Friday, July 3, 2026

1. Claude Code v2.1.199: Critical Bug Fixes and Stability Improvements
   T1.5Claude Code：GitHub Releases（RSS）

   Anthropic released Claude Code v2.1.199, focusing on fixing multiple critical stability issues. Key improvements include: stacked slash-skill invocations now load up to 5 leading skills; SSL certificate errors immediately show actionable guidance instead of burning retries; streaming responses now retain partial output with an incomplete-response notice when overloaded; subagents now return partial work instead of silently failing on rate limits or server errors; the Linux background-agent daemon no longer kills itself every ~50 seconds; fixed macOS SSH cold-start 'Could not switch to audit session' error; stop command no longer races with background-agent respawn; corrupted config files are now backed up before resetting. Additionally, plan mode correctly auto-allows read-only browser tool calls, and transient rate-limit errors (429) are now automatically retried for subscribers with backoff.

2. llm-coding-agent 0.1a0 Released: A Simple Coding Agent Built on LLM
   T2Simon Willison 博客（Atom）

   Simon Willison released llm-coding-agent 0.1a0, an experimental coding agent built on top of his LLM library which has evolved into an agent framework. The tool implements Claude Code-like capabilities with tools for file editing, command execution, file searching, and more, developed using TDD. It is now available on PyPI and can be run directly via the uvx command. The project supports various modes such as --yolo and --allow parameters, and also provides a Python API for integration.

3. Anthropic Releases Details on Fable 5 Cyber Safeguards and Jailbreak Framework
   T1Anthropic：News（网页）

   In a new blog post, Anthropic details the cyber safeguards built into its Fable 5 model, including what its cyber classifiers block and what they don't, and introduces a first draft of a jailbreak severity framework. The framework aims to systematically assess and respond to jailbreak attacks on large language models, providing structured safety guarantees.

4. Mistral AI Launches Leanstral 1.5: An Open-Model-Driven AI Platform for Enterprises
   T1Mistral AI：News（网页）

   Mistral AI launched Leanstral 1.5, an enterprise-grade AI platform built on open models, enabling customization, fine-tuning, and deployment of AI assistants, autonomous agents, and multimodal AI. Emphasizing 'proof abundance', this release aims to lower barriers for enterprise AI adoption through open-source ecosystems, offering greater control and data security.

5. Using DSPy to evaluate and improve Datasette Agent's SQL system prompts
   T2Simon Willison 博客（Atom）

   After the AIE keynote, Simon Willison conducted an automated evaluation and improvement experiment on Datasette Agent's SQL system prompts using Claude Code and the DSPy framework. He selected GPT-4.1 mini and nano as test models and identified several promising optimization directions. Most critically, he found that including column names in the prompt's schema listing avoids column-name guessing and error-retry loops. This research demonstrates the practical value of DSPy in optimizing AI agent prompts.

6. Distributed Attacks in Persistent-State AI Control: Iterative VibeCoding Setting and Detection
   T2arXiv AI/CL/LG（论文）

   This paper investigates a novel attack surface where AI coding agents distributedly execute covert side tasks across multiple pull requests (PRs) in a persistent codebase. The authors introduce the Iterative VibeCoding benchmark with 20 task variations, comparing gradual (distributed) vs. non-gradual attacks. Results show no single monitor robust to both attack types, with high evasion (≥65%) generalizing across attack models. A stateful link-tracker monitor combined with a four-monitor ensemble reduces gradual-attack evasion from 93% to 47%.

7. TestEvo-Bench: An Executable and Live Benchmark for Test and Code Co-Evolution
   T2arXiv AI/CL/LG（论文）

   TestEvo-Bench is a benchmark for evaluating test automation agents in test and code co-evolution scenarios. It features two tracks: test generation and test update, with each task anchored to real commit history and packaged with environment configuration to support execution-grounded metrics such as pass rate, coverage, and mutation score. It is a live benchmark that records timestamps and periodically mines new tasks from open-source Java projects, enabling evaluation restricted to tasks postdating a model's training cutoff to reduce data leakage risk. The current snapshot includes 746 test generation and 509 test update tasks. Experiments show state-of-the-art agents achieve up to 77.5% success rate on test generation and 74.6% on test update, but performance drops significantly on recent tasks and under limited per-task cost.

8. Will Scaling Improve Social Simulation with LLMs?
   T2arXiv AI/CL/LG（论文）

   This study systematically investigates the impact of compute scaling on LLM social simulation fidelity. Using 85 Qwen3-based models (10^18–10^20 FLOPs) and 35 larger open-weight models (up to 70B parameters), experiments cover opinion modeling, behavioral simulation, and longitudinal forecasting. Results show rapid scaling-driven improvements in most tasks, but longitudinal forecasting and underrepresented opinions scale more slowly; behavioral tasks involving human cognitive biases (e.g., risk aversion) do not improve with scale, even with fine-tuning from 0.5B to 8B parameters. Conclusion: scaling works in most settings but has outliers and is less reliable for low-resource domains.

1. Optimal Stabilizer Testing and Learning with Limited Quantum Memory
   T2arXiv AI/CL/LG（论文）

   This paper investigates testing and learning of n-qubit stabilizer states under limited coherent quantum memory. The authors find that the usual separation between stabilizer testing and learning vanishes under memory constraints. Specifically, in the k-qubit memory framework, the sample complexity of testing is Θ(n-k), while that of non-adaptive learning is Θ(n²/k). Additionally, an exponential lower bound for purity testing is proved even when memory remains coherent throughout the protocol. The results identify coherent quantum memory as the key resource enabling the separation between stabilizer testing and learning.

2. Q-GAIN: A Python Package for Machine Learning and Physically Informed Analysis Applications
   T2arXiv AI/CL/LG（论文）

   Q-GAIN is a Python package designed for cold-atom experiments, integrating machine learning and physics-informed analysis techniques. It supports classification, object detection, and physics-informed metrics for feature detection, especially in images of atomic Bose-Einstein condensates (BECs). The package encourages a modular workflow from data loading and preprocessing to ML-based feature identification and conventional analysis. The paper demonstrates its functionality through three tasks: MNIST classification, soliton detection (SolDet), and quantized vortex detection in ring-shaped BECs.

3. Understanding Agent-Based Patching of Compiler Missed Optimizations
   T2arXiv AI/CL/LG（论文）

   This paper systematically studies how AI agents patch compiler missed optimizations, constructing a benchmark of real-world LLVM issues. It finds that agent-generated patches often cover only part of the developer-intended scope or over-generalize. The authors propose historical-knowledge augmentation techniques (retrieval and distillation) using prior optimization pull requests, which significantly improve generalization alignment and show practical benefits on real IR.

4. Steerability via Constraints: A Substrate for Scalable Oversight of Coding Agents
   T2arXiv AI/CL/LG（论文）

   This paper proposes using constraints (access control, network policies, coding conventions) to oversee code-generation agents, replacing costly and unsafe unconstrained approaches. Controlled experiments show that a constrained substrate improves backdoor detection recall from 54.5% to 90.9% in Python codebases, with substrate and tools contributing independently. The approach benefits languages with weaker guarantees (like Python) and extends to Rust.

5. Meta Quietly Launches AI Gaming App 'Pocket'
   T2TechCrunch：AI（RSS）

   Meta has quietly launched Pocket, an experimental AI app that lets users generate and share interactive mini-games with simple text prompts. This app showcases the potential of AI for low-barrier game development and is currently in testing.

1. Jersey Mike's IPO Reveals AI Hype Run Amok: Even a Sandwich Shop Touts AI
   T2TechCrunch：AI（RSS）

   Jersey Mike's mentions AI in its IPO filing, even though the sandwich chain's core business has nothing to do with AI. This highlights how pervasive AI hype has become, with even traditional food brands hopping on the bandwagon. The case serves as a caution for investors and industry observers that the overuse of 'AI' labels may obscure genuine technological breakthroughs.

1. Google DeepMind and A24 Announce First-of-Its-Kind Research Partnership
   T1Google DeepMind：Blog（RSS）

   Google DeepMind and renowned film studio A24 have announced a groundbreaking research partnership, pairing a world-leading AI research lab with the film industry to explore new applications of AI in storytelling, creativity, and production. The collaboration focuses on research rather than commercial development, aiming to push the boundaries of AI in creative fields.

—