Together AI presents nine research papers at ICML 2026 covering agents, long-context training, and inference optimization

Together AI published a preview of its ICML 2026 research contributions ahead of the International Conference on Machine Learning, running July 6-11 in Seoul. The company is presenting nine papers spanning agent serving infrastructure, model shaping, systems optimization, and GPU kernel benchmarking.

What's new

Together's ICML 2026 papers fall into four categories:

Frontier agents:

• DSGym — a framework covering 1,000+ tasks across 10+ domains for training data-science agents without human labeling
• ThunderAgent — a serving system that treats agent workflows as first-class objects, achieving "1.5 to 3.6x higher serving throughput for agent workloads, with three lines of code to adopt"
• TTT-Discover — a test-time training method for scientific discovery, reaching state-of-the-art results "across four fields, mathematics, GPU kernels, competitive algorithms, and biology, all with open models"

Model shaping:

• RARO — adversarial training without verifiers, achieving a 25% win rate against expert responses
• V1 — answer selection method delivering up to 10% more correct answers from existing model generations

Systems and algorithms:

• Aurora — adaptive speculative decoding that delivers 1.25x speedup as traffic shifts, learning production patterns in real time
• Untied Ulysses — long-context training enabling "5M-token context training on a single 8xH100 node, with up to 87.5% less attention memory"
• OEA — expert assignment optimization for mixture-of-experts models that reduces decode latency by up to 39% without retraining

Kernels:

• ParallelKernelBench — a benchmark with 87 distributed workloads for evaluating multi-GPU kernel generation

Context

Together AI operates one of the larger shared inference platforms for open models. Its research output bridges academic work and production deployment—papers like ThunderAgent and Aurora address real-world serving constraints rather than synthetic benchmarks. Nine accepted papers at a top-tier ML venue is a significant output for a company Together's size and reflects its dual role as a research organization and commercial inference provider.

Why it matters

The breadth of this research slate reflects where production AI workloads are heading: multi-step agent workflows, very long context windows, and efficiency improvements that directly reduce per-token inference cost.

ThunderAgent's throughput gains address one of the costliest bottlenecks in production agent systems—the overhead from managing tool calls, partial outputs, and context accumulation across long workflows. Untied Ulysses' 5M-token training result on a single 8xH100 node matters because it makes long-context research more accessible to labs that do not have multi-cluster budgets. OEA's 39% MoE latency reduction applies across the growing class of MoE-architecture models deployed in production.

For practitioners following the inference efficiency and open-model research space, the ICML 2026 papers represent a useful signal of where applied inference research is moving in mid-2026.