This week’s AI research highlights breakthroughs in model training and materials discovery, alongside strategic insights into AI’s global trajectory. Together, these studies underscore a dual focus: advancing technical frontiers while addressing the reliability and societal impact of artificial intelligence.
by Anthropic
Anthropic’s announcement presents Claude Sonnet 4.5, their latest frontier model featuring significant improvements in reasoning, mathematics and computer usage, achieving state-of-the-art results on real benchmarks such as SWE-bench Verified and OSWorld.
by Adithya Bhaskar, Xi Ye, Danqi Chen (Princeton University)
Preprint paper revolutionising the training of Language Models by introducing RLMT, a methodology that forces them to generate extended reasoning (CoT), thereby improving their chat and general reasoning capabilities.
by Dawei Li, Zhen Tan, Chengshuai Zhao et al. (Arizona State Universuty,Emory University)
To address concerns about bias in judgements generated by LLMs, this preprint study formalises the judgement detection task and introduces J-Detector, a lightweight neural detector that can accurately identify and interpret that kind of judgments.
by Srishti Gureja, Elena Tommasone, Jingyi He et al. (Cohere)
The study identifies the verification ceiling as the main bottleneck limiting the quality and diversity of training data for code LLMs due to the rigidity of synthetic verifiers, proponing to overcome it through recalibrated verifications.
by Anthropic
Anthropic introduces context engineering, an evolution of prompt engineering that optimizes the limited context available to LLMs. The report outlines advanced strategies, allowing LLMs to maintain consistency and focus on tasks that extend beyond their context window.
by Ryotaro Okabe, Mouyang Cheng, Abhijatmedhi Chotrattanapituk et al. (Nature)
Nature study introducing SCIGEN, a diffusion-based generative model that integrates explicit geometric constraints to guide the discovery of new stable candidates for quantum materials.
by Bengüsu Özcan, Jakob Graabak, Daan Juijn, Sam Bogerd (CFG report)
The CFG article explores how the transition to advanced AI could unfold through five main scenarios, crossing the speed of growth of AI capabilities (gradual or rapid) with the concentration of its development (centralised or decentralised).
by Zhen Zhang, Zhichu Ren, Chia-Wei Hsu et al. (Nature)
Nature-published paper which introduces CRESt, a system that fuses multimodal AI models with bayesian optimization and robotic automation to transform experimental research, helping scientists drive more reliable breakthroughs in laboratories.
by LearnLM Team, Google et al.
This paper presents Learn Your Way, an approach based on Generative Artificial Intelligence that overcomes the limitations of “one-size-fits-all” textbooks by transforming and enhancing them with layers of multiple representations and personalisation.
by RTD Team
RDT2 introduces the first robotics foundation model enabling zero-shot cross-embodiment generalization by scaling up Universal Manipulation Interface (UMI) data collection to 10,000+ hours across 100+ real-world scenes, combined with VLA pretraining and diffusion distillation for ultra-fast inference.