Paper review 2026-07-11: 13 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. From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization

Ying Chang, Jiahang Xu, Xuan Feng, et al. · arXiv 2026/07 · paper
proposed: software_testing · tags: none

STRACE is a framework that filters noisy agent execution traces and performs causal localization over a trajectory dependency graph to build high-signal optimization contexts for reflection-based agent improvement, demonstrated on a formal verification benchmark.
_{Its evaluated benchmark (VeruSAGE-Bench) is formal verification, so per benchmark-routing it belongs to software_testing.}

2. ArtisanCAD: An Industrial-Level CAD Agent with Expert-Grounded Knowledge Distillation

Yunhan Xu, Qifeng Wu, Xunjin Li, et al. · arXiv 2026/07 · paper
proposed: cad · tags: none

ArtisanCAD is an industrial CAD agent that distills expert CATIA procedural knowledge into reusable parameterized skills via an executable intermediate representation, turning ambiguous prompts into production-grade B-Rep models.
_{The task's purpose is producing 3D/CAD program artifacts from natural language, matching the cad leaf.}

3. Disentangling Ambiguity from Instability in Large Language Models: A Clinical Text-to-SQL Case Study

Angelo Ziletti, Leonardo D'Ambrosi · arXiv 2026/05 · paper
proposed: database · tags: none

CLUES decomposes semantic uncertainty in clinical Text-to-SQL into ambiguity (needing clarification) and instability (needing review) scores via a bipartite semantic graph. It improves failure prediction over prior uncertainty methods and enables targeted interventions for query generation errors.
_{The downstream deliverable is a SQL query, so this Text-to-SQL uncertainty framework belongs to database per the query-synthesis boundary.}

4. What Predicts Correctness in Text-to-SQL? A Selective-Prediction Study

Robert Richardson · arXiv 2026/07 · paper
proposed: database · tags: empirical

This selective-prediction study evaluates which signals (self-consistency, log-probability, LLM judges, trained verifiers) best predict correctness of text-to-SQL queries on BIRD and Spider, finding LLM-judge ensembles generalize best.
_{An empirical study confined to one activity (text-to-SQL) routes to that activity's leaf plus empirical tag, per the single-activity survey/study boundary.}

5. Breaking Database Lock-in: Agentic Regeneration of High Performance Storage Readers for Database Bypass

Victor Giannakouris, Immanuel Trummer · arXiv 2026/07 · paper
proposed: database · tags: none

Jailbreak uses LLM-assisted code synthesis to regenerate high-performance storage readers directly from database file-format documentation/source, bypassing JDBC/ODBC drivers for bulk analytical access with up to 27x speedups.
_{The produced code (storage-format readers) is deliverable database-access code, placing it in the database artifact leaf.}

6. SQuaD-SQL: Efficient Text-to-SQL with Small Language Models via LLM-Guided Knowledge Distillation

Wangyu Wu, Xiaojian Lin, Rong Fu, et al. · arXiv 2026/07 · paper
proposed: database · tags: model

SQuaD-SQL uses LLM-guided knowledge distillation and synthetic data generation to train small language models for Text-to-SQL, approaching LLM performance with much greater efficiency. It achieves 86.9% execution accuracy on WikiSQL while enabling single-GPU training and deployment.
_{A specialized model for producing SQL queries routes to database per the task-served rule.}

7. An Online Reference-Free Evaluation Framework for Flowchart Image-to-Code Generation

Giang Son Nguyen, Zi Pong Lim, Sarthak Ketanbhai Modi, et al. · arXiv 2026/07 · paper
proposed: graphics · tags: none

Proposes a reference-free evaluation framework for flowchart image-to-code (e.g., Mermaid) generation using OCR-based recall and visual-entailment-based precision metrics, validated with strong correlation to ground-truth metrics on FlowVQA.
_{Serves the downstream task of generating diagram/visualization code from a specification, matching the graphics artifact leaf.}

8. fog: Expressing Motion and Emotion through Function Composition of AI-Generated Code

Vivian Liu, Lydia Chilton · arXiv 2026/07 · paper
proposed: graphics · tags: none

Introduces fog, a function-composition framework for AI-generated motion functions that express verbs, gestures, and emotions in Heider-Simmel style animations, paired with an editor for direct manipulation. Perceptual and user studies show generated motions are recognizable and support rapid iteration.
_{The framework's deliverable is generated animation code, matching graphics/animation-from-specification.}

9. HiFuzz: Hierarchical Reinforcement Learning for Semantic-Aware and Adaptive CPU Fuzzing

Ya Wang, Hanwei Fan, Zhenguo Liu, et al. · arXiv 2026/07 · paper
proposed: hardware · tags: none

HiFuzz is a hierarchical RL fuzzing framework with separate program- and basic-block-level agents that generates test programs to reach deep architectural states in RISC-V CPU verification.
_{Fuzzing for processor/CPU verification is hardware verification via code, matching the hardware leaf.}

10. Evaluating SageMath-Augmented LLM Agents for Computational and Experimental Mathematics

Pavel Snopov, German Magai · arXiv 2026/07 · paper
proposed: world_research · tags: none

This work evaluates ReAct-style LLM agents augmented with the SageMath computer algebra system on research-level mathematics problems from RealMath, showing substantial gains from CAS tool access toward automated computational math exploration.
_{Code (SageMath) is used as an instrument for mathematical discovery and conjecture exploration, matching world_research.}

11. Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems

Kalle Kujanpää, Ning Liu, Shahnawaz Alam, et al. · arXiv 2026/07 · paper
proposed: world_terminal · tags: none

Presents a self-evolving LLM agent pipeline that compiles repeated procedural steps into validated tools instead of regenerating code at inference time, deployed for alarm-triage diagnosis in a Fulfillment Center. Production results show reduced latency and error rate in this SRE-style incident-diagnosis system.
_{The served task is system operations/incident triage (SRE), matching world_terminal's system-administration examples.}

12. EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI

Xinjie Wang, Liu Liu, Taojun Ding, et al. · arXiv 2026/07 · paper
proposed: world_physical · tags: training-data

EmbodiedGen V2 is a generative 3D world engine that assembles sim-ready, policy-ready environments (assets, affordances, multi-room scenes) for embodied AI, enabling closed-loop RL training that transfers to real robots.
_{Resource paper synthesizing simulation environments/trajectories whose downstream task is embodied robot policy training, per agency/world_physical.}

13. Prismata: Confining Cross-Site Prompt Injection in Web Agents

Corban Villa, Alp Eren Ozdarendeli, Sijun Tan, et al. · arXiv 2026/07 · paper
proposed: world_browser · tags: none

Presents Prismata, a defense that enforces contextual least privilege for web agents by dynamically deriving trust labels for page content and mechanically confining agent capabilities, mitigating cross-site prompt injection. It substantially reduces attack success across published web-agent attacks while preserving task utility.
_{The agent acts on live webpages via a browser, so securing that action loop belongs to world_browser rather than software_security.}

machine payload (do not edit)

[{"paper": {"id": "2607.07702", "title": "From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization", "authors": ["Ying Chang", "Jiahang Xu", "Xuan Feng", "Chenyuan Yang", "Peng Cheng", "Yuqing Yang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-08", "links": {"paper": "https://arxiv.org/abs/2607.07702", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "STRACE is a framework that filters noisy agent execution traces and performs causal localization over a trajectory dependency graph to build high-signal optimization contexts for reflection-based agent improvement, demonstrated on a formal verification benchmark.", "reason": "Its evaluated benchmark (VeruSAGE-Bench) is formal verification, so per benchmark-routing it belongs to software_testing."}, {"paper": {"id": "2607.05750", "title": "ArtisanCAD: An Industrial-Level CAD Agent with Expert-Grounded Knowledge Distillation", "authors": ["Yunhan Xu", "Qifeng Wu", "Xunjin Li", "Yuanwei Bin", "Qingsong Yao", "Jianghang Gu", "Guan Wang", "Weihao Lv", "Huiyu Yang", "Wenfa Luo", "Jiao Xiang", "Yuntian Chen", "Shiyi Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-08", "links": {"paper": "https://arxiv.org/abs/2607.05750", "github": "", "website": ""}}, "category": "cad", "tags": [], "summary": "ArtisanCAD is an industrial CAD agent that distills expert CATIA procedural knowledge into reusable parameterized skills via an executable intermediate representation, turning ambiguous prompts into production-grade B-Rep models.", "reason": "The task's purpose is producing 3D/CAD program artifacts from natural language, matching the cad leaf."}, {"paper": {"id": "2602.12015", "title": "Disentangling Ambiguity from Instability in Large Language Models: A Clinical Text-to-SQL Case Study", "authors": ["Angelo Ziletti", "Leonardo D'Ambrosi"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-18", "links": {"paper": "https://arxiv.org/abs/2602.12015", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "CLUES decomposes semantic uncertainty in clinical Text-to-SQL into ambiguity (needing clarification) and instability (needing review) scores via a bipartite semantic graph. It improves failure prediction over prior uncertainty methods and enables targeted interventions for query generation errors.", "reason": "The downstream deliverable is a SQL query, so this Text-to-SQL uncertainty framework belongs to database per the query-synthesis boundary."}, {"paper": {"id": "2607.06799", "title": "What Predicts Correctness in Text-to-SQL? A Selective-Prediction Study", "authors": ["Robert Richardson"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-07", "links": {"paper": "https://arxiv.org/abs/2607.06799", "github": "", "website": ""}}, "category": "database", "tags": ["empirical"], "summary": "This selective-prediction study evaluates which signals (self-consistency, log-probability, LLM judges, trained verifiers) best predict correctness of text-to-SQL queries on BIRD and Spider, finding LLM-judge ensembles generalize best.", "reason": "An empirical study confined to one activity (text-to-SQL) routes to that activity's leaf plus empirical tag, per the single-activity survey/study boundary."}, {"paper": {"id": "2607.07696", "title": "Breaking Database Lock-in: Agentic Regeneration of High Performance Storage Readers for Database Bypass", "authors": ["Victor Giannakouris", "Immanuel Trummer"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-08", "links": {"paper": "https://arxiv.org/abs/2607.07696", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "Jailbreak uses LLM-assisted code synthesis to regenerate high-performance storage readers directly from database file-format documentation/source, bypassing JDBC/ODBC drivers for bulk analytical access with up to 27x speedups.", "reason": "The produced code (storage-format readers) is deliverable database-access code, placing it in the database artifact leaf."}, {"paper": {"id": "2607.08161", "title": "SQuaD-SQL: Efficient Text-to-SQL with Small Language Models via LLM-Guided Knowledge Distillation", "authors": ["Wangyu Wu", "Xiaojian Lin", "Rong Fu", "Zaiyang Yu", "Xuhang Chen", "Wenjun Yu", "Zhenhong Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-09", "links": {"paper": "https://arxiv.org/abs/2607.08161", "github": "", "website": ""}}, "category": "database", "tags": ["model"], "summary": "SQuaD-SQL uses LLM-guided knowledge distillation and synthetic data generation to train small language models for Text-to-SQL, approaching LLM performance with much greater efficiency. It achieves 86.9% execution accuracy on WikiSQL while enabling single-GPU training and deployment.", "reason": "A specialized model for producing SQL queries routes to database per the task-served rule."}, {"paper": {"id": "2602.13376", "title": "An Online Reference-Free Evaluation Framework for Flowchart Image-to-Code Generation", "authors": ["Giang Son Nguyen", "Zi Pong Lim", "Sarthak Ketanbhai Modi", "Yon Shin Teo", "Wenya Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-09", "links": {"paper": "https://arxiv.org/abs/2602.13376", "github": "", "website": ""}}, "category": "graphics", "tags": [], "summary": "Proposes a reference-free evaluation framework for flowchart image-to-code (e.g., Mermaid) generation using OCR-based recall and visual-entailment-based precision metrics, validated with strong correlation to ground-truth metrics on FlowVQA.", "reason": "Serves the downstream task of generating diagram/visualization code from a specification, matching the graphics artifact leaf."}, {"paper": {"id": "2607.07952", "title": "fog: Expressing Motion and Emotion through Function Composition of AI-Generated Code", "authors": ["Vivian Liu", "Lydia Chilton"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-08", "links": {"paper": "https://arxiv.org/abs/2607.07952", "github": "", "website": ""}}, "category": "graphics", "tags": [], "summary": "Introduces fog, a function-composition framework for AI-generated motion functions that express verbs, gestures, and emotions in Heider-Simmel style animations, paired with an editor for direct manipulation. Perceptual and user studies show generated motions are recognizable and support rapid iteration.", "reason": "The framework's deliverable is generated animation code, matching graphics/animation-from-specification."}, {"paper": {"id": "2607.06619", "title": "HiFuzz: Hierarchical Reinforcement Learning for Semantic-Aware and Adaptive CPU Fuzzing", "authors": ["Ya Wang", "Hanwei Fan", "Zhenguo Liu", "Xiaofeng Zhou", "Yangdi Lyu", "Jiang Xu", "Wei Zhang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-07", "links": {"paper": "https://arxiv.org/abs/2607.06619", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "HiFuzz is a hierarchical RL fuzzing framework with separate program- and basic-block-level agents that generates test programs to reach deep architectural states in RISC-V CPU verification.", "reason": "Fuzzing for processor/CPU verification is hardware verification via code, matching the hardware leaf."}, {"paper": {"id": "2607.06820", "title": "Evaluating SageMath-Augmented LLM Agents for Computational and Experimental Mathematics", "authors": ["Pavel Snopov", "German Magai"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-07", "links": {"paper": "https://arxiv.org/abs/2607.06820", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "This work evaluates ReAct-style LLM agents augmented with the SageMath computer algebra system on research-level mathematics problems from RealMath, showing substantial gains from CAS tool access toward automated computational math exploration.", "reason": "Code (SageMath) is used as an instrument for mathematical discovery and conjecture exploration, matching world_research."}, {"paper": {"id": "2607.08010", "title": "Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems", "authors": ["Kalle Kujanpää", "Ning Liu", "Shahnawaz Alam", "Yeshwanth Reddy Sura", "Tianyu Yang", "Kristina Klinkner", "Shervin Malmasi"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-09", "links": {"paper": "https://arxiv.org/abs/2607.08010", "github": "", "website": ""}}, "category": "world_terminal", "tags": [], "summary": "Presents a self-evolving LLM agent pipeline that compiles repeated procedural steps into validated tools instead of regenerating code at inference time, deployed for alarm-triage diagnosis in a Fulfillment Center. Production results show reduced latency and error rate in this SRE-style incident-diagnosis system.", "reason": "The served task is system operations/incident triage (SRE), matching world_terminal's system-administration examples."}, {"paper": {"id": "2607.07459", "title": "EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI", "authors": ["Xinjie Wang", "Liu Liu", "Taojun Ding", "Andrew Choi", "Chaodong Huang", "Mengao Zhao", "Ziang Li", "Jackson Jiang", "Chunlei Yu", "Shengxiang Liu", "Wei Xu", "Zhizhong Su"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-08", "links": {"paper": "https://arxiv.org/abs/2607.07459", "github": "", "website": ""}}, "category": "world_physical", "tags": ["training-data"], "summary": "EmbodiedGen V2 is a generative 3D world engine that assembles sim-ready, policy-ready environments (assets, affordances, multi-room scenes) for embodied AI, enabling closed-loop RL training that transfers to real robots.", "reason": "Resource paper synthesizing simulation environments/trajectories whose downstream task is embodied robot policy training, per agency/world_physical."}, {"paper": {"id": "2607.08147", "title": "Prismata: Confining Cross-Site Prompt Injection in Web Agents", "authors": ["Corban Villa", "Alp Eren Ozdarendeli", "Sijun Tan", "Raluca Ada Popa"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-09", "links": {"paper": "https://arxiv.org/abs/2607.08147", "github": "", "website": ""}}, "category": "world_browser", "tags": [], "summary": "Presents Prismata, a defense that enforces contextual least privilege for web agents by dynamically deriving trust labels for page content and mechanically confining agent capabilities, mitigating cross-site prompt injection. It substantially reduces attack success across published web-agent attacks while preserving task utility.", "reason": "The agent acts on live webpages via a browser, so securing that action loop belongs to world_browser rather than software_security."}]

[Zhaoyang-Chu]

/reject 13 web agent acts via GUI primitives, not code as action