Paper review 2026-07-24: 25 candidates

[github-actions]

Reply with commands, one per line:
/approve all · /approve 1,3-5 · /reject 2 · /edit 3 category=world_terminal tags=benchmark (edit implies approve; tags=- clears tags). Valid category keys: see taxonomy.json.

1. Improving LLM-Driven Test Generation by Learning from Mocking Information

Jamie Lee, Flynn Teh, Hengcheng Zhu, et al. · arXiv 2026/04 · paper
proposed: software_testing · tags: none

MOCKMILL is an LLM-based tool that generates unit tests by exploiting developer-written mocking information from existing test suites, using iterative generation-and-repair to ensure executable tests. Evaluation on Java projects shows improved coverage and mutation kills over baselines.
_{Proactive unit test generation tool, matching software_testing's test-generation scope.}

2. Symbolon: Symbolic Execution by Learning Code Transformation

Jie Zhu, Penghui Li, Zhongxuan Li, et al. · arXiv 2026/07 · paper
proposed: software_testing · tags: none

Symbolon learns code transformations offline, distills them into reusable agent skills, and uses an agent to apply them to repo-level targets to improve symbolic execution scalability. It boosts KLEE's coverage substantially and uncovers 21 new Linux kernel bugs.
_{An agentic tool improving proactive defect-search via symbolic execution, fitting software_testing's verification/fuzzing scope.}

3. Context Matters: Improving the Practical Reliability of LLM-Based Unit Test Generation

Junjie Chen, Ziqi Wang, Lin Yang, et al. · arXiv 2026/07 · paper
proposed: software_testing · tags: none

CATGen is a context-aware, multi-stage workflow for LLM-based unit test generation that combines structured context retrieval, deterministic scaffolding, and static-analysis post-processing to improve compilation success and coverage. It is evaluated on industrial projects and Defects4J.
_{Proactive unit test generation to assure correctness independent of a reported defect fits software_testing.}

4. Towards Automated Formal Verification of zkEVMs Using LLM-Guided Constraint Synthesis

Shichen Huang, Zhenghe Jiang, Yi Jiang, et al. · arXiv 2026/07 · paper
proposed: software_testing · tags: none

VeriSynth is an LLM-guided framework that synthesizes executable Python/Z3 formal verification models from Rust zkEVM opcode handler code, using SMT solving as a correctness arbiter within a closed-loop repair pipeline to detect subtle implementation bugs.
_{Formal verification of software correctness via LLM-generated verification models falls under software_testing's formal-verification scope.}

5. Case study: solving P-99 with LPTP and an LLM

Fred Mesnard, Thierry Marianne, Étienne Payet, et al. · arXiv 2026/07 · paper
proposed: software_testing · tags: none

A case study using Claude to generate Prolog code and tests for 33 P-99 exercises, then formally prove type, termination, and correctness properties using the LPTP theorem prover. Demonstrates 'vericoding': LLM-generated code paired with machine-checked formal proofs.
_{Formal verification of generated software correctness falls under software_testing's proof-assisted correctness scope.}

6. From Resource Flow to Executable Tests: Petri-Net-Guided LLM Test Generation for Concurrent Stateful Rust APIs

Kaiwen Zhang, Guanjun Liu · arXiv 2026/07 · paper
proposed: software_testing · tags: none

Proposes a Petri-net-guided methodology that derives concurrent state scenarios for stateful Rust APIs and uses them to constrain LLM-based synthesis of executable concurrency tests, with a repair loop and semantic oracle to ensure fidelity.
_{Proactive test generation for correctness of concurrent APIs, independent of any specific reported defect, fits software_testing.}

7. EvoSQL: Memory-Augmented Critic-Generator Co-Evolution for Text-to-SQL

Jiawei Zhou, Jianwei Wang, Chenyu Zhou, et al. · arXiv 2026/06 · paper
proposed: database · tags: none

EvoSQL is a co-evolution framework pairing a SQL generator with a critic, using execution feedback and a contextualized candidate memory to iteratively refine Text-to-SQL outputs. It also introduces an SDPO fine-tuning stage and shows gains on Spider and BIRD benchmarks.
_{The task's deliverable is a SQL query, so it routes to the database artifact leaf.}

8. Demonstrating GenDB: Instance-Optimized and Customized Query Processing Code Generation via LLM Agents

Jiale Lao, Immanuel Trummer · arXiv 2026/07 · paper
proposed: database · tags: none

GenDB is a generative query engine where LLM agents produce instance-optimized query execution code tailored to specific data, workloads, and hardware, replacing hand-engineered DBMS internals. A demo lets users explore the workload analysis, code generation, and optimization pipeline on TPC-H and a custom benchmark.
_{The agent's ultimate deliverable is database query execution code, placing it in the database artifact leaf.}

9. Bespoke OLAP: Synthesizing Workload-Specific One-size-fits-one Database Engines

Johannes Wehrstein, Timo Eckmann, Matthias Jasny, et al. · arXiv 2026/07 · paper
proposed: database · tags: none

Bespoke OLAP is a fully autonomous LLM-driven synthesis pipeline that generates workload-specific database engines from scratch, using iterative performance feedback and validation to guide storage and execution design. It achieves order-of-magnitude speedups over general-purpose engines like DuckDB.
_{An autonomous system that produces database-engine code specialized to a workload, so the deliverable is database-centric program code.}

10. Flint: A Semantics-Driven Data Visualization Intermediate Language

Yunhai Wang, Kecheng Lu, Junhao Chen, et al. · arXiv 2026/07 · paper
proposed: graphics · tags: none

Flint is a semantics-driven intermediate language that compiles concise structural data specifications into optimized, library-agnostic visualization configurations, translating them into executable Vega-Lite, ECharts, or Chart.js code. It is designed as an authoring target usable by both humans and AI agents generating visualizations.
_{Serves chart-from-spec visualization code generation, matching the graphics leaf's definition of producing rendering code from a specification.}

11. JAXBench: Benchmarking Autonomous TPU Kernel Optimization

Arya Tschand, Charles Hong, Julian Walker, et al. · arXiv 2026/05 · paper
proposed: systems · tags: benchmark

Introduces JAXBench, a TPU-native benchmark of 50 JAX/Pallas kernel workloads for evaluating AI-driven kernel optimization on Google Cloud TPUs. Finds that curated TPU documentation and search-based methods substantially improve correctness and speedup over XLA baselines.
_{Benchmark for autonomous generation/optimization of low-level kernel code, matching the systems leaf (KernelBench-style routing).}

12. MKEvolve: A Modular Multi-Agent Framework for Kernel Code Generation

Jason Yoo, Rajarshi Saha, Shaowei Zhu, et al. · arXiv 2026/06 · paper
proposed: systems · tags: none

MKEvolve is a multi-agent framework that co-evolves modular decompositions of PyTorch modules with LLM-generated Triton kernels for each submodule, using beam search per subkernel. It improves correctness and speedup over end-to-end synthesis on KernelBench while cutting token usage.
_{The deliverable is systems-level GPU kernel code, matching the systems leaf per KernelBench precedent.}

13. Multi-turn RL with Structural and Performance Aware Rewards for CUDA Kernel Generation

Quazi Ishtiaque Mahmud, Nesreen K. Ahmed, Ali Jannesari · arXiv 2026/07 · paper
proposed: systems · tags: training-data

CudaPerf is a reflective multi-turn RL framework for CUDA kernel generation that combines verifiable execution rewards with structural code-aware rewards (memory coalescing, occupancy, arithmetic intensity) and iterative refinement, released alongside a dataset of ~3.9k C/PyTorch-to-CUDA program pairs. It achieves large speedup and correctness gains over strong baselines on C-to-CUDA and PyTorch-to-CUDA transformation benchmarks.
_{Produces optimized GPU kernel code, matching the systems leaf's scope of systems-level code generation, with a released training dataset.}

14. Symbol and Footprint Database for Electronic Components by Agentic Recognition and Generation

Yichen Shi, Yuzhi Liu, Zhuofu Tao, et al. · arXiv 2026/07 · paper
proposed: hardware · tags: none

SFgen is an agentic MLLM-based recognition and generation pipeline that automatically creates symbol and footprint representations for electronic components, achieving 86% and 80% accuracy respectively, used to build the growing SFnet component database for PCB design.
_{An agent that generates hardware design artifacts (component symbols/footprints) for PCB design, matching the hardware artifact leaf.}

15. EvoDRC: A Self-Evolving Agentic Framework for Automated DRC Violation Repair

Bing-Yue Wu, Chia-Tung Ho, Haoyu Yang, et al. · arXiv 2026/07 · paper
proposed: hardware · tags: none

EvoDRC is a self-evolving agentic framework where LLM repair agents fix design rule check (DRC) violations in block-level chip layouts, using local analysis tools and evolving repair skills from experience.
_{The agent repairs physical chip design layout code to close DRC violations, a hardware design task under artifact/hardware.}

16. Ripple: An Open, AI-Formalized Lean 4 Framework for Computing with CRNs

Ho-Lin Chen, Xiang Huang · arXiv 2026/07 · paper
proposed: world_research · tags: none

Ripple is a machine-checked Lean 4 framework unifying results on chemical reaction network (CRN) computation, majority algorithms, and Turing-completeness proofs, produced predominantly by AI agents. It repaired gaps in published proofs and surfaced a new open problem in the mathematics of CRNs.
_{An AI-agent-driven Lean formalization whose purpose is mathematical knowledge, not software, matches world_research per the theorem-proving boundary.}

17. Euclean: Automated Geometry Problem Formalization with Unified Verification in Lean

Linbin Tang, Jingyan You, Zilin Kang, et al. · arXiv 2026/06 · paper
proposed: world_research · tags: training-data

Euclean is a four-stage agentic pipeline that autoformalizes geometry problems into native Mathlib Lean, producing the large OMNI-Geometry and Numina-Geometry datasets and improving downstream theorem-proving accuracy.
_{Autoformalization for mathematical theorem proving serves mathematical knowledge, so it routes to world_research per the pure-math boundary, with training-data tag for the released formalization datasets.}

18. Self-Evolving Recommendation System: End-To-End Autonomous Model Optimization With LLM Agents

Haochen Wang, Yi Wu, Daryl Chang, et al. · arXiv 2026/07 · paper
proposed: world_research · tags: none

A self-evolving system where LLM agents act as machine learning engineers, autonomously generating, training, and deploying recommendation model improvements validated against production business metrics at YouTube.
_{Agents conduct machine-learning engineering through interaction (training/tuning models), matching the world_research MLE-bench-style definition where the model is the product of experimental interaction.}

19. OPTScientist: Multi-Agent Discovery of Typed Optimizer Programs for Transformer Pretraining

Zhongzheng Li, Tiancan Feng, Wenhao Li, et al. · arXiv 2026/06 · paper
proposed: world_research · tags: none

OPTScientist is a multi-agent framework where Theorist, Designer, Engineer, and Reviewer agents collaboratively search a typed DSL to discover novel optimizer programs for transformer pretraining, combining evolutionary search with compiler validation. It discovers RS-MR, a reduced-state matrix optimizer that improves pretraining over baselines.
_{The agent writes and executes code as the instrument to pursue an ML/scientific discovery (a better optimizer), matching world_research's MLE/AI-scientist pattern rather than a code-artifact deliverable.}

20. LeanFlow: A Case Study in Workflow-Driven Lean Autoformalization

Lazar Milikic, Simon Guilloud, Khanh Nguyen, et al. · ICML 2026 · paper
proposed: world_research · tags: none

LeanFlow is an LLM agent system for translating unformalized mathematical papers into buildable Lean projects, studied via ablations of model, workflow, and toolset on two case-study papers. It also reports strong results on RLM25 and an ICML AI-for-Math challenge.
_{Autoformalization for mathematical knowledge is theorem-proving work, which the taxonomy routes to world_research.}

21. Scaling Closed-Loop Feature Channel Configuration with LLMs

Tolgay Atinc Uzun, Radu Timofte, Dmitry Ignatov · arXiv 2026/07 · paper
proposed: world_research · tags: none

This study scales closed-loop LLM-based neural-architecture channel-width search to 250 candidates per cycle, generating and evaluating 2000 executable network configurations on CIFAR-100. It finds improving accuracy trends, better parameter efficiency, and emergent architectural regularities like non-power-of-two channel widths.
_{Code is the instrument for an ML-engineering search whose product is a trained/optimized model, matching world_research's MLE-bench-style scope.}

22. Autonomous disproofs of the sum-product conjecture over $\mathbb R$ with GPT-5.5 Pro

Yichen Huang · arXiv 2026/07 · paper
proposed: world_research · tags: none

The authors build a simple three-stage prompting agent on GPT-5.5 Pro that autonomously generates proofs disproving the Erdős–Szemerédi sum-product conjecture over R, succeeding in 7 of 8 trials. They release code, intermediate outputs, and the generated proofs as a reproducible case study.
_{An autonomous agent pursuing mathematical discovery/theorem-proving belongs to world_research per the taxonomy's explicit Lean/proof example.}

23. Scalable Low-Cost Laboratory Automation: A Digital Twin-Integrated Robotic Platform for Autonomous Liquid Handling (RAINBOTTM)

Mohamed Rami Ayeche, Souhil Sid, Ahyen Mostofa, et al. · arXiv 2026/07 · paper
proposed: world_research · tags: none

RAINBOT is a low-cost, G-code/Python-controlled liquid-handling robot built from a 3D printer, paired with a digital-twin web interface for remote supervision and coupled to an inverse-design search framework for autonomous experimentation. It demonstrates automated liquid mixing validated against expected chemical behavior.
_{Code drives physical lab automation toward goal-directed scientific discovery, matching the self-driving-lab example under world_research.}

24. Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks

Mack Nixon, Liam Wright, Yevgeniya Kovalchuk, et al. · arXiv 2026/07 · paper
proposed: world_research · tags: benchmark

Introduces an open-source framework and benchmark evaluating open-weight LLM agents that write R code to perform longitudinal cohort-study data preparation tasks like harmonization and merging. Locally deployable 31-35B models nearly saturate the benchmark, enabling AI-assisted data prep under governance restrictions.
_{An LLM agent uses code as a tool to prepare and clean research data, matching data-science agent tasks in world_research.}

25. OpenSkillRisk: Benchmarking Agent Safety When Using Real-World Risky Third-Party Skills

Qiyuan Liu, Tingfeng Hui, Kun Zhan, et al. · arXiv 2026/07 · paper
proposed: world_terminal · tags: benchmark

OpenSkillRisk is a safety benchmark of 263 real-world risky third-party skills used by CLI agent frameworks, evaluating whether LLM-based agents recognize and avoid latent risks during execution. It tests thirteen state-of-the-art LLMs across three mainstream CLI agent systems, revealing frequent unsafe executions.
_{Evaluates CLI/shell-based agent frameworks acting through executable skills, making the terminal the agent's world of action.}

machine payload (do not edit)

[{"paper": {"id": "2604.19315", "title": "Improving LLM-Driven Test Generation by Learning from Mocking Information", "authors": ["Jamie Lee", "Flynn Teh", "Hengcheng Zhu", "Mengzhen Li", "Mattia Fazzini", "Valerio Terragni"], "venue": "arXiv 2026/04", "category": "", "published": "2026-04-21", "links": {"paper": "https://arxiv.org/abs/2604.19315", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "MOCKMILL is an LLM-based tool that generates unit tests by exploiting developer-written mocking information from existing test suites, using iterative generation-and-repair to ensure executable tests. Evaluation on Java projects shows improved coverage and mutation kills over baselines.", "reason": "Proactive unit test generation tool, matching software_testing's test-generation scope.", "source": "crawl"}, {"paper": {"id": "2606.29108", "title": "Symbolon: Symbolic Execution by Learning Code Transformation", "authors": ["Jie Zhu", "Penghui Li", "Zhongxuan Li", "Chihao Shen", "Ziyang Li", "Yizheng Chen", "Kexin Pei"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2606.29108", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "Symbolon learns code transformations offline, distills them into reusable agent skills, and uses an agent to apply them to repo-level targets to improve symbolic execution scalability. It boosts KLEE's coverage substantially and uncovers 21 new Linux kernel bugs.", "reason": "An agentic tool improving proactive defect-search via symbolic execution, fitting software_testing's verification/fuzzing scope.", "source": "crawl"}, {"paper": {"id": "2607.19682", "title": "Context Matters: Improving the Practical Reliability of LLM-Based Unit Test Generation", "authors": ["Junjie Chen", "Ziqi Wang", "Lin Yang", "Chen Yang", "Xiao Chu", "Jianyi Zhou", "Guangtai Liang", "Qianxiang Wang", "Dong Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.19682", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "CATGen is a context-aware, multi-stage workflow for LLM-based unit test generation that combines structured context retrieval, deterministic scaffolding, and static-analysis post-processing to improve compilation success and coverage. It is evaluated on industrial projects and Defects4J.", "reason": "Proactive unit test generation to assure correctness independent of a reported defect fits software_testing.", "source": "crawl"}, {"paper": {"id": "2607.19795", "title": "Towards Automated Formal Verification of zkEVMs Using LLM-Guided Constraint Synthesis", "authors": ["Shichen Huang", "Zhenghe Jiang", "Yi Jiang", "Ling-I Wu", "Jingyang Li", "Guoqiang Li"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.19795", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "VeriSynth is an LLM-guided framework that synthesizes executable Python/Z3 formal verification models from Rust zkEVM opcode handler code, using SMT solving as a correctness arbiter within a closed-loop repair pipeline to detect subtle implementation bugs.", "reason": "Formal verification of software correctness via LLM-generated verification models falls under software_testing's formal-verification scope.", "source": "crawl"}, {"paper": {"id": "2607.21196", "title": "Case study: solving P-99 with LPTP and an LLM", "authors": ["Fred Mesnard", "Thierry Marianne", "Étienne Payet", "Wim Vanhoof"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.21196", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "A case study using Claude to generate Prolog code and tests for 33 P-99 exercises, then formally prove type, termination, and correctness properties using the LPTP theorem prover. Demonstrates 'vericoding': LLM-generated code paired with machine-checked formal proofs.", "reason": "Formal verification of generated software correctness falls under software_testing's proof-assisted correctness scope.", "source": "crawl"}, {"paper": {"id": "2607.21530", "title": "From Resource Flow to Executable Tests: Petri-Net-Guided LLM Test Generation for Concurrent Stateful Rust APIs", "authors": ["Kaiwen Zhang", "Guanjun Liu"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.21530", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "Proposes a Petri-net-guided methodology that derives concurrent state scenarios for stateful Rust APIs and uses them to constrain LLM-based synthesis of executable concurrency tests, with a repair loop and semantic oracle to ensure fidelity.", "reason": "Proactive test generation for correctness of concurrent APIs, independent of any specific reported defect, fits software_testing.", "source": "crawl"}, {"paper": {"id": "2607.20489", "title": "EvoSQL: Memory-Augmented Critic-Generator Co-Evolution for Text-to-SQL", "authors": ["Jiawei Zhou", "Jianwei Wang", "Chenyu Zhou", "Chaojian Shi", "Ming Dong", "Kai Wang"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-04", "links": {"paper": "https://arxiv.org/abs/2607.20489", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "EvoSQL is a co-evolution framework pairing a SQL generator with a critic, using execution feedback and a contextualized candidate memory to iteratively refine Text-to-SQL outputs. It also introduces an SDPO fine-tuning stage and shows gains on Spider and BIRD benchmarks.", "reason": "The task's deliverable is a SQL query, so it routes to the database artifact leaf.", "source": "crawl"}, {"paper": {"id": "2607.20630", "title": "Demonstrating GenDB: Instance-Optimized and Customized Query Processing Code Generation via LLM Agents", "authors": ["Jiale Lao", "Immanuel Trummer"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20630", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "GenDB is a generative query engine where LLM agents produce instance-optimized query execution code tailored to specific data, workloads, and hardware, replacing hand-engineered DBMS internals. A demo lets users explore the workload analysis, code generation, and optimization pipeline on TPC-H and a custom benchmark.", "reason": "The agent's ultimate deliverable is database query execution code, placing it in the database artifact leaf.", "source": "crawl"}, {"paper": {"id": "2603.02001", "title": "Bespoke OLAP: Synthesizing Workload-Specific One-size-fits-one Database Engines", "authors": ["Johannes Wehrstein", "Timo Eckmann", "Matthias Jasny", "Carsten Binnig"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2603.02001", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "Bespoke OLAP is a fully autonomous LLM-driven synthesis pipeline that generates workload-specific database engines from scratch, using iterative performance feedback and validation to guide storage and execution design. It achieves order-of-magnitude speedups over general-purpose engines like DuckDB.", "reason": "An autonomous system that produces database-engine code specialized to a workload, so the deliverable is database-centric program code.", "source": "crawl"}, {"paper": {"id": "2607.20775", "title": "Flint: A Semantics-Driven Data Visualization Intermediate Language", "authors": ["Yunhai Wang", "Kecheng Lu", "Junhao Chen", "Alper Sarikaya", "Chenglong Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20775", "github": "", "website": ""}}, "category": "graphics", "tags": [], "summary": "Flint is a semantics-driven intermediate language that compiles concise structural data specifications into optimized, library-agnostic visualization configurations, translating them into executable Vega-Lite, ECharts, or Chart.js code. It is designed as an authoring target usable by both humans and AI agents generating visualizations.", "reason": "Serves chart-from-spec visualization code generation, matching the graphics leaf's definition of producing rendering code from a specification.", "source": "crawl"}, {"paper": {"id": "2607.20466", "title": "JAXBench: Benchmarking Autonomous TPU Kernel Optimization", "authors": ["Arya Tschand", "Charles Hong", "Julian Walker", "Nina Cai", "Shangkun Wang", "Suvinay Subramanian", "Sundar Dev", "Vijay Janapa Reddi", "Amir Yazdanbakhsh", "Sethu Sankaran"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-19", "links": {"paper": "https://arxiv.org/abs/2607.20466", "github": "", "website": ""}}, "category": "systems", "tags": ["benchmark"], "summary": "Introduces JAXBench, a TPU-native benchmark of 50 JAX/Pallas kernel workloads for evaluating AI-driven kernel optimization on Google Cloud TPUs. Finds that curated TPU documentation and search-based methods substantially improve correctness and speedup over XLA baselines.", "reason": "Benchmark for autonomous generation/optimization of low-level kernel code, matching the systems leaf (KernelBench-style routing).", "source": "crawl"}, {"paper": {"id": "2607.20501", "title": "MKEvolve: A Modular Multi-Agent Framework for Kernel Code Generation", "authors": ["Jason Yoo", "Rajarshi Saha", "Shaowei Zhu", "Tao Yu", "Wei Tang", "Youngsuk Park"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-20", "links": {"paper": "https://arxiv.org/abs/2607.20501", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "MKEvolve is a multi-agent framework that co-evolves modular decompositions of PyTorch modules with LLM-generated Triton kernels for each submodule, using beam search per subkernel. It improves correctness and speedup over end-to-end synthesis on KernelBench while cutting token usage.", "reason": "The deliverable is systems-level GPU kernel code, matching the systems leaf per KernelBench precedent.", "source": "crawl"}, {"paper": {"id": "2607.20908", "title": "Multi-turn RL with Structural and Performance Aware Rewards for CUDA Kernel Generation", "authors": ["Quazi Ishtiaque Mahmud", "Nesreen K. Ahmed", "Ali Jannesari"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.20908", "github": "", "website": ""}}, "category": "systems", "tags": ["training-data"], "summary": "CudaPerf is a reflective multi-turn RL framework for CUDA kernel generation that combines verifiable execution rewards with structural code-aware rewards (memory coalescing, occupancy, arithmetic intensity) and iterative refinement, released alongside a dataset of ~3.9k C/PyTorch-to-CUDA program pairs. It achieves large speedup and correctness gains over strong baselines on C-to-CUDA and PyTorch-to-CUDA transformation benchmarks.", "reason": "Produces optimized GPU kernel code, matching the systems leaf's scope of systems-level code generation, with a released training dataset.", "source": "crawl"}, {"paper": {"id": "2607.19767", "title": "Symbol and Footprint Database for Electronic Components by Agentic Recognition and Generation", "authors": ["Yichen Shi", "Yuzhi Liu", "Zhuofu Tao", "Li Huang", "Yuhao Gao", "Ting-Jung Lin", "Lei Hel"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.19767", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "SFgen is an agentic MLLM-based recognition and generation pipeline that automatically creates symbol and footprint representations for electronic components, achieving 86% and 80% accuracy respectively, used to build the growing SFnet component database for PCB design.", "reason": "An agent that generates hardware design artifacts (component symbols/footprints) for PCB design, matching the hardware artifact leaf.", "source": "crawl"}, {"paper": {"id": "2607.20019", "title": "EvoDRC: A Self-Evolving Agentic Framework for Automated DRC Violation Repair", "authors": ["Bing-Yue Wu", "Chia-Tung Ho", "Haoyu Yang", "Brucek Khailany", "Vidya A. Chhabria"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20019", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "EvoDRC is a self-evolving agentic framework where LLM repair agents fix design rule check (DRC) violations in block-level chip layouts, using local analysis tools and evolving repair skills from experience.", "reason": "The agent repairs physical chip design layout code to close DRC violations, a hardware design task under artifact/hardware.", "source": "crawl"}, {"paper": {"id": "2607.13531", "title": "Ripple: An Open, AI-Formalized Lean 4 Framework for Computing with CRNs", "authors": ["Ho-Lin Chen", "Xiang Huang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.13531", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "Ripple is a machine-checked Lean 4 framework unifying results on chemical reaction network (CRN) computation, majority algorithms, and Turing-completeness proofs, produced predominantly by AI agents. It repaired gaps in published proofs and surfaced a new open problem in the mathematics of CRNs.", "reason": "An AI-agent-driven Lean formalization whose purpose is mathematical knowledge, not software, matches world_research per the theorem-proving boundary.", "source": "crawl"}, {"paper": {"id": "2607.19374", "title": "Euclean: Automated Geometry Problem Formalization with Unified Verification in Lean", "authors": ["Linbin Tang", "Jingyan You", "Zilin Kang", "Hanzhang Liu", "Sophia Zhang", "Zenan Li", "Chenrui Cao", "Liangcheng Song", "Jiaao Wu", "Xian Zhang", "Fan Yang"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-17", "links": {"paper": "https://arxiv.org/abs/2607.19374", "github": "", "website": ""}}, "category": "world_research", "tags": ["training-data"], "summary": "Euclean is a four-stage agentic pipeline that autoformalizes geometry problems into native Mathlib Lean, producing the large OMNI-Geometry and Numina-Geometry datasets and improving downstream theorem-proving accuracy.", "reason": "Autoformalization for mathematical theorem proving serves mathematical knowledge, so it routes to world_research per the pure-math boundary, with training-data tag for the released formalization datasets.", "source": "crawl"}, {"paper": {"id": "2602.10226", "title": "Self-Evolving Recommendation System: End-To-End Autonomous Model Optimization With LLM Agents", "authors": ["Haochen Wang", "Yi Wu", "Daryl Chang", "Li Wei", "Lukasz Heldt"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2602.10226", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "A self-evolving system where LLM agents act as machine learning engineers, autonomously generating, training, and deploying recommendation model improvements validated against production business metrics at YouTube.", "reason": "Agents conduct machine-learning engineering through interaction (training/tuning models), matching the world_research MLE-bench-style definition where the model is the product of experimental interaction.", "source": "crawl"}, {"paper": {"id": "2607.20486", "title": "OPTScientist: Multi-Agent Discovery of Typed Optimizer Programs for Transformer Pretraining", "authors": ["Zhongzheng Li", "Tiancan Feng", "Wenhao Li", "Qingsong Ran", "Shikun Feng", "Xiaoyuan Zhang", "Yue Wang", "Xiaoguang Zhao"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-02", "links": {"paper": "https://arxiv.org/abs/2607.20486", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "OPTScientist is a multi-agent framework where Theorist, Designer, Engineer, and Reviewer agents collaboratively search a typed DSL to discover novel optimizer programs for transformer pretraining, combining evolutionary search with compiler validation. It discovers RS-MR, a reduced-state matrix optimizer that improves pretraining over baselines.", "reason": "The agent writes and executes code as the instrument to pursue an ML/scientific discovery (a better optimizer), matching world_research's MLE/AI-scientist pattern rather than a code-artifact deliverable.", "source": "crawl"}, {"paper": {"id": "2607.20503", "title": "LeanFlow: A Case Study in Workflow-Driven Lean Autoformalization", "authors": ["Lazar Milikic", "Simon Guilloud", "Khanh Nguyen", "Viktor Kuncak"], "venue": "ICML 2026", "category": "", "published": "2026-06-26", "links": {"paper": "https://arxiv.org/abs/2607.20503", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "LeanFlow is an LLM agent system for translating unformalized mathematical papers into buildable Lean projects, studied via ablations of model, workflow, and toolset on two case-study papers. It also reports strong results on RLM25 and an ICML AI-for-Math challenge.", "reason": "Autoformalization for mathematical knowledge is theorem-proving work, which the taxonomy routes to world_research.", "source": "crawl"}, {"paper": {"id": "2607.20516", "title": "Scaling Closed-Loop Feature Channel Configuration with LLMs", "authors": ["Tolgay Atinc Uzun", "Radu Timofte", "Dmitry Ignatov"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-07", "links": {"paper": "https://arxiv.org/abs/2607.20516", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "This study scales closed-loop LLM-based neural-architecture channel-width search to 250 candidates per cycle, generating and evaluating 2000 executable network configurations on CIFAR-100. It finds improving accuracy trends, better parameter efficiency, and emergent architectural regularities like non-power-of-two channel widths.", "reason": "Code is the instrument for an ML-engineering search whose product is a trained/optimized model, matching world_research's MLE-bench-style scope.", "source": "crawl"}, {"paper": {"id": "2607.20525", "title": "Autonomous disproofs of the sum-product conjecture over $\\mathbb R$ with GPT-5.5 Pro", "authors": ["Yichen Huang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-09", "links": {"paper": "https://arxiv.org/abs/2607.20525", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "The authors build a simple three-stage prompting agent on GPT-5.5 Pro that autonomously generates proofs disproving the Erdős–Szemerédi sum-product conjecture over R, succeeding in 7 of 8 trials. They release code, intermediate outputs, and the generated proofs as a reproducible case study.", "reason": "An autonomous agent pursuing mathematical discovery/theorem-proving belongs to world_research per the taxonomy's explicit Lean/proof example.", "source": "crawl"}, {"paper": {"id": "2607.20662", "title": "Scalable Low-Cost Laboratory Automation: A Digital Twin-Integrated Robotic Platform for Autonomous Liquid Handling (RAINBOTTM)", "authors": ["Mohamed Rami Ayeche", "Souhil Sid", "Ahyen Mostofa", "Rehaan Hussain", "Ali Shayesteh", "Fadwa El Mellouhi"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20662", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "RAINBOT is a low-cost, G-code/Python-controlled liquid-handling robot built from a 3D printer, paired with a digital-twin web interface for remote supervision and coupled to an inverse-design search framework for autonomous experimentation. It demonstrates automated liquid mixing validated against expected chemical behavior.", "reason": "Code drives physical lab automation toward goal-directed scientific discovery, matching the self-driving-lab example under world_research.", "source": "crawl"}, {"paper": {"id": "2607.21482", "title": "Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks", "authors": ["Mack Nixon", "Liam Wright", "Yevgeniya Kovalchuk", "Alison Fang-Wei Wu", "Martin Danka", "Andy Boyd", "David Bann"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.21482", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "Introduces an open-source framework and benchmark evaluating open-weight LLM agents that write R code to perform longitudinal cohort-study data preparation tasks like harmonization and merging. Locally deployable 31-35B models nearly saturate the benchmark, enabling AI-assisted data prep under governance restrictions.", "reason": "An LLM agent uses code as a tool to prepare and clean research data, matching data-science agent tasks in world_research.", "source": "crawl"}, {"paper": {"id": "2607.20121", "title": "OpenSkillRisk: Benchmarking Agent Safety When Using Real-World Risky Third-Party Skills", "authors": ["Qiyuan Liu", "Tingfeng Hui", "Kun Zhan", "Kaike Zhang", "Ning Miao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.20121", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "OpenSkillRisk is a safety benchmark of 263 real-world risky third-party skills used by CLI agent frameworks, evaluating whether LLM-based agents recognize and avoid latent risks during execution. It tests thirteen state-of-the-art LLMs across three mainstream CLI agent systems, revealing frequent unsafe executions.", "reason": "Evaluates CLI/shell-based agent frameworks acting through executable skills, making the terminal the agent's world of action.", "source": "crawl"}]