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

Together AI has announced nine papers accepted at ICML 2026, which it says span the full AI stack from frontier agents down to GPU kernels. The company said the research becomes part of the Together platform, and that production workloads point toward the next research problem. It cited Aurora, its paper on adaptive speculative decoding, as work that ships today as its ATLAS speculator in production.

In the frontier agents layer, DSGym is described as a framework for evaluating and training data-science agents, with more than 100 tasks across 10+ domains and additions including 90 expert bioinformatics tasks and 92 end-to-end Kaggle-style modeling competitions. Together AI said the work was used to train a 4B model into a state-of-the-art open-source data-science agent with no human labeling. ThunderAgent is presented as delivering 1.5 to 3.6x higher serving throughput for agent workloads. Learning to Discover at Test Time is described as reaching state-of-the-art discoveries across mathematics, GPU kernels, competitive algorithms, and biology using the open gpt-oss-120b model.

On model shaping, RARO is described as achieving a 25% win rate against expert responses versus 5.9% for supervised fine-tuning, without a verifier. V1 is described as unifying generation and self-verification for parallel reasoners.

Algorithmic and systems papers include Aurora, UPipe (Untied Ulysses) for memory-efficient context parallelism, and Opportunistic Expert Activation for batch-aware expert routing in Mixture-of-Experts models.

In kernels, ParallelKernelBench is described as the first benchmark for multi-GPU kernel generation, with 87 real distributed workloads drawn from codebases including Megatron-LM, DeepSpeed, NeMo-RL, and TensorRT-LLM. Together AI said today's best frontier model solves fewer than a third zero-shot, and that Gemini 3 Pro climbs from 24 to 35 correct in an agentic loop.

The papers are at ICML 2026, July 6 to 11, in Seoul, where Together AI said it will be at booth B714. The company also said it is hiring researchers and research engineers across the stack.