Together AI at ICML 2026: frontier research across the full stack

Together AI at ICML 2026: frontier research across the full stack

🚀 Now serving MiniMax-M3 for efficient inference →

⚡ On-demand B200s now available on Together GPU Clusters →

📊 Delivering 31% more TPS than the next-fastest OSS engine for production coding agent workloads →

💬 How Together built the world's fastest speech-to-text stack →

🇫🇷 Join us at RAISE 2026 in Paris →

All blog posts

Research

Published 6/30/2026

Together AI at ICML 2026: frontier research across the full stack

Eight papers from Together AI and our peers were accepted to ICML 2026 in Seoul. If you are attending come by booth B714. We’d love to connect.

Authors

Together Research

Table of contents

40+ Models Chosen for Production...40+ Models Chosen for Production...40+ Models Chosen for Production...

Links in this article

DSGym paper ThunderAgent paper TTT-Discover paper RARO paper V1 paper Aurora paper Untied Ulysses paper Opportunistic Expert Activation paper

Eight papers is a lot to take in as a list. The better way to read them is by where they sit in the stack. Frontier AI is not built at a single layer. It is the product of research that runs from the agent down to the GPU kernel, and a gain at any one layer is wasted if the layers around it cannot keep up. This is at the core of how we work. From frontier agents at the top to kernels at the bottom, our research touches each one, and each layer feeds the next. The research becomes part of the Together platform, and the production workloads running on that platform point us to the next research problem. Aurora, our ICML paper on adaptive speculative decoding, is a clear example: the same line of work ships today as our ATLAS speculator in production.

Here is this year's work, layer by layer, top down.

01 Frontier agents Agents that do real work — measured on tasks you can't fake your way through.

DSGym 1,000+ tasks across 10+ domains

ThunderAgent Up to 3.6× faster agent inference

TTT-Discover Beats best human, open model

02 Model shaping Turning a base model into a reasoner — even where there's no answer key to check.

RARO 25% win rate, no verifier

V1 Up to 10% more correct answers

03 Algorithmic optimizations Cutting the cost of every token — speculative decoding that adapts to live traffic.

Aurora 1.25× as traffic shifts

04 Systems optimizations Fitting more on the same GPUs — longer context, bigger batches, faster MoE decode.

Untied Ulysses 5M-token context on one node

OEA Up to 39% faster MoE decode

05 Kernels The GPU kernels it all runs on — where microseconds compound across the whole stack.

No new ICML paper at this layer this year, but it is core to the same flywheel. You can read about our kernels research work on the Together blog.

Frontier agents The top of the stack: building agents that do real work, and measuring them honestly. Paper 1 DSGym: A Holistic Framework for Evaluating and Training Data Science Agents

1,000+ tasks 10+ domains, one API

More than 100 data-science tasks across 10+ domains, unified behind a single evaluation and training API.  Data-science agents have been hard to measure fairly, every benchmark has its own interface and many of their tasks can be solved without ever opening the data. DSGym standardizes the measurement and closes that loophole. It puts diverse evaluation suites behind one API with shared abstractions for datasets, agents, and metrics, and runs each task in a self-contained execution environment where the agent has to work with the data rather than recall an answer. On top of the refined existing suites, it adds 90 expert bioinformatics tasks grounded in academic literature and 92 end-to-end Kaggle-style modeling competitions. The same environment then runs in reverse as a training engine: trajectory generation and synthetic query pipelines produce execution-verified data, which we used to train a 4B model into a state-of-the-art open-source data-science agent, with no human labeling. Evaluate honestly, synthesize from the same harness, fine-tune, re-evaluate, all in one framework. Authors: Fan Nie, Junlin Wang, Harper Hua, Federico Bianchi, Yongchan Kwon, Zhenting Qi, Owen Queen, Shang Zhu, James Zou. With collaborators at Stanford, Duke, and Harvard. Paper: arXiv:2601.16344 Paper 2 ThunderAgent: A Simple, Fast and Program-Aware Agentic Inference System

up to 3.6× agent throughput

Achieving 1.5 to 3.6x higher serving throughput for agent workloads, with three lines of code to adopt. Parallel agent workloads stop collapsing under load, and the fix is not a faster model. The problem is the inference engine has no idea it is running an agent. It treats each step of a multi-turn workflow as an isolated request, inflating latency by up to 7.14x under load. ThunderAgent makes the workflow itself a first-class object the scheduler can reason about end to end. Alongside the throughput gain, it delivers 1.8 to 3.9x faster RL rollouts and up to 4.2x disk savings over state-of-the-art systems. Authors: Hao Kang, Ziyang Li, Xinyu Yang, Weili Xu, Yinfang Chen, Junxiong Wang, Beidi Chen, Tushar Krishna, Chenfeng Xu, Simran Arora. With collaborators at Georgia Tech, CMU, and UIUC. Paper: arXiv:2602.13692 Paper 3 Learning to Discover at Test Time (TTT-Discover)

Beats best human open model, ~$500

State-of-the-art discoveries across four fields, mathematics, GPU kernels, competitive algorithms, and biology, all with open models. Every prior result at this level relied on closed frontier models behind an API you cannot inspect or run. TTT-Discover reaches it with an open 120B model and one unchanged method, for a few hundred dollars per problem. The usual recipe for AI discovery is search: prompt a frozen model thousands of times and keep the best sample. TTT-Discover instead runs reinforcement learning at test time on the single problem in front of it, so every attempt becomes training data for the next one and the model improves as it works, and with the same sampling budget plain best-of-N never catches up. The same setup, unchanged, set a tighter bound on a 60-year-old Erdős problem in mathematics (past the previous AI record, which used closed models), discovered a GPU kernel faster than the best prior submission on the GPUMode leaderboard, produced a first-place-level finish on a competitive-programming contest, and set a new high on a single-cell denoising benchmark in biology, all on the open gpt-oss-120b. The code and the record-setting kernels are public....