Tag
Andrej Karpathy built HackerTube—a YouTube clone featuring videos recommended by Hacker News users—with Claude Opus 5.5 in about two hours without writing any code, demonstrating a purely conversational AI development workflow.
提出一个基于LLM代理的多语言ASR数据采样框架，通过语言学感知的错误驱动优化在12种Common Voice语言上以更少数据和算力超越全数据训练效果，并显著加速收敛。
论文提出 LegoSR，一个统一的符号回归与微分方程发现框架，通过共享的方程中间表示整合七个求解器，并引入 LegoSR Agent 基于求解器知识和历史经验自适应选择求解器，在含噪 ODE/PDE 基准上接近七个专家调优求解器的并集性能，较通用 agent 最高提升 29 个百分点。
本文介绍 PaperFormalizer——一个基于 Lean 的作者参与式自动形式化流程，可从 LaTeX 论文出发构建语句依赖图、审计语句并将其形式化证明，同时在语义存在歧义时让作者做出决策，从而产出经 Lean 验证的论文版本。
开源了 PuedoBot，一个基于确定性扫描加双 pass LLM 流水线的 CLI 安全审计工具，第二个 pass 专门反驳第一个 pass 的发现以减少 SAST 误报。
WEDGE 是一个弱点引导的导航学习框架，利用 LLM 从预训练策略的成功与失败轨迹中诊断失败模式，并据此生成针对性训练环境，在 ProcTHOR 和 HM3D 上显著优于随机与难例挖掘基线。
LongHarness 是一个模型无关的智能体框架，通过「证据-结论协同演化」和「证据生命周期管理」机制，在超过 1 亿 token 的文档语料上进行有状态的证据管理，显著提升了长上下文推理任务的准确率（如 OfficeQA Pro V1 上从 15.8% 提升至 60.2%），并超越了 Claude Code 和 Codex 等现有智能体框架。
该论文提出Image Factuality Assessment任务及EviScene多模态ReAct智能体，通过视觉检查、外部检索与证据分析的交替推理来判定图像所描绘场景是否事实成立，并发布了包含10,174张图像的EviLens专家审阅基准，在Qwen3.5-9B上实现了79.05%的平衡准确率。
SWE-Journey 提出一个更贴近真实场景的编码助手评测基准，通过弱到强合成流水线构建长时程编码任务，并利用从真实交互数据挖掘的用户画像与模拟智能体复现多轮交互。实验显示模型面对软件架构师时通过率超75%，但面对非程序员时不足25%。
Goldsmith, an agentic annotation pipeline presented in an arXiv paper, converts small expert-annotated gold sets into structured annotation definitions optimized via an executable loss-guided loop, outperforming direct rewriting, OPRO, APE, and PromptBreeder.
Agentic-GER proposes an LLM-based agent for correcting domain-specific terminology in long-form speech transcripts using global context and selective re-transcription, achieving significant improvements in ASR accuracy for Chinese and English.
A tweet promotes a 'Jev' prompt for LLMs and AI agents that analyzes users to optimize their AI setup and save time and money.
ClinAgent is a ReAct-based agent system using agentic RAG to enable natural language querying of clinical trial information, with evaluation across multiple LLM backends showing complementary strengths.
CityPlanner introduces a unified sandbox environment and atomic-task reinforcement learning for executable urban planning, outperforming heuristic, task-specific RL, and general LLM-agent baselines on a real-world benchmark.
A developer describes building a read-only analytics agent for restaurant POS systems using Node.js, TypeScript, and Gemini Flash, seeking advice on tool selection, grounding techniques, and patterns for implementing write actions.
This paper diagnoses algorithmic mode collapse in code-level autonomous research loops and proposes Diversity-Aware Proposal Sampling (DAPS) as a lightweight mitigation to preserve semantic diversity and improve generalization.
MacroAgent introduces a novel framework using LLMs to design contour algorithms for macro legalization in VLSI circuits, achieving significant improvements in layout regularity and performance.
VortexChat is an LLM-based agentic framework that automates the inverse design of integrated photonic devices from natural language specifications, demonstrated by autonomously fabricating a terahertz multiplexer with high performance.
Task-CoEvolve is a novel approach for efficient LLM agent harness optimization that adaptively selects validation tasks to reduce evaluation costs while maintaining performance, achieving an 80% reduction in evaluations on benchmarks.
This paper presents an automated agentic approach using Large Language Models to synthesize interpretable Python feature extractors for algorithm selection in constraint satisfaction problems, outperforming expert-curated methods.