Loading post…
Together AIAnnouncement
Together AI at ICML 2026: frontier research across the full stack
Summary
Together AI says nine papers from its researchers and peers were accepted to ICML 2026 in Seoul, spanning frontier agents, model shaping, algorithmic and systems optimization, and kernels. The collection is presented as a full-stack view of AI research, with work ranging from DSGym’s unified evaluation and training API for more than 1,000 data-science tasks to ThunderAgent’s program-aware scheduling, TTT-Discover’s test-time reinforcement learning, and RARO’s verifier-free adversarial training. Reported results include up to 3.6× agent throughput, state-of-the-art discoveries from an open 120B model, a 25% expert-response win rate without a verifier, and 5M-token context on one node for Untied Ulysses. Aurora illustrates the research-to-production loop: its adaptive speculative-decoding work ships as the ATLAS speculator on the Together platform.
Context
The papers were accepted to ICML 2026 in Seoul. The post invites attendees to visit booth B714 and presents the work as research that feeds into the Together platform and is informed by production workloads.
Approach / What changed
The work is organized top down by stack layer: frontier agents, model shaping, algorithmic optimizations, systems optimizations, and kernels. The framing connects evaluation, training, inference, hardware utilization, and production deployment, with Aurora given as an example of research becoming the ATLAS speculator.
Takeaways
DSGym unifies more than 1,000 data-science tasks across 10+ domains behind one evaluation and training API. Its self-contained environments require agents to work with data, and the same harness produces execution-verified training data without human labeling.
ThunderAgent treats an agent’s multi-turn workflow as a first-class scheduling object instead of handling each step as an isolated request. The system reports 1.5–3.6× higher serving throughput, 1.8–3.9× faster RL rollouts, and up to 4.2× disk savings.
RARO replaces a ground-truth verifier with an adversarial policy-and-critic game using pairwise comparisons and a tie option. It reports a 25% win rate against expert responses versus 5.9% for supervised fine-tuning, while reaching 54.4% on Countdown without a verifier.
FAQ
How does DSGym make data-science agent evaluation more reliable?
DSGym places diverse evaluation suites behind one API with shared abstractions for datasets, agents, and metrics. Each task runs in a self-contained environment, requiring the agent to work with the data rather than recall an answer.
What does ThunderAgent change in agent inference?
It makes the complete multi-turn agent workflow visible to the scheduler, rather than treating every step as an independent request. The post reports up to 3.6× serving throughput and says adoption requires three lines of code.
How can RARO train reasoning without a verifier?
RARO uses one model as both a policy producing responses and a relativistic critic choosing between two responses. Pairwise comparison with a tie option provides training feedback when tasks such as poetry or financial analysis lack a correctness checker.
Source and freshness
- Source
- together.ai
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.