GRPO, Dr. GRPO, and DAPO Are All the Same Operation: Group-Standard-Deviation IdentityRareDxR1: Autonomous Medical Reasoning for Rare Disease Diagnosis Surpasses Human AnnotationSNAP-FM: Sparse Nonlinear Accelerated Projection Enforces Physics Laws in Generative ModelsByteDance Releases Seed2.0 Model Series for Complex Real-World TasksMnemosyne: Agentic Transaction Processing Validates and Repairs AI-Generated WorkflowsMaking Failure Safe: A Constrained, Verifiable Agent Framework for Web Data CollectionVercel's Andrew Qu: Agents Are a New Kind of Software Requiring Skills and SandboxesGitHub Reaches Inbox Zero on 20,000+ Secret Scanning Alerts in 9 MonthsSkill Engineering: Against One-Shot AI Design; Agents Still Need Human GuidanceGeoffrey Litt: 'Understand to Participate' Is Key When Collaborating with Coding AgentsSimon Willison Releases llm-coding-agent 0.1a0: Simple Coding Agent Built on LLM LibraryFable Returns with New Sonnet VersionAdobe Experiments with 'Agentic Sites' That Auto-Generate Pages Per Visitor IntentUsing DSPy to Evaluate and Improve Datasette Agent's SQL System PromptsGRPO, Dr. GRPO, and DAPO Are All the Same Operation: Group-Standard-Deviation Identity

A new paper introduces RareDxR1, an autonomous medical reasoning system for rare disease differential diagnosis. It identifies precise phenotypes from unstructured patient symptoms and performs complex reasoning, outperforming existing AI and human annotation levels.

A new arXiv paper introduces SNAP-FM, an accelerated projection method that enforces conservation laws, boundary conditions, and nonlinear invariants in generative models. It uses sparse nonlinear accelerated projection to ensure outputs automatically satisfy underlying physics while retaining scalability. Compared to traditional physics-informed neural networks (PINNs), SNAP-FM significantly improves convergence speed and accuracy on multiple physical simulation benchmarks. The work directly benefits scientific computing and engineering simulation requiring physical explainability.

ByteDance released the Seed2.0 model series, designed to tackle complex real-world tasks. The paper starts from identifying genuine user needs and builds a reliable, forward-looking evaluation system via selection and abstraction. Seed is ByteDance’s foundational LLM family, and Seed2.0 is optimized for task complexity and scenario diversity. The paper is public, but no model weights or API release timeline is provided. It offers reference for developers deploying AI in unstructured, multi-step scenarios.

A new arXiv paper proposes Mnemosyne, an agentic transaction processing framework for validating and repairing AI-generated workflows. LLMs, solvers, and agent teams increasingly generate workflow actions, repairs, and plans, but generated actions can be syntactically valid yet stale, infeasible, conflicting, or destructive. Mnemosyne introduces structured validation at the transaction level to detect and fix these issues. The work directly applies to teams relying on AI-automated business processes, code repair, or experiment steps.

A new arXiv paper proposes a constrained, verifiable agent framework for generating reliable web scrapers from natural-language requirements. Direct generation with LLMs remains unreliable due to dependency errors, broken selectors, schema mismatches, and heterogeneous page structures. The framework introduces constraints and verification to prevent common errors, ensuring robustness in open-web environments. It practically assists analysts and developers needing automated data collection.

Vercel’s Chief of Software, Andrew Qu, discussed how its agent framework ‘eve’ was built and argued that agents represent a new software paradigm requiring skills, sandboxes, and agent-readable websites. He emphasized the shift in how software will be created and interacted with.

GitHub achieved inbox zero on over 20,000 secret scanning alerts across 15,000 repositories in nine months. The approach involved separating signal from noise, building remediation workflows, and streamlining alert handling.

Paul Bakaus discussed ‘skill engineering’ and the importance of human judgment in the ‘loopmaxxing’ era. He argued against one-shot AI design, stating that agents still need human steering and continuous guidance.

Geoffrey Litt introduced the concept ‘Understand to participate’ at AIE. He argued that as coding agents construct larger and more sophisticated changes, developers must actively understand the agent’s logic to avoid the challenge of passively managing tickets.

Simon Willison released llm-coding-agent 0.1a0, a simple coding agent built on his evolving LLM framework (now an agent framework). Part of the Fable 5 experiment, the code is open-sourced on GitHub using the python-lib-template-repository template.

Fable is back and has introduced a new Sonnet version. Further details are pending, but the announcement has generated community interest.

Adobe is experimenting with ‘agentic sites’ that generate pages around individual users’ intent on the fly. Carlos Sanchez discussed this future of the Web at the AIEWF conference, suggesting a shift from static pages to intent-driven assembly.

Simon Willison used the DSPy framework to evaluate and improve the SQL system prompts for Datasette Agent. Inspired by an AIE keynote, he initiated an async research task to test if DSPy could enhance the agent’s SQL generation performance.