A paper suggests that large language models can interpret raw neural network weights without code execution. Qwen 3.8 27B running on a laptop successfully identifies game types from model weights in four classic games.
From Simon Willison
This paper explores the idea of 'regeneration'—replacing existing code with new implementations generated from requirements and tests. It questions whether this approach could reduce technical debt but also highlights risks of losing unrecorded knowledge and behavior.
From arXiv
StateBench is a new benchmark that tests how well language models track state changes in long documents, revealing differences in performance across architectures and difficulty levels.
From Hugging Face
DSL-Bench is a new benchmark framework that evaluates how well large language models generate code for domain-specific languages, revealing significant gaps between syntax and semantic correctness across different DSLs.
From arXiv
A study of 84,200 GitHub skills reveals patterns in their creation, evolution, and maintenance, offering insights for improving agent reliability and security.
From arXiv
A study shows that separating planning and coding tasks between different LLMs can be as effective as using a single model for both, with stronger combinations often performing better.
From arXiv
A study compares Robotic Process Automation (RPA) and Autonomous Agents (APA) for retrieving medical reports in healthcare, finding RPA to be faster, cheaper, and more consistent for stable interfaces.
From arXiv
Google Antigravity, equipped with new Cheminformatics Skills, streamlines drug discovery by automating data curation, molecular analysis, and hypothesis generation, reducing a multi-day process to under an hour.
From Google DeepMind
CWE-bench v2 is a public benchmark with 100 security-related tasks covering 91 CWEs, designed to evaluate models' ability to find and fix security vulnerabilities. It includes rigorous quality checks and open-source transparency.
From Hugging Face
A study of a spec-driven coding workflow using Claude Code shows that structured, gated processes can achieve high task completion with minimal undeclared changes, though the orchestrator has notable limitations.
From arXiv
A study shows that increasing the strength of verifiers in LLM-based automated research does not improve discovery quality. Instead, the quality of proposals depends more on the diversity of the proposer pool.
From arXiv
This paper explores how LLM-proposed experiments require protocol repairs, or 'patches,' and introduces a framework for categorizing and auditing these changes.
From arXiv
EvoML is a new agentic framework for automated ML pipeline construction that uses validation gates and evolving expert skills to improve reliability and performance across diverse tasks.
From arXiv
A paper claims that frontier LLMs like GPT-6 Astra can directly interact with neural network weights, but the author argues that this is not as impressive as it seems and can be done by human researchers.
From Hacker News
A study finds that 34.82% of issues in IRBL benchmarks are mislabeled, significantly impacting performance metrics across multiple techniques.
From arXiv
A study shows that different data labeling strategies for vulnerability prediction models significantly impact their performance on classification, prioritization, and localization tasks.
From arXiv
A study shows coding agents do not consistently follow policies in Agent Context Files (ACFs). The effectiveness of policy wording in influencing compliance varied by style and model.
From arXiv
A study of 4,790 SE research papers and 244 researcher responses shows that AI is reshaping software engineering research, but the field is reorganizing around AI without critically questioning its impact.
From arXiv