Paper review 2026-07-31: 16 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. GVR-Coder: A Visual-Feedback Framework for Structured SVG Generation in Complex Document and Meeting Scenarios

Yiming Xu, Jihua Kang, Chunsai Du, et al. · arXiv 2026/07 · paper
proposed: graphics · tags: none

GVR-Coder is a Text-to-SVG framework with curriculum fine-tuning, RL from dual rendering feedback, and a generate-verify-repair agent loop to produce logically coherent diagrams from long documents. It is paired with a new large-scale dataset, DocMeetSVG-100K.
_{The agent's deliverable is SVG diagram code, matching the graphics artifact leaf.}

2. KernelGenBench: A Multi-Source and Multi-Chip Benchmark for LLM-based Kernel Generation

Peiyu Zang, Jian Tao, Jialing Zhang, et al. · arXiv 2026/07 · paper
proposed: systems · tags: benchmark

KernelGenBench is a multi-source, multi-chip benchmark evaluating LLM- and agent-generated Triton accelerator kernels across operator sources and heterogeneous hardware platforms, revealing cross-platform performance degradation and high generation cost.
_{Benchmark for generating low-level systems/kernel code, so it routes to the systems leaf via benchmark routing.}

3. Orchard: An Open-Source Agentic Modeling Framework

Baolin Peng, Wenlin Yao, Qianhui Wu, et al. · arXiv 2026/07 · paper
proposed: systems · tags: model

Orchard is an open-source agentic modeling framework built on a Kubernetes-native environment service and RL training recipes, producing specialized agents for software engineering (SWE-bench SOTA among open models), GUI computer-use, and personal-assistant tasks.
_{Like the PithTrain precedent, its deliverable is the training/environment system itself (systems-level software), not a single task performed in a world.}

4. Specification-Guided Synthesis of Deadlock-Free Communication Protocol Refinements with Large Language Models

Yang Li, Ping Hou, Nobuko Yoshida · arXiv 2026/07 · paper
proposed: systems · tags: none

Syntropy is an LLM-guided synthesis framework that generates deadlock-free communication protocol refinements using multiparty session type specifications as correctness constraints. It achieves high validity and syntactic correctness while producing diverse protocol variants across multiple LLMs.
_{The deliverable is distributed-systems/network protocol code, matching the systems leaf's 'network protocol implementations' example.}

5. CircuitProver: Agentic Lean 4 Theorem Proving with Reusable Circuit Proof Library for Hardware Verification

Ziyi Yang, Wenji Fang, Chen Chen, et al. · arXiv 2026/07 · paper
proposed: hardware · tags: benchmark

CircuitProver is an agentic Lean 4 framework that translates parameterized hardware designs and specifications into machine-checked proofs, building a reusable proof library and introducing the first benchmark suite for agentic hardware theorem proving.
_{Formal verification of hardware description code via agentic theorem proving falls under hardware's 'hardware verification via code' scope.}

6. CHARGE: Leveraging CWE Hierarchies for Hardware Security SystemVerilog Assertion Generation

Xiao Tan, Cynthia Sturton · arXiv 2026/07 · paper
proposed: hardware · tags: none

CHARGE automatically generates SystemVerilog security assertions for unverified RTL modules by reasoning over CWE hierarchies with LLMs. Evaluated on Hack@DAC SoC designs, it detects known bugs and even finds a previously unidentified vulnerability.
_{Produces hardware verification code (SVAs) for RTL, matching the hardware leaf's 'hardware verification via code' example.}

7. ARES: Adaptive Reasoning-Effort Steering for PPA- and Cost-Aware RTL Optimization with LLM Agents

Stef Cuyckens, Mihaela Jivanescu, Jun Yin, et al. · arXiv 2026/07 · paper
proposed: hardware · tags: none

ARES is an LLM-agent framework for RTL PPA optimization that adaptively escalates reasoning effort per edit-synthesis-analysis iteration based on progress stalls, tracked against a normalized dollar cost. It closes up to 83% of the gap to hand-optimized designs at a fraction of the token cost of prior methods.
_{The served task is producing and optimizing RTL hardware description code, matching the hardware leaf.}

8. LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents

Jingya Wang, Yuyang Gao, Liuzhenghao Lv, et al. · arXiv 2026/07 · paper
proposed: world_research · tags: none

LabEvolver is a training-free framework giving wet-lab robotic agents episodic memory distilled from execution trajectories, combining a grounded trial loop with an evolving skill/safety library. It improves robotic solution-preparation efficiency and safety, and generalizes to ALFWorld tasks.
_{A self-driving-lab style agent performing scientific experimentation through interaction fits world_research's self-driving-labs example.}

9. Albilich: Steerable Proof-State Orchestration for LLM-Based Mathematical Research with CAS Integration

Ting Gong, Michael Ruofan Zeng, Yong Yang · arXiv 2026/07 · paper
proposed: world_research · tags: none

Albilich is an agentic harness for AI-assisted mathematical research combining long-horizon reasoning, computer-algebra-system (CAS) integration, literature retrieval, and persistent context management. It solves RealMath problems and produces novel results on open Kourovka Notebook group-theory problems.
_{An AI-scientist-style research agent using CAS execution to pursue mathematical discovery fits world_research.}

10. Vibe-FDTR: An agent-oriented framework for reproducible frequency-domain thermoreflectance data analysis

Fuwei Yang, Weiheng Li, Bai Song · arXiv 2026/07 · paper
proposed: world_research · tags: none

Vibe-FDTR is an agent framework coupling a configuration-driven code package with LLM agent skills to autonomously perform reproducible frequency-domain thermoreflectance thermal analysis from natural-language requests.
_{An autonomous data-analysis agent for scientific measurement whose deliverable is domain insight, matching world_research.}

11. Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Junlin Yang, Che Jiang, Yu Fu, et al. · arXiv 2026/07 · paper
proposed: world_research · tags: model training-data

Introduces OpenMLE, a full-stack system with verifiable task environments, RL, and long-horizon search for training Frontis-MA1, a 35B model that performs machine-learning-engineering tasks via program-evolution operators (Draft, Improve, Debug, Crossover). It substantially improves MLE-Bench Lite scores over its base model, approaching much larger frontier models.
_{An MLE agent that trains and evolves ML pipelines through execution feedback matches world_research's MLE-engineering agent example.}

12. Open Security Benchmark: Towards Autonomous Enterprise Cyber Defense

Gal Engelberg, Michael Arenzon, Leon Goldberg · arXiv 2026/07 · paper
proposed: world_terminal · tags: benchmark

Open Security Benchmark (OSB) evaluates agentic AI on autonomous enterprise security posture investigation over a frozen, curated enterprise environment, using text-to-SQL and vendor APIs as investigation modalities with ground-truth scoring.
_{Security posture investigation and defense operations across enterprise systems is a system-administration/security-operations task, matching world_terminal.}

13. Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Training

Zhihong Pan, Jiyuan He, Kai Zhang, et al. · arXiv 2026/07 · paper
proposed: world_terminal · tags: training-data

Meta-Task reframes terminal task synthesis itself as a terminal task, using an agent operating inside real containers to iteratively generate, execute, and verify diverse terminal tasks for training data. Fine-tuning on the synthesized trajectories improves Terminal-Bench 2.0 performance with far less data than prior approaches.
_{The synthesized trajectories serve terminal agents acting via shell commands, so it routes to world_terminal with the training-data tag.}

14. ORCA-bench: How Ready Are Language Model Agents for Oncall?

Albert Gong, Kyuseong Choi, Abhineet Agarwal, et al. · arXiv 2026/07 · paper
proposed: world_terminal · tags: benchmark

ORCA-bench evaluates general-purpose coding agents on production-fidelity oncall root-cause-analysis tasks over live telemetry (metrics, logs, traces) and source code across a microservice testbed. Frontier agents achieve only 10-25% accuracy, exposing large gaps for real production reliability work.
_{SRE/oncall incident diagnosis on a live system is DevOps/operations work, matching the world_terminal boundary example of incident mitigation.}

15. Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3

Jens Lehmann, Andrei Aioanei, Sahar Vahdati · arXiv 2026/07 · paper
proposed: world_game · tags: none

Introduces Tycho, a coding-agent system for ARC-AGI-3 that constructs, tests, repairs, or bypasses executable game-specific world models to plan actions efficiently. Evaluated across 25 public games, it achieves near-optimal human-relative action efficiency using orchestration policies over frontier LLMs.
_{The agent uses executable code-based world models as a policy to play and win games, matching world_game's code-as-policy definition.}

16. Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Yash Pandya, Sahil Gupta, Sarthak Harne, et al. · arXiv 2026/07 · paper
proposed: world_apps · tags: training-data benchmark

Echoverse compiles specifications into stateful, login-gated applications with graded tasks and a co-evolution loop that jointly repairs environments and trains computer-use agents. Training on twelve such environments substantially improves a 9B model's accuracy across evaluation splits, and environments are released as a benchmark.
_{Trains and evaluates agents operating stateful software applications, matching world_apps; resource paper follows its served task.}

machine payload (do not edit)

[{"paper": {"id": "2607.28073", "title": "GVR-Coder: A Visual-Feedback Framework for Structured SVG Generation in Complex Document and Meeting Scenarios", "authors": ["Yiming Xu", "Jihua Kang", "Chunsai Du", "Qifan Zhang", "Wangqiu Zhou", "Yiting Wu", "Tianqi Li", "Qi Song"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28073", "github": "", "website": ""}}, "category": "graphics", "tags": [], "summary": "GVR-Coder is a Text-to-SVG framework with curriculum fine-tuning, RL from dual rendering feedback, and a generate-verify-repair agent loop to produce logically coherent diagrams from long documents. It is paired with a new large-scale dataset, DocMeetSVG-100K.", "reason": "The agent's deliverable is SVG diagram code, matching the graphics artifact leaf.", "source": "crawl"}, {"paper": {"id": "2607.27231", "title": "KernelGenBench: A Multi-Source and Multi-Chip Benchmark for LLM-based Kernel Generation", "authors": ["Peiyu Zang", "Jian Tao", "Jialing Zhang", "Yichen Yuan", "Wentao Zhang", "Guang Liu", "Yonghua Lin"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.27231", "github": "", "website": ""}}, "category": "systems", "tags": ["benchmark"], "summary": "KernelGenBench is a multi-source, multi-chip benchmark evaluating LLM- and agent-generated Triton accelerator kernels across operator sources and heterogeneous hardware platforms, revealing cross-platform performance degradation and high generation cost.", "reason": "Benchmark for generating low-level systems/kernel code, so it routes to the systems leaf via benchmark routing.", "source": "crawl"}, {"paper": {"id": "2605.15040", "title": "Orchard: An Open-Source Agentic Modeling Framework", "authors": ["Baolin Peng", "Wenlin Yao", "Qianhui Wu", "Hao Cheng", "Xiao Yu", "Rui Yang", "Tao Ge", "Alessandro Sordoni", "Xingdi Yuan", "Yelong Shen", "Pengcheng He", "Tong Zhang", "Zhou Yu", "Jianfeng Gao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2605.15040", "github": "", "website": ""}}, "category": "systems", "tags": ["model"], "summary": "Orchard is an open-source agentic modeling framework built on a Kubernetes-native environment service and RL training recipes, producing specialized agents for software engineering (SWE-bench SOTA among open models), GUI computer-use, and personal-assistant tasks.", "reason": "Like the PithTrain precedent, its deliverable is the training/environment system itself (systems-level software), not a single task performed in a world.", "source": "crawl"}, {"paper": {"id": "2607.27964", "title": "Specification-Guided Synthesis of Deadlock-Free Communication Protocol Refinements with Large Language Models", "authors": ["Yang Li", "Ping Hou", "Nobuko Yoshida"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.27964", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "Syntropy is an LLM-guided synthesis framework that generates deadlock-free communication protocol refinements using multiparty session type specifications as correctness constraints. It achieves high validity and syntactic correctness while producing diverse protocol variants across multiple LLMs.", "reason": "The deliverable is distributed-systems/network protocol code, matching the systems leaf's 'network protocol implementations' example.", "source": "crawl"}, {"paper": {"id": "2607.27259", "title": "CircuitProver: Agentic Lean 4 Theorem Proving with Reusable Circuit Proof Library for Hardware Verification", "authors": ["Ziyi Yang", "Wenji Fang", "Chen Chen", "Zhiyao Xie", "Hongce Zhang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-29", "links": {"paper": "https://arxiv.org/abs/2607.27259", "github": "", "website": ""}}, "category": "hardware", "tags": ["benchmark"], "summary": "CircuitProver is an agentic Lean 4 framework that translates parameterized hardware designs and specifications into machine-checked proofs, building a reusable proof library and introducing the first benchmark suite for agentic hardware theorem proving.", "reason": "Formal verification of hardware description code via agentic theorem proving falls under hardware's 'hardware verification via code' scope.", "source": "crawl"}, {"paper": {"id": "2607.27776", "title": "CHARGE: Leveraging CWE Hierarchies for Hardware Security SystemVerilog Assertion Generation", "authors": ["Xiao Tan", "Cynthia Sturton"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.27776", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "CHARGE automatically generates SystemVerilog security assertions for unverified RTL modules by reasoning over CWE hierarchies with LLMs. Evaluated on Hack@DAC SoC designs, it detects known bugs and even finds a previously unidentified vulnerability.", "reason": "Produces hardware verification code (SVAs) for RTL, matching the hardware leaf's 'hardware verification via code' example.", "source": "crawl"}, {"paper": {"id": "2607.27879", "title": "ARES: Adaptive Reasoning-Effort Steering for PPA- and Cost-Aware RTL Optimization with LLM Agents", "authors": ["Stef Cuyckens", "Mihaela Jivanescu", "Jun Yin", "Chao Fang", "Marian Verhelst"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.27879", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "ARES is an LLM-agent framework for RTL PPA optimization that adaptively escalates reasoning effort per edit-synthesis-analysis iteration based on progress stalls, tracked against a normalized dollar cost. It closes up to 83% of the gap to hand-optimized designs at a fraction of the token cost of prior methods.", "reason": "The served task is producing and optimizing RTL hardware description code, matching the hardware leaf.", "source": "crawl"}, {"paper": {"id": "2607.27690", "title": "LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents", "authors": ["Jingya Wang", "Yuyang Gao", "Liuzhenghao Lv", "Yonghong Tian", "Yuyang Liu"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.27690", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "LabEvolver is a training-free framework giving wet-lab robotic agents episodic memory distilled from execution trajectories, combining a grounded trial loop with an evolving skill/safety library. It improves robotic solution-preparation efficiency and safety, and generalizes to ALFWorld tasks.", "reason": "A self-driving-lab style agent performing scientific experimentation through interaction fits world_research's self-driving-labs example.", "source": "crawl"}, {"paper": {"id": "2607.27705", "title": "Albilich: Steerable Proof-State Orchestration for LLM-Based Mathematical Research with CAS Integration", "authors": ["Ting Gong", "Michael Ruofan Zeng", "Yong Yang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.27705", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "Albilich is an agentic harness for AI-assisted mathematical research combining long-horizon reasoning, computer-algebra-system (CAS) integration, literature retrieval, and persistent context management. It solves RealMath problems and produces novel results on open Kourovka Notebook group-theory problems.", "reason": "An AI-scientist-style research agent using CAS execution to pursue mathematical discovery fits world_research.", "source": "crawl"}, {"paper": {"id": "2607.28200", "title": "Vibe-FDTR: An agent-oriented framework for reproducible frequency-domain thermoreflectance data analysis", "authors": ["Fuwei Yang", "Weiheng Li", "Bai Song"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28200", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "Vibe-FDTR is an agent framework coupling a configuration-driven code package with LLM agent skills to autonomously perform reproducible frequency-domain thermoreflectance thermal analysis from natural-language requests.", "reason": "An autonomous data-analysis agent for scientific measurement whose deliverable is domain insight, matching world_research.", "source": "crawl"}, {"paper": {"id": "2607.28568", "title": "Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering", "authors": ["Junlin Yang", "Che Jiang", "Yu Fu", "Tianwei Luo", "Can Ren", "Weizhi Wang", "Kaikai Zhao", "Hongyi Liu", "Yuxin Zuo", "Yuru Wang", "Yuchen Fan", "Kai Tian", "Zhenzhao Yuan", "Xiaojian Lin", "Li Sheng", "Rushi Qiang", "Guoli Jia", "Xingtai Lv", "Ermo Hua", "Dianqiao Lei", "Youbang Sun", "Ning Ding", "Bowen Zhou", "Kaiyan Zhang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28568", "github": "", "website": ""}}, "category": "world_research", "tags": ["model", "training-data"], "summary": "Introduces OpenMLE, a full-stack system with verifiable task environments, RL, and long-horizon search for training Frontis-MA1, a 35B model that performs machine-learning-engineering tasks via program-evolution operators (Draft, Improve, Debug, Crossover). It substantially improves MLE-Bench Lite scores over its base model, approaching much larger frontier models.", "reason": "An MLE agent that trains and evolves ML pipelines through execution feedback matches world_research's MLE-engineering agent example.", "source": "crawl"}, {"paper": {"id": "2607.27288", "title": "Open Security Benchmark: Towards Autonomous Enterprise Cyber Defense", "authors": ["Gal Engelberg", "Michael Arenzon", "Leon Goldberg"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-29", "links": {"paper": "https://arxiv.org/abs/2607.27288", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "Open Security Benchmark (OSB) evaluates agentic AI on autonomous enterprise security posture investigation over a frozen, curated enterprise environment, using text-to-SQL and vendor APIs as investigation modalities with ground-truth scoring.", "reason": "Security posture investigation and defense operations across enterprise systems is a system-administration/security-operations task, matching world_terminal.", "source": "crawl"}, {"paper": {"id": "2607.27929", "title": "Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Training", "authors": ["Zhihong Pan", "Jiyuan He", "Kai Zhang", "Yupeng Han", "Ze Liu", "Yuze Zhao", "Yongcong Ye", "Zhaohua Yang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.27929", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["training-data"], "summary": "Meta-Task reframes terminal task synthesis itself as a terminal task, using an agent operating inside real containers to iteratively generate, execute, and verify diverse terminal tasks for training data. Fine-tuning on the synthesized trajectories improves Terminal-Bench 2.0 performance with far less data than prior approaches.", "reason": "The synthesized trajectories serve terminal agents acting via shell commands, so it routes to world_terminal with the training-data tag.", "source": "crawl"}, {"paper": {"id": "2607.28545", "title": "ORCA-bench: How Ready Are Language Model Agents for Oncall?", "authors": ["Albert Gong", "Kyuseong Choi", "Abhineet Agarwal", "Jason Schechner", "Ryan Huang", "Raj Agrawal", "Anish Agarwal", "Raaz Dwivedi"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28545", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "ORCA-bench evaluates general-purpose coding agents on production-fidelity oncall root-cause-analysis tasks over live telemetry (metrics, logs, traces) and source code across a microservice testbed. Frontier agents achieve only 10-25% accuracy, exposing large gaps for real production reliability work.", "reason": "SRE/oncall incident diagnosis on a live system is DevOps/operations work, matching the world_terminal boundary example of incident mitigation.", "source": "crawl"}, {"paper": {"id": "2607.28287", "title": "Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3", "authors": ["Jens Lehmann", "Andrei Aioanei", "Sahar Vahdati"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28287", "github": "", "website": ""}}, "category": "world_game", "tags": [], "summary": "Introduces Tycho, a coding-agent system for ARC-AGI-3 that constructs, tests, repairs, or bypasses executable game-specific world models to plan actions efficiently. Evaluated across 25 public games, it achieves near-optimal human-relative action efficiency using orchestration policies over frontier LLMs.", "reason": "The agent uses executable code-based world models as a policy to play and win games, matching world_game's code-as-policy definition.", "source": "crawl"}, {"paper": {"id": "2607.28074", "title": "Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale", "authors": ["Yash Pandya", "Sahil Gupta", "Sarthak Harne", "Archana Yadav", "Kavyansh Chourasia", "Hussein Mozannar", "Vibhav Vineet", "Sara Abdali", "Corby Rosset", "Yash Lara", "Ahmed Awadallah", "Ece Kamar", "Akshay Nambi"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28074", "github": "", "website": ""}}, "category": "world_apps", "tags": ["training-data", "benchmark"], "summary": "Echoverse compiles specifications into stateful, login-gated applications with graded tasks and a co-evolution loop that jointly repairs environments and trains computer-use agents. Training on twelve such environments substantially improves a 9B model's accuracy across evaluation splits, and environments are released as a benchmark.", "reason": "Trains and evaluates agents operating stateful software applications, matching world_apps; resource paper follows its served task.", "source": "crawl"}]