Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Can Large Language Models Integrate Spatial Data? Empirical Insights into Reasoning Strengths and Computational Weaknesses

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Pre-computed geometry lifts LLM spatial-data accuracy to 99.5 percent

desk verdict The core result is real and useful—LLMs jump from ~55% to ~95% accuracy on spatial join/union once given three precomputed features—but the headline 98-99% numbers come from a test set that excludes the hard ambiguous cases, so treat them as an upper bound on the easy subset. read the letter →

arxiv 2508.05009 v1 pith:CGW5K3LX submitted 2025-08-07 cs.AI cs.CL

Bin Han , Robert Wolfe , Anat Caspi , Bill Howe This is my paper
classification cs.AIcs.CL
keywords spatialdataintegrationlargelanguagemodelsreasoningjoinunionheuristicpromptingself-correctioncomputationalgeometry
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

This paper asks whether a large language model can perform spatial data integration—deciding whether a sidewalk segment runs alongside a road (spatial join) or whether two sidewalk records describe the same real-world path (spatial union)—when the task is specified in plain English instead of as a tuned algorithm. The authors find that LLMs alone are poor at this: with raw GeoJSON coordinates and natural-language instructions, five models average below 60% accuracy and often reason incoherently about geometry. When the prompt instead supplies three pre-computed geometric quantities—minimum angle, minimum distance, and buffered overlap area—the strongest models reach 98.4% join and 96.0% union accuracy, close to or above the best hand-tuned heuristic baselines. A two-step review-and-refine prompting scheme pushes the best results to 99.5% and 96.5%, and the paper argues that LLMs can pick appropriate thresholds per example, removing the need for a user to tune features and thresholds by hand. If this holds, spatial integration becomes accessible to domain experts without labeled training data or threshold engineering; the authors themselves caution that LLMs remain unreliable at the underlying computational geometry.

What carries the argument

The load-bearing device is heuristic prompting with pre-computed geometric features: for each candidate pair the paper computes three quantities—min angle (the smallest acute angle between sub-segments, an alignment measure), min distance (nearest-point separation), and max area (the maximum overlap percentage under a 10-meter buffer)—and hands these numbers to the LLM along with natural-language descriptions of what they mean. This converts a computational-geometry problem into a threshold-judgment problem the LLM can solve with world knowledge. The second mechanism is review-and-refine: an initial answer (from a heuristic, a random guess, or another method) is shown to the model, which fir

What would settle it

Take the same feature-prompting and review-and-refine protocols and run them (1) on the ambiguous Bellevue pairs the authors excluded from the labeled sets, (2) on street networks in a city with materially different road widths or sidewalk configurations, or (3) on join/union pairs synthesized to defeat the three features (e.g., parallel sidewalks separated from the road by a median). A drop toward the below-60% plain-prompting range, or a sharp per-city variance in chosen thresholds, would show the reported ceiling is an artifact of the feature set rather than a general LLM capability.

Watch

Extended reading notes

Core claim

The paper's central claim is a division of labor: LLMs have enough real-world understanding to interpret human-meaningful spatial relations (adjacency, parallelism, sameness), but they cannot reliably turn that understanding into correct computational geometry from coordinates alone. Evidence comes from chain-of-thought outputs, which mention proximity and alignment in nearly every sample yet commit errors the authors classify as wrong logic, unclear calculation, or no calculation—and even a reasoning-specialized model performs accurate calculations but draws logically flawed conclusions. When the geometry is pre-digested into three features (smallest angle between segments, minimum distance

Load-bearing premise

The load-bearing premise is that the three hand-picked geometric features—minimum angle, minimum distance, and 10-meter-buffer overlap—are sufficient to express what 'runs alongside' and 'represents the same sidewalk' mean, so the LLM only has to choose thresholds; if those features miss the geometry that matters for other cities, tasks, or data formats, the high accuracies will not transfer.

Editorial extensions

If this is right

  • Domain experts could integrate noisy urban GIS layers such as road and sidewalk networks by writing a sentence and running cheap pre-computed features, without collecting task-specific labeled training data.
  • Feature prompting removes threshold tuning: the best model's 98.4% join accuracy beats 93% of the 215 heuristic feature/threshold combinations the authors enumerated, and the model sets thresholds case-by-case rather than globally.
  • Review-and-refine can be bolted onto any initial predictor—random, heuristic, or LLM—and consistently raises low accuracy while retaining high accuracy, reaching 99.5% (join) and 96.5% (union).
  • Plain-language LLM integration without features stays below 60% accuracy on these tasks, so LLMs are not yet a drop-in replacement for computational geometry.
  • The same feature-prompting pipeline is likely to extend to other spatial predicates beyond join and union, since the models generalize the threshold-selection behavior across feature combinations.

Reading between the lines

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

  • Editorial inference: the reported 98–99% figures would probably shrink on the ambiguous pairs the paper explicitly excluded as requiring case-by-case discussion; a stronger test is to run the same feature prompts on exactly those pairs.
  • Editorial inference: because the features and the 10-meter buffer were chosen for Bellevue's road types and widths, transfer to cities with wider roads, medians, or non-linestring geometries (polygons, points) is untested; the weakest link is the sufficiency of the feature set, not the LLM.
  • Editorial inference: the per-example threshold variation the paper observes suggests the LLM is blending data evidence with prior world knowledge—useful for robustness, but a source of unpredictable decisions on out-of-distribution geometry; one could expose the chosen thresholds and compare them with per-example oracle thresholds.
  • Editorial inference: the 'pre-compute geometry, let the LLM judge thresholds' pattern is testable on other natural-language spatial predicates such as 'crosses,' 'is adjacent to,' and 'partially overlaps,' and on other data encodings, which would reveal whether the bottleneck is spatial semantics or computation.
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

4 major / 4 minor

Summary. The paper studies whether LLMs can perform two spatial data integration tasks—spatial join (does a sidewalk run alongside a road?) and spatial union (do two sidewalk annotations represent the same entity?)—using GeoJSON LineStrings. It compares plain zero-/few-shot prompting, prompting with natural-language heuristic hints, and prompting with precomputed geometric features (min angle, min distance, max area under a 10-m buffer), across five LLMs. It also proposes a review-and-refine method in which the LLM critiques and corrects an initial heuristic or random answer. The central empirical result is that plain prompting is poor (~54–60% accuracy on join) while feature-based prompting reaches 94–98%, and review-and-refine can reach 99.5% (join) and 96.5% (union), sometimes exceeding hand-tuned heuristic baselines. Qualitative chain-of-thought analysis shows models can name relevant spatial concepts but often fail at computational geometry. The paper concludes that LLMs are promising for spatial integration when given precomputed features, but are unreliable when left to derive the geometry themselves.

Significance. If the headline results were representative of real integration workloads, the paper would be a useful empirical contribution: it demonstrates a consistent and large plain-vs-features accuracy gap across five models, and it documents concrete failure modes of LLM spatial reasoning on real geometries. The review-and-refine result is interesting and actionable. The paper also benefits from using real urban data (Bellevue sidewalks and roads) rather than synthetic textbook geometry. However, the headline accuracy claims rest on a test set built by excluding thousands of candidate pairs that the authors judged 'ambiguous'. Because those excluded pairs are precisely the edge cases that the paper's motivation says LLMs should help address, the reported 98–99% accuracies may substantially overstate performance on the full candidate population. The paper also reports no confidence intervals or repeated runs, which matters for the narrow margin by which some LLM configurations are said to surpass the best heuristics.

major comments (4)
  1. [Appendix A; §§3, 5.2] The evaluation set excludes the ambiguous cases most relevant to the paper's motivation. For the join task, 14,556 candidate pairs were generated but only 10,240 labeled; the remaining ~4,316 were 'deemed ambiguous and required case-by-case discussion; they were thus excluded from this study.' For the union task, 5,496 candidates were generated but only 1,329 labeled, excluding ~4,167 pairs. Section 3 motivates the study by saying that 'in cases where heuristics are insufficient and do not generalize well, we can assess LLMs ability on those specific cases,' but the test sets omit exactly such cases. The headline accuracies (98.4%, 99.5%, 96.0%, 96.5%) therefore characterize an 'easy' subset of the candidate pool, not the full pool a deployed system would encounter. This is load-bearing for the claim that LLMs 'eliminate the need for feature selection and threshold tuning' (T2 finding 3)
  2. [§5.2, Tables 9–10; §5.4] All accuracies are point estimates without confidence intervals, repeated runs, or significance tests. The test sets are small (1,000 join pairs; 399 union pairs), so a difference of 0.4–0.5 percentage points (e.g., 4o achieving 99.4% vs. the best trio heuristic at 99.0%; qwen-plus 96.0% vs. the best duo heuristic 96.2%) is within a handful of label flips. The claim that LLMs 'surpass the most effective heuristic baselines' in §5.4 is therefore not well supported at the reported granularity. Please provide bootstrap CIs or standard errors over test samples, and where possible run each configuration multiple times (even with temperature 0, API decoding can be non-deterministic) and report mean and variance.
  3. [§5.2, Finding 3; §3; Appendix A] The claim that LLMs 'eliminate the need for feature selection and threshold tuning' overstates what the experiments show. The prompt designer selects the specific features (min angle, min distance, max area) and the 10-meter buffer radius that defines max area; the LLM only chooses threshold values for that fixed feature set. The 10-meter buffer is justified by 'typical road widths' in Seattle/Bellevue, which is itself a task-specific design choice. Thus the method reduces threshold tuning but does not eliminate feature selection—it moves it into prompt construction. The paper should be more precise: the empirical contribution is that LLMs can select reasonable per-instance thresholds once a small, task-relevant feature set is supplied. This is still valuable, but the current phrasing invites a broader interpretation that the experiments do not support.
  4. [Appendix F, Tables 20 and 23] The union-task feature prompts are internally inconsistent with the rest of the paper. They refer to 'sidewalk, road' instead of 'sidewalk 1, sidewalk 2,' and Table 20 includes 'min distance' and a three-key-statistics block even though the union task uses only min angle and max area. These inconsistencies are minor on their own, but they matter because the paper's central claim is that the supplied features are exactly what the LLM needs; if the actual prompt text includes extra or mislabeled features, the reader cannot fully reconstruct the experiments from the appendix.
minor comments (4)
  1. [Table 7] Caption says 'The best single heuristic configuration is parallel (max area=0.8),' but the row shown is the overlap heuristic. Should read 'overlap (max area=0.8).'
  2. [Table 10] The zero-shot hints row for 'parallel (p)' is incomplete: it shows entries '0.544, 0.43' and then stops. Full values for all models should be reported.
  3. [Abstract and §5.2] The abstract says LLMs 'surpassing the most effective heuristic baselines,' but Table 9 shows the best feature-prompt result (qwen-plus 98.4%) is below the best trio heuristic (99.0%); only the review-and-refine result (99.5%) surpasses it. Please qualify the claim or point explicitly to the review-and-refine configuration.
  4. [General] No code, data, or prompt templates are provided in a public repository. Given the sensitivity of the results to the exact prompt wording and to the ambiguous-pair exclusion, releasing the preprocessing pipeline, the full candidate pool, and the prompts would substantially strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the claims are empirical accuracy measurements on human-labeled data, not derivations reducible to their inputs.

full rationale

This paper is an empirical benchmark rather than a derivation chain. The central claims—that LLMs with precomputed geometric features reach certain accuracies and that review-and-refine improves them—are evaluated against human visual labels, not against the features themselves. The LLMs are zero-shot or few-shot prompted and are not trained or fitted on the test labels. The geometric features (min angle, min distance, max area) are inputs to the prompt, but the labels come from manual visual inspection, so there is no equation-level equivalence between input and output. The paper's self-citations (e.g., Han et al., 2024; Wolfe & Mitra, 2024) appear in related work and are not load-bearing for the empirical results. The 10-meter buffer and the three chosen features are feature-engineering choices, and the exclusion of ambiguous candidate pairs in Appendix A is a dataset-construction limitation; both affect generalization and may bias the reported accuracy, but they do not make the results circular. No fitted parameter is renamed as a prediction, and no derived quantity reduces by construction to an input. Thus the appropriate circularity score is 0.

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

The central claim rests on a hand-chosen feature set, a hand-chosen buffer, private manual labels, and deliberate exclusion of ambiguous pairs. No new entities are introduced.

free parameters (3)
  • 10-meter buffer radius = 10 m
    Used to generate candidate pairs for the join task and to compute the max area overlap feature. Chosen from typical road widths, not fitted, but it determines which pairs are labeled and which features the LLM sees.
  • Heuristic threshold grids = angles {1,2,5,10,20} join, {1,2,3,4,5} union; distances {1,2,3,4,5}; overlap {0.1,...,0.5} join, {0.5,...,0.9} union
    Define the baseline heuristic accuracies and are also the quantities the LLM is asked to infer. The choice of grid affects the best-heuristic comparison.
  • Exclusion of ambiguous pairs = approx. 4,316 join / 4,167 union pairs removed
    The authors manually exclude pairs that require case-by-case discussion, which changes test difficulty and positive/negative balance.
assumptions (4)
  • domain assumption The semantic relationship 'runs alongside' (join) and 'represents the same sidewalk' (union) is adequately captured by min angle, min distance, and max area computed with a 10-meter buffer.
    All high-performing LLM prompts feed these exact features; if the feature set is incomplete, results do not generalize. Stated in Section 3 and Appendix A.
  • domain assumption Candidate pairs generated by 10-meter buffer (join) or intersection (union) contain all relevant matches.
    Appendix A: under the assumption that a sidewalk running alongside a road should fall within the buffer zone.
  • domain assumption Manual visual inspection labels are correct ground truth.
    Appendix A: manually labeled through visual inspection, with no inter-annotator agreement or adjudication reported.
  • ad hoc to paper Accuracy on this curated test set transfers to real integration workloads.
    The conclusion positions LLMs as a flexible alternative based on one city's data and two tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Large Language Models Integrate Spatial Data? Empirical Insights into Reasoning Strengths and Computational Weaknesses." pith.science (2026). https://pith.science/paper/CGW5K3LX

@misc{pith2026250805009,
  author       = {Pith},
  title        = {Pith review of: Can Large Language Models Integrate Spatial Data? Empirical Insights into Reasoning Strengths and Computational Weaknesses},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CGW5K3LX}},
  note         = {Machine review of arXiv:2508.05009}
}
read the original abstract

We explore the application of large language models (LLMs) to empower domain experts in integrating large, heterogeneous, and noisy urban spatial datasets. Traditional rule-based integration methods are unable to cover all edge cases, requiring manual verification and repair. Machine learning approaches require collecting and labeling of large numbers of task-specific samples. In this study, we investigate the potential of LLMs for spatial data integration. Our analysis first considers how LLMs reason about environmental spatial relationships mediated by human experience, such as between roads and sidewalks. We show that while LLMs exhibit spatial reasoning capabilities, they struggle to connect the macro-scale environment with the relevant computational geometry tasks, often producing logically incoherent responses. But when provided relevant features, thereby reducing dependence on spatial reasoning, LLMs are able to generate high-performing results. We then adapt a review-and-refine method, which proves remarkably effective in correcting erroneous initial responses while preserving accurate responses. We discuss practical implications of employing LLMs for spatial data integration in real-world contexts and outline future research directions, including post-training, multi-modal integration methods, and support for diverse data formats. Our findings position LLMs as a promising and flexible alternative to traditional rule-based heuristics, advancing the capabilities of adaptive spatial data integration.

Figures

Figures reproduced from arXiv: 2508.05009 by the authors.

Figure 1
Figure 1. Positive and negative examples for the spatial join and union task. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Plain prompting (left) and our heuristic prompting (right) for the spatial integration [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Average results of heuristic prompting method for the spatial join and union task. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Accuracies of review-and-refine method on both spatial join and union task. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Left: Distribution of min angle in positive and negative spatial join data. Right: train accuracy using different feature thresholds for min angle, for spatial join task. Clearance Heuristic: (min distance) [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Left: Distribution of min distance in positive and negative spatial join data. Right: train accuracy using different feature thresholds for min distance, for spatial join task. Overlap Heuristic (max area) [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Left: Distribution of max area in positive and negative spatial join data. Right: train accuracy using different feature thresholds for max area, for spatial join task. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Left: Distribution of min angle in positive and negative spatial union data. Right: train accuracy using different feature thresholds for min angle, for spatial union task. Overlap Heuristic (max area) [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Left: Distribution of max area in positive and negative spatial union data. Right: train accuracy using different feature thresholds for max area, for spatial union task. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Cheap Expertise: Mapping and Challenging Industry Perspectives in the Expert Data Gig Economy

    cs.CY 2026-05 unverdicted novelty 5.0 of 10

    AI data firms view human expertise as an extractable, low-cost resource to feed AI systems while treating institutional expertise as something needing liberation or reform to fit this model.

Reference graph

Works this paper leans on

30 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [5]

    Mediating culture: Cultivating socio-cultural understanding of ai in children through participatory design

    Aayushi Dangol, Michele Newman, Robert Wolfe, Jin Ha Lee, Julie A Kientz, Jason Yip, and Caroline Pitt. Mediating culture: Cultivating socio-cultural understanding of ai in children through participatory design. In Proceedings of the 2024 ACM Designing Interactive Systems Conference, pp. 1805–1822,

    work page 2024

  2. [6]

    Documenting large webtext corpora: A case study on the colossal clean crawled corpus

    Jesse Dodge, Maarten Sap, Ana Marasovi´c, William Agnew, Gabriel Ilharco, Dirk Groen- eveld, Margaret Mitchell, and Matt Gardner. Documenting large webtext corpora: A case study on the colossal clean crawled corpus. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 1286–1305,

    work page 2021

  3. [10]

    Cross-lingual text classification with large language models

    Bin Han, Sean T Yang, and Christopher LuVogt. Cross-lingual text classification with large language models. In Companion Proceedings of the ACM on Web Conference 2025, pp. 1005–1008,

    work page 2025

  4. [11]

    Edwin H Jacox and Hanan Samet

    Accessed: 2025-03-27. Edwin H Jacox and Hanan Samet. Spatial join techniques. ACM Transactions on Database Systems (TODS), 32(1):7–es,

    work page 2025

  5. [13]

    Transllama: Llm-based simulta- neous translation system

    Roman Koshkin, Katsuhito Sudoh, and Satoshi Nakamura. Transllama: Llm-based simulta- neous translation system. arXiv preprint arXiv:2402.04636,

    arXiv

  6. [14]

    Rlaif vs

    Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267,

    arXiv

  7. [15]

    From generation to judgment: Opportunities and challenges of llm-as-a-judge

    Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, et al. From generation to judgment: Opportunities and challenges of llm-as-a-judge. arXiv preprint arXiv:2411.16594, 2024a. Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun...

    arXiv

  8. [16]

    MoT: Memory-of-thought enables ChatGPT to self-improve

    Xiaonan Li and Xipeng Qiu. MoT: Memory-of-thought enables ChatGPT to self-improve. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 6354–6374, Singapore, December

    work page 2023

Show all 30 references
  1. [17]

    doi: 10.18653/v1/2023.emnlp-main.392

    Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.392. URL https://aclanthology.org/2023.emnlp-main.392/. Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in ...

    2023 arXiv

  2. [18]

    When llm-based code generation meets the software develop- ment process

    Feng Lin, Dong Jae Kim, et al. When llm-based code generation meets the software develop- ment process. arXiv preprint arXiv:2403.15852,

    arXiv

  3. [19]

    URL https://www

    doi: 10.1016/j.compenvurbsys.2021.101606. URL https://www. sciencedirect.com/science/article/abs/pii/S0198971521000259. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Trainin...

    2021

  4. [21]

    Map stitcher: Graph sampling-based map conflation

    Erfan Hosseini Sereshgi and Carola Wenk. Map stitcher: Graph sampling-based map conflation. In Proceedings of the 3rd ACM SIGSP ATIAL International Workshop on Spatial Big Data and AI for Industrial Applications (GeoIndustry 2024), pp. 5–15. ACM,

    2024

  5. [22]

    URL https://dl.acm.org/doi/10.1145/3681766.3699604

    doi: 10.1145/3681766.3699604. URL https://dl.acm.org/doi/10.1145/3681766.3699604. Abraham C´ardenas Trist´an, Eduardo Javier Trevi˜no Garza, Oscar Alberto Aguirre Calder´on, Javier Jim´enez P´erez, Marco Aurelio Gonz´alez Tagle, and Xanat Antonio N´emiga. Spatial technologies ...

    2013

  6. [23]

    Large language models as urban residents: An llm agent framework for personal mobility generation

    Jiawei Wang, Renhe Jiang, Chuang Yang, Zengqing Wu, Makoto Onizuka, Ryosuke Shibasaki, Noboru Koshizuka, and Chuan Xiao. Large language models as urban residents: An llm agent framework for personal mobility generation. arXiv preprint arXiv:2402.14744,

    arXiv

  7. [24]

    Under review

    13 Preprint. Under review. Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V . Le. Finetuned language models are zero-shot learners.ArXiv, abs/2109.01652,

    arXiv

  8. [25]

    Robert Wolfe and Tanushree Mitra

    URL https://api.semanticscholar.org/CorpusID:237416585. Robert Wolfe and Tanushree Mitra. The impact and opportunities of generative ai in fact- checking. In The 2024 ACM Conference on Fairness, Accountability, and Transparency, pp. 1531–1543,

    2024

  9. [26]

    Bloomberggpt: A large language model for finance

    Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. Bloomberggpt: A large language model for finance. arXiv preprint arXiv:2303.17564,

    arXiv

  10. [27]

    Do language models mirror human confidence? exploring psychological insights to address overconfidence in llms

    Chenjun Xu, Bingbing Wen, Bin Han, Robert Wolfe, Lucy Lu Wang, and Bill Howe. Do language models mirror human confidence? exploring psychological insights to address overconfidence in llms. arXiv preprint arXiv:2506.00582,

    arXiv

  11. [28]

    Simulating classroom education with llm-empowered agents

    Zheyuan Zhang, Daniel Zhang-Li, Jifan Yu, Linlu Gong, Jinchang Zhou, Zhiyuan Liu, Lei Hou, and Juanzi Li. Simulating classroom education with llm-empowered agents. arXiv preprint arXiv:2406.19226,

    arXiv

  12. [29]

    How well do llms generate code for different application domains? benchmark and evaluation

    Dewu Zheng, Yanlin Wang, Ensheng Shi, Hongyu Zhang, and Zibin Zheng. How well do llms generate code for different application domains? benchmark and evaluation. arXiv preprint arXiv:2412.18573,

    arXiv

  13. [2004]

    Leveraging large language models for learning complex legal concepts through storytelling

    Hang Jiang, Xiajie Zhang, Robert Mahari, Daniel Kessler, Eric Ma, Tal August, Irene Li, Alex’Sandy’ Pentland, Yoon Kim, Deb Roy, et al. Leveraging large language models for learning complex legal concepts through storytelling. arXiv preprint arXiv:2402.17019,

    arXiv

  14. [2008]

    Under review

    14 Preprint. Under review. A Data Processing Spatial Join Dataset We first filter roads from Bellevue-City, selecting only those with highway types in {secondary, residential, tertiary, primary, livingstreet} (types likely to have paired sidewalks), resulting in 17,800 road an...

    2017

  15. [2010]

    Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William H

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Pond ´e, Jared Kaplan, Harrison Edwards, Yura Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryd...

    arXiv

  16. [2017]

    How contextual are contextualized word representations? comparing the geometry of bert, elmo, and gpt-2 embeddings

    Kawin Ethayarajh. How contextual are contextualized word representations? comparing the geometry of bert, elmo, and gpt-2 embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural...

    2019

  17. [2019]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948,

    arXiv

  18. [2021]

    Yu Ying Chiu, Ashish Sharma, Inna Wanyin Lin, and Tim Althoff

    URL https://api.semanticscholar.org/ CorpusID:235755472. Yu Ying Chiu, Ashish Sharma, Inna Wanyin Lin, and Tim Althoff. A computational frame- work for behavioral assessment of llm therapists. arXiv preprint arXiv:2401.00820,

    arXiv

  19. [2022]

    Mrg-ser: Self-supervised spatial entity resolution based on multi-relational graph

    Hanchen Qiu, Haojia Zhu, Zhicheng Li, and Jiahui Jin. Mrg-ser: Self-supervised spatial entity resolution based on multi-relational graph. Proceedings of the VLDB Endowment. ISSN, 2150:8097, . Giridhar Kaushik Ramachandran, Yujuan Fu, Bin Han, Kevin Lybarger, Nicholas J Dobbins...

    arXiv

  20. [2023]

    org/wp-content/uploads/2023/12/MP-IST-SET-126-07.pdf

    URL https://innovationhub-act. org/wp-content/uploads/2023/12/MP-IST-SET-126-07.pdf . Hainan Chen and Volker Walter. Hierarchical quality inspection of spatial data by data integration. In Proceedings of the ASPRS 2010 Annual Conference, San Diego, CA, USA, pp. 26–30. Citeseer,

    2023

  21. [2024]

    Lm vs lm: Detecting factual errors via cross examination

    Roi Cohen, May Hamri, Mor Geva, and Amir Globerson. Lm vs lm: Detecting factual errors via cross examination. arXiv preprint arXiv:2305.13281,

    arXiv

  22. [2025]

    Huskyscribe at mediqa-sum 2023: Summarizing clinical dialogues with transformers

    Bin Han, Haotian Zhu, Sitong Zhou, Sofia Ahmed, Md Mushfiqur Rahman, Fei Xia, and Kevin Lybarger. Huskyscribe at mediqa-sum 2023: Summarizing clinical dialogues with transformers. In CLEF (Working Notes), pp. 1488–1509,

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.