Dharma Field Extractor is a 2-billion-parameter model that extracts structured fields from documents into JSON, outperforming general-purpose LLMs in quality and cost.
From Hugging Face
EPISTEME is a new framework that requires coding agents to explain their tool calls, enabling better auditing and preventing cheating in automated software development.
From arXiv
Trace2Eval is a new framework that evaluates coding agents by reconstructing real-world user interactions, allowing for more realistic and dynamic benchmarking of AI models.
From arXiv
LIBRA is a new framework that transforms flat tool catalogs into executable libraries by dynamically composing and testing tool interactions, improving retrieval performance across multiple benchmarks.
From arXiv
Datasette Agent 0.5a0 introduces a 'panel' feature that allows AI to chat alongside Datasette pages, enabling the creation and modification of database-backed apps through natural language prompts.
From Simon Willison
OpenAI developed an autonomous system that identified and fixed 948 local privilege-escalation vulnerabilities in the Linux kernel.
From OpenAI
This paper explores new approaches to access control for AI agents, focusing on how to manage permissions as agents make autonomous decisions. It highlights key areas of progress and remaining challenges in securing agent behavior.
From arXiv
HELIX is a new framework that dynamically partitions LLM inference workloads across different accelerators, improving performance by adapting to hardware and model architecture.
From arXiv
EvoCodeAgent is a new framework for generating complete software projects by evolving both architecture and code together, using experience to improve over time. It outperforms existing methods in test pass rates.
From arXiv
A new benchmark, CASE-Bench, evaluates how repository context influences the security of code generated by LLMs, revealing significant gaps in secure code production.
From arXiv
SBVUL is a new approach that uses two LLM agents to detect exploitable statements in C/C++ functions, achieving better performance than previous methods in some cases.
From arXiv
This paper explores how LLM-based testers can retain useful testing experience while avoiding outdated or irrelevant information, using a maintenance contract for reusable test routes and locators in mobile GUI testing.
From arXiv
DARE is a framework that measures and predicts the difficulty of coding tasks to improve automated program repair by matching support to task-specific features.
From arXiv
A study evaluates the construct validity of coding benchmarks used to assess LLM-based agentic coding tools, finding significant issues with measurement and standardization.
From arXiv
This paper introduces ABE, a methodology for managing AI agent behavior in enterprise settings by treating observable behavior as the primary engineering artifact, rather than the model or prompt.
From arXiv
Ripple-65B Preview, a new open-weight model, achieves state-of-the-art performance on agentic coding tasks, outperforming larger models with significantly lower training costs.
From Hacker News
Datasette Apps 0.3a0 introduces improved integration with LLMs via Datasette Agent, allowing users to tag apps, manage stored queries, and create apps with permissions checks.
From Simon Willison
OpenAI plans to acquire Pierre Computer Company, Inc., aiming to enhance Codex's capabilities in code building, understanding, and review.
From OpenAI
TACT is a new method for analyzing qualitative data using an AI council to improve accuracy and control. It offers two variants and shows strong performance on real-world data.
From arXiv