Apple at ICML 2026: 28 Papers, On-Device MLX and a Negative Result on Multi-Agents

On the US national holiday, when almost the entire AI industry was quiet, Apple published an overview of its presence at the upcoming ICML 2026. This is a curated list of accepted papers and demos, not an announcement of a new model or dataset — but it provides a useful window into the direction of Apple’s machine learning research.

Pre-Seoul Overview

ICML 2026 takes place in Seoul from July 6 to 11, and Apple arrives with 28 accepted papers spanning computer vision, natural language processing, methods and algorithms, and data science. In addition to the main program, the team has 9 workshop papers. The post is a classic pre-conference roundup: it gathers in one place everything Apple’s researchers will present over the coming week.

What Is Apple Presenting at ICML 2026?

The emphasis is on several thematic lines. In the domain of generative video, VideoFlexTok stands out — a flexible video tokenization method earning an Oral presentation and an Expo talk on July 6. On the language model side, the paper “Learning Unmasking Policies for Diffusion Language Models” (Oral, July 7) addresses token unmasking policies for diffusion language models — an area drawing increasing attention in recent months as an alternative to autoregressive generation.

Memory and long-term context form another core pillar. MemoryLLM introduces a plug-in feed-forward memory for transformers, while EpiCache introduces an episodic KV-cache for long-running conversations, targeting the problem of context growth in repeated interactions.

Highlighted Papers

Particularly notable is “Multi-Agent Teams Hold Experts Back” — a negative result that directly challenges the current wave of enthusiasm around multi-agent systems. Rather than confirming that more agents means better performance, the paper shows that in certain settings multi-agent scaling hinders the most capable individual models. Such findings are valuable because they discipline expectations: not every task is suited to being broken across multiple agents, and orchestration itself carries a cost.

The breadth of topics — from video tokenization to diffusion language models to memory mechanisms — shows that Apple is not targeting one narrowly defined problem, but covering a broad front of foundational research.

On-Device Agentic Coding with MLX

Apple’s booth features a demo of “Local agentic coding with MLX” — a showcase of on-device inference tooling through Apple’s MLX framework. This aligns with Apple’s consistent strategy of moving AI computation to local hardware, where sensitive data never leaves the device. Agentic coding running locally, without relying on the cloud, is a practical direction that combines privacy with responsiveness.

Apple in Conference Leadership

Beyond papers, Apple is also present in ICML’s organizational structure. Several senior Apple researchers serve as Area Chairs and Senior Area Chairs, including Samy Bengio, Vladlen Koltun, Marco Cuturi, Jiatao Gu, and Tatiana Likhomanenko. Such a role in the review and program leadership of one of the world’s largest ML conferences signals how deeply Apple is integrated into the broader research community, not just commercial product development.

Overall, the announcement is not breaking news, but it is a useful snapshot: it shows that even on the quietest day of the year, foundational research flows on, and Apple enters ICML 2026 with a broad portfolio and a clear emphasis on memory, diffusion models, and on-device inference.