§ Coding

Probe-driven test-time RL method for code generation via behavioral agreement on synthetic inputs instead of surface-form voting.

ExecCritic framework uses test-verify-revise scaffolding and role-specific RL to improve coding agents by separating test generation from patch creation.

MIT researcher uses GPT-5.6 Sol with Codex to autonomously run quantum computing experiments, analyze results, and calibrate qubits.

1Password reports 21% engineering productivity gain using OpenAI Codex for feature development and internal tooling.

ProcArena benchmark evaluates LLMs on multi-scenario PL/SQL development including direct generation, modification, debugging, and interactive multi-turn workflows.

DFlow enables verifier representation reuse in block diffusion speculative decoding, improving LLM inference by recovering computation from rejected tokens.

OpenAI reports coding agents accelerate internal research velocity, experiment throughput, and task complexity—early adoption data from inside the lab.

PROOF system enables LLM-driven codebase maintenance via verified natural-language specifications rather than direct code manipulation, abstracting topology into hierarchical representations.

Coupled-noise distillation enabling single-pass block generation in diffusion language models via noise-conditioned masking.

Simon Willison demonstrates using Blender's Python API with ChatGPT Codex on macOS to generate images via coding agents.

KOPA-Bench: 145 Korean public API tasks; EDGE synthesis method closes open-source LLM gap in multi-step tool-calling for on-premise agents.

LLM-based decompilers optimize for recompilability over semantic accuracy, potentially masking vulnerabilities that traditional decompilers expose as placeholders.

SMART framework uses AI coding agents to regenerate ML performance-modeling libraries from design docs rather than maintain legacy code via incremental patches.

BUGSTONE-E2E framework mines CVE patch history to detect unknown vulnerabilities by extracting and generalizing code fixes from public databases.

Speculative Uncertainty method recovers failure signals for coding agents using draft-model cross-likelihoods, enabling safe execution without logit access.

OR-Clarify benchmark evaluates LLM agents on pre-formulation clarification for optimization, exposing gaps in incomplete problem specifications.

Verifier-guided framework combines QLoRA, MoE routing, and symbolic solvers (Z3, FOL) for explainable LLM reasoning in education.

AxQM benchmark: 1,019 formal proof-synthesis tasks for autoformalization in finite-dimensional quantum mechanics.

OpenAI releases GPT-6 Astra with SOTA computer use and coding; 2.5x higher token cost but lower per-task cost despite reduced interpretability.

SWE-Gate benchmark evaluates coding agents on review-constraint compliance beyond functional correctness in repository-level tasks.

Terminal-Universe reconstructs executable terminal environments from agent trajectory tool-execution history for scalable post-training.

Two-stage OPD-then-RL sequential scheme outperforms joint on-policy distillation and RLVR fusion for reasoning LLM post-training on logic and math tasks.

LLMs over-edit code during repair; study of 400 BigCodeBench problems shows widespread over-editing even in frontier models.

Instruction duplication: repeating procedural instructions at inference time improves deterministic compliance across 7 instruction-tuned models without retraining.

Playco reduced manual fixes by 50% in game prototyping using GPT-6 Astra, building three themed variants from shared foundation.

OpenAI releases GPT-6 Astra with improved computer use, coding, cybersecurity, and science capabilities.

llm-anthropic 0.28 displays Claude reasoning traces by default and adds refusal exception handling.

Nemotron-3-Nano/Ultra-CC models trained on 22k curated competitive programming problems with SFT, RL, and GenCorrect test-time strategy for IOI/ICPC performance.

CodePoisonRAG demonstrates black-box knowledge poisoning attacks on retrieval-augmented code generation systems, revealing supply-chain vulnerabilities.

Repo-To-Skill distills GitHub repositories into compact verified skills for autonomous ML research agents; operational knowledge layer.

NE-R1: adaptive retrieval-augmented NER with on-demand mechanism for long-tail and domain-specific entity recognition.

SALA framework improves in-context learning for complex reasoning by learning task-specific logical alignment instead of matching fixed demonstration patterns.

DiffIE uses conditional discrete diffusion for multi-output open information extraction, replacing autoregressive generation with parallel trajectory sampling.

Paint.NET developer credits Claude AI for reverse-engineering Direct2D wrapper to run on WINE; demonstrates LLM utility for systems programming.

Claude Fable 5.1 achieves 52.6% on Terminal-Bench-Science 0.1; Willison reports improved coding/creative task performance vs. prior versions.

Adaptive retrieval method for repository-level code generation identifies critical tokens needing fine-grained context via token-aware RAG.

Empirical study on using LLMs to design video coding tools via generation-evaluation loops, tested on Planar mode intra prediction.

Harness-of-Harness: framework for multi-day autonomous software development with continual improvement via iterative planning-coding-testing loops.

Self-Routing post-training framework routes LLM samples to GRPO, distillation, or regularization based on rollout correctness without external teachers.

Empirical study on Qwen2.5-3B and Phi-3.5-mini shows hints in code generation provide steering, not missing information; unrelated hints rescue 40-50% as many failures.

Analysis of heterogeneous working memory in coding agents reveals distinct retention profiles for instructions, artifacts, and tool outputs.

GMTS: gradient magnitude-based token selection improves RL training for LLM reasoning tasks.

Discrete diffusion language model framework for targeted PDE solver code repair via localized re-masking and infilling.

Empirical study of Claude Code plugin marketplace structure and maintenance dynamics, examining co-evolution of NL instructions, scripts, and configs vs. traditional packages.

Sequential test-time scaling outperforms parallel sampling for LLM machine translation, offering higher ceilings and sample efficiency on smaller budgets.

NL2AGBench evaluates LLM auto-formalization of natural-language geometry problems into AlphaGeometry's domain-specific language for theorem proving.

Prove2Me platform enables AI agents and humans to collaboratively formalize mathematics in Lean 4, lowering barriers to proof verification at scale.

Systematic cost-accuracy ablation of 17 text-to-SQL pipeline configurations reveals marginal contributions of individual components in ICL models.

LoopArena benchmarks models as runtime loop controllers for coding agents, isolating loop guidance quality from agent capability.

PASK optimizes KV persistence for structured generation by tracking parser transitions to preserve schema-critical decoding signals.