MASS lets a base model propose, execute and grade its own multi-agent workflows, then fine-tunes on the best traces in an evolutionary loop — removing the external-verifier bottleneck. Two cycles on Qwen3.6-27B lifted performance per output token from 1.2x to 1.6x across four open-ended benchmarks.
Drones hit Yandex's Kaluga facility — Russia's largest at roughly 1.4 million square feet and 63 MW — and state media admitted multiple modules were completely taken out of service. It was the second Yandex data center struck in two days.
The early inference numbers cover attention, MoE and speculative verification for models like Kimi K3 NVFP4, plus end-to-end RL training setups pairing SGLang rollouts with CPU-based agent sandboxes.
Agents trained to inspect schemas, run probe queries and iterate inside isolated executable databases — SFT plus execution-reward RL — took Qwen3-14B from 22.2% to 54.8% on Spider 2.0-SQLite, with transfer to BIRD-Interact and LiveSQLBench.
Makes the grader the evolving object in self-improvement loops: inspectable expressions over small deterministic fault detectors, synthesized from clustered failures and gated against reward hacking.
Streams structure-aware detection over agent trajectories and intervenes before irreversible actions, instead of paying a safeguard agent's cost per step or issuing post-hoc verdicts after the damage is done.
A 3,862-question benchmark in an observe-transform-infer framework: build a scene from observations, anticipate how an intervention changes it, and infer unseen outcomes — rather than reading off visible relations.
Treats the embodied world as a Turing machine — tape is robot-environment state, rules are policy — with decision-making written entirely in code and tracked from camera images and proprioception.
A menu bar app for running multiple coding agents at once; it topped AI-adjacent Show HN launches on October 11.
Launched on Product Hunt, aimed at research agents that need literature and evidence retrieval.
Corbenic AI's memory layer for AI agents is available on GitHub.
An open-source toolkit of evaluation skills that let coding agents improve automatically.
Microsoft's OS search now takes actions on results instead of only finding files.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
