Together AI details nine ICML 2026 papers on serving, training, and agents

Together AI had nine papers accepted to ICML 2026, held in Seoul, and showcased them at booth B714. The research spans serving, training, and evaluation, with several results aimed at practical inference workloads.

Agent serving and data science

DSGym provides over 1,000 data-science tasks across more than ten domains through a single API. It adds 90 expert bioinformatics tasks and 92 Kaggle-style modeling competitions. The dataset was used to train a 4 B parameter open-source data-science agent without human labeling.

ThunderAgent targets agent workloads, delivering 1.5 to 3.6 times higher serving throughput. It cuts latency inflation of multi-turn workflows by up to 7.14 times under load. For reinforcement learning, ThunderAgent provides 1.8 to 3.9 times faster rollouts and up to 4.2 times disk-space savings.

Test-time reinforcement learning and verifiers

TTT-Discover applies test-time reinforcement learning, letting each attempt become training data for the next. It outperforms the best human and open-model baselines on four fields using a 120 B open model, costing roughly $500 per problem.

RARO attains a 25% win rate against expert responses, versus 5.9% for supervised fine-tuning, without any verifier. It reaches 54.4% accuracy on the Countdown benchmark, close to the 57.7% achieved with a ground-truth verifier. V1 improves correct-answer rates by up to 10% using the same compute via a Swiss-tournament verifier.

Long context and MoE latency

Aurora gives a 1.5 times speedup on day 0 for new frontier models, including MiniMax M2.1 229B and Qwen3-Coder-Next 80B, and an extra 1.25 times as traffic shifts. It trains the speculative decoder online with asynchronous reinforcement learning and hot-swaps weights with zero downtime.

Untied Ulysses enables 5 M-token context training on a single 8xH100 node with up to 87.5% less attention memory. It extends context length about 25% beyond prior methods and reaches 8 M tokens on two nodes with comparable throughput.

Opportunistic Expert Activation reduces Mixture-of-Experts decode latency by up to 39% without retraining or architectural changes. It preserves accuracy across AIME24, GPQA, LiveCodeBench, and MATH 500 benchmarks.

ParallelKernelBench benchmarks 87 multi-GPU kernel tasks.