§ Coding

Ruff v0.16.0 enables 413 default linting rules (up from 59), breaking existing CI pipelines and catching syntax/runtime errors previously uncaught.

Anthropic releases Claude Opus 5 with improvements in agent execution, coding, and professional tasks.

Petri-net-guided LLM test generation for concurrent Rust APIs addresses shallow test synthesis by integrating formal models with executable test concretization.

Analysis of code model representations shows Qwen2.5-Coder and DeepSeek-Coder align on grammatical concepts across Python/Rust, with task-driven specialization.

Open-source evaluation framework for open-weight LLM agents on longitudinal data tasks, addressing privacy constraints in research deployments.

TTEL: inference-time algorithm using token-level error localization and environment feedback for efficient test-time scaling.

Linear probes on hidden states detect early non-convergence in chain-of-thought reasoning; DeepSeek-R1-Distill-Qwen-7B shows 90.3% converged vs 6.6% non-converged AIME accuracy.

LLMs extract natural-language abstractions from solution traces to improve problem-solving via retrieval and RL-augmented training on math tasks.

PyroDash enables cost-efficient SLM-LLM collaborative inference by training SLM to emit control tokens requesting frozen LLM handoff during generation.

NTT DATA deploys ChatGPT Enterprise and Codex across 9,000 employees to reduce incident analysis time to 30 minutes.

CodeRescue optimizes cost-aware routing for coding agents, determining when to retry vs. escalate after execution failures.

MaLoRA and MaLOR introduce dynamic, recurrent state-space adapters for task and token-level LLM adaptation, improving reasoning via selective modulation.

Off-Context GRPO uses privileged training information (solution prefixes) to enable RL with verifiable rewards on hard reasoning problems avoiding zero-gradient plateaus.

Anthropic engineers discuss Claude Code, Claude Tag Slack integration, coding agent security, and internal tool usage in fireside chat.

Coding agents lower ROI threshold for reverse-engineering home automation, shifting economics of personal automation projects despite maintenance risk.

SWE-Pruner Pro: prunes long context in coding LLMs by extracting internal relevance signals, improving efficiency over external classifiers.

PPL-Factory: task-aware perplexity-based data selection for efficient LLM fine-tuning across reasoning and language tasks.

OR Else: smooth one-sided saturation rule replaces PPO clipping for stable LLM post-training with reduced gradient discontinuities.

TRIM reduces verbose AI-generated code by minimizing agent trajectory artifacts through search-process cleanup.

SGA module detects and fixes geometric errors in LLM-generated pedagogical animation code via symbolic scene graphs.

BayesPO formulates prompt optimization as Bayesian posterior sampling over discrete tokens using parallel-tempered MCMC.

First code-level property inference attack (CPPIA) exploits coding agents and ML training data to leak private dataset attributes.

GPT-5.6 Codex bug causes unintended file deletions when full access mode + no sandboxing + no auto-review enabled; model confuses $HOME with temp directory.

MM-IssueLoc benchmark isolates visual evidence impact in multimodal software repository issue localization across 23 languages and 652 instances.

OpenAI documents internal use of Codex for custom tool development, prototyping, and creative ideation workflows.

Simon Willison ports Grok's Rust Mermaid-to-Unicode renderer to WebAssembly for browser use via Claude Code.

Generative Compilation: decoder-level compiler feedback for LLM code generation without white-box model access.

Placebo-controlled methodology (PoPE) to measure whether small code LLMs can actually use execution error feedback to repair code.

FileMark VSCode extension uses line-anchored feedback to reduce token generation in Claude Opus (22%) and Sonnet (58%), cutting code-editing latency and cost.

OpenAI integrates Codex capabilities into ChatGPT, shifting toward a multi-purpose platform rather than pure chat interface.

Codex usage grew 10x to 7M users in 6 months; article questions whether it has outpaced Claude Code amid sparse adoption metrics.

Simon Willison documents productivity spike in Datasette project correlating with Opus 4.8 and GPT-5.5 releases via GitHub commit frequency analysis.

LLMs applied to EDA front-end design via HDL generation and agentic AI frameworks like OpenClaw for chip design automation.

Tokenizer vocabulary transplantation fixes Bengali ASR failure in Moonshine by replacing English-centric byte tokenizer with BanglaBERT.

Large-scale study of LLM CLI coding-agent failure trajectories reveals onset patterns and recovery mechanisms, treating failure as temporal process not final outcome.

Pipeline combines Ghidra reverse engineering, anchor-based retrieval, and LLM reasoning to recover source code from stripped binaries.

Cohere introduces Dynamic Speculative Decoding (DSD) that optimizes K parameter selection based on hardware constraints to improve inference efficiency.

ProjAgent uses procedural similarity retrieval to improve repo-level code generation by matching functional logic across codebases.

llm-meta-ai 0.1 released, enabling CLI access to Meta's muse-spark-1.1 model via Simon Willison's llm tool.

llm 0.31.1 patches tool call JSON error in OpenAI Chat Completion endpoints with empty arguments.

Kenton Varda banned AI-generated PR/commit messages after finding them missed high-level context while restating obvious code details.

Pyligent training framework treats reasoning as validated search with backtracking, enabling models to correct mid-inference and recover from failed branches.

OpenAI identifies validity issues in SWE-Bench Pro coding benchmark, questioning reliability of popular AI model evaluation metric.

RuBench 1.0 repository-level coding benchmark with 25 natively-authored Russian task specifications across Python, PHP, TypeScript, JavaScript projects.

Code agent framework for automated software verification outperforms fixed proof strategies, proving larger fraction of Coq theorems than prior LLM approaches.

UI2App benchmark evaluates LLM-generated web applications on interaction fidelity, not just visual rendering, using image-driven inputs.

Spider 2.0-AIFunc benchmark of 465 text-to-SQL instances covering AI-native SQL functions on Snowflake platform.

Empirical study on fine-tuning and metrics for neural decompilation of Dart AOT binaries with 154-task HumanEval-Dart benchmark.

TREK uses distillation for exploration support expansion in GRPO, enabling reasoning on hard prompts via black-box or white-box teacher trajectories.

Iterative pseudo-labeling improves Mandarin-English code-switching ASR by leveraging unlabeled data via semi-supervised training.