pith. sign in

arxiv: 2606.05920 · v1 · pith:4TN2BK5Jnew · submitted 2026-06-04 · 💻 cs.SE · cs.CL

Asuka-Bench: Benchmarking Code Agents on Underspecified User Intent and Multi-Round Refinement

Xin Wang , Liangtai Sun , Yaoming Zhu , Shuang Zhou , Jiaxing Liu , Fengjiao Chen , Lin Qiu , Xuezhi Cao
show 3 more authors
Xunliang Cai Licheng Zhang Zhendong Mao
This is my paper

Pith reviewed 2026-06-28 00:24 UTC · model grok-4.3

classification 💻 cs.SE cs.CL
keywords code agentsbenchmarksweb developmentmulti-round refinementunderspecified intentLLM evaluationiterative feedback
0
0 comments X

The pith

Asuka-Bench shows code agents complete only 52 percent of web projects after three rounds of feedback on underspecified tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Asuka-Bench to measure how code agents perform when user intent starts vague and must be clarified over multiple rounds through feedback on browser-rendered output. It replaces one-shot prompt-to-code evaluation with a closed loop in which a code agent builds a web project, a UI agent runs tests on the live site, and a user LLM converts test results into natural-language comments for the next round. Across 50 tasks, 784 criteria, and eight models in two frameworks, the benchmark finds a 38-point spread in weighted task pass rate and shows that even the strongest model finishes just 52 percent of projects after three rounds. This setup matters because it mirrors how actual web development proceeds when requirements emerge from seeing intermediate results rather than from an exhaustive initial specification.

Core claim

Asuka-Bench demonstrates that current code agents vary widely in their capacity to incorporate natural-language feedback and repair web projects over successive rounds, with weighted Task Pass Rate differing by 38 percentage points and the top model reaching only 52 percent completion after three iterations.

What carries the argument

The closed loop of Code Agent generating web projects, UI Agent executing test cases on the deployed site, and User LLM converting evaluation outcomes into natural-language feedback for the next round.

If this is right

  • Models differ substantially in their ability to repair projects from feedback.
  • Performance gaps between models remain large even after multiple refinement rounds.
  • The benchmark leaves substantial headroom, since the strongest model reaches only 52 percent completion.
  • Weighted Task Pass Rate provides a finer separation than simple pass/fail metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Agents that treat feedback as an explicit state update rather than a new prompt may close part of the observed gap.
  • Extending the same closed-loop structure to non-web domains would test whether the refinement difficulty is web-specific or general.
  • Task Pass Rate after one round versus after three rounds could serve as a direct measure of an agent's feedback incorporation skill.

Load-bearing premise

The closed loop of code agent, UI agent test execution, and user LLM feedback accurately captures real-world underspecified user intent and multi-round refinement in web development.

What would settle it

An experiment in which human users replace the User LLM and produce materially different feedback that changes model ranking or completion rates on the same 50 tasks.

Figures

Figures reproduced from arXiv: 2606.05920 by Fengjiao Chen, Jiaxing Liu, Liangtai Sun, Licheng Zhang, Lin Qiu, Shuang Zhou, Xin Wang, Xuezhi Cao, Xunliang Cai, Yaoming Zhu, Zhendong Mao.

Figure 1
Figure 1. Figure 1: Comparison between existing paradigm and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Asuka-Bench. Top: Dataset construction—raw queries from three sources are rewritten into a Clarified PRD, from which an underspecified query, simulated data, and DAG-organized evaluation criteria are derived. Bottom-left: Evaluation loop—a Code Agent generates a web project from the underspecified query; a UI Agent executes criteria as browser-based test cases; a User LLM synthesizes pass/fail … view at source ↗
Figure 3
Figure 3. Figure 3: Dataset overview. Left: Distribution of benchmark tasks across six application categories. Right: Summary statistics of the Asuka-Bench dataset. ments, retaining only a high-level description of the desired application. This underspecified query serves as the sole input to the Code Agent during evaluation. Second, since each benchmark task re￾quires a fully functional frontend application with￾out backend … view at source ↗
Figure 5
Figure 5. Figure 5: Weighted Task Pass Rate per Claude Code model under four settings: a single round on the Clari￾fied PRD, and three Rounds on the underspecified query Saturation Analysis. To verify that the evalua￾tion tasks are inherently solvable rather than ill￾defined, we extend the SOTA model (Claude-4.6- Sonnet) to 8 interaction rounds. As shown in Fig￾ure 4, both Project Completion and weighted Task Pass Rate conver… view at source ↗
Figure 4
Figure 4. Figure 4: Cumulative Project Completion Rate and weighted Task Pass Rate for Claude-4.6-Sonnet across 8 feedback rounds, both reaching 100%. 20 30 40 50 60 70 80 90 Weighted Task Pass Rate (%) +8.7 +15.2 +21.7 +26.6 +20.1 +21.2 24.7 51.8 41.3 70.1 39.2 76.4 40.8 76.6 49.7 88.1 55.7 89.4 Seed-2.0-Pro Qwen3.5-Plus MiniMax-M2.7 GLM-5 GPT-5.4 Claude-4.6-Sonnet Unspecified Query R1/R2/R3 PRD R1 [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 6
Figure 6. Figure 6: Cumulative weighted Implemented (translucent outer bar; tasks reachable under the DAG evaluation protocol) and Pass (solid inner bar; tasks fully passed) rates per round, under the underspecified-query setting. Top: Claude Code; bottom: OpenHands. The Implemented−Pass gap measures evaluable tasks the model failed to complete, and how it shrinks across rounds reflects each model’s repair-from-feedback abili… view at source ↗
Figure 7
Figure 7. Figure 7: The web-based annotation console used by human annotators. The deployed project under evaluation is [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Rule for human annotators [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt used to synthesize the underspecified user query from a Clarified PRD. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Vague user input of proj_34 PRD — Supplier Resource Evaluation Visualization Ranking Page 1. Project Overview. A visualization ranking page for supplier management and procurement evaluation. Suppliers are scored on resource-type and quality-type indicators and ranked by composite score; a radar chart shows each supplier’s per-dimension score structure for quick strength / weakness diagnosis. 2. Core Func… view at source ↗
Figure 11
Figure 11. Figure 11: Full PRD doc of proj_34 [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Task-dependency DAG for the case study ( [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Successful implementation of the case-study PRD by Claude-4.6-Sonnet after three rounds of feedback. [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Progressive repair on the Rural-Stay Revenue Strategy Tool (Claude-4.6-Sonnet). v1 (left) shows four [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Progressive repair on the Visual Social Media Platform (Claude-4.6-Sonnet). v1 (left) shows four [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
read the original abstract

Existing code-generation benchmarks score a single mapping from a complete prompt to a one-shot output. However, real web development is different. Users seldom write a full spec at the start; many requirements only become clear once they look at an intermediate result and react to it. We present Asuka-Bench, a benchmark that pairs underspecified user intent with multi-round refinement, grounded in browser-rendered behavior. Each task is resolved through a closed loop: a Code Agent generates a web project, a UI Agent executes test cases on the deployed site, and a User LLM turns evaluation outcomes into natural-language feedback for the next round. The benchmark comprises 50 web tasks with 784 evaluation criteria and 2402 expected outcomes. We benchmark 8 LLMs across 2 agent frameworks. The results separate models clearly: weighted Task Pass Rate varies by 38 percentage points and models also differ substantially in their ability to repair from feedback. Asuka-Bench is also far from saturated: even the strongest model completes only 52% of projects after three rounds.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Asuka-Bench, a benchmark of 50 web development tasks (784 criteria, 2402 expected outcomes) that evaluates code agents on underspecified intents via a closed-loop protocol: Code Agent generates a project, UI Agent runs browser tests, and User LLM converts outcomes into natural-language feedback for up to three refinement rounds. Benchmarking eight LLMs across two frameworks shows clear separation, with weighted Task Pass Rate spanning 38 percentage points and the strongest model reaching only 52% project completion after three rounds.

Significance. If the simulation protocol is shown to be a faithful proxy, the benchmark would be a useful addition to code-agent evaluation by moving beyond one-shot generation to iterative, feedback-driven refinement grounded in rendered behavior. The reported performance gaps and low absolute scores provide concrete evidence that current agents struggle with repair from underspecified feedback, which could guide future work on multi-turn agent architectures.

major comments (2)
  1. [§3, §4] §3 (Benchmark Construction) and §4 (Evaluation Protocol): The central claim that Asuka-Bench measures real-world multi-round refinement rests on the User LLM converting UI Agent test outcomes into natural-language feedback. No human baseline, inter-rater agreement study, or ablation comparing LLM-generated feedback distributions to human developer responses after viewing rendered pages is reported. Without this, both absolute completion rates and model rankings risk being artifacts of the specific LLM feedback style rather than evidence about underspecified intent.
  2. [§5] §5 (Results): The headline separation of 38 pp in weighted Task Pass Rate and the 52% ceiling for the best model after three rounds are presented as evidence of benchmark utility, yet these quantities inherit the unvalidated feedback mechanism; any systematic difference in feedback precision or scope between the User LLM and actual users would directly alter both the reported numbers and the cross-model ordering.
minor comments (2)
  1. [Abstract, §2] The abstract and §2 would benefit from an explicit statement of how the 50 tasks were sampled and how the 784 criteria were derived to ensure they are representative rather than hand-curated.
  2. [Table 1] Table 1 (model results) should include per-round breakdown and variance across the 50 tasks to allow readers to assess whether the 38 pp gap is driven by a few outlier tasks.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the importance of validating the User LLM feedback mechanism. We address the two major comments point by point below, maintaining that the benchmark provides a reproducible evaluation of multi-round refinement while acknowledging the lack of human comparison data.

read point-by-point responses

  1. Referee: [§3, §4] §3 (Benchmark Construction) and §4 (Evaluation Protocol): The central claim that Asuka-Bench measures real-world multi-round refinement rests on the User LLM converting UI Agent test outcomes into natural-language feedback. No human baseline, inter-rater agreement study, or ablation comparing LLM-generated feedback distributions to human developer responses after viewing rendered pages is reported. Without this, both absolute completion rates and model rankings risk being artifacts of the specific LLM feedback style rather than evidence about underspecified intent.

    Authors: We agree that no human baseline or inter-rater study is present in the manuscript. The protocol deliberately employs an LLM for feedback generation to ensure full reproducibility and scalability across all 50 tasks and 784 criteria; human feedback would introduce non-deterministic variability and prohibitive cost for repeated evaluations. Feedback generation is constrained to the concrete outcomes returned by the UI Agent's browser tests (e.g., specific assertion failures), rather than open-ended interpretation of rendered pages. This design choice allows the benchmark to isolate differences in how code agents incorporate natural-language signals over multiple rounds. While a human validation study would strengthen claims of ecological validity, its absence does not render the reported separations meaningless within the defined closed-loop setting. revision: no

  2. Referee: [§5] §5 (Results): The headline separation of 38 pp in weighted Task Pass Rate and the 52% ceiling for the best model after three rounds are presented as evidence of benchmark utility, yet these quantities inherit the unvalidated feedback mechanism; any systematic difference in feedback precision or scope between the User LLM and actual users would directly alter both the reported numbers and the cross-model ordering.

    Authors: The 38-point spread and 52% ceiling are results obtained under the fixed Asuka-Bench protocol, including the specific User LLM. They demonstrate that even with consistent, test-derived feedback, current agents exhibit substantial differences in repair capability and remain far from saturation. We do not claim the numbers generalize directly to human users; any change in feedback style would affect absolute scores. However, because the same feedback generator is applied uniformly to all models, the relative ordering reflects differences in agent architectures' ability to handle iterative, underspecified signals. We will add a short paragraph in the limitations section noting this design decision and its implications for interpreting absolute performance. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical benchmark results are direct measurements

full rationale

The paper defines Asuka-Bench explicitly with 50 tasks, 784 criteria, and a closed-loop protocol using Code Agent, UI Agent, and User LLM. Reported metrics (weighted Task Pass Rate varying 38 pp, max 52% completion) are computed outcomes from executing the 8 LLMs on these fixed tasks across rounds. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the derivation of the separation claims; results follow from running the stated protocol on the stated tasks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central contribution is the new benchmark and its evaluation loop; no free parameters, invented physical entities, or ad-hoc mathematical axioms are described in the abstract.

axioms (1)
  • domain assumption The closed loop using UI Agent test execution and User LLM feedback accurately represents real user refinement of underspecified web intents.
    This premise is required for the benchmark to claim relevance beyond synthetic tasks.

pith-pipeline@v0.9.1-grok · 5751 in / 1219 out tokens · 27421 ms · 2026-06-28T00:24:13.902252+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 14 canonical work pages · 7 internal anchors

  1. [8]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Webmmu: A benchmark for multimodal multilingual website understanding and code generation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  2. [9]

    Advances in neural information processing systems , volume=

    Web2code: A large-scale webpage-to-code dataset and evaluation framework for multimodal llms , author=. Advances in neural information processing systems , volume=

  3. [10]

    Design2code: Benchmarking multimodal code generation for automated front-end engineering , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  4. [11]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Webuibench: a comprehensive benchmark for evaluating multimodal large language models in webui-to-code , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  5. [13]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Webvoyager: Building an end-to-end web agent with large multimodal models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  6. [16]

    MiniMax-01: Scaling Foundation Models with Lightning Attention

    MiniMax-01: Scaling Foundation Models with Lightning Attention , author=. arXiv preprint arXiv:2501.08313 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [19]

    2025 , howpublished=

    Introducing GPT-5.4 , author=. 2025 , howpublished=

    2025

  8. [20]

    2025 , howpublished=

    Gemini 3.1 Pro , author=. 2025 , howpublished=

    2025

  9. [21]

    2025 , howpublished=

    Introducing Claude Sonnet 4.6 , author=. 2025 , howpublished=

    2025

  10. [22]

    2025 , howpublished=

    Seed 2.0 , author=. 2025 , howpublished=

    2025

  11. [23]

    2025 , howpublished=

    MiniMax-M2.7 , author=. 2025 , howpublished=

    2025

  12. [24]

    NeurIPS , year=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. NeurIPS , year=

  13. [25]

    NeurIPS , year=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. NeurIPS , year=

  14. [26]

    Zhang, Chenchen and Li, Yuhang and Xu, Can and others , booktitle=

  15. [27]

    Anthropic . 2025. Introducing claude sonnet 4.6. https://www.anthropic.com/news/claude-sonnet-4-6. Accessed: 2026

    2025

  16. [28]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. Program synthesis with large language models. arXiv preprint arXiv:2108.07732

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  17. [29]

    Rabiul Awal, Mahsa Massoud, Aarash Feizi, Zichao Li, Suyuchen Wang, Christopher Pal, Aishwarya Agrawal, David Vazquez, Siva Reddy, Juan A Rodriguez, and 1 others. 2025. Webmmu: A benchmark for multimodal multilingual website understanding and code generation. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages ...

    2025

  18. [30]

    ByteDance Seed Team . 2025. Seed 2.0. https://seed.bytedance.com/en/seed2. Accessed: 2026

    2025

  19. [31]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, and 1 others. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  20. [32]

    Lingyue Fu, Bolun Zhang, Hao Guan, Yaoming Zhu, Lin Qiu, Weiwen Liu, Xuezhi Cao, Xunliang Cai, Weinan Zhang, and Yong Yu. 2025. Automatically benchmarking llm code agents through agent-driven annotation and evaluation. arXiv preprint arXiv:2510.24358

    work page arXiv 2025

  21. [33]

    Google DeepMind . 2025. Gemini 3.1 pro. https://deepmind.google/models/gemini/pro/. Accessed: 2026

    2025

  22. [34]

    Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. 2024. Webvoyager: Building an end-to-end web agent with large multimodal models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6864--6890

    2024

  23. [35]

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. 2023. Swe-bench: Can language models resolve real-world github issues? arXiv preprint arXiv:2310.06770

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  24. [36]

    Kimi Team . 2025. https://arxiv.org/abs/2507.04205 Kimi k2: A mixture-of-experts model with 1 trillion total parameters . arXiv preprint arXiv:2507.04205

    work page arXiv 2025

  25. [37]

    Zhiyu Lin, Zhengda Zhou, Zhiyuan Zhao, Tianrui Wan, Yilun Ma, Junyu Gao, and Xuelong Li. 2025. Webuibench: a comprehensive benchmark for evaluating multimodal large language models in webui-to-code. In Findings of the Association for Computational Linguistics: ACL 2025, pages 15780--15797

    2025

  26. [38]

    Zimu Lu, Yunqiao Yang, Houxing Ren, Haotian Hou, Han Xiao, Ke Wang, Weikang Shi, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. 2025. Webgen-bench: Evaluating llms on generating interactive and functional websites from scratch. arXiv preprint arXiv:2505.03733

    work page arXiv 2025

  27. [39]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-refine: Iterative refinement with self-feedback. In NeurIPS

    2023

  28. [40]

    MiniMax . 2025. Minimax-m2.7. https://huggingface.co/MiniMaxAI/MiniMax-M2.7. Accessed: 2026

    2025

  29. [41]

    OpenAI . 2025. Introducing gpt-5.4. https://openai.com/index/introducing-gpt-5-4/. Accessed: 2026

    2025

  30. [42]

    Qwen Team . 2025. https://arxiv.org/abs/2505.09388 Qwen3 technical report . arXiv preprint arXiv:2505.09388

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [43]

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. In NeurIPS

    2023

  32. [44]

    Chenglei Si, Yanzhe Zhang, Ryan Li, Zhengyuan Yang, Ruibo Liu, and Diyi Yang. 2025. Design2code: Benchmarking multimodal code generation for automated front-end engineering. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pag...

    2025

  33. [45]

    Team GLM . 2024. https://arxiv.org/abs/2406.12793 Chatglm: A family of large language models from glm-130b to glm-4 all tools . arXiv preprint arXiv:2406.12793

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [46]

    OpenHands: An Open Platform for AI Software Developers as Generalist Agents

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H. Tran, Fuqiang Li, Ren Ma, Mingzhang Zheng, Bill Qian, Yanjun Shao, Niklas Muennighoff, Yizhe Zhang, Binyuan Hui, and 5 others. 2024. https://arxiv.org/abs/2407.16741 Openhands: An open platform for ai software developer...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [47]

    Kai Xu, YiWei Mao, XinYi Guan, and ZiLong Feng. 2025. Web-bench: A llm code benchmark based on web standards and frameworks. arXiv preprint arXiv:2505.07473

    work page arXiv 2025

  36. [48]

    John Yang, Carlos E Jimenez, Alex L Zhang, Kilian Lieret, Joyce Yang, Xindi Wu, Ori Press, Niklas Muennighoff, Gabriel Synnaeve, Karthik R Narasimhan, and 1 others. 2024. Swe-bench multimodal: Do ai systems generalize to visual software domains? arXiv preprint arXiv:2410.03859

    work page arXiv 2024

  37. [49]

    Sukmin Yun, Haokun Lin, Rusiru Thushara, Mohammad Q Bhat, Yongxin Wang, Zutao Jiang, Mingkai Deng, Jinhong Wang, Tianhua Tao, Junbo Li, and 1 others. 2024. Web2code: A large-scale webpage-to-code dataset and evaluation framework for multimodal llms. Advances in neural information processing systems, 37:112134--112157

    2024

  38. [50]

    Chenchen Zhang, Yuhang Li, Can Xu, and 1 others. 2025. ArtifactsBench : Bridging the visual-interactive gap in LLM code generation evaluation. In arXiv preprint arXiv:2507.04952

    work page arXiv 2025

  39. [51]

    Hongda Zhu, Yiwen Zhang, Bing Zhao, Jingzhe Ding, Siyao Liu, Tong Liu, Dandan Wang, Yanan Liu, and Zhaojian Li. 2025. Frontendbench: A benchmark for evaluating llms on front-end development via automatic evaluation. arXiv preprint arXiv:2506.13832

    work page arXiv 2025