#AI编程
9361 related articles
Should Coding Agents Follow Precedent or Task Intent?
A study of 5,625 trials reveals coding agents prioritize repository precedents over task intent, but written priority rules (59/60) outperform code examples (38/60) when intent conflicts.
CopyLoop: A Training-Free Decoding Acceleration Method for Looped Transformers
Looped Transformer decoding is slow? CopyLoop introduces a training-free context-copy proposal and shallow-layer screening method, achieving up to 10.70x decoding speedup across multiple model families, with strong performance on code generation and math reasoning tasks.
The Digital Trap of AI-Written Papers: Provenance Auditing Matters More Than Fabrication Detection
Re-analysis of 540 AI research agent paper judgments finds numerical errors mostly misattribution, not fabrication. Deterministic checker certifies claims with >90% precision, urging provenance-first audit frameworks.
ForeSCITE: Using Temporal Trajectories to Predict Future Biomedical Discoveries
ForeSCITE learns semantic temporal trajectories from ~3M biomedical conclusions, generates future hypotheses via projection and decoding, beats five baseline types in retrospective evaluation, with 37 CAR-T hypotheses endorsed by experts.
QUANTA: Agentic AI Enables Autonomous Measurement of Quantum Materials
QUANTA is an agentic AI framework for quantum materials characterization that combines retrieval-augmented literature with dedicated agents to enable closed-loop autonomous measurement—experiment planning, troubleshooting, and data interpretation—on the PPMS platform, validated on TaP and Co2MnGa.
From Grind to Grasp: Rethinking Agent Work and Depth
The study *From Grind to Grasp* proposes rethinking intelligence through work budgets and information allowances, defines a grasp metric to distinguish grind from grasp, and warns of organizational risks from agent collectives.
Claude Moves Global Law into Git: A New Paradigm for Semantic, Versioned Legal Queries
Developers use Claude Code to turn global law into Git repositories, enabling semantic, versioned, and Agent-queryable legal data. The project topped Hacker News, is open-sourced under CC0, and provides a reliable data source for AI legal Q&A.
AI Agent Performs Mouse Brain Surgery: Claude Sonnet 5 Succeeds and Patient Recovers
Reddit buzz: a Claude Sonnet 5-powered AI agent completed mouse brain surgery with recovery. We analyze the technical background, multi-source information, and potential impact.
Soei: Building Observability of Evidence and Intent for Multi-Agent Programming
Soei studies the reliability of the open-source audit tool Agenttrail, quantifying evidence gaps across 45 real programming agent sessions: 23.4% of edit segments lack independent follow-up checks, 57.8% of completion markers lack evidence, and it proposes a hash-chained ledger and evidence gating.
Self-Evolving Agent ExLabel: Improving SWE Dataset Quality Assessment
ExLabel is a self-evolving, experience-guided LLM assistant for assessing task clarity and test coverage in SWE datasets, with significant F1 gains on the SWE-TQ benchmark; the downstream model solves 66.2% of SWE-bench Verified problems.
Resume Guide for Programmers Transitioning to AI: From an HR Perspective to Job Requirement Matching
How should programmers transitioning to AI write a resume? This Bilibili tutorial gives five key preparation actions: set aside two or three days to review skills and projects, choose a simple template, reverse-engineer the tech stack from job sites, emphasize LLM capabilities, and do a final check from an HR perspective.
Building a RAG Knowledge Base: A Complete Guide from Architecture to Retrieval Evaluation
A systematic overview of the full RAG knowledge base workflow: from overall architecture and mature retrieval solutions to key details of retrieval evaluation and who should learn it, helping developers build a path from beginner to hands-on practice.
Turn Screenshots into Web Code in Seconds: Anyone Can Write HTML with AI
A Bilibili creator demonstrates using AI to turn webpage screenshots directly into HTML and CSS code without manual coding. The article analyzes the tool's usage, new features like video-based app cloning, iterative refinement, and the real boundaries of AI giving zero-foundation users programming ability.
WorkBuddy vs Codex vs Claude Code: AI Agent Selection and Office Automation in Practice
Compare WorkBuddy, Codex, and Claude Code across positioning, thresholds, and target users; learn three selection criteria for office automation and WorkBuddy's practical path.
Human-in-the-Loop in LangGraph: Where Should interrupt() Go? Six do/don't Pairs
A practical Reddit note on LangGraph human-in-the-loop: where to place interrupt(), how to write approval payloads, why side effects must come after interrupt(), checkpointer and thread_id pairing, and why approval rate isn't a success metric.
Android Tracing 2.0: A Tracing Revolution with 7x Speedup and Context Propagation
Android Tracing 2.0 explained: 7x faster tracing API with cross-coroutine and thread context propagation, custom metadata, pluggable backends. Wyze sped up startup 45%, R8 25%, Compose 1.13 recomposition tracing now available.
Tech FrontiersWhat Should CLAUDE.md Contain? The Truth About "Don't Make Mistakes" Instructions, Tested
Testing what CLAUDE.md should contain: vague rules fail, "don't make mistakes" saves 27% tokens inexplicably. Three AI coding tools reveal six practical principles.
Reddit Discussion: Real-World AI Agent Use Cases Shared for Business School Classrooms
A Reddit business school lecturer seeks real AI Agent use cases for classroom teaching. This article compiles typical AI Agent applications in professional and personal scenarios and offers three analytical dimensions for selecting classroom cases, helping readers understand the practical value and limitations of AI Agents.
Node.js 26 LTS Lands on Vercel: HTTP/2 by Default and the Temporal API
Vercel makes Node.js 26 LTS generally available for builds and Vercel Functions, with HTTP/2 enabled by default in built-in fetch (Undici 8), the Temporal API, and V8 14.6's Map.prototype.getOrInsert() — plus automatic and manual upgrade paths and an HTTP/1.1 fallback.
Agentic Engineering with Claude Code and Android CLI
Google's Android team shares how to do agentic engineering with Claude Code and Android CLI: build the right environment, keep context around 150K tokens, inject info precisely via hooks and skills, and use layered verification loops to raise benchmark pass rates from 22% to 90%.