Microsoft Research Asia introduced Harnessed Agentic RL by open-sourcing Agent Lightning v1.0, where the same agent harness used in deployment participates directly in reinforcement learning without being reimplemented in the training loop. The paper reports raising Qwen3.5-9B Pass@1 on SWE-bench Verified from 41.8% to 56.4% using about 6,000 training samples. Agent training shifts from tokenizing a rebuilt interaction loop to observing real harness model calls via a proxy, enabling end-to-end RL control on standard Kubernetes jobs with a lightweight ~3,500-line codebase.
Rivercell, a Paris-based biotech, raised a $25m seed round to build an AI model aimed at understanding and predicting human cell behavior for drug discovery. The funding amount is $25m. The company will use the money to develop and refine its human-cell “world model” to support faster drug discovery.
Nemotron Labs fine-tuned Nemotron models for IOI 2026 and IMO 2026 using supervised fine-tuning, reinforcement learning, and a generate-evaluate-refine inference loop, producing gold-level performance. In IOI 2026, Nemotron-3-Ultra-CC scored 535.4 out of 600, above the 361.12 gold threshold. The work is packaged for reuse on Hugging Face with released checkpoints, training datasets, and an IMO benchmark, changing focus from one-off results to a repeatable specialization recipe.
Ai2 published the Bolmo research in Nature and released new Hugging Face checkpoints for byte-level language models beyond the original Olmo-based versions. The release includes Stage 1 checkpoints that freeze the global model while training Bolmo’s byte-level components. This enables faster retrofitting of subword models into competitive byte-level models and supports broader experimentation without retraining from scratch.
Caterpillar and CoreWeave are partnering to speed up training for physical AI systems used in construction machines. They report that the feedback loop for using newly annotated field data in simulations has been cut from months or weeks to hours within a workday. As a result, physical AI development relies on much larger data and updated AI cloud infrastructure to handle terabytes of sensor and control data per machine per day.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
