Paper review 2026-08-05: 20 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. TraceCAD: Trace-Guided Repair for Agentic CAD Generation

Fengxiao Fan, Jingzhe Ni, Fan Sang, et al. · arXiv 2026/08 · paper
proposed: cad · tags: none

TraceCAD is a trace-guided recovery layer for LLM-based CAD agents that links requirements, modeling steps, and failure evidence as persistent state to diagnose and repair faulty parametric CAD programs. It improves geometric quality and repair reliability over DeepCAD-derived benchmarks.
_{An agent produces executable parametric CAD programs, matching the cad artifact leaf.}

2. LiveEvalBench: Toward Open-World Evaluation for Web Generation

Yiyao Wang, Zhen Wen, Yinghao Tang, et al. · arXiv 2026/08 · paper
proposed: web · tags: benchmark

LiveEvalBench is an agentic, adaptive evaluation framework for frontend web generation, using collaborative Build/Code/UI-Tester roles to assess deployment, code, and browser interaction. It aligns well with human judgment and reveals fine-grained gaps in frontier models' web-generation abilities.
_{Serves evaluation of web application code generation, matching the artifact/web leaf as a benchmark resource.}

3. Pattern over Pixels: Measuring Pattern Completion Bias in Multimodal Code Generation

Khai-Nguyen Nguyen, Oscar Chaparro, Antonio Mastropaolo · arXiv 2026/08 · paper
proposed: web · tags: benchmark

Introduces a benchmark measuring pattern-completion bias in multimodal LLMs performing screenshot-to-code fill-in-the-blank tasks, where repeated UI patterns bias models toward visually incorrect but pattern-consistent outputs. Evaluation across five frontier MLLMs shows strong bias and low accuracy on perturbed webpage elements.
_{Benchmark for screenshot-to-code generation, a task under the web artifact leaf per the taxonomy's explicit example.}

4. BAP-SQL: Budget-Aware Observation Planning for Agentic Text-to-SQL

Chong Peng, Pin Qian, Su Wang, et al. · arXiv 2026/08 · paper
proposed: database · tags: none

BAP-SQL is a budget-aware observation-planning method for agentic text-to-SQL that estimates query risk, rewrites SQL, and enforces runtime limits to improve tight-budget query success.
_{The agent's task ultimately produces SQL queries, matching the database artifact leaf.}

5. Evaluating LLMs in Database Scenarios: A Lifecycle Benchmark for Assessing Their Potential in Core Database Tasks

Shunfan Zheng, Dongsheng Shi, Yue Li, et al. · arXiv 2026/08 · paper
proposed: database · tags: benchmark

DBLifeBench evaluates LLMs across five database lifecycle phases (design, implementation, operation, debugging, maintenance), moving beyond narrow Text-to-SQL evaluation. It finds specialized Text-to-SQL models suffer catastrophic forgetting on non-coding phases like design and maintenance.
_{Benchmark for LLM/agent database tasks routes to the database leaf regardless of its lifecycle breadth.}

6. Automated Visualization Code Synthesis via Multi-Path Reasoning and Feedback-Driven Optimization

Wonduk Seo, Daye Kang, Hyunjin An, et al. · arXiv 2026/08 · paper
proposed: graphics · tags: none

VisPath is a multi-path reasoning and feedback-driven framework that generates and executes visualization code from underspecified natural-language requests, aggregating diverse candidate scripts to synthesize an optimal chart. It outperforms prior methods on MatPlotBench and the Qwen-Agent Code Interpreter Benchmark.
_{Chart-from-spec code generation with the visual artifact's code as the deliverable falls under artifact/graphics.}

7. SeaSlides: Semantic Abstraction Layer for Agentic Slide Generation

Shengjun Fang, Chenyang Wu, Zongzhang Zhang · arXiv 2026/08 · paper
proposed: graphics · tags: benchmark

SeaSlides is an agentic slide-generation framework where the model authors structured slide content through reusable components and capability modules (routing equations, code, and charts to dedicated renderers) rather than raw SVG/HTML geometry, improving readability of rich technical decks. The authors introduce SeaSlidesBench-Rich to evaluate content-heavy slide generation.
_{The agent writes template/component code (HTML/Typst) to render visual presentation artifacts, matching graphics' chart/visual-artifact-from-specification scope.}

8. KernelBrain: Coarse-to-Fine, Budget-Aware Search for Agentic GPU Kernel Optimization

Shuai Che, Gang Peng · arXiv 2026/05 · paper
proposed: systems · tags: none

KernelBrain is an agentic optimization system that uses LLM-guided mutation, adaptive resource allocation, and profiler-informed diagnosis to search for high-performance GPU kernel variants. It achieves substantial speedups over PyTorch and prior kernel-optimization agents on Triton kernel tasks.
_{The agent produces and optimizes GPU kernel code, which is systems-level code per the KernelBench-to-systems precedent.}

9. Don't Regenerate, Debug: A Domain-Specific Agent for Repairing Near-Miss Hardware Operators

Yansong Sun, Shenxiu Wu, Siyuan Chen, et al. · arXiv 2026/08 · paper
proposed: systems · tags: none

This paper presents a domain-specific debug agent that repairs near-miss GPU/NPU accelerator kernels that compile and run but fail numerical validation, using retrieved patterns and diagnostic instrumentation instead of regenerating from scratch. It achieves much higher pass rates at far lower token cost than regeneration-based pipelines.
_{The task is repairing hardware-accelerator kernel code, a systems-level code artifact, matching the systems leaf per the KernelBench precedent.}

10. VeriTrace: Human-Like Temporal Exploration Completes Agentic Action Space

Yu-Tung Liu, Cunxi Yu · arXiv 2026/08 · paper
proposed: hardware · tags: none

VeriTrace is a multi-agent system with an Inspector agent that explores signals and time-windows in waveforms to debug Verilog RTL, achieving 100% Pass@1 on VerilogEval-V2.
_{Task centers on generating/debugging hardware description (Verilog) code, matching the hardware artifact leaf.}

11. LACE: Large Language Model Aided Multi-Agent Framework for Agile RISC-V Instruction Extension

Pingqing Zheng, Jiayin Qin, Fuqi Zhang, et al. · arXiv 2026/08 · paper
proposed: hardware · tags: none

LACE is an LLM-aided multi-agent workflow translating natural-language RISC-V instruction extension intents into a compact IR, performing retrieval-guided RTL edits, and validating via a formal checking flow across four embedded cores.
_{Multi-agent system producing and validating hardware description (RTL) code, matching the hardware leaf.}

12. MDArena: Evaluating Coding Agents on Realistic Molecular Dynamics Workflows

Nithishwer Mouroug Anand, Wei-Tse Hsu, Kyle Vaccaro, et al. · arXiv 2026/07 · paper
proposed: world_research · tags: benchmark

MDArena is a benchmark of 50 containerized molecular dynamics tasks drawn from real biomolecular simulation projects, used to evaluate coding agents (Codex, OpenCode) on realistic scientific workflows like trajectory analysis and free-energy protocols. Results show agents make partial progress but struggle with reproducible scientific rigor.
_{Coding agents act through code to conduct scientific simulation research, matching world_research's scientific-discovery agent scope.}

13. OR-Agent: Bridging Evolutionary Search and Structured Research for Automated Algorithm Discovery

Qi Liu, Ruochen Hao, Can Li, et al. · arXiv 2026/08 · paper
proposed: world_research · tags: none

OR-Agent is a multi-agent framework that organizes heuristic algorithm discovery for combinatorial optimization (TSP, CVRP, bin packing) as tree-structured research with reflection-based learning from past trials. It outperforms evolutionary search baselines on classical optimization and simulation-based cooperative driving tasks.
_{Agent conducts research-style experimentation to discover algorithms/heuristics, matching the discovery-through-interaction pattern of world_research.}

14. SPEAR: Code-Augmented Agentic Prompt Optimization

Mengyin Lu, Cong Feng, Huimin Han, et al. · arXiv 2026/08 · paper
proposed: world_research · tags: none

SPEAR is an agentic prompt optimizer that ports the code-as-action paradigm to automatic prompt engineering, using a Python sandbox for structural error analysis (confusion matrices, error clustering) to iteratively improve prompts. It outperforms prior optimizers on industrial LLM-judge tasks and BBH benchmarks.
_{The agent writes and executes code as the instrument for iterative experiment analysis and tuning, matching world_research's ML-engineering/data-analysis scope rather than producing a code artifact.}

15. Traceable Multi-Agent System for Knowledge-Based Forecasting

Junhyeok Kang, Sangjun Han, Hyeokjun Choe, et al. · arXiv 2026/08 · paper
proposed: world_research · tags: none

TraceMAS is a multi-agent forecasting system that interprets documents, searches data, and generates code to build traceable causal-loop-grounded forecasting models, demonstrated on crude oil price forecasting. It focuses on making the agents' evidence-to-forecast pipeline inspectable.
_{Agents generate and revise code as an instrument to produce forecasts/insight, matching world_research's data-science/ML engineering scope.}

16. DataSpace: Benchmarking Data Agents for Verifiable Analytics over Heterogeneous Workspaces

Boyan Li, Zhuowen Liang, Yupeng Xie, et al. · arXiv 2026/08 · paper
proposed: world_research · tags: benchmark

DataSpace is a benchmark of 410 cross-language analytics tasks over heterogeneous workspaces (databases, documents, multimedia) requiring data agents to produce verifiable tabular results, evaluated with a deterministic execution-grounded evaluator. It served as the official KDD Cup 2026 competition benchmark and reveals unsaturated performance across frontier models and agent harnesses.
_{Benchmarks data agents whose deliverable is analytical insight/tabular results, not code itself, fitting world_research.}

17. Permission Denied: Policy-Graded Evaluation of Coding Agents in Hardened Environments

Dotan Davidovich, Yair Amar, Hai Rozencwajg, et al. · arXiv 2026/08 · paper
proposed: world_terminal · tags: empirical benchmark

Evaluates 12 coding agents on Terminal-Bench 2.1 under nested enterprise security policies (restricted credentials, egress, filesystems), finding hardening degrades success and inflates cost unevenly across models. Releases Boundary-Bench, an open-source plugin for policy-constrained agent evaluation.
_{Coding agents complete terminal tasks under real-world OS/security constraints, matching world_terminal's system-administration/hardened-environment scope.}

18. Tiny Enough to Break In: Agentic Remote Access Trojans Powered by Small Language Models

Yuhan You, Suhas Adavelly, Victoria Lovelace, et al. · arXiv 2026/08 · paper
proposed: world_terminal · tags: none

Implements an 'agentic RAT' where a small locally-deployed language model autonomously reasons over reconnaissance data and issues shell commands to exploit vulnerable services in a lab, achieving root-shell access with imperfect reliability.
_{An offensive-security agent that acts via shell/terminal commands to exploit systems, matching the CTF/penetration-testing inclusion under world_terminal.}

19. ETA: A New Agentic Paradigm for Embodied Tasks

Yitong Chen, Zezheng Huai, Sixian Li, et al. · arXiv 2026/08 · paper
proposed: world_physical · tags: none

ETA introduces an agentic paradigm extending digital agents into the physical world via a Planner-Tool-Interface-World loop, with OpenETA providing replaceable planners, tools, and auditable trajectories for robots. It integrates with Codex as a lightweight plugin exposing observation and motion primitives.
_{Code-driven agentic framework controlling robots places it in the physical-world acting leaf.}

20. UrbanAgent: A Tool-Augmented Agent for Cross-System Urban Tasks

Jiayu Cao, Xingyuan Zeng, feiyu Li, et al. · arXiv 2026/08 · paper
proposed: world_apps · tags: benchmark

Proposes UrbanAgent, a tool-augmented LLM agent that converts natural-language requests into cross-system urban workflows using code execution, API calls, and MCP, plus the Urban-Eval benchmark for evaluating task success and execution quality.
_{An agent that acts on real-world fragmented digital city services via code/API calls, best matching operating professional/enterprise applications under world_apps.}

machine payload (do not edit)

[{"paper": {"id": "2608.03062", "title": "TraceCAD: Trace-Guided Repair for Agentic CAD Generation", "authors": ["Fengxiao Fan", "Jingzhe Ni", "Fan Sang", "Xiaolong Yin", "Yu Liu", "Ruofeng Tong", "Min Tang", "Peng Du"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03062", "github": "", "website": ""}}, "category": "cad", "tags": [], "summary": "TraceCAD is a trace-guided recovery layer for LLM-based CAD agents that links requirements, modeling steps, and failure evidence as persistent state to diagnose and repair faulty parametric CAD programs. It improves geometric quality and repair reliability over DeepCAD-derived benchmarks.", "reason": "An agent produces executable parametric CAD programs, matching the cad artifact leaf.", "source": "crawl"}, {"paper": {"id": "2608.03689", "title": "LiveEvalBench: Toward Open-World Evaluation for Web Generation", "authors": ["Yiyao Wang", "Zhen Wen", "Yinghao Tang", "Yixiao Fu", "Lin Yuan", "Xiaolau Zhang", "Jun Zhou", "Wei Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03689", "github": "", "website": ""}}, "category": "web", "tags": ["benchmark"], "summary": "LiveEvalBench is an agentic, adaptive evaluation framework for frontend web generation, using collaborative Build/Code/UI-Tester roles to assess deployment, code, and browser interaction. It aligns well with human judgment and reveals fine-grained gaps in frontier models' web-generation abilities.", "reason": "Serves evaluation of web application code generation, matching the artifact/web leaf as a benchmark resource.", "source": "crawl"}, {"paper": {"id": "2608.03691", "title": "Pattern over Pixels: Measuring Pattern Completion Bias in Multimodal Code Generation", "authors": ["Khai-Nguyen Nguyen", "Oscar Chaparro", "Antonio Mastropaolo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03691", "github": "", "website": ""}}, "category": "web", "tags": ["benchmark"], "summary": "Introduces a benchmark measuring pattern-completion bias in multimodal LLMs performing screenshot-to-code fill-in-the-blank tasks, where repeated UI patterns bias models toward visually incorrect but pattern-consistent outputs. Evaluation across five frontier MLLMs shows strong bias and low accuracy on perturbed webpage elements.", "reason": "Benchmark for screenshot-to-code generation, a task under the web artifact leaf per the taxonomy's explicit example.", "source": "crawl"}, {"paper": {"id": "2608.02876", "title": "BAP-SQL: Budget-Aware Observation Planning for Agentic Text-to-SQL", "authors": ["Chong Peng", "Pin Qian", "Su Wang", "Yihang Chen", "Varun Sah"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.02876", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "BAP-SQL is a budget-aware observation-planning method for agentic text-to-SQL that estimates query risk, rewrites SQL, and enforces runtime limits to improve tight-budget query success.", "reason": "The agent's task ultimately produces SQL queries, matching the database artifact leaf.", "source": "crawl"}, {"paper": {"id": "2608.03794", "title": "Evaluating LLMs in Database Scenarios: A Lifecycle Benchmark for Assessing Their Potential in Core Database Tasks", "authors": ["Shunfan Zheng", "Dongsheng Shi", "Yue Li", "Xin Yi", "Linlin Wang", "Gerard de Melo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03794", "github": "", "website": ""}}, "category": "database", "tags": ["benchmark"], "summary": "DBLifeBench evaluates LLMs across five database lifecycle phases (design, implementation, operation, debugging, maintenance), moving beyond narrow Text-to-SQL evaluation. It finds specialized Text-to-SQL models suffer catastrophic forgetting on non-coding phases like design and maintenance.", "reason": "Benchmark for LLM/agent database tasks routes to the database leaf regardless of its lifecycle breadth.", "source": "crawl"}, {"paper": {"id": "2502.11140", "title": "Automated Visualization Code Synthesis via Multi-Path Reasoning and Feedback-Driven Optimization", "authors": ["Wonduk Seo", "Daye Kang", "Hyunjin An", "Taehan Kim", "Soohyuk Cho", "Seungyong Lee", "Minhyeong Yu", "Jian Park", "Yi Bu", "Seunghyun Lee"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2502.11140", "github": "", "website": ""}}, "category": "graphics", "tags": [], "summary": "VisPath is a multi-path reasoning and feedback-driven framework that generates and executes visualization code from underspecified natural-language requests, aggregating diverse candidate scripts to synthesize an optimal chart. It outperforms prior methods on MatPlotBench and the Qwen-Agent Code Interpreter Benchmark.", "reason": "Chart-from-spec code generation with the visual artifact's code as the deliverable falls under artifact/graphics.", "source": "crawl"}, {"paper": {"id": "2608.03298", "title": "SeaSlides: Semantic Abstraction Layer for Agentic Slide Generation", "authors": ["Shengjun Fang", "Chenyang Wu", "Zongzhang Zhang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03298", "github": "", "website": ""}}, "category": "graphics", "tags": ["benchmark"], "summary": "SeaSlides is an agentic slide-generation framework where the model authors structured slide content through reusable components and capability modules (routing equations, code, and charts to dedicated renderers) rather than raw SVG/HTML geometry, improving readability of rich technical decks. The authors introduce SeaSlidesBench-Rich to evaluate content-heavy slide generation.", "reason": "The agent writes template/component code (HTML/Typst) to render visual presentation artifacts, matching graphics' chart/visual-artifact-from-specification scope.", "source": "crawl"}, {"paper": {"id": "2608.02611", "title": "KernelBrain: Coarse-to-Fine, Budget-Aware Search for Agentic GPU Kernel Optimization", "authors": ["Shuai Che", "Gang Peng"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-19", "links": {"paper": "https://arxiv.org/abs/2608.02611", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "KernelBrain is an agentic optimization system that uses LLM-guided mutation, adaptive resource allocation, and profiler-informed diagnosis to search for high-performance GPU kernel variants. It achieves substantial speedups over PyTorch and prior kernel-optimization agents on Triton kernel tasks.", "reason": "The agent produces and optimizes GPU kernel code, which is systems-level code per the KernelBench-to-systems precedent.", "source": "crawl"}, {"paper": {"id": "2608.02712", "title": "Don't Regenerate, Debug: A Domain-Specific Agent for Repairing Near-Miss Hardware Operators", "authors": ["Yansong Sun", "Shenxiu Wu", "Siyuan Chen", "Runlin Hou", "Junhao Qiu", "Junming Cao", "Shudi Shao", "Zhichao Lu", "Qingfu Zhang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.02712", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "This paper presents a domain-specific debug agent that repairs near-miss GPU/NPU accelerator kernels that compile and run but fail numerical validation, using retrieved patterns and diagnostic instrumentation instead of regenerating from scratch. It achieves much higher pass rates at far lower token cost than regeneration-based pipelines.", "reason": "The task is repairing hardware-accelerator kernel code, a systems-level code artifact, matching the systems leaf per the KernelBench precedent.", "source": "crawl"}, {"paper": {"id": "2608.02878", "title": "VeriTrace: Human-Like Temporal Exploration Completes Agentic Action Space", "authors": ["Yu-Tung Liu", "Cunxi Yu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.02878", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "VeriTrace is a multi-agent system with an Inspector agent that explores signals and time-windows in waveforms to debug Verilog RTL, achieving 100% Pass@1 on VerilogEval-V2.", "reason": "Task centers on generating/debugging hardware description (Verilog) code, matching the hardware artifact leaf.", "source": "crawl"}, {"paper": {"id": "2608.02915", "title": "LACE: Large Language Model Aided Multi-Agent Framework for Agile RISC-V Instruction Extension", "authors": ["Pingqing Zheng", "Jiayin Qin", "Fuqi Zhang", "Zishen Wan", "Shang Wu", "Yu Cao", "Caiwen Ding", "Yang Katie Zhao"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.02915", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "LACE is an LLM-aided multi-agent workflow translating natural-language RISC-V instruction extension intents into a compact IR, performing retrieval-guided RTL edits, and validating via a formal checking flow across four embedded cores.", "reason": "Multi-agent system producing and validating hardware description (RTL) code, matching the hardware leaf.", "source": "crawl"}, {"paper": {"id": "2608.02642", "title": "MDArena: Evaluating Coding Agents on Realistic Molecular Dynamics Workflows", "authors": ["Nithishwer Mouroug Anand", "Wei-Tse Hsu", "Kyle Vaccaro", "Eden James Gage", "Jonathan David Colburn", "Linda Xi Phan", "Minjoon Seo", "Kevin Guan", "Philip C. Biggin"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-31", "links": {"paper": "https://arxiv.org/abs/2608.02642", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "MDArena is a benchmark of 50 containerized molecular dynamics tasks drawn from real biomolecular simulation projects, used to evaluate coding agents (Codex, OpenCode) on realistic scientific workflows like trajectory analysis and free-energy protocols. Results show agents make partial progress but struggle with reproducible scientific rigor.", "reason": "Coding agents act through code to conduct scientific simulation research, matching world_research's scientific-discovery agent scope.", "source": "crawl"}, {"paper": {"id": "2602.13769", "title": "OR-Agent: Bridging Evolutionary Search and Structured Research for Automated Algorithm Discovery", "authors": ["Qi Liu", "Ruochen Hao", "Can Li", "Wanjing Ma"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2602.13769", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "OR-Agent is a multi-agent framework that organizes heuristic algorithm discovery for combinatorial optimization (TSP, CVRP, bin packing) as tree-structured research with reflection-based learning from past trials. It outperforms evolutionary search baselines on classical optimization and simulation-based cooperative driving tasks.", "reason": "Agent conducts research-style experimentation to discover algorithms/heuristics, matching the discovery-through-interaction pattern of world_research.", "source": "crawl"}, {"paper": {"id": "2605.26275", "title": "SPEAR: Code-Augmented Agentic Prompt Optimization", "authors": ["Mengyin Lu", "Cong Feng", "Huimin Han", "Guangming Lu", "Yu Sun", "Xiaonan Ding", "Shihui Long", "Fengyi Li", "Tanvi Motwani"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2605.26275", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "SPEAR is an agentic prompt optimizer that ports the code-as-action paradigm to automatic prompt engineering, using a Python sandbox for structural error analysis (confusion matrices, error clustering) to iteratively improve prompts. It outperforms prior optimizers on industrial LLM-judge tasks and BBH benchmarks.", "reason": "The agent writes and executes code as the instrument for iterative experiment analysis and tuning, matching world_research's ML-engineering/data-analysis scope rather than producing a code artifact.", "source": "crawl"}, {"paper": {"id": "2608.03339", "title": "Traceable Multi-Agent System for Knowledge-Based Forecasting", "authors": ["Junhyeok Kang", "Sangjun Han", "Hyeokjun Choe", "Soonyoung Lee"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03339", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "TraceMAS is a multi-agent forecasting system that interprets documents, searches data, and generates code to build traceable causal-loop-grounded forecasting models, demonstrated on crude oil price forecasting. It focuses on making the agents' evidence-to-forecast pipeline inspectable.", "reason": "Agents generate and revise code as an instrument to produce forecasts/insight, matching world_research's data-science/ML engineering scope.", "source": "crawl"}, {"paper": {"id": "2608.03451", "title": "DataSpace: Benchmarking Data Agents for Verifiable Analytics over Heterogeneous Workspaces", "authors": ["Boyan Li", "Zhuowen Liang", "Yupeng Xie", "Xiaotian Lin", "Tianqi Luo", "Xinyu Liu", "Yizhang Zhu", "Zhangyang Peng", "Yuan Li", "Zhengxuan Zhang", "Jiayi Zhang", "Nan Tang", "Guoliang Li", "Yuyu Luo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03451", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "DataSpace is a benchmark of 410 cross-language analytics tasks over heterogeneous workspaces (databases, documents, multimedia) requiring data agents to produce verifiable tabular results, evaluated with a deterministic execution-grounded evaluator. It served as the official KDD Cup 2026 competition benchmark and reveals unsaturated performance across frontier models and agent harnesses.", "reason": "Benchmarks data agents whose deliverable is analytical insight/tabular results, not code itself, fitting world_research.", "source": "crawl"}, {"paper": {"id": "2608.02670", "title": "Permission Denied: Policy-Graded Evaluation of Coding Agents in Hardened Environments", "authors": ["Dotan Davidovich", "Yair Amar", "Hai Rozencwajg", "Or Hiltch"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2608.02670", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["empirical", "benchmark"], "summary": "Evaluates 12 coding agents on Terminal-Bench 2.1 under nested enterprise security policies (restricted credentials, egress, filesystems), finding hardening degrades success and inflates cost unevenly across models. Releases Boundary-Bench, an open-source plugin for policy-constrained agent evaluation.", "reason": "Coding agents complete terminal tasks under real-world OS/security constraints, matching world_terminal's system-administration/hardened-environment scope.", "source": "crawl"}, {"paper": {"id": "2608.03009", "title": "Tiny Enough to Break In: Agentic Remote Access Trojans Powered by Small Language Models", "authors": ["Yuhan You", "Suhas Adavelly", "Victoria Lovelace", "Cameron Berryman", "Joel Sadler", "Daniel Graham"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03009", "github": "", "website": ""}}, "category": "world_terminal", "tags": [], "summary": "Implements an 'agentic RAT' where a small locally-deployed language model autonomously reasons over reconnaissance data and issues shell commands to exploit vulnerable services in a lab, achieving root-shell access with imperfect reliability.", "reason": "An offensive-security agent that acts via shell/terminal commands to exploit systems, matching the CTF/penetration-testing inclusion under world_terminal.", "source": "crawl"}, {"paper": {"id": "2608.03924", "title": "ETA: A New Agentic Paradigm for Embodied Tasks", "authors": ["Yitong Chen", "Zezheng Huai", "Sixian Li", "Yubang Wang", "Haozhe Zhang", "Yifei Zhang", "Hechang Chen", "Jingjing Gong", "Yu-Gang Jiang", "Xipeng Qiu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03924", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "ETA introduces an agentic paradigm extending digital agents into the physical world via a Planner-Tool-Interface-World loop, with OpenETA providing replaceable planners, tools, and auditable trajectories for robots. It integrates with Codex as a lightweight plugin exposing observation and motion primitives.", "reason": "Code-driven agentic framework controlling robots places it in the physical-world acting leaf.", "source": "crawl"}, {"paper": {"id": "2608.03018", "title": "UrbanAgent: A Tool-Augmented Agent for Cross-System Urban Tasks", "authors": ["Jiayu Cao", "Xingyuan Zeng", "feiyu Li", "Zhijing Huang", "Xujie Yuan", "Rongxiang Chen", "Shimin Di", "Libin Zheng", "Jian Yin"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.03018", "github": "", "website": ""}}, "category": "world_apps", "tags": ["benchmark"], "summary": "Proposes UrbanAgent, a tool-augmented LLM agent that converts natural-language requests into cross-system urban workflows using code execution, API calls, and MCP, plus the Urban-Eval benchmark for evaluating task success and execution quality.", "reason": "An agent that acts on real-world fragmented digital city services via code/API calls, best matching operating professional/enterprise applications under world_apps.", "source": "crawl"}]