Beyond Next-Token: How diffu-GRPO Brings Reinforcement Learning to Diffusion LLMs
By solving the intractable log-probability barrier with mean-field estimation, d1 proves that non-autoregressive models can unlock DeepSeek-style reasoning leaps.
The day's biggest AI story, explained at length with its sources. Page 3 of 12.
By solving the intractable log-probability barrier with mean-field estimation, d1 proves that non-autoregressive models can unlock DeepSeek-style reasoning leaps.
Moving past sanitized bug-ticket benchmarks, FDE-Gym pits top LLMs against unstated constraints, impatient stakeholders, and live business metrics—with sobering results.
By treating experiment cycles as discrete credit windows, MLExpert bridges the long-horizon gap in autonomous machine learning engineering, hitting a 72.7% medal rate on…
NVIDIA and ChipAgents put frontier models through the entire 12-step silicon flow. The verdict: Claude Fable leads at 16.6%, while GPT-6 Astra collapses on physical cons…
Frontier Red Team research shows sub-30B models are months from autonomous propagation across IoT hardware, behaving less like fast cyberweapons and more like digital sl…
By replacing discrete word tokens with raw hidden-state embeddings, latent reasoning architectures unlock breadth-first planning while slashing inference overhead.
Incident-Arena tests frontier models on live Kubernetes clusters under load, revealing that applying the right patch is not enough to survive real production outages.
R-Zero copies one model into a question-writing Challenger and an answering Solver, lifting a small model's math scores by about 6.5 points, with caveats the paper itsel…
DeepMind's flagship leads 13 benchmark rows and rewrites Google's core C++ in Rust, but general access remains locked behind Fairwind cybersecurity vetting.
By training agents inside live SaaS sandboxes against strict logic rubrics, OpenAI pushes GPT-6 Astra past basic GUI navigation into complex, multi-stakeholder business…
Google DeepMind's first Gemini 4 release blows past output context ceilings, tops DeepSWE at 77.9%, and introduces a radical gatekeeper rollout model for frontier AI.
TypeSafe AI's Jev drops text generation for parallel, typed decisions; the company claims it is 40x to 200x faster on quick decision tasks, based on its own tests.