Pith. sign in

REVIEW 4 major objections 4 minor 25 references

On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures

T0 review · 4 major / 4 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 This paper shows that LLMs are measurably less accurate when asked for facts in measurement systems other than their default, and that the reasoning fix for large models raises inference costs by 180% to 300%.

desk verdict Solid, useful study of a real bias — the currency ground-truth selection needs transparency before the exact magnitudes are quoted. read the letter →

arxiv 2506.02591 v1 pith:ZH4ITJGB submitted 2025-06-03 cs.CL

Minh Duc Bui , Kyung Eun Park , Goran Glavaš , Fabian David Schmidt , Katharina von der Wense This is my paper
classification cs.CL
keywords LLMfactualretrievalmeasurementsystemsculturalbiasunitconversionmulti-hopreasoningchain-of-thoughtinferencecostgeographic
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

LLMs are measurably less accurate when asked to state a fact in a measurement system other than the one they default to, even when the underlying fact is the same. The paper builds new datasets of GDP, food prices, and city distances and shows that models perform best with their default system and drop significantly otherwise. Large models can recover the gap with chain-of-thought reasoning, but that reasoning increases test-time compute and inference cost by 180% to 300%. The authors argue this is a cultural-equity problem: users from backgrounds whose units are underrepresented pay more and get less reliable answers.

What carries the argument

The argument is carried by three purpose-built datasets: fiscal GDP-per-capita figures for 148 countries in 2021, converted across 112 currencies; food prices per weight from 76 metric-system countries plus US pound-based prices; and city distances computed for metric countries with straight-line Haversine distances and for the US with shortest road distances. Performance is measured by inverse mean absolute percentage deviation (MAPD), with deviations capped at 100%. The paper frames a non-default measurement query as a multi-hop question—retrieve the fact in the default system, retrieve the conversion rate, then combine them with arithmetic—and tests two reasoning strategies: explicit sequential single hops and chain-of-thought prompting. The cost comparison uses API pricing for Llama 3.3 70B to translate longer responses into percentage cost increases.

What would settle it

Re-run the currency evaluation using the average 2021 exchange rate for each currency instead of the most favorable rate; if the high-income versus low-income accuracy gap and the default-USD advantage largely persist, the finding is robust, and if they shrink or reverse, the ground-truth choice carried the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that current LLMs are not stable across measurement systems: they yield the best performance when used with their default system and generally suffer significant, pronounced drops for alternative systems. Across 35 model–dataset combinations, the default system outperforms all others in 28 cases, with statistically significant gains over the second-best system in 22 of those. Reasoning strategies such as chain-of-thought stabilize the large models to near-default accuracy, but the models start from the default system (USD in 91% of currency reasoning traces) and the extra tokens raise test-time compute and expense by 180% to 300%. The paper further shows that currency accuracy tracks country income, with large models scoring 43% on high-income countries' currencies versus 9% on low-income countries' currencies, and that aligning prompt language with the target system does not remove the drop in nearly all cases.

Load-bearing premise

The load-bearing premise is that the ground-truth values are correct and unbiased: the 'most favorable' exchange rates used to define currency values must not systematically favor some countries, and the US road distances must be comparable to the straight-line distances used elsewhere.

Editorial extensions

If this is right

  • Users who request facts in a non-default measurement system get measurably less accurate answers on average across the seven tested open-source models.
  • The available fix, reasoning prompts, is not free: inference expenses rise by roughly 180% to 300%, so the cost burden falls on those who need conversions.
  • Even matching the prompt language to the target system (e.g., Korean prompts with KRW) does not eliminate the performance drop in almost all tested cases.
  • Currency accuracy correlates with country income, so the bias has a socioeconomic dimension: low-income countries' currencies are answered far less accurately.
  • Reasoning rescues large models (70B and above) but does not stabilize small models, suggesting the gap depends on scale.

Reading between the lines

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

  • The cost asymmetry identified for measurement systems likely generalizes to other culturally bound formats—dates, address formats, number notation, temperature—where the model's canonical rendering also differs from the user's.
  • A testable extension would compare accuracy and cost when the model is first instructed to output in its default system and then convert explicitly, versus when the user requests the target system directly; the paper's sequential-hop results suggest the former may be both cheaper and more accurate.
  • The ground-truth construction is the main vulnerability: if the choice of the 'most favorable' 2021 exchange rate was partly tuned toward model outputs, the headline currency gaps could be inflated; re-running with average rates would confirm the claim.
  • The length task mixes road distances for the US with straight-line distances elsewhere, so asking for the same geometry in both systems would isolate the measurement effect from the distance-type confound.
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 investigates whether open-source LLMs can answer factual questions in measurement systems other than their default. The authors compile three new datasets: GDP per capita for 148 countries in 112 currencies; food prices per kilogram and per pound (plus traditional weight units); and city distances in kilometers and miles (plus historical length units). Across seven open-source LLMs, they find that models perform best in their default measurement system, that accuracy drops for alternative systems, and that explicit reasoning (sequential hops or chain-of-thought) largely stabilizes large models but increases test-time compute and inference cost by roughly 180% to 300% for alternative systems.

Significance. If the results hold, this is a timely contribution to the study of cultural biases in LLMs. The paper introduces three new datasets, evaluates seven models, and connects measurement-system bias to test-time cost, a practical concern for users from underrepresented cultural backgrounds. The multi-hop framing of unit conversion is appropriate, and the release of code and datasets is a strength. The main direction is consistent across datasets and model families, but several ground-truth construction choices need clarification and robustness checks before the reported magnitudes can be taken at face value.

major comments (4)
  1. [§2.2 (Fiscal Data)] The currency ground truth is built by converting World Bank USD figures with 'the most favorable rate within that year, ensuring optimal performance' (Section 2.2). The selection rule is not defined: it is unclear from whose perspective the rate is favorable, and the phrase 'ensuring optimal performance' suggests the rate may have been chosen to fit model outputs. Because the USD labels are the original World Bank values while all alternative-currency labels are derived through this selected rate, the default-vs-alternative comparison is asymmetric. The exact magnitudes in Section 3.3 and the income-group gap in Figure 3 (34 points) are not supported until the authors specify the selection rule and show robustness to using, for example, average, median, or end-of-year rates.
  2. [§2.2 (City Distances)] For the kilometer dataset, distances are Haversine straight-line distances, while for the US mile dataset the text states that 'shortest road distances' are used. This confounds the length-unit comparison with the distance definition: US ground truth is road distance, while German, Russian, Chinese, and Japanese ground truth is great-circle distance. Consequently, the default-vs-alternative length results in Figure 2 and Table 6 may reflect the model's ability to estimate road versus straight-line distances rather than its ability to convert units. The authors should compute the same distance type for all countries or otherwise control for this factor in the analysis.
  3. [§2.2 (Food Prices)] The food-price dataset is described as spanning 2010–2021, but the paper does not state the reference year (or years) used for evaluation. The prompt template in Appendix B.1 includes 'in year <YEAR>', yet the dataset description and the examples in Table 1 for the weight task omit the year. If the ground truth is a single historical price while the prompt does not fix the year, the measured error conflates unit-conversion failures with temporal mismatch. Please state the reference year for each sample and, if multiple years are pooled, explain how the query specifies the year.
  4. [§4 (Cost Analysis)] The headline cost increase of 180–300% is computed from Groq API pricing for Llama 3.3 70B, but the paper reports neither the underlying token counts nor the cost formula, and it is unclear whether input tokens and multiple API calls in the sequential strategy are included. Since the title and abstract frame the result as an increase in test-time compute, the authors should report per-condition output (and input) token statistics and a reproducible cost calculation, ideally accompanied by latency measurements.
minor comments (4)
  1. [§3.3 (Evaluation Metric)] The term 'inverse MAPD' is a misnomer; the metric is 100 × (1 − mean capped APD), which is a monotone transformation rather than an inverse. Consider renaming it for clarity.
  2. [§3.3 (Discussion)] The statement that 'LLMs yield the best performance when used with their default system' is too strong, since the preceding paragraph reports that this holds in 28 of 35 model–dataset combinations and Table 6 contains exceptions (e.g., Llama 3.3 70B in the Default=Pound weight setting, where Kilogram gives a higher MAPD). Please qualify the claim.
  3. [Appendix B.2] The appendix states that outputs are generated with a limit of 40 new tokens per request and that the limit is increased for CoT, but the increased limit is not reported. Providing the exact token limits would improve reproducibility.
  4. [Figure 2] The legend distinguishes filled and half-filled points for statistical significance, but the distinction is difficult to see in grayscale; please use distinct colors or shapes in addition to the fill.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical evaluation with externally sourced ground truth, and its central claims do not reduce to their inputs by construction.

full rationale

The paper's central claims are empirical measurements of LLM behavior. The ground-truth data come from external sources (World Bank, IMF, World Food Program, U.S. Bureau of Labor Statistics, SimpleMaps), and the measurement conversions in Table 7 are standard physical or currency definitions, not quantities fitted to the models. RQ1's 'default system' is determined by observed free-form model output frequencies before any accuracy comparison; RQ2 then compares accuracy against externally fixed ground truths, so the statement that default systems achieve the highest MAPD is a contingent empirical finding, not a consequence of how 'default' was defined. RQ3's cost estimates are computed from token counts and API pricing and are not derived from the accuracy results. The self-citations in the paper, such as Bui et al. 2024 in Related Work, are contextual and do not carry any load-bearing premise; no uniqueness theorem or ansatz is imported from prior author work. The only passage that warrants scrutiny is Section 2.2's statement that the most favorable 2021 IMF exchange rate was selected 'ensuring optimal performance.' This selection is underspecified and could affect the magnitude of currency-specific gaps and the income-group analysis, but the paper does not disclose that the rate was chosen using model outputs, and no equation or construction exhibits a fitted parameter renamed as a prediction. Even under an unfavorable reading, such a selection would be an evaluation-construction bias rather than a circular derivation, and it would tend to make alternative-currency results look better, not force the paper's default-vs-alternative finding. Therefore the derivation chain is self-contained and no circular step is identifiable.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The paper contributes benchmarks and measurements rather than a derivation, so the ledger is mostly domain assumptions about data sources and evaluation. The two hand-chosen evaluation constants (deviation cap, significance threshold) are listed as free parameters because they shape the reported scores. No new entities are introduced.

free parameters (2)
  • APD cap (maximum per-sample deviation) = 1.0 (100%)
    Chosen by hand to bound extreme predictions; all MAPD numbers depend on this cap, and no sensitivity analysis is reported.
  • Wilcoxon significance threshold = p = 0.001
    Chosen by hand; multiple comparisons across many currencies and systems are not corrected, so some significant differences may be false positives.
assumptions (6)
  • domain assumption GDP per capita (current USD) from the World Bank for 2021 accurately represents the fiscal facts being queried.
    Used as ground truth for the currency task (Section 2.2).
  • domain assumption IMF 2021 exchange rates provide valid conversion factors, and choosing the most favorable monthly rate is a valid way to define the target value.
    Ground-truth construction for fiscal data (Section 2.2); the 'most favorable' choice is a red flag because it may bias evaluation.
  • domain assumption The same 'distance' quantity is measured across countries.
    Metric countries use straight-line Haversine distances while the US is described as using shortest road distances (Section 2.2); inconsistency could confound length-system comparisons.
  • domain assumption LLM outputs under a 40-new-token limit can be parsed reliably into the intended numeric answers.
    Evaluation pipeline parses generated numbers; the paper does not report parsing failures or truncation rates (Appendix B.2).
  • domain assumption Groq API pricing for Llama 3.3 70B is representative of LLM inference cost.
    Cost-increase percentages are derived from this pricing page (footnote 1, Section 4), not from measured energy or hardware costs.
  • domain assumption The classification of units as High, Medium, or Low usage reflects real-world cultural prevalence.
    Appendix Table 5 labels usage levels; these labels feed the 'underrepresented systems' interpretation but are not sourced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures." pith.science (2026). https://pith.science/paper/ZH4ITJGB

@misc{pith2026250602591,
  author       = {Pith},
  title        = {Pith review of: On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZH4ITJGB}},
  note         = {Machine review of arXiv:2506.02591}
}
read the original abstract

Measurement systems (e.g., currencies) differ across cultures, but the conversions between them are well defined so that humans can state facts using any measurement system of their choice. Being available to users from diverse cultural backgrounds, large language models (LLMs) should also be able to provide accurate information irrespective of the measurement system at hand. Using newly compiled datasets we test if this is the case for seven open-source LLMs, addressing three key research questions: (RQ1) What is the default system used by LLMs for each type of measurement? (RQ2) Do LLMs' answers and their accuracy vary across different measurement systems? (RQ3) Can LLMs mitigate potential challenges w.r.t. underrepresented systems via reasoning? Our findings show that LLMs default to the measurement system predominantly used in the data. Additionally, we observe considerable instability and variance in performance across different measurement systems. While this instability can in part be mitigated by employing reasoning methods such as chain-of-thought (CoT), this implies longer responses and thereby significantly increases test-time compute (and inference costs), marginalizing users from cultural backgrounds that use underrepresented measurement systems.

Figures

Figures reproduced from arXiv: 2506.02591 by the authors.

Figure 1
Figure 1. We test whether LLMs can translate facts [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Disparity and Performance of LLMs Across Different Measurement Systems: We report the performance in terms of MAPD (cf. §3.3). The best-performing system (higher is better) for each model is highlighted with bold points. Fully filled points indicate statistically significant improvement over the second-best system, whereas half-filled points do not. We report the raw numbers and their significance in Appendix [PITH… view at source ↗
Figure 3
Figure 3. Currency Performance Mapped and Categorized by Income Groups: We visualize the performance for all currencies—averaged across large models (over 70B parameters)—on a world map, with countries categorized by income levels. tems; LLMs yield the best performance when used with their default system and generally suffer significant and pronounced drops in per￾formance for alternative systems. Such issues have severe impl… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of Multi-Hop Reasoning: An￾swering w.r.t. an alternative measurement system re￾quires multi-hop reasoning, consisting of the following steps: (i) retrieving the answer in the default system, (ii) retrieving the conversion rate between the default and alte…
Figure 5
Figure 5. Figure 5: Frequency of Measurement Systems as Starting Points in CoT Reasoning. We count how often each measurement system is chosen as the starting point for the model’s CoT reasoning and report the proportion. We average this proportion across all models. 5 Related Work Biases…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 20 canonical work pages

  1. [2]

    InProceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing, pages 14113–14130, Miami, Florida, USA

    Hopping too late: Explor- ing the limitations of large language models on multi- hop queries. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing, pages 14113–14130, Miami, Florida, USA. As- sociation for Computational Linguistics. Minh Duc Bui, Katharina von der Wense, and Anne Lauscher

    work page 2024

  2. [3]

    Multi3Hate: Multimodal, Multilingual, and Multicultural Hate Speech Detection with Vision-Language Models

    Multi3hate: Multimodal, multilingual, and multicultural hate speech de- tection with vision-language models.Preprint, arXiv:2411.03888. Yong Cao, Yova Kementchedjhieva, Ruixiang Cui, An- tonia Karamolegkou, Li Zhou, Megan Dare, Lucia Donatelli, and Daniel Hershcovich

    work page Pith review arXiv

  3. [5]

    Georgiana Dinu, Prashant Mathur, Marcello Federico, Stanislas Lauly, and Yaser Al-Onaizan

    Aya expanse: Combin- ing research breakthroughs for a new multilingual frontier.Preprint, arXiv:2412.04261. Georgiana Dinu, Prashant Mathur, Marcello Federico, Stanislas Lauly, and Yaser Al-Onaizan

    arXiv

  4. [6]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8823–8838, Online

    Hierarchical graph network for multi-hop question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8823–8838, Online. Association for Computa- tional Linguistics. Neil Fantom and Umar Serajuddin

    work page 2020

  5. [8]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu

    The llama 3 herd of models.Preprint, arXiv:2407.21783. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu

    arXiv

  6. [9]

    Accessed: 2024- 02-13

    International Fi- nancial Statistics: Exchange Rates. Accessed: 2024- 02-13. Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel

    work page 2024

  7. [10]

    Exploring cross-cultural differences in English hate speech annotations: From dataset construction to analysis. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies (Volume 1: Long Papers), pages 4205–4224, Mexico City, Mexico. Association for Computational Lingu...

    work page 2024

  8. [11]

    InFindings of the Association for Computational Linguistics: ACL 2024, pages 9668–9688, Bangkok, Thailand

    Understanding and patch- ing compositional reasoning in LLMs. InFindings of the Association for Computational Linguistics: ACL 2024, pages 9668–9688, Bangkok, Thailand. Associ- ation for Computational Linguistics. Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi

    work page 2024

Show all 25 references
  1. [12]

    Mazda Moayeri, Elham Tabassi, and Soheil Feizi

    Multi-hop question answering.Preprint, arXiv:2204.09140. Mazda Moayeri, Elham Tabassi, and Soheil Feizi

    arXiv

  2. [13]

    InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Trans- parency, FAccT ’24, page 1211–1228, New York, NY , USA

    Worldbench: Quantifying geographic disparities in llm factual recall. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Trans- parency, FAccT ’24, page 1211–1228, New York, NY , USA. Association for Computing Machinery. Junho Myung, Nayeon Lee, Yi Zhou,...

    2024

  3. [16]

    InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 5687–5711, Singa- pore

    Measuring and narrowing the compositionality gap in language mod- els. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 5687–5711, Singa- pore. Association for Computational Linguistics. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, ...

    2023

  4. [18]

    InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 4996–5025, Miami, Florida, USA

    CultureBank: An online community-driven knowledge base towards cultur- ally aware language technologies. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 4996–5025, Miami, Florida, USA. Asso- ciation for Computational Linguistics. SimpleMaps

    2024

  5. [19]

    https:// simplemaps.com/data/world-cities

    World cities database. https:// simplemaps.com/data/world-cities. Accessed: 2025-05-28. Kai Sun, Yifan Xu, Hanwen Zha, Yue Liu, and Xin Luna Dong

    2025

  6. [20]

    Head-to-tail: How knowledgeable are large language models (LLMs)? A.K.A. will LLMs replace knowledge graphs? InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technologies (Volume 1: Long Pa- p...

    2024

  7. [21]

    Accessed: 2024-02-13

    GDP per capita (current US). Accessed: 2024-02-13. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal

    2024

  8. [22]

    Accessed: 2025-02-13

    Consumer price index (cpi). Accessed: 2025-02-13. Frank Wilcoxon

    2025

  9. [23]

    Accessed: 2025-02-13

    Consumer price index (cpi). Accessed: 2025-02-13. Sohee Yang, Elena Gribovskaya, Nora Kassner, Mor Geva, and Sebastian Riedel

    2025

  10. [64]

    For CoT prompting, we increase the token limit, which extends the processing time to nearly 24 hours. Llama 3.3 Llama 3.1 Qwen2.5 Aya Average 70B 70B 8B 72B 7B 32B 8B Currency (Fiscal Data) Default = USD 71.51* 75.99 31.58* 76.51* 65.01* 63.08* 54.05* 62.53* EUR 62.45 75.48 2....

  11. [2018]

    InProceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium

    HotpotQA: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 Conference on Empiri- cal Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium. Association for Com- putational Linguistics. A Dataset A.1 Measurement Sys...

    2018

  12. [2019]

    Association for Computational Linguistics

    Language models as knowl- edge bases? InProceedings of the 2019 Confer- ence on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463–2473, Hong Kong, China. Association for Compu...

    2019

  13. [2020]

    InFindings of the Asso- ciation for Computational Linguistics: EMNLP 2020, pages 1026–1036, Online

    Hy- bridQA: A dataset of multi-hop question answering over tabular and textual data. InFindings of the Asso- ciation for Computational Linguistics: EMNLP 2020, pages 1026–1036, Online. Association for Computa- tional Linguistics. John Dang, Shivalika Singh, Daniel D’souza, Ara...

    2020

  14. [2022]

    Association for Com- putational Linguistics

    Do language models understand measurements? In Findings of the Association for Computational Lin- guistics: EMNLP 2022, pages 1782–1792, Abu Dhabi, United Arab Emirates. Association for Com- putational Linguistics. Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewi...

    2022

  15. [2023]

    InProceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12216–12235, Singapore

    Dissecting recall of factual associa- tions in auto-regressive language models. InProceed- ings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12216–12235, Singapore. Association for Computational Linguis- tics. Aaron Grattafiori, Abhimanyu D...

    2023

  16. [2024]

    Eden Biran, Daniela Gottesman, Sohee Yang, Mor Geva, and Amir Globerson

    Physics of language models: Part 3.1, knowledge storage and extraction.Preprint, arXiv:2309.14316. Eden Biran, Daniela Gottesman, Sohee Yang, Mor Geva, and Amir Globerson

    arXiv

  17. [2025]

    Weiyan Shi, Ryan Li, Yutong Zhang, Caleb Ziems, Sunny Yu, Raya Horesh, Rogério Abreu De Paula, and Diyi Yang

    Qwen2.5 technical report.Preprint, arXiv:2412.15115. Weiyan Shi, Ryan Li, Yutong Zhang, Caleb Ziems, Sunny Yu, Raya Horesh, Rogério Abreu De Paula, and Diyi Yang

    arXiv

Pith tools

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