trajectory descriptions, constructs a global spatial model. In a 100x100 grid with 1024 POIs and 200 roads, Qwen2.5-0.5B is trained in two stages (relational, then trajectory) and evaluated on explicit prediction of unseen POI-pair distances/azimuths, probing of absolute coordinates in hidden states, shortest-path planning to unseen pairs, and robustness to perturbations. The headline results are distance MRPE 0.11% and azimuth MRPE 0.79% at an 80:20 split; probe MAE 0.78/0.71 km for x/y coordinates; and 83.63% exact shortest-path accuracy in the Bridged Exposure setting. The paper concludes that LLMs can develop global spatial understanding from local, relative observations.

Significance. If the result holds, the paper makes a useful empirical contribution to the study of emergent world models in LLMs: it provides a controlled, separable testbed (perception vs. navigation vs. latent representation vs. robustness) and a clear falsifiable prediction (unseen POI-pair distances and held-out probe coordinates). The two-stage continual pre-training design and the inclusion of a perturbation analysis are strengths, as is the honest Limitations section that acknowledges the loss of general language capabilities and the lack of mechanism analysis. The significance is limited by the synthetic, noise-free, templated nature of the data, the small model scale (0.5B), and the absence of released code/data; the results should therefore be read as an existence proof in a toy environment rather than as evidence about real-world spatial language.

major comments (5)
  1. [Section 2 (Data) and Section 3.1 (Setting)] The relational training set's sampling scheme is not reported, and the connectivity of the graph formed by training POI pairs is never checked. Global layout recovery from pairwise distances and azimuths is only possible if this graph is connected; if it has multiple components, distances and azimuths between components are unconstrained and no model could generalize to test pairs across components. The paper must state how many (pi,pj) pairs were generated, how they were sampled (uniform over all pairs? local neighborhoods?), verify that the training-pair graph is connected, and confirm that held-out test pairs are within the same component. Without this, the near-zero MRPE in Table 1 and the coherent probe coordinates in Table 2 are not interpretable as evidence of a genuinely global map.
  2. [Section 2 (Simulation Environment)] The description of traversal weights is internally inconsistent: 'traversal weights w randomly sampled from [0.8, 1.2] ... higher weights indicate faster travel' implies that shortest-path cost should be distance/w, but the paper states that shortest paths are computed with Dijkstra's algorithm, whose standard interpretation of edge weights is as costs (higher = slower). The robustness analysis in Section 4.3 repeatedly refers to 'high-speed roads' and 'higher traversal speed,' suggesting w is intended as speed. Please clarify the exact cost function used in Dijkstra and reconcile the terminology; this affects the validity of the trajectory dataset and the interpretation of Figure 5 and Table 6.
  3. [Tables 1, 2, 4, 5, 6 and appendix tables] All experiments report single point estimates without standard deviations, confidence intervals, or the number of random seeds. The splits are described only as 'randomly assigned' or 'randomly select,' so it is unclear whether the claims (e.g., 0.11% MRPE, R²=1.00, 83.63% SPA) are stable across runs. Please run each configuration with at least 3-5 seeds and report mean and standard deviation (or a variance summary), and additionally report the numbers of training pairs and trajectories for each split.
  4. [Section 4.1 (Bridged Exposure)] The description 'generalizing shortest-path planning to unseen POI pairs' is stronger than what the Bridged Exposure setting actually tests. In this setting, Pheldout POIs do appear in training trajectories as origins or destinations paired with Pmain POIs; only paths between two Pheldout POIs are unseen. The model therefore has partial exposure to the roadmap locations of Pheldout. The genuinely unseen-region results (No-Exposure) show SPA=0.00% and SPD=5.33 km, which supports directional awareness but not exact path planning. Please rephrase the claim and make the distinction between 'unseen pairs' and 'unseen regions' explicit throughout the section and the abstract.
  5. [Reproducibility (Appendix B and Ethics Statement)] No code, data-generation scripts, or trained models are released, despite the Appendix stating that 'all models and tools are publicly available' without providing a link. Because the entire study depends on the exact templated string formats, the random POI placement, and the specific pair sampling, the results cannot be verified or extended without these artifacts. I request release of the synthetic data generator, the training scripts, and the evaluation code.
minor comments (6)
  1. [Figure 1 caption] The caption contains a typo: 'Firsr' should be 'First'.
  2. [Throughout] The model names are inconsistently typeset: 'QWEN 2.5-0.5B' and 'LlaMA-3.2-1B' should be 'Qwen2.5-0.5B' and 'LLaMA-3.2-1B'.
  3. [Table 3] The notation in the header (e.g., 'VMR (↑1.0)', 'FD (↓0.0)') is not defined in the table; add a note that ↑ indicates higher is better and ↓ indicates lower is better.
  4. [Section 2 and Table 8] Distances are expressed in kilometers in the text and in meters in the example data format (Table 8); standardize the units to avoid ambiguity.
  5. [Appendix E.2, Table 22] The Qwen2.5-1.5B row reports R²=1.61 for distance prediction, which is mathematically impossible (R² cannot exceed 1); this entry should be corrected.
  6. [References] The reference 'Qwen An Yang, Baosong Yang, ...' appears malformed; it should be attributed to the Qwen Team (e.g., 'Qwen Team, An Yang, ...').

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported predictions use genuine held-out splits, no parameters are fitted to test targets, and no load-bearing self-citation chain is present.

full rationale

The derivation chain is self-contained with respect to circularity. The central perception results (Table 1) are obtained by continual pre-training on pairwise distance/azimuth descriptions and evaluating on pairs not seen during training; no parameter is fitted to the held-out targets. The latent-coordinate probe (Table 2) trains an MLP regressor on 90% of POIs and evaluates on the remaining 10%, and the latent consistency correlations (Figure 2) and the compositional geometry probe (Figure 3) use direct geometry or held-out POIs, so the probed quantities are not built into the training objective by construction. The navigation result (Table 4) holds out paths between Pheldout POIs while exposing those POIs through bridging trajectories, and exact-match accuracy is measured against Dijkstra ground truth rather than against training labels. There is no load-bearing self-citation: the reference list contains no prior work by the present authors, and no uniqueness theorem or ansatz is imported from the authors' own papers. The skeptic's concern about connectivity of the relational-pair graph is a legitimate precondition for the generalization claim, but it is a completeness condition rather than circularity: if the graph were disconnected, the reported test errors would be unexplained, not forced by construction. Likewise, sharing the synthetic coordinate frame between training and evaluation affects external validity but does not make the predictions equivalent to their inputs. No circular step is therefore identified.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or ontological entities; the simulated POIs, roads, and traversal weights are experimental affordances. The main ledger entries are the hand-chosen environment and training parameters plus the domain assumptions needed to extrapolate from a noise-free synthetic grid to human-like spatial cognition.

free parameters (5)
  • Traversal weight range = w ~ U[0.8,1.2] per road
    Defines road costs used to generate shortest-path labels; the paper says higher weights mean faster travel, which conflicts with Dijkstra cost semantics, leaving the optimized quantity ambiguous.
  • Environment scale = 100x100 grid, 1024 POIs, 200 roads
    Chosen by hand; scale affects how much relational integration is needed and how easy token embeddings are to probe.
  • Relational train/test split = 80:20 (also 60:40, 40:60)
    Directly controls generalization performance (distance MRPE ranges from 2.63% to 0.11% across splits); the connectivity of the resulting graph is not reported.
  • Probe MLP configuration = Two hidden layers, 128 and 64 units, Adam lr 1e-3, alpha 1e-4
    Probing results are probe-dependent; linear probes fail on composition and step-wise coordinates, so the representational claims are tied to this probe family.
  • Perturbation frequency threshold = tau from 1000 to 60000
    Used to select high-frequency turning points in the robustness analysis; the finding that recovery depends on training frequency is measured across this hand-chosen threshold grid.
assumptions (5)
  • standard math Dijkstra's algorithm returns the intended optimal paths for the generated trajectory labels
    Trajectory labels are produced by Dijkstra on a weighted grid graph; if edge weights are speeds rather than costs, the labels may not optimize travel time, so the navigation ground truth is convention-dependent.
  • domain assumption The synthetic grid and templated language are a valid proxy for human local spatial observations
    Distances and azimuths are exact, noise-free, and generated from one consistent coordinate frame; real human descriptions are approximate and referentially ambiguous, so external validity is assumed rather than demonstrated.
  • domain assumption Unique POI and road tokens carry no semantic prior
    The model treats arbitrary placeholders as anchors and learns coordinates from relational text; this prevents real-world geospatial leakage but limits relevance to natural place names.
  • domain assumption MLP probing is an appropriate instrument to reveal the geometry of hidden states
    The claim that hidden states encode absolute coordinates is established through probes; linear probes only partially succeed, so the representation may be nonlinear and probe-relative.
  • ad hoc to paper The internal absolute coordinate frame corresponds to the simulation's origin and axes
    Pairwise distances and azimuths determine layout only up to translation, yet the probe predicts absolute coordinates; the paper does not explain how the model anchors the origin, so 'absolute' alignment is with the simulation's arbitrary frame.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can LLMs Learn to Map the World from Local Descriptions?." pith.science (2026). https://pith.science/paper/QYNGLJ5M

@misc{pith2026250520874,
  author       = {Pith},
  title        = {Pith review of: Can LLMs Learn to Map the World from Local Descriptions?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QYNGLJ5M}},
  note         = {Machine review of arXiv:2505.20874}
}
read the original abstract

Recent advances in Large Language Models (LLMs) have demonstrated strong capabilities in tasks such as code and mathematics. However, their potential to internalize structured spatial knowledge remains underexplored. This study investigates whether LLMs, grounded in locally relative human observations, can construct coherent global spatial cognition by integrating fragmented relational descriptions. We focus on two core aspects of spatial cognition: spatial perception, where models infer consistent global layouts from local positional relationships, and spatial navigation, where models learn road connectivity from trajectory data and plan optimal paths between unconnected locations. Experiments conducted in a simulated urban environment demonstrate that LLMs not only generalize to unseen spatial relationships between points of interest (POIs) but also exhibit latent representations aligned with real-world spatial distributions. Furthermore, LLMs can learn road connectivity from trajectory descriptions, enabling accurate path planning and dynamic spatial awareness during navigation.

Figures

Figures reproduced from arXiv: 2505.20874 by the authors.

Figure 1
Figure 1. Summary of our research framework. Firsr, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Consistency between POI latent representa [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Latent spatial composition evaluation. An MLP predicts distance and azimuth between POI pairs using [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Heatmap of turning point frequencies. The [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: POIs Distribution. C.2 Data Format We provide examples of the data format used for training, as shown in [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Consistency of POI hidden state vectors with [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Consistency of POI point last hidden state [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Heatmap of turning point frequencies in the [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 11
Figure 11. Figure 11: Case study on the model’s behavior under interference during navigation at different statistical frequen [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 11 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  5. [5]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. arXiv preprint arXiv:2309.16609

    arXiv 2023

  6. [6]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374

    arXiv 2021

  7. [7]

    E. W. Dijkstra. 1959. A note on two problems in connexion with graphs. Numerische Mathematik, 1(1)

    work page 1959

  8. [8]

    Jie Feng, Yuwei Du, Tianhui Liu, Siqi Guo, Yuming Lin, and Yong Li. 2024 a . Citygpt: Empowering urban spatial cognition of large language models. arXiv preprint arXiv:2406.13948

    arXiv 2024

Show all 29 references
  1. [9]

    Jie Feng, Jun Zhang, Junbo Yan, Xin Zhang, Tianjian Ouyang, Tianhui Liu, Yuwei Du, Siqi Guo, and Yong Li. 2024 b . Citybench: Evaluating the capabilities of large language model as world model. CoRR

    2024

  2. [10]

    Language models represent space and time

    Wes Gurnee and Max Tegmark. Language models represent space and time. In The Twelfth International Conference on Learning Representations

  3. [11]

    Linear latent world models in simple transformers: A case study on othello-gpt

    Dean Hazineh, Zechen Zhang, and Jeffrey Chiu. Linear latent world models in simple transformers: A case study on othello-gpt. In Socially Responsible Language Modelling Research

  4. [12]

    Charles Jin and Martin Rinard. 2024. https://proceedings.mlr.press/v235/jin24e.html Emergent representations of program semantics in language models trained on programs . In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Mach...

    2024

  5. [13]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199--22213

    2022

  6. [14]

    Hopkins, David Bau, Fernanda B

    Kenneth Li, Aspen K. Hopkins, David Bau, Fernanda B. Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. https://openreview.net/forum?id=DeG07_TcZvT Emergent world representations: Exploring a sequence model trained on a synthetic task . In Proceedings of the 11th Internat...

    2023

  7. [15]

    Wenbin Li, Di Yao, Ruibo Zhao, Wenjie Chen, Zijie Xu, Chengxue Luo, Chang Gong, Quanliang Jing, Haining Tan, and Jingping Bi. 2024. Stbench: Assessing the ability of large language models in spatio-temporal analysis. arXiv preprint arXiv:2406.19065

    2024 arXiv

  8. [16]

    Bastien Li \'e tard, Mostafa Abdou, and Anders S gaard. 2021. Do language models know the way to rome? In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 510--517

    2021

  9. [17]

    Nicolas Martorell. 2025. http://arxiv.org/abs/2502.16690 From text to space: Mapping abstract spatial models in llms during a grid-world navigation task

    2025 arXiv

  10. [18]

    Ida Momennejad, Hosein Hasanbeig, Felipe Vieira Frujeri, Hiteshi Sharma, Nebojsa Jojic, Hamid Palangi, Robert Ness, and Jonathan Larson. 2023. Evaluating cognitive maps and planning in large language models with cogeval. Advances in Neural Information Processing Systems, 36:69...

    2023

  11. [19]

    Neel Nanda, Andrew Lee, and Martin Wattenberg. 2023. https://doi.org/10.18653/v1/2023.blackboxnlp-1.2 Emergent linear representations in world models of self-supervised sequence models . In Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks...

    2023 doi

  12. [20]

    Fabian Pedregosa, Ga \"e l Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. 2011. Scikit-learn: Machine learning in python. the Journal of machine Learning research, 12:2825--2830

    2011

  13. [21]

    Santhosh Kumar Ramakrishnan, Erik Wijmans, Philipp Kraehenbuehl, and Vladlen Koltun. 2024. Does spatial cognition emerge in frontier models? arXiv preprint arXiv:2410.06468

    2024 arXiv

  14. [22]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  15. [23]

    Johannes Treutlein, Dami Choi, Jan Betley, Samuel Marks, Cem Anil, Roger B Grosse, and Owain Evans. 2024. Connecting the dots: Llms can infer and verbalize latent structure from disparate training data. Advances in Neural Information Processing Systems, 37:140667--140730

    2024

  16. [24]

    Keyon Vafa, Justin Chen, Ashesh Rambachan, Jon Kleinberg, and Sendhil Mullainathan. 2024. Evaluating the world model implicit in a generative model. Advances in Neural Information Processing Systems, 37:26941--26975

    2024

  17. [25]

    Mind's eye of llms: Visualization-of-thought elicits spatial reasoning in large language models

    Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. Mind's eye of llms: Visualization-of-thought elicits spatial reasoning in large language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

  18. [26]

    Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. 2024 a . Thinking in space: How multimodal large language models see, remember, and recall spaces. arXiv preprint arXiv:2412.14171

    2024 arXiv

  19. [27]

    Qwen An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxin Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, ...

    2024 arXiv

  20. [28]

    Dazhou Yu, Riyang Bao, Gengchen Mai, and Liang Zhao. 2025. Spatial-rag: Spatial retrieval augmented generation for real-world spatial reasoning questions. arXiv preprint arXiv:2502.18470

    2025 arXiv

  21. [29]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. https://doi.org/10.18653/v1/2024.acl-demos.38 L lama F actory: Unified efficient fine-tuning of 100+ language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin...

    2024 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.