Together AI brings nine research papers to ICML 2026 in Seoul

On June 30, 2026, Together AI announced that nine research papers from the company and its partners had been accepted to ICML 2026 in Seoul. The company is presenting them at booth B714, and together they span the entire AI infrastructure stack, from agents down to GPU kernels.

How the research stack fits together

Together AI argues that frontier AI isn't built at a single layer. A gain in one layer is useless if the neighboring layers can't keep up. The company sorted its nine papers into five layers of the stack, with agents at the top and GPU kernels at the base. It shows how each layer feeds the next, while production workloads on the Together platform point to which research problem to tackle next.

ICML (International Conference on Machine Learning) ranks among the world's top machine learning conferences, alongside NeurIPS and ICLR, and getting a paper into its program has traditionally been seen across the industry as a strong signal of research quality. In 2026, the conference takes place in Seoul, where Together AI will have its own booth, B714, for meeting attendees.

What's new in agents and model training

Three papers focus on agents that do real work and are evaluated on tasks they can't fake their way through. DSGym is a framework for evaluating and training data agents: more than 1,000 tasks across 10+ domains behind a single API. ThunderAgent speeds up agent inference by up to 3.6x. TTT-Discover is an open model that, according to its authors, beats the best human result in its task category.

Two more papers tackle turning a base model into a reasoner even when there's no reference answer to check against. RARO trains a model without a verifier and achieves a 25% win rate against the baseline. V1 adds up to 10% more correct answers. Both papers address the same practical problem: how to improve a model on tasks where correctness can't be checked by simple comparison against an "answer key."

How inference and GPU kernels get faster

At the algorithmic optimization layer, Aurora implements adaptive speculative decoding. In this technique, a lighter model predicts several tokens ahead and the main model verifies them, which speeds up text generation. Aurora's algorithm adapts to live traffic and delivers up to a 1.25x speedup. According to Together AI, this same line of research already runs in production on its platform as the ATLAS speculator, meaning the path from an ICML paper to a real service took not years but a single development cycle.

Key facts at the systems and hardware layers:

• Untied Ulysses handles a 5-million-token context on a single compute node
• Opportunistic Expert Activation (OEA) speeds up MoE model decoding by up to 39%
• ParallelKernelBench is a suite of 87 multi-GPU kernel tasks, the foundation of the stack, where raw hardware turns into usable speed
• In total, 9 papers from Together AI and its partners were accepted to ICML 2026
• The company is presenting them at booth B714 in Seoul

Why it matters to the company

Together AI ties its research directly to production: the Aurora work is already used on its platform as the ATLAS speculator. The company describes the loop this way: research results become part of the Together platform, and production workloads on that platform reveal which research problem to solve next.

What it means

ICML 2026 shows that the race for faster, cheaper AI inference is playing out not only at the model level but also in agent frameworks, decoding algorithms, and GPU kernels. Companies like Together AI are building out research across that entire stack at once rather than at a single point.