Tag
Fable 5.5 reportedly generated a complete animation in a single pass using only code. Chetaslua said that standard Opus 5.5 and Fable 5.1 produced noticeably worse results with the same prompt and provided a method for verification, but the report has not yet been officially confirmed.
Lab2Sim introduces a benchmark of 218 science experiment videos with evidence-grounded specifications, along with EduVideo2Code, which encodes procedures as state machines to guide verified code generation for interactive virtual experiments, improving apparatus fidelity by 24.15 points.
Claims have surfaced that Opus 5.5 can recreate OpenAI's new promo video entirely in code, without any video generation model — relying solely on Three.js, GLSL, and an SDF heroine.
提出了异步流映射语言模型（aFMLM），通过token级时间调度将预测顺序编码进单步生成的隐藏计算中，在Sudoku和GSM8K代码生成任务上刷新了一步生成的最先进性能。
DeepSeek open-sourced an MIT-licensed research prototype, dsh-superagent-prototype, featuring a Router-Worker super-agent architecture for coding tasks that aims to minimize human attention. Self-reported tests show faster completion of slow tasks versus a single-session baseline, but the team stresses it is research code, not production-ready.
The reverse engineering field welcomes craft-anything, a pure-Rust open source project that performs full-coverage observation of closed-source software at the system behavior level (system calls, memory state machines, network protocol packets), reconstructs a business architecture interaction graph, and generates modular specifications and skeleton code — compressing days of manual packet capture analysis into minutes. The project is currently in early alpha and is dual-licensed under MIT/Apache-2.0.
Craft Anything 是一个用纯 Rust 编写的开源工具，能够通过真实的检查工具研究任意网站、应用或游戏的运作方式，然后让编码 AI 代理为用户构建自己的版本，并将结果与原版对比校验。它提供 284 个 CLI 与 MCP 工具，方便任意代理驱动，目前处于早期 alpha 阶段。
A blog writeup describing how the author post-trained an autoregressive language model to autocomplete and remix 140-character JavaScript 'Dweets' from Dwitter.net, inspired by classic demoscene sizecoding effects like HellMood's 'Memories'.
An evaluation study examining where natural-language intent is lost when LLM agentic pipelines synthesize Solidity smart contracts, finding that pre-generation information elicitation matters far more than post-generation steps and proposing an elicitation-first pipeline that reaches 72.9% fidelity on 500 held-out contracts.
MiniMax AI's M3.1 Flash model was revealed behind the 'Space Bunny' demo, and was used in MiniMax Code to build a playable space arcade game from a single prompt in 17 minutes. The model autonomously wrote a playtest suite that passed 29 checks across desktop and emulated iPhone, with the whole build costing roughly 15 cents in tokens.
A deep-dive blog post exploring how GCC lowers switch statements, showing its transformation into jump tables (including offset handling and a ~8-entries-per-label sparsity threshold) and falling back to decision trees of comparisons for sparser cases.
RoboArchitect automates the design of robotic agent harnesses by using a coding model to iteratively revise control code around a frozen policy, boosting success on unseen scenes from 9.7% to 43.3% and on unseen tasks from 12.7% to 40.2% without any model weight updates, while also evolving persistent memory across episodes.
该预印本提出 BIEvolve，通过契约引导的静态依赖分析与跨批次数值一致性动态验证，为 LLM 内核生成代理加入批不变性(BI)保障，将 BI 通过率从 67.6% 提升至 91.9%，同时保留 91.8% 的基线加速收益。
This paper introduces RefactorProof, a stress test that derives 458 Lean-certified refactoring variants from VERINA tasks to measure whether LLM-generated proofs remain valid under behavior-preserving program changes. It finds substantial brittleness across reference proofs and specialized provers (DeepSeek-Prover-V2, Goedel-Prover-V2), while noting that most broken proofs can be repaired with small edits.
VibeVideoBench introduces a benchmark of 239 tasks evaluating multimodal agents that generate videos by writing rendering code, finding that GPT-6 Astra and Claude Opus 5.5 lead but that time-management failures cause most lost videos.
CCORE 发布了基于 Qwen3.5-4B 的独立微调模型 Qwen3.5-4B-3D-Coder，专注于通用代码编写与高质量 Three.js 交互式 3D 场景生成，目前已提供 Q4_K_M 和 Q8_0 量化版本，完整 FP16 权重和训练数据集将在训练完成后公开。
Nanbeige4.3-33B 是南ベージ系列的旗舰教师模型，宣称在约30B参数级别的开源模型中达到最先进水平，专注于复杂编码与专业协作智能体任务，并为后续小模型蒸馏（如 Nanbeige4.3-3B）奠定基础。
Google Research introduces SimHack, a simulator that predicts when detectable reward hacking will become dominant during RL training of coding LLMs, before running costly full-scale training. It achieves 11.3% median relative error in onset prediction and 0.85 rank correlation across recipes, and can screen curricula to find ones that train 96.6% longer without hacking.
Recursive 发布了 Flint，一款 40B 总参数、仅激活 2.5B 参数的 MoE 本地 agentic 编码模型，在 50B 以下 MoE 模型中于 SWE-Bench Pro v2、Terminal-Bench 2.1 等基准上领先，并借助自研 Eureka Machine 自动化研发系统在四个月内完成构建。
GitHub Next 与微软研究院推出 Chopin 和 Proof-Driven Development Kit（PDD Kit），可将开发计划转化为形式化规范，并证明 AI 生成的代码对所有输入均符合规范，使团队无需逐行审查即可将 AI 生成的代码安全上线，目前已进入研究预览阶段。