The digest for 2 November 2026
Download it as an EPUB, or keep a permanent link to it
Frontier LLMs like Astra can interpret and modify neural network weights without training or gradients, improving performance in games and math tasks. This suggests a new paradigm for weight-space programming.
From Lobsters
LLM 0.37 includes new models, improved embedding support, and AI-assisted PR triage, with a new default model that's half the cost of the previous one.
From Simon Willison
A new technique allows LLMs to directly read and edit neural network weights, bypassing traditional code-based training. This could enable autonomous model improvement without human intervention, though it's limited to simple tasks for now.
From Latent Space
VeriHarm-Bench is a new benchmark that measures harmful success in LLM coding agents, where agents complete tasks correctly but cause unintended harm. It shows that even high-performing models like GPT-5.6-Sol and Claude-Opus-5 struggle with safety.
From arXiv
COVE is a new tool that improves the accuracy of LLM outputs by preventing errors before generation and validating results afterward, without requiring extra LLM calls.
From arXiv
CR-Bench-v2 is a new benchmark for evaluating AI code review tools, showing that even the best systems detect only about one-third of known defects.
From arXiv
APG is a new tool that enables LLM-based coding agents to perform static analysis on C/C++ code by leveraging LLVM-IR graph databases. It provides 38 atomic capabilities for analysis and repair, outperforming existing methods in benchmark tests.
From arXiv
A new benchmark shows coding agents struggle with chained requests, completing only 51% of tasks when requests are linked, compared to 95% when handled individually.
From arXiv
SWE-Trail is a new method for collecting coding agent trajectories that uses past data to guide future collection, improving performance without exposing history to the model during training.
From arXiv
VeriBase is a framework that automatically builds reusable theorem libraries for software verification by generalizing code and theorems from successful verification tasks. It improves success rates and reduces proof completion time.
From arXiv
CrossSWE is a new benchmark testing coding agents' ability to make coordinated changes across multiple repositories. It shows that even the best models struggle with cross-repository consistency, achieving only 58.48% Pass@1.
From arXiv
Momentic is an AI QA tool that tests apps end-to-end by using the app itself to verify changes, reducing manual testing time and improving reliability.
From Hacker News
Kyle Kingsbury argues that the use of large language models (LLMs) undermines trust in software documentation and systems. Maintainers must find ways to ensure accuracy and demonstrate quality in an era where documentation can be easily faked.
From Simon Willison
OpenAI developed Pando, a Git-compatible backend system to handle the scale of development required by Codex.
From OpenAI
Qodo enhances AI code review by incorporating company-specific knowledge and coding standards into Codex, enabling developers to catch critical issues before merging code.
From OpenAI
A new benchmark called DistDebug-Bench evaluates LLM agents' ability to diagnose bugs in distributed systems, revealing significant gaps in their current capabilities.
From arXiv
A study shows that coding agents can lose performance on earlier tasks when evolving to new ones, highlighting the challenge of continual self-evolution without catastrophic forgetting.
From arXiv
A new study explores how LLMs can repair outdated tests when source code evolves, showing that combining test failure symptoms, code changes, and context improves repair success. The research also introduces a reasoning taxonomy that enhances LLM-based test repair efficiency.
From arXiv
IssueLighter is a new LLM-based tool that improves the accuracy of identifying invalid issue reports in open source projects by considering issue types and providing explainable rationales.
From arXiv