Kaggle's AgenticRL-Bench Reveals Top AI Agents Solve Only 22.8% of Broken RL Systems
Kaggle and Eigent AI launch a benchmark where coding agents debug real RL training stacks, and the top model clears just 22.8 percent of tasks.
20 stories
Kaggle and Eigent AI launch a benchmark where coding agents debug real RL training stacks, and the top model clears just 22.8 percent of tasks.
Sakana AI spins up an independent cyber defense lab in Tokyo, betting that quality engineering upstream beats patch-chasing downstream.
A new pair of unadjusted Langevin samplers cuts gradient evaluations by up to 17x on constrained distributions while staying stable at larger steps.
OpenAI patched an autonomous weapons alignment gap flagged by Vals AI, and the fix shipped in its newest frontier model Astra.
OpenAI's Responses API now preserves reasoning across turns when the conversation prefix stays intact, but edits silently drop prior thinking.
Cognition's FrontierCode 2.0 benchmark puts Google's new mid-tier model within a point of Claude Opus 5.5 while costing roughly a third per task.
A GGUF-quantized build of Google's UMT5-XXL encoder is quietly powering most local Wan video pipelines, shrinking a 22GB model to under 4GB.
Nous Research is embedding its open-source Hermes agent inside Teams, the Microsoft Store and Azure, with skills that teammates can share and improve.
Cognition's new Devin Security Threats feature turns fresh CVE disclosures into one-click scans and auto-generated pull requests across your repos.
Replit's agent swarm cloned a popular personal AI assistant in three days, releasing Amuse as a free, self-hosted alternative that keeps your data local.
Google's Gemini app for Mac can now click, type, and navigate apps in a background window while you keep using your screen.
Google Quantum AI measured contextuality on Willow, winning the magic square game 98.3% of the time and sustaining quantum advantage across 105 qubits.
OpenAI brings its ChatGPT voice model to the API with cheaper idle time, steadier accents, and better handling of letters and numbers.
Gemini's new Tool Retrieval loads function schemas on demand, letting agents scale to hundreds of tools without blowing up context windows.
A community-trained LoRA adapter pushes MiniMax H3 Image toward Seedream 5.5's cinematic photographic look, with a 1,134-image dataset and 6,000 training steps.
Kyutai's AURA distills a bidirectional video diffusion teacher into a causal student that streams lip-synced talking heads live on one H100.
Google's Genie 4 world model is now open to paid Gemini API users. It streams 720p worlds with native audio that you can steer live from code, at about $0.10 per second.
Goodfire mapped every neuron in Qwen3.6, arguing that individual units, not learned features, are the right basis for interpretability.
Google's Antigravity desktop app can now open a remote Linux machine over SSH, so agents, terminals and Git worktrees run on the server and keep going after you close the window.
Parcha Labs open-sourced AgentRun, a layer that sits around Claude Code, Codex, Pi or OpenCode and adds learning, crash recovery and swappable computers, so repeated agent work gets cheaper.