Pith. sign in

REVIEW 4 major objections 5 minor 23 references

WuYu-EnvLE-Bench: A Benchmark for Evaluating Large Language Models in Environmental Law Enforcement

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Current LLMs handle rule-bounded enforcement tasks but remain unreliable at evidence-chain construction, contradiction detection, and procedural judgment — and scaling past medium size yields diminishing returns.

desk verdict A genuinely useful new benchmark whose headline numbers rest on an unnamed LLM judge; worth refereeing, but the judge must be identified and validated. read the letter →

arxiv 2607.17745 v1 pith:OWIJCGNU submitted 2026-07-20 cs.AI

Ziliang Yang , Yi Zhang , Kaijun Lin , Jiachao Ke , Haihong Xu , Zongguo Wen This is my paper
classification cs.AI
keywords largelanguagemodelsenvironmentallawenforcementbenchmarkevidence-chainreasoningcontradictiondetectionmodelscalingLLM-as-a-Judgeadministrativepenalty
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 tries to establish that large language models can participate in environmental law enforcement only within strict boundaries. Using a benchmark built from real enforcement cases, it shows that LLMs score well on rule-bounded tasks — penalty-type identification and penalty-result reasoning stay above 80 — but collapse on evidence-chain reasoning, where contradiction monitoring scores hover between 22 and 32 and multi-evidence integration peaks near 49. It further argues that model scaling shows diminishing returns: moving from small to medium models brings large gains, but further parameter growth yields flat or negative deployment value once resource cost is factored in. The conclusion is that enforcement LLMs should be optimized for evidence-grounded, rule-aware, task-adaptive reasoning rather than raw scale.

What carries the argument

The carrying mechanism is the benchmark's two-dimensional task–domain matrix (14 tasks across three enforcement stages and 12 pollution-medium subdomains) combined with two metrics: AES, a stage-macro-averaged capability score, and IEI, a deployment score that multiplies a weighted capability-and-quality base by a resource-efficiency factor. Open-ended tasks are scored by an LLM-as-a-Judge framework with a strict pass rule, which is the instrument that produces the sharp task-level distinctions.

What would settle it

Have a panel of experienced enforcement officers independently score a stratified sample of open-ended responses using the same rubrics, then compare with the LLM judge's pass/fail calls; low agreement or a different judge model reordering the models would refute the evidence-bottleneck and diminishing-returns conclusions.

Watch

Extended reading notes

Core claim

The paper's central discovery is a capability gradient, not a simple pass/fail. Across a task–domain matrix covering pre-, in-, and post-enforcement stages and water, air, and solid-waste subdomains, model performance is determined by task structure and evidentiary complexity rather than pollution medium. LLMs reliably perform stable rule-matching and standardized output generation; they fail when the task requires coordinating heterogeneous evidence, detecting cross-source contradictions, or applying procedural constraints. The paper documents this through large score gaps within the same models — multi-evidence extraction reaches 97 for one model while contradiction monitoring stays below

Load-bearing premise

The open-ended task scores depend on a single, unnamed LLM judge applying a strict pass rule; if that judge is biased by style or model family, the reported stage profiles and scaling conclusions would shift.

Editorial extensions

If this is right

  • Deployment guidelines should treat LLMs as assistance for rule-bounded and fact-explicit steps (complaint screening, penalty classification, result reasoning), not as autonomous fact-finders.
  • Procurement decisions should compare medium-scale open-source models with frontier closed-source models on task-level scores; on structured tasks the gap is small, and IEI can reverse the ranking.
  • Evidence-chain tasks — contradiction monitoring, multi-evidence integration, and procedure-constrained judgment — define the current capability ceiling; investment should target these before more scaling.
  • Reasoning mode should be selected per task; explicit 'thinking' helps in gap-identification tasks but degrades faithfulness in extraction and source-attribution tasks.
  • Regulatory use should separate capability ceiling from operational efficiency; a model can have the highest accuracy but lower practical value once inference cost is counted.

Reading between the lines

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

  • A direct extension of the paper's results: evaluation budgets should be allocated by task class rather than by average AES, since the average conceals the 22–32% contradiction floor.
  • The score pattern suggests a testable mechanism — LLMs perform local pairwise consistency checks but lack global source-credibility calibration; a probe could isolate this by giving models evidence items with conflicting reliability labels and asking for a ranked fact chain.
  • Because the open-ended scores rely on one LLM judge with a strict pass rule and no reported human-expert agreement, the absolute numbers should be treated as provisional until judge identity and sensitivity are reported; this is my inference, not the paper's claim.
  • The task–medium interaction predicts that cross-domain generalization holds only when legal rules are explicit; if evidence organization is scrambled within the same pollution medium, performance should drop — a testable manipulation the paper did not run.
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 / 5 minor

Summary. The paper introduces WuYu-EnvLE-Bench, a Chinese environmental-law-enforcement benchmark built from real case materials, with 2,521 instances across 14 tasks, three enforcement stages, and 12 pollution-medium subdomains. It defines a two-level evaluation: AES, a macro-average of task scores over stages, and IEI, which combines AES, an LLM-judged Quality Factor, and a resource-efficiency adjustment. The authors evaluate a range of open- and closed-source LLMs (plus SLMs) and report that models are strong on rule-bounded tasks (e.g., penalty-type identification, penalty-result reasoning) but weak on evidence-chain construction, contradiction detection, multi-source integration, and procedural judgment, and that scaling yields diminishing returns once medium-scale models are reached.

Significance. If its measurement assumptions hold, this is a useful and timely contribution. The task taxonomy, case-to-task pipeline, gold standards, rubrics, and appendices are unusually detailed, and the separation of rule-based, structured, and open-ended scoring is sensible. The paper also makes an honest, falsifiable claim: current LLMs are useful but bounded tools for environmental enforcement, with a capability ceiling in evidence-centric reasoning. The main strengths are the domain grounding, the explicit metric formulas, and the publicly documentable benchmark protocol. However, the central quantitative claims rest on an LLM-as-a-Judge framework whose validity is not established. The findings are plausible, but the absence of judge validation, judge-sensitivity analysis, and human agreement makes the headline numbers—especially the low scores on Contradiction Monitoring and Multi-Evidence Integration—not yet fully load-bearing.

major comments (4)
  1. [§3.1, App C.3.2] All open-ended task scores and the Quality Factor Q_m are produced by an unidentified LLM judge using a strict pass rule. No judge model is named, no human-expert agreement is reported, and no judge-family or prompt-sensitivity analysis is given. This is load-bearing because the paper's main findings—Contradiction Monitoring 22–32, Multi-Evidence Integration max ~49, Inspection Planning gaps, and the 'medium models approach frontier' comparison—are computed from these judge scores. I request: (1) identify the judge model and version; (2) report agreement on a sample against at least two domain experts; (3) recompute headline task scores with an alternative judge or a non-LLM extraction-based metric and show rank stability.
  2. [§B.2.4.2, §B.3.3.2, §4.1] GPT-5.5 was used to generate contradictory distractor evidence (Contradiction Monitoring) and perturbed legal provisions (Legal Change Disturbance), and GPT-5.5 is itself among the evaluated models. Although the paper says outputs were manually reviewed, this creates a potential same-model contamination: the evaluated model may be indirectly tested on content generated by itself, which can bias relative rankings. This is especially relevant because Contradiction Monitoring is one of the weakest tasks for all models, including GPT-5.5. Please clarify how much GPT-5.5 output survives after human review, and provide a robustness check that excludes or separately analyzes tasks with GPT-5.5-authored items.
  3. [§C.6–C.8, §4.1, §5 (RQ2)] The diminishing-returns conclusion is partly driven by IEI, whose construction uses several free parameters: lambda_p=0.2, lambda_c=1.0, w_a=w_q=0.5, equal stage weights, equal quality-dimension weights, and the 1.5 characters/token conversion. These choices are transparent but not derived from data or sensitivity analysis. A reader cannot tell whether 'further parameter expansion does not bring proportional enforcement gains' is a property of the models or of the chosen resource-penalty coefficients. Please add a sensitivity analysis over lambda_p, lambda_c, and the quality/capability weight, and show AES-only scaling curves alongside the IEI curves.
  4. [§4.2, Table 1] The paper reports task-level differences of a few points (e.g., 22.5 vs 29.1 on Contradiction Monitoring) as meaningful capability conclusions, but no confidence intervals or significance tests are provided. With task sample sizes around 158–193, such small differences may be within sampling noise. I ask the authors to report standard errors or confidence intervals for the main task-level and stage-level scores, at least for the headline comparisons.
minor comments (5)
  1. [Abstract / §4.1] The abstract and Section 4.1 use different AES values for the same models (e.g., GPT-5.5 AES appears as 69.3/53.5/76.4 stage scores in §4.1 but the overall AES is not explicitly listed in the text; the tables in Appendix D are on a 0–100 scale). Please ensure the main-text AES values are consistent with the appendix tables and clearly state the mapping from normalized AES to the 0–100 scale.
  2. [Figure 2] The donut-chart percentages for the 14 task types sum to 100% but several are rounded to one decimal and one reads 6.3% for Legal Change Disturbance. Please check rounding and label the figure with exact counts to avoid apparent inconsistencies.
  3. [App C.3.2] The text says the unified judge prompt is 'provided in Appendix C.3.2,' but the appendix only includes a figure reference (Figure C.1) without the full prompt text. Since the benchmark is meant to be reproducible, include the complete judge prompt and output schema in the appendix.
  4. [§3.3 / Eq. (3)] The definition of AES uses A_ES^norm with stage weights alpha_g, and the text says equal weights are used. Please state explicitly that alpha_g = 1/3 in the equation area, rather than only in prose, so the formula is self-contained.
  5. [General] Some model names are written inconsistently (e.g., 'GPT 5.5' in Table 1 vs 'GPT-5.5' elsewhere; 'Qwen3.5 397B-A17B' vs 'Qwen3.5-397B-A17B'). Please standardize model naming throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: AES/IEI are transparent aggregations of task scores and resource proxies; the scaling and evidence-chain findings are empirical, not definitionally forced.

full rationale

The paper's derivation chain is definitionally transparent. TaskScorem,t is normalized and then macro-averaged into stage scores and AES (§3.3, App C.4); no task score is defined in terms of AES or IEI. IEI is explicitly a weighted product of capability/quality/resource proxies (App C.8: CB_m = w_a·AES_m + w_q·Q_m; IEI_m = CB_m·RE_m), so the observation that very large models have lower IEI when capability gains are flat is a stated design property of the metric, not a hidden re-import of the RQ2 conclusion. The empirical scaling claim in RQ2 is separately grounded in AES values (Tables 1/D), which are macro-averages independent of parameter count. There are no load-bearing self-citations: the references are external benchmarks, legal datasets, and methods, and no author-overlapping citation is invoked to force a conclusion. The LLM-as-a-Judge protocol for open-ended tasks and the Quality Factor is a measurement instrument; while the judge identity and human agreement are not reported (an internal-validity limitation), nothing in the scoring equations makes a model's score equal to an input or to the judge's own output by construction. The use of GPT-5.5 to generate distractor texts for Contradiction Monitoring and perturbed provisions for Legal Change Disturbance is a data-construction contamination risk, but the gold labels are independently anchored: distractors are 'manually reviewed and mixed with valid evidence' (App B.2.4) and gold answers/rubrics are used by the judge. No quoted equation or citation exhibits a reduction of a claimed result to its own inputs. Therefore, no significant circularity is found.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claims are empirical, not derived from first principles. They depend on task scores that require trusting expert gold standards, LLM-as-a-Judge validity, and the representativeness of rewritten/simulated case materials. The composite indices AES/IEI add several hand-set weights (stage weights, quality weights, penalty coefficients) that influence the scaling and efficiency conclusions. No new physical entities are introduced.

free parameters (8)
  • Equal stage weights alpha_g = 1/3 (AES) = 1/3 each
    Hand-set so pre/in/post-enforcement contribute equally; changing this would change AES rankings.
  • Capability base weights w_a = w_q = 0.5 = 0.5, 0.5
    Equal weighting of AES and Quality Factor in IEI; not derived from data.
  • Open-source parameter penalty lambda_p = 0.2
    Controls Resource Efficiency penalty in IEI; chosen to 'limit impact' per C.6; directly affects diminishing-returns conclusion for open models.
  • Closed-source cost penalty lambda_c = 1.0
    Controls Cost Burden in IEI; hand-set in C.7.
  • Multi-step rule-based weights alpha=beta=0.5 = 0.5, 0.5
    Final answer vs process equally weighted; affects Penalty Decision and Penalty Result Reasoning task scores.
  • Structured-recognition sub-objective weights gamma_k = equal unless rubric states otherwise
    Equal weights for sub-objectives in tasks like Anomaly Detection; affects structured task scores.
  • Quality Factor dimension weights beta_R=beta_E=beta_P=1/3 = 1/3 each
    Reliability, Explainability, and Professionalism equally weighted in Qm.
  • Token-to-character conversion factor = 1.5 chars/token
    Assumed for closed-source price conversion in C.7; not empirically derived in this paper.
assumptions (5)
  • domain assumption LLM-as-a-Judge outputs are a valid measure of enforcement response quality.
    Open-ended task scores and Quality Factor are based on judge model scores (Sec 3.1, App C.3.2); no human agreement or judge-model sensitivity analysis is reported.
  • domain assumption Gold standards and rubrics reviewed by experts are correct.
    The benchmark rests on manual expert review (Sec 3.1), but no inter-annotator reliability is reported.
  • domain assumption Rewritten/simulated task inputs preserve the difficulty distribution of real enforcement.
    Complaints, public sentiment threads, inquiry dialogues, and anomalies are generated/simulated from case briefs (App B), not collected verbatim.
  • domain assumption Chinese environmental law and MEE procedure is the target domain.
    Legal bases and case materials are Chinese; conclusions about 'environmental enforcement' are scoped to this jurisdiction.
  • ad hoc to paper One output token approximately equals 1.5 output characters.
    Used in C.7 to convert OpenRouter token prices to character-level costs; a rough empirical convention.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WuYu-EnvLE-Bench: A Benchmark for Evaluating Large Language Models in Environmental Law Enforcement." pith.science (2026). https://pith.science/paper/OWIJCGNU

@misc{pith2026260717745,
  author       = {Pith},
  title        = {Pith review of: WuYu-EnvLE-Bench: A Benchmark for Evaluating Large Language Models in Environmental Law Enforcement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWIJCGNU}},
  note         = {Machine review of arXiv:2607.17745}
}
read the original abstract

Large language models (LLMs) are increasingly considered for environmental enforcement, but their ability to produce traceable enforcement decisions remains unclear. We introduce WuYu-EnvLE-Bench, a benchmark built from real enforcement cases, regulatory standards, and expert review. It contains 2,521 benchmark instances, 14 tasks, and 12 pollution-medium subdomains across pre-enforcement, in-enforcement, and post-enforcement workflows. Using Absolute Environmental Enforcement Score (AES) and Intelligent Enforcement Index (IEI), we evaluate open-source and closed-source LLMs across capability, response quality, and resource efficiency. Results show that LLMs perform well on rule-bounded tasks but remain unreliable in evidence-chain construction, contradiction detection, multi-source integration, and procedural judgment. Model scaling also shows diminishing returns: medium-sized models approach leading models in structured tasks, while larger models do not reliably overcome evidence-reasoning bottlenecks. WuYu-EnvLE-Bench highlights the need for evidence-grounded, rule-aware, and task-adaptive enforcement reasoning.

Figures

Figures reproduced from arXiv: 2607.17745 by the authors.

Figure 1
Figure 1. Overview of WuYu-EnvLE-Bench. Panel (a) shows the data-to-benchmark pipeline with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Data statistics of WuYu-EnvLE-Bench. Panel (a) summarizes the distribution across [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Overall AES and dimension-level performance across models, enforcement stages, and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Model Intelligent Enforcement Index (IEI) comparison. Panel (a) ranks open-source models [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Cross-analysis of enforcement tasks and pollution-medium subdomains. Panel (a) reports [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Effect of Thinking mode on overall AES by models [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Task- and domain-level AES changes under Thinking mode. Panels (a)-(c) show score [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Representative cases across WuYu-EnvLE-Bench tasks [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 4 linked inside Pith

  1. [1]

    Gray and Jay P

    Wayne B. Gray and Jay P. Shimshack. The effectiveness of environmental monitoring and enforcement: A review of the empirical evidence.Review of Environmental Economics and Policy, 5(1):3–24, 2011

    2011

  2. [2]

    Ensuring compliance with environmental regulations in the east- ern partnership countries

    OECD. Ensuring compliance with environmental regulations in the east- ern partnership countries. https://www.oecd.org/content/dam/oecd/ en/topics/policy-sub-issues/ensuring-environmental-compliance/ brochure-environmental-compliance-assurance-eap-2024.pdf, 2024

    2024

  3. [3]

    Conley K. Hurst. The scope of evidentiary review in constitutional challenges to agency action. The University of Chicago Law Review, 88(6):1511–1554, 2021

    2021

  4. [4]

    Ho, Christopher Ré, Adam Chilton, Aditya Narayana, Alex Chohlas-Wood, Austin Peters, et al

    Neel Guha, Julian Nyarko, Daniel E. Ho, Christopher Ré, Adam Chilton, Aditya Narayana, Alex Chohlas-Wood, Austin Peters, et al. LegalBench: A collaboratively built benchmark for measuring legal reasoning in large language models. InAdvances in Neural Information Processing Systems, volume 36, pages 44123–44279, 2023

    2023

  5. [5]

    Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang

    Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R. Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. Scibench: Evaluating college-level scientific problem-solving abilities of large language models. InProceedings of the 41st International Conference on Machine Learning, 2024

    2024

  6. [6]

    A benchmark of expert-level academic questions to assess AI capabilities.Nature, 649:1139–1146, 2026

    Center for AI Safety, Scale AI, and HLE Contributors Consortium. A benchmark of expert-level academic questions to assess AI capabilities.Nature, 649:1139–1146, 2026

    2026

  7. [7]

    Manning, and Daniel E

    Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning, and Daniel E. Ho. Hallucination-free? assessing the reliability of leading AI legal research tools. Journal of Empirical Legal Studies, 22(2):216–242, 2025

    2025

  8. [8]

    ESGe- nius: Benchmarking LLMs on environmental, social, and governance (ESG) and sustainability knowledge

    Chaoyue He, Xin Zhou, Yi Wu, Xinjia Yu, Yan Zhang, Lei Zhang, Di Wang, et al. ESGe- nius: Benchmarking LLMs on environmental, social, and governance (ESG) and sustainability knowledge. InProceedings of the 2025 Conference on Empirical Methods in Natural Lan- guage Processing, pages 14612–14653, Suzhou, China, 2025. Association for Computational Linguistics

    2025

Show all 23 references
  1. [9]

    Fine-tuning large language models for interdisciplinary environmental challenges.Environmen- tal Science and Ecotechnology, 27:100608, 2025

    Yuanxin Zhang, Sijie Lin, Yaxin Xiong, Nan Li, Lijin Zhong, Longzhen Ding, and Qing Hu. Fine-tuning large language models for interdisciplinary environmental challenges.Environmen- tal Science and Ecotechnology, 27:100608, 2025

    2025

  2. [10]

    Enforcing environmental regulation.Journal of Environmental Law, 23(2): 169–201, 2011

    Neil Gunningham. Enforcing environmental regulation.Journal of Environmental Law, 23(2): 169–201, 2011

    2011

  3. [11]

    Densing law of LLMs.Nature Machine Intelligence, 7(11): 1823–1833, 2025

    Chaojun Xiao, Jie Cai, Weilin Zhao, Biyuan Lin, Guoyang Zeng, Jie Zhou, Zhi Zheng, Xu Han, Zhiyuan Liu, and Maosong Sun. Densing law of LLMs.Nature Machine Intelligence, 7(11): 1823–1833, 2025

    2025

  4. [12]

    Lawbench: Benchmarking legal knowledge of large language models

    Zhiwei Fei, Xiaoyu Shen, Dawei Zhu, Fengzhe Zhou, Zhuo Han, Alan Huang, Songyang Zhang, et al. Lawbench: Benchmarking legal knowledge of large language models. InProceedings of the 2024 conference on empirical methods in natural language processing, pages 7933–7962, 2024

    2024

  5. [13]

    Lexeval: A comprehensive chinese legal benchmark for evaluating large language models.Advances in Neural Information Processing Systems, 37:25061–25094, 2024

    Haitao Li, You Chen, Qingyao Ai, Yueyue Wu, Ruizhe Zhang, and Yiqun Liu. Lexeval: A comprehensive chinese legal benchmark for evaluating large language models.Advances in Neural Information Processing Systems, 37:25061–25094, 2024

    2024

  6. [14]

    Beyond guilt: Legal judgment prediction with trichotomous reasoning

    Kepu Zhang, Haoyue Yang, Xu Tang, Weijie Yu, and Jun Xu. Beyond guilt: Legal judgment prediction with trichotomous reasoning. arXiv preprint arXiv:2412.14588, 2024

    2024 arXiv

  7. [15]

    LawShift: Benchmarking legal judgment prediction under statute shifts

    Zhuo Han, Yi Yang, Yi Feng, Wanhong Huang, Xuxing Ding, Chuanyi Li, Jidong Ge, and Vincent Ng. LawShift: Benchmarking legal judgment prediction under statute shifts. In Advances in Neural Information Processing Systems, volume 38, 2025. 18

    2025

  8. [16]

    Legala- gentbench: Evaluating LLM agents in legal domain

    Haitao Li, Junjie Chen, Jingli Yang, Qingyao Ai, Wei Jia, Youfeng Liu, Kai Lin, et al. Legala- gentbench: Evaluating LLM agents in legal domain. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 2322–2344, 2025

    2025

  9. [17]

    LLM-based HSE compliance assessment: Benchmark, performance, and advancements

    Jianwei Wang, Mengqi Wang, Yinsi Zhou, Zhenchang Xing, Qing Liu, Xiwei Xu, Wenjie Zhang, and Liming Zhu. LLM-based HSE compliance assessment: Benchmark, performance, and advancements. arXiv preprint arXiv:2505.22959, 2025

    2025 arXiv

  10. [18]

    Benchmarking LLMs for environmental review and permitting

    Rounak Meyur, Hung Phan, Koby Hayashi, Ian Stewart, Shivam Sharma, Sarthak Chaturvedi, Mike Parker, et al. Benchmarking LLMs for environmental review and permitting. arXiv preprint arXiv:2407.07321, 2024

    2024 arXiv

  11. [19]

    Environmental large language model evaluation (ELLE) dataset: A benchmark for evaluating generative AI applications in eco-environment domain

    Jing Guo, Nan Li, and Ming Xu. Environmental large language model evaluation (ELLE) dataset: A benchmark for evaluating generative AI applications in eco-environment domain. arXiv preprint arXiv:2501.06277, 2025

    2025 arXiv

  12. [20]

    Measures for ecological and environmental administrative penalties

    Ministry of Ecology and Environment of the People’s Republic of China. Measures for ecological and environmental administrative penalties. https://www.mee.gov.cn/gzk/gz/ 202305/P020230516552750932005.pdf, 2023

    2023

  13. [21]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. InProceedings of the 9th International Conference on Learning Representations (ICLR 2021), 2021

    2021

  14. [22]

    Holistic evaluation of language models.Transactions on Machine Learning Research, 2023

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, et al. Holistic evaluation of language models.Transactions on Machine Learning Research, 2023

    2023

  15. [23]

    it’s not what you see but what you hear

    Jim Curtis and Stefan Kaufman. "it’s not what you see but what you hear...": Understand- ing environment protection officers’ responsive decision making.Journal of Environmental Management, 262:110336, 2020. 19 A Ethical Considerations The source materials used to construct Wu...

    2020

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.