@MSFTResearch: Among Microsoft’s 100+ accepted papers, 3 oral presentations, and 1 expo demo at ICML in Seoul, highlights include Fara…

X AI KOLs Following News
microsoft-research icml machine-learning computer-use-agents benchmarking protein-ml llm-reasoning

Summary

Microsoft Research highlights from ICML 2026 include the Fara 1.5 computer-use agent family, critique-resilient benchmarking, expanded protein ML benchmarks with FLIP2, and improved LLM reasoning stability, with over 100 accepted papers.

Among Microsoft’s 100+ accepted papers, 3 oral presentations, and 1 expo demo at ICML in Seoul, highlights include Fara 1.5 computer-use agents, critique-resilient benchmarking, expanded protein ML benchmarks with FLIP2, and improved LLM reasoning stability. https://msft.it/6019vaCHU
Original Article
View Cached Full Text

Cached at: 07/07/26, 06:21 PM

Among Microsoft’s 100+ accepted papers, 3 oral presentations, and 1 expo demo at ICML in Seoul, highlights include Fara 1.5 computer-use agents, critique-resilient benchmarking, expanded protein ML benchmarks with FLIP2, and improved LLM reasoning stability. https://msft.it/6019vaCHU

Highlights from ICML 2026

Source: https://www.linkedin.com/pulse/highlights-from-icml-2026-microsoftresearch-alcoe/?trackingId=A7KG%2BeKSQaO%2BP2vkqRe3Og%3D%3D Microsoft Research

Microsoft Research

We advance science and technology to benefit humanity.

Published Jul 7, 2026

Microsoft is sponsoringThe International Conference on Machine Learning (ICML), a premier gathering dedicated to the advancement of machine learning related fields including statistics and data science, as well as machine vision, computational biology, speech recognition, and robotics.

More than 100 papers from Microsoft were accepted for this year’s conference in Seoul, Korea. This edition of Research Focus offers an overview of work selected for an expo demonstration and three oral presentations.

EXPO DEMONSTRATION

Fara1.5 is a family of lightweight computer-use agent (CUA) models at three scales—4 billion, 9 billion, and 27 billion parameters—each achieving state-of-the-art results on browser-use benchmarks among models of comparable size. Fara1.5 models are the next evolution of Fara-7B and make many advancements, such as more robust user interaction/oversight, more efficient task execution, and overall stronger task completion performance. These models are powered by a synthetic data generation pipeline, FaraGen1.5, that builds synthetic environments and tasks at scale, has agents attempt them to collect demonstrations, and filters the resulting trajectories with a robust verifier. Overall, the Fara1.5 family of models delivers strong performance while being cost-effective and capable of running on-device or modest hardware.

ORAL PRESENTATIONS

As frontier large language models (LLMs) routinely achieve near perfect scores on new benchmarks shortly after they are published, benchmarking itself could become ineffective. If frontier models keep improving, it will become increasingly difficult for humans to generate discriminative tasks, provide accurate ground-truth answers, or evaluate complex solutions. This threatens the ability to measure AI progress—a scenario the authors call thepost-comprehension regime.

To address this, this research introduces Critique-Resilient Benchmarking, an adversarial framework for comparing models even when humans cannot fully understand the tasks or solutions. It definescritique-resilient correctness: an answer is considered correct unless an adversary can convincingly prove otherwise. Rather than fully evaluating solutions, humans act as bounded verifiers, judging localized claims and critiques.

Using an itemized bipartite Bradley-Terry model, the framework ranks LLMs based on both their ability to solve hard problems and to generate difficult but solvable questions. Evaluated on mathematical tasks across eight frontier models, the resulting rankings show stability and correlation with independent measures of model capability. The framework reframes benchmarking as an adversarial generation-evaluation game in which humans remain the final arbiters.

Machine learning (ML) methods that predict a protein’s fitness from its amino acid sequence are sensitive to changes in data distributions, which limits generalization across common conditions encountered in protein engineering. This reduces the credibility of ML tools among protein engineers. The FLIP benchmark established protocols for testing generalization under some domain shifts, but was limited to measurements of stability, binding, and viral capsid viability. This research introduces FLIP2, a protein fitness benchmark spanning seven new datasets, including enzymes, protein-protein interactions, and light-sensitive proteins, as well as splits that measure generalization relevant to real-world protein engineering campaigns. Evaluating a suite of benchmark models across these datasets and suites reveals that simpler models often matched or outperformed fine-tuned protein language models on FLIP2, challenging the utility of existing transfer learning techniques.

In closed-loop multi-turn agent reinforcement learning, LLM agents exhibit reasoning collapse, where reasoning shifts toward generic templates, weakly coupled to the inputs. This research introduces a filtering method that preserves input-specific reasoning and improves agent stability and performance.

The authors show that reasoning collapse is easy to miss with entropy or surface diversity metrics, since reasoning text still varies but becomes input-agnostic. They propose an information-theoretic framework that decomposes variation in reasoning traces into conditional entropy (randomness given the same input) and mutual information (dependence on the input). Template collapse occurs when reasoning appears diverse but has become disconnected from the input and relies on generic patterns. To detect this, the authors introduce a method that measures how strongly a model’s reasoning depends on the input and shows that this dependence drops during collapse. They explain that collapse occurs when learning signals become weak, allowing generic optimization pressures to override input-specific behavior. To counter this, they prioritize training updates that contain stronger learning signals. Across multiple environments, model sizes, and modalities, this improves input-specific reasoning, stability, and overall performance.

Research Focus

Research Focus

143,820 followers

More articles by Microsoft Research

Explore content categories

Similar Articles