Archives

mai 2026

3148 articles

GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> Imbad0202 /</span> academic-research-skills

Dépôt GitHub proposant un workflow de recherche académique intégré à Claude Code. Cycle complet : recherche → rédaction → révision → finalisation. Outil pour structurer la production de contenu académique avec assistance IA.

Claude CodeOutilsPrompt engineering
SIG
35
HYP
45
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> obra /</span> superpowers

Superpowers est un framework et une méthodologie de développement logiciel basée sur des agents autonomes. Le projet propose une approche structurée pour construire des systèmes multi-agents.

Agents IAMulti-agentsOpen source
SIG
35
HYP
45
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> rtk-ai /</span> rtk

rtk est un proxy CLI en Rust qui réduit la consommation de tokens LLM de 60-90% sur les commandes dev courantes. Binaire unique, zéro dépendance.

OutilsInfrastructureGénération de code
SIG
72
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> rohitg00 /</span> agentmemory

AgentMemory : système de mémoire persistante pour agents IA de codage, basé sur des benchmarks réels. Repo GitHub trending proposant une architecture de stockage et récupération de contexte pour améliorer la continuité des agents autonomes.

Agents IAGénération de codeOutils
SIG
45
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> anthropics /</span> claude-plugins-official

Anthropic publie un répertoire officiel de plugins Claude Code de haute qualité, gérés directement par l'entreprise.

ClaudeClaude CodeOutils
SIG
65
HYP
15
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> dmtrKovalenko /</span> fff

fff est un toolkit de recherche de fichiers haute performance conçu pour les agents IA, Neovim, Rust, C et NodeJS. Optimisé pour la vitesse et la précision.

Agents IAOutilsOpen source
SIG
45
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> screenpipe /</span> screenpipe

Screenpipe (YC S26) enregistre localement et en continu tout ce que l'utilisateur fait, dit et entend pour donner aux IA accès à son expérience. Traitement privé et sécurisé sur l'appareil.

Agents IAOutilsOpen source
SIG
65
HYP
55
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> 0xPlaygrounds /</span> rig

Rig est un framework Rust pour construire des applications LLM modulaires et scalables. Le projet gagne en popularité sur GitHub Trending.

Open sourceInfrastructureGénération de code
SIG
45
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> rtk-ai /</span> rtk

rtk est un proxy CLI en Rust qui réduit la consommation de tokens LLM de 60-90% sur les commandes dev courantes. Binary unique, zéro dépendance.

OutilsGénération de codeInfrastructure
SIG
65
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> git-ai-project /</span> git-ai

git-ai est une extension Git pour tracer le code généré par IA dans les dépôts. Outil open-source permettant d'identifier et de documenter les contributions IA.

Génération de codeOutilsOpen source
SIG
45
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> Michael-A-Kuykendall /</span> shimmy

Shimmy : serveur d'inférence Rust compatible API OpenAI, sans dépendances Python. Supporte GGUF et SafeTensors, échange de modèles à chaud, découverte automatique, binaire unique. Gratuit.

Open sourceInfrastructureGénération de code
SIG
72
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> vercel-labs /</span> agent-browser

Vercel Labs publie agent-browser, un outil CLI d'automatisation de navigateur conçu pour les agents IA. Permet aux agents d'interagir avec des interfaces web.

Agents IAOutilsOpen source
SIG
65
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> KeygraphHQ /</span> shannon

Shannon Lite est un pentesteur IA autonome et transparent pour applications web et APIs. Il analyse le code source, identifie les vecteurs d'attaque et exécute des exploits réels pour prouver les vulnérabilités avant la production.

Agents IASécurité IAGénération de code
SIG
45
HYP
55
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> tech-leads-club /</span> agent-skills

Registre de compétences sécurisé et validé pour agents IA de codage professionnels. Permet d'étendre Antigravity, Claude Code, Cursor, Copilot et autres avec confiance.

Agents IAGénération de codeOutils
SIG
45
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> rohitg00 /</span> agentmemory

AgentMemory : système de mémoire persistante pour agents IA de codage, basé sur des benchmarks réels. Repo GitHub visant à améliorer la rétention d'informations et la continuité des agents autonomes.

Agents IAGénération de codeBenchmarks
SIG
45
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> humanlayer /</span> 12-factor-agents

Humanlayer publie un framework de 12 principes pour construire des agents IA en production. Inspiré de la méthodologie 12-factor app, il couvre architecture, observabilité, et gestion des erreurs pour des systèmes fiables.

Agents IAInfrastructureOutils
SIG
65
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> colbymchenry /</span> codegraph

Codegraph : graphe de connaissances pré-indexé pour Claude Code, Codex, Cursor et OpenCode. Réduit les tokens et appels d'outils, fonctionne 100% en local.

Claude CodeGénération de codeRAG
SIG
65
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> heygen-com /</span> hyperframes

Hyperframes est un framework permettant aux agents IA de générer du contenu vidéo via HTML. Outil conçu pour automatiser la création vidéo dans les workflows d'agents.

Agents IAGénération de vidéosOutils
SIG
45
HYP
55
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> diegosouzapw /</span> OmniRoute

OmniRoute est une passerelle IA gratuite unifiant 160+ fournisseurs via un seul endpoint. Compression RTK+Caveman jusqu'à ~95% d'économies de contexte, auto-fallback intelligent, support MCP/A2A, APIs multimodales, versions Desktop/PWA.

Agents IAMCPInfrastructure
SIG
45
HYP
65
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> n8n-io /</span> n8n

n8n est une plateforme d'automatisation de workflows en fair-code avec capacités IA natives. Combine construction visuelle et code personnalisé, 400+ intégrations, auto-hébergement ou cloud.

Agents IAGénération de codeOpen source
SIG
65
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> K-Dense-AI /</span> scientific-agent-skills

K-Dense-AI publie scientific-agent-skills, une collection de compétences prêtes à l'emploi pour agents IA dans la recherche, sciences, ingénierie, analyse, finance et rédaction.

Agents IAOutilsOpen source
SIG
45
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> HKUDS /</span> CLI-Anything

CLI-Anything transforme les interfaces en agents natifs via CLI. Projet open-source visant à rendre tous les logiciels compatibles avec les agents IA.

Agents IAOpen sourceOutils
SIG
45
HYP
55
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> HKUDS /</span> ViMax

ViMax est un système de génération vidéo agentique intégrant directeur, scénariste, producteur et générateur vidéo. Le projet GitHub propose une architecture multi-agents pour orchestrer la création vidéo de bout en bout.

Agents IAMulti-agentsGénération de vidéos
SIG
45
HYP
65
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> anthropics /</span> claude-plugins-official

Anthropic publie un répertoire officiel de plugins Claude Code de haute qualité, gérés directement par l'entreprise.

ClaudeClaude CodeOutils
SIG
65
HYP
20
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> NVlabs /</span> Sana

NVIDIA Labs publie Sana, un modèle de diffusion transformer linéaire pour la synthèse d'images haute résolution. Architecture efficace réduisant la complexité computationnelle tout en maintenant la qualité visuelle.

Génération d'imagesOpen sourcePapers
SIG
75
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> BigBodyCobain /</span> Shadowbroker

Shadowbroker agrège données publiques (jets privés, satellites espions, séismes) dans une interface unifiée. Permet de connecter des agents IA pour identifier corrélations non détectées. Agrégation open-source de renseignements d'origine ouverte.

Agents IAOpen sourceOutils
SIG
45
HYP
65
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> Alishahryar1 /</span> free-claude-code

Outil permettant d'utiliser Claude Code gratuitement en ligne de commande, extension VSCode ou Discord avec support vocal, inspiré par OpenClaw.

Claude CodeOutilsOpen source
SIG
35
HYP
55
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> tirth8205 /</span> code-review-graph

Code-Review-Graph crée un graphe de connaissances local pour Claude Code. Il construit une carte persistante de la base de code, réduisant les tokens de 6,8× sur les reviews et jusqu'à 49× sur les tâches quotidiennes.

Claude CodeRAGGénération de code
SIG
72
HYP
45
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> Imbad0202 /</span> academic-research-skills

Dépôt GitHub proposant un workflow de recherche académique intégré à Claude Code. Cycle complet : recherche → rédaction → révision → finalisation. Outil pour structurer la production de contenu académique avec IA.

Claude CodeOutilsPrompt engineering
SIG
35
HYP
25
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> anthropics /</span> skills

Anthropic publie un dépôt public pour les Agent Skills, des composants réutilisables destinés à étendre les capacités des agents IA.

AnthropicAgents IAOpen source
SIG
45
HYP
20
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> unslothai /</span> unsloth

Unsloth Studio est une interface web pour entraîner et exécuter localement des modèles ouverts (Gemma 4, Qwen 3.6, DeepSeek, gpt-oss). Plateforme de fine-tuning et inférence sans infrastructure cloud.

Fine-tuningOpen sourceOutils
SIG
65
HYP
35
GitHub Trending·

<svg aria-hidden="true" data-component="Octicon" height="16" viewBox="0 0 16 16" version="1.1" width="16" data-view-component="true" class="octicon octicon-repo mr-1 tmp-mr-1 color-fg-muted"> <path d="M2 2.5A2.5 2.5 0 0 1 4.5 0h8.75a.75.75 0 0 1 .75.75v12.5a.75.75 0 0 1-.75.75h-2.5a.75.75 0 0 1 0-1.5h1.75v-2h-8a1 1 0 0 0-.714 1.7.75.75 0 1 1-1.072 1.05A2.495 2.495 0 0 1 2 11.5Zm10.5-1h-8a1 1 0 0 0-1 1v6.708A2.486 2.486 0 0 1 4.5 9h8ZM5 12.25a.25.25 0 0 1 .25-.25h3.5a.25.25 0 0 1 .25.25v3.25a.25.25 0 0 1-.4.2l-1.45-1.087a.249.249 0 0 0-.3 0L5.4 15.7a.25.25 0 0 1-.4-.2Z"></path> </svg> <span data-view-component="true" class="text-normal"> ZhuLinsen /</span> daily_stock_analysis

Système d'analyse boursière alimenté par LLM pour marchés chinois, hongkongais et américains. Agrège données de marché en temps réel, actualités et génère décisions via tableau de bord LLM. Exécution automatisée sans frais.

Agents IARAGOutils
SIG
45
HYP
35
arXiv cs.AI·

Spherical VAE with Cluster-Aware Feasible Regions: Guaranteed Prevention of Posterior Collapse

Nouvelle méthode VAE utilisant la géométrie sphérique et des contraintes cluster-aware pour éliminer mathématiquement l'effondrement postérieur. Transforme les données en coque sphérique, applique K-means, définit une région faisable entre variance intra-cluster et perte d'effondrement. Garantit 100% de prévention d'effondrement sans surcharge computationnelle ni restrictions sur le décodeur.

PapersÉvaluations
SIG
75
HYP
25
arXiv cs.AI·

The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level

Étude comparative de l'interprétabilité des architectures Mixture-of-Experts (MoE) vs réseaux denses. Les experts MoE montrent moins de polysémantique neuronale que les FFN denses, particulièrement avec routage sparse. Les experts fonctionnent comme des spécialistes de tâches linguistiques fines (ex: fermeture de crochets LaTeX), pas comme des spécialistes de domaines larges.

SIG
78
HYP
15
arXiv cs.AI·

CounterRefine: Answer-Conditioned Counterevidence Retrieval for Inference-Time Knowledge Repair in Factual Question Answering

CounterRefine est une couche de réparation légère pour RAG qui traite la première réponse comme une hypothèse à tester. Le système émet des requêtes d'expansion conditionnées par la réponse pour récupérer des preuves spécifiques, puis applique une étape de raffinement KEEP/REVISE validée déterministiquement. Sur SimpleQA, il améliore un baseline RAG de 5,8 points de taux correct.

RAGRaisonnementÉvaluations
SIG
75
HYP
15
arXiv cs.AI·

SkillMOO: Multi-Objective Optimization of Agent Skills for Software Engineering

SkillMOO optimise les bundles de compétences pour agents de codage via recherche multi-objectif (NSGA-II) sur taux de réussite et coût d'inférence. Sur 16 tâches SkillsBench, le framework atteint le meilleur rang de pass rate sur 11/12 tâches non-nulles, réduisant les coûts jusqu'à 31,7% et gagnant jusqu'à 21 points de pourcentage en taux de réussite.

Agents IAGénération de codeBenchmarks
SIG
78
HYP
15
arXiv cs.AI·

Identifying Latent Actions and Dynamics from Offline Data via Demonstrator Diversity

Article théorique sur la récupération d'actions latentes et de dynamiques d'environnement à partir de trajectoires hors ligne sans observations d'actions. Les auteurs exploitent la diversité des démonstrateurs (chacun suivant une politique distincte) pour identifier les noyaux de transition latents via factorisation matricielle non-négative. Identifiabilité prouvée sous conditions de rang et diversité de politique.

Reinforcement learningPapers
SIG
75
HYP
15
arXiv cs.AI·

Herding CATs: ALARA for Agent Harness Engineering in Portable Composable Multi-Agent Teams

Papier présentant CAT (Context-Agent-Tool), une couche de données pour gérer les équipes multi-agents. Applique le principe ALARA (exposition minimale raisonnable) au contexte. Évalue 22 modèles (0.6B–35B paramètres) sur 115 tâches pratiques via npcsh, un shell CLI. ~2500 exécutions testent opérations fichiers, recherche web, scripting multi-étapes, chaînage d'outils et délégation inter-agents.

Multi-agentsAgents IAOutils
SIG
75
HYP
15
arXiv cs.AI·

DSPR: Dual-Stream Physics-Residual Networks for Trustworthy Industrial Time Series Forecasting

DSPR (Dual-Stream Physics-Residual Networks) propose un cadre de prévision de séries temporelles industrielles qui découple les patterns temporels stables des dynamiques résiduelles dépendantes du régime. Via un module de fenêtre adaptative et un graphe dynamique guidé par la physique, le modèle atteint 99% de précision de conservation et 97,2% de ratio de variation totale sur quatre benchmarks industriels.

BenchmarksRaisonnementInfrastructure
SIG
78
HYP
25
arXiv cs.AI·

ECHO: Efficient Chest X-ray Report Generation with One-step Block Diffusion

ECHO est un modèle de diffusion pour la génération de rapports de radiographie thoracique. Il utilise une distillation conditionnelle directe et un entraînement asymétrique pour générer des rapports en une seule étape par bloc, réduisant la latence d'inférence de 8× tout en améliorant RaTE (+64,33%) et SemScore (+60,58%) par rapport aux méthodes autorégressives.

VisionGénération de codeBenchmarks
SIG
75
HYP
25
arXiv cs.AI·

Limitations of Sequence-Based Protein Representations for Parkinson's Disease Classification: A Leakage-Free Benchmark

Étude comparative de représentations protéiques (composition acides aminés, k-mers, ProtBERT, descripteurs physicochimiques) pour classifier la maladie de Parkinson. ProtBERT + MLP atteint F1=0.704±0.028, ROC-AUC=0.748±0.047. Les séquences primaires seules montrent un pouvoir discriminant limité; des features structurelles/fonctionnelles sont nécessaires.

BenchmarksPapersEmbeddings
SIG
72
HYP
15
arXiv cs.CL·

HINT-SD: Targeted Hindsight Self-Distillation for Long-Horizon Agents

HINT-SD propose une auto-distillation ciblée pour entraîner des agents LLM sur des horizons longs. La méthode utilise l'historique complet de trajectoire pour identifier les actions pertinentes aux échecs et applique la distillation conditionnée par feedback uniquement sur ces spans. Sur BFCL v3 et AppWorld, elle améliore les baselines de 18,80% tout en réduisant le temps par étape d'entraînement de 2,26×.

Agents IAReinforcement learningRaisonnement
SIG
72
HYP
18
arXiv cs.AI·

Investigation into In-Context Learning Capabilities of Transformers

Étude empirique systématique des capacités d'apprentissage en contexte des transformers sur des tâches de classification binaire Gaussienne. Les auteurs analysent comment la précision dépend de la dimension d'entrée, du nombre d'exemples en contexte et de la diversité des tâches d'entraînement. Ils caractérisent l'émergence du surapprentissage bénin et identifient les régions paramétriques critiques.

RaisonnementBenchmarksPapers
SIG
72
HYP
15
arXiv cs.AI·

MindMelody: A Closed-Loop EEG-Driven System for Personalized Music Intervention

MindMelody est un système fermé en temps réel qui génère de la musique personnalisée à partir de signaux EEG. Un modèle Transformer-GNN décode l'EEG en états émotionnels (Valence-Arousal), un LLM avec RAG formule des plans d'intervention, et un contrôleur EEG hiérarchique guide la synthèse audio. Une boucle de rétroaction continue adapte les paramètres selon l'EEG de l'utilisateur.

RaisonnementRAGGénération de code
SIG
72
HYP
35
arXiv cs.AI·

Experiment-as-Code Labs: A Declarative Stack for AI-Driven Scientific Discovery

Experiment-as-Code Labs propose un paradigme où les expériences scientifiques sont encodées en configurations déclaratives compilables vers des APIs d'instruments. Les agents IA formulent des hypothèses, un système effectue analyses de programme et orchestration, puis les expériences s'exécutent via contrôle d'équipements physiques. Stack générique science/lab/instrument-agnostique.

Agents IAPapersRaisonnement
SIG
72
HYP
28
arXiv cs.CL·

SD-Search: On-Policy Hindsight Self-Distillation for Search-Augmented Reasoning

SD-Search propose une auto-distillation par hindsight pour améliorer les agents de raisonnement augmentés par recherche. Un modèle unique joue deux rôles (étudiant et enseignant) : l'enseignant a accès aux résultats des requêtes passées et guide l'étudiant via divergence Jensen-Shannon au niveau des tokens. Aucun modèle externe ni annotation supplémentaire requise.

RaisonnementReinforcement learningAgents IA
SIG
78
HYP
15
arXiv cs.AI·

When Does Non-Uniform Replay Matter in Reinforcement Learning?

Étude sur l'efficacité du replay non-uniforme en RL off-policy. Les auteurs identifient trois facteurs clés : volume de replay, récence des transitions et entropie de la distribution d'échantillonnage. Ils proposent Truncated Geometric replay, qui privilégie les expériences récentes tout en maintenant haute entropie, améliorant l'efficacité d'échantillonnage en régimes bas-volume.

Reinforcement learningBenchmarksPapers
SIG
72
HYP
15
arXiv cs.CL·

CasualSynth: Generating Structurally Sound Synthetic Data

CausalSynth est un framework qui génère des données synthétiques respectant les mécanismes causaux du domaine cible. Il combine un Structural Causal Model (SCM) pour générer des squelettes causaux, un LLM comme réalisateur contraint, et une vérification itérative pour corriger les violations structurelles. Testé sur ASIA, ALARM et MIMIC-Struct, il atteint 96% de réalisabilité avec des taux de faux positifs à α=0.05.

PapersRaisonnementBenchmarks
SIG
78
HYP
15
arXiv cs.CL·

Self-Improving CAD Generation Agents with Finite Element Analysis as Feedback

Agents CAD auto-améliorants utilisant l'analyse par éléments finis (FEA) comme feedback. Les modèles Codex (GPT-5.5) et Claude Code (Opus-4.7) ne produisent aucun artefact valide en première tentative ; seuls ~20% des exigences sont respectées. Deux signaux de supervision (schéma blueprint texte et rendu 21-vues) améliorent la boucle itérative : Box-IoU passe de 0.444 à 0.592 sur S2O.

Agents IAGénération de codeRaisonnement
SIG
78
HYP
25
arXiv cs.CL·

MemRepair: Hierarchical Memory for Agentic Repository-Level Vulnerability Repair

MemRepair est un framework d'agent augmenté par mémoire pour la réparation de vulnérabilités au niveau du dépôt. Il combine trois couches mémoire (History-Fix, Security-Pattern, Refinement-Trajectory) avec une boucle de raffinement itérative. Évalué sur SEC-Bench, PatchEval et Multi-SWE-bench, MemRepair atteint 58.0%, 58.2% et 30.58% de taux de résolution, surpassant OpenHands, SWE-agent et InfCode-C++.

Agents IAGénération de codeSécurité IA
SIG
82
HYP
18
arXiv cs.CL·

QQJ: Quantifying Qualitative Judgment for Scalable and Human-Aligned Evaluation of Generative AI

QQJ est un cadre d'évaluation des systèmes IA génératifs qui combine jugement humain et LLM. Il utilise des rubriques multi-dimensionnelles conçues par experts et calibre les évaluateurs LLM sur un petit ensemble d'annotations de haute qualité. Expériences sur texte et images montrent meilleure alignement avec le jugement humain que les métriques automatiques traditionnelles.

ÉvaluationsLlamaVision
SIG
72
HYP
28
arXiv cs.CL·

CyberCorrect: A Cybernetic Framework for Closed-Loop Self-Correction in Large Language Models

CyberCorrect formalise l'auto-correction des LLM comme système de contrôle en boucle fermée. Un détecteur d'erreurs tri-modal (auto-cohérence, confiance verbalisée, vérification logique) alimente un contrôleur de correction type-dirigé. Sur CyberCorrect-Bench (440 tâches), le framework atteint 79,8% de précision (+6,2pp vs SOTA) et réduit les sur-corrections de 41%.

RaisonnementÉvaluationsPapers
SIG
78
HYP
25