§ Coding

LLM CLI 0.32 ships reasoning trace visibility, OpenAI Responses API support, server-side tools, and improved logging for inference workflows.

LLMs recover missing semantic optimizations in C/C++ code that compilers overlook, evaluated on SeGaBench with 120 cases.

LLM capability for testing Terminal User Interfaces: benchmark across ratatui/Rust, bubbletea/Go, textual/Python showing 12% coverage in real applications.

Pattern Completion Bias benchmark measuring how repeated UI patterns degrade multimodal LLM accuracy on screenshot-to-code fill-in-the-blank tasks.

LiveEvalBench: agentic, adaptive evaluation framework treating web generation as interactive problem with diverse valid implementations.

CausalOPD distills step-dependent causal reasoning into smaller models using online process distillation with first-wrong-step supervision.

Multi-teacher distillation method decouples language-specific knowledge in multilingual LLM-based automatic speech recognition.

Survey of LLM applications across PDE workflows: discovery, numerical solver generation, diagnostics, and scientific feedback loops.

Alibaba releases Qwen 3.8 Max (2.4T params) and 27B open-weight models optimized for coding and collaboration tasks.

Steve Yegge reports Claude Opus 4.7 introduced behavioral regression—excessive self-modification cycles—that broke his Gas Town code-generation project.

OpenAI launches education plugins for ChatGPT Work and Codex targeting K-12 and higher education institutions.

SWE-Touch benchmark tests coding agents in shared workspaces with user code edits, exposing weaknesses in handling conflicting modifications during task execution.

David Crawshaw proposes using LLM-based automation for nightly software rebasing and testing via cron jobs.

Systematic comparison of training-free vs training-based intent classification methods for LLM prompt routing and domain specialization.

Antares: compact LLMs (350M–3B) for agentic vulnerability localization via SFT and RL on cybersecurity reasoning over code.

Simon Willison argues LLMs make open-source dev tools more practical by lowering code-reading barriers through AI-assisted comprehension.

Circles deploys OpenAI API and Codex for telecom personalization, achieving 22% ARPU lift and 9% churn reduction.

Cloud-ScPO derives preference pairs from LLM hidden-state geometry to improve mathematical reasoning without human annotation.

RefactorAssist agentic system improves LLM code refactoring reliability by detecting and correcting functional behavior changes pre-deployment.

FinHardBench evaluates LLM-generated FPGA hardware for financial trading (latency-critical 5-10ns); 33 tasks span module generation, pipeline tuning, adaptation.

Simon Willison releases smevals, an open eval framework for benchmarking models, prompts, and inference harnesses across configurations.

MOT-SR: LLM-based symbolic regression with data analysis mechanisms and multi-objective optimization for scientific equation discovery.

ARCTIC AI code critique system prioritizes correctness and security over style via intent prediction and drift detection.

AgenticRepair framework augments agentic AI program repair with security-specific context engineering for vulnerability patching.

Benchmark comparing 10 LLMs and 6 prompting strategies for automated code generation in fluid system simulation (WNTR, Modelica).

Change2Task generates executable coding agent tasks from merged pull requests with verified state reconstruction and validation.

PAIChecker reveals 13.6% misalignment between PR-issue pairs in SWE-bench-like benchmarks, questioning LLM code-resolution eval validity.

Empirical study of LLM ability to synthesize microservice architectures from natural-language requirements including service definitions and interactions.

HARGO framework applies heterogeneity-aware RL post-training to optimize LLM behavior on HPC tasks like data-race detection beyond SFT accuracy.

Locksmith Loop uses agentic test synthesis to validate COBOL-to-Java migrations via iterative branch penetration and witness search.

AgenticASR refines speech transcription via agent-based approach that revises text when context clarifies from later speech.

D. Richard Hipp (SQLite creator) draws historical parallel between SQL's labor displacement and current AI impact on programming roles.

DenseOn and LateOn: open-source 149M-parameter retrieval models trained on 1.88M supervised pairs; competitive on multilingual and code search.

SpecFirst framework improves LLM agent program synthesis from scratch by separating behavioral specification elicitation from code synthesis.

MindForge trains small LMs on whole-lifecycle program synthesis from scratch via source-free generation across full software development cycle.

Study shows LLM-generated Kubernetes security patches improve when prompted with runtime service topology context rather than isolated KSPM findings.

OpenAI field report documents how AI coding agents accelerate scientific computing workflows in genomics and adjacent domains.

RL for code optimization via learnable execution time rewards and DMC-Optim framework to overcome measurement noise and sparsity.

Custom harness engineering distributes security controls for AI coding agents without vendor lock-in, enabling organizational scaling.

Empirical analysis of attention patterns in LLM-based automated program repair to explain inconsistent patch generation performance.

Claude Opus 4.7 generates shuttling compilers for trapped-ion quantum computers from specifications, handling linear, junction, and general graph architectures.

Study of correctness degradation in code repair agents under forced revision loops; 30 HumanEval benchmarks show revision doesn't guarantee reliability.

User study (n=34) on explainable AI impact for LLM code review: XAI effects on developer trust in automated review reasoning.

Robotics learns from bitter lesson scaling; AI agents complete week-long programming tasks; OpenAI discovers accidental vulnerability in reasoning model.

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.

MineValiCoder addresses LLM-based TDD by mining test quality and using bipartite graphs to validate code without human test cases.

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.