Toggle light / dark theme

Abstract: Large Reasoning Models (LRMs) represent a breakthrough in AI problem-solving capabilities, but their effectiveness in interactive environments can be limited. This paper introduces and analyzes overthinking in LRMs. A phenomenon where models favor extended internal reasoning chains over environmental interaction. Through experiments on software engineering tasks using SWE Bench Verified, we observe three recurring patterns: Analysis Paralysis, Rogue Actions, and Premature Disengagement. We propose a framework to study these behaviors, which correlates with human expert assessments, and analyze 4,018 trajectories. We observe that higher overthinking scores correlate with decreased performance, with reasoning models exhibiting stronger tendencies toward overthinking compared to non-reasoning models. Our analysis reveals that simple efforts to mitigate overthinking in agentic environments, such as selecting the solution with the lower overthinking score, can improve model performance by almost 30% while reducing computational costs by 43%. These results suggest that mitigating overthinking has strong practical implications. We suggest that by leveraging native function-calling capabilities and selective reinforcement learning overthinking tendencies could be mitigated. We also open-source our evaluation framework and dataset to facilitate research in this direction at this https URL.

From: Alejandro Cuadron [view email].

A team of researchers has developed a “gut-on-chip” (a miniature model of the human intestine on a chip-sized device) capable of reproducing the main features of intestinal inflammation and of predicting the response of melanoma patients to immunotherapy treatment. The results have just been published in Nature Biomedical Engineering.

The interaction between microbiota and has long been known. It is the result of both systemic effects, i.e., the elicited in the entire body by immunotherapy, and local processes, especially in the gut, where most of the bacteria that populate our body live. However, the latter can only be studied in animal models, with all their limitations.

Indeed, there is no clinical reason to subject a patient receiving immunotherapy for melanoma to colonoscopy and colon biopsy. Yet intestinal inflammation is one of the main side effects of this treatment, often forcing the therapy to be discontinued.

Attosecond time-resolved experiments have revealed the increasing importance of electronic correlations in the collective plasmon response as the size of the system decreases to sub-nm scales.

The study, published in the journal Science Advances, was led by the University of Hamburg and DESY as part of a collaboration with Stanford, SLAC National Accelerator Laboratory, Ludwig-Maximilians-Universität München, Northwest Missouri State University, Politecnico di Milano and the Max Planck Institute for the Structure and Dynamics of Matter.

Plasmons are collective electronic excitations that give rise to unique effects in matter. They provide a means of achieving extreme light confinement, enabling groundbreaking applications such as efficient solar energy harvesting, ultrafine sensor technology, and enhanced photocatalysis.

Dr. Ben Allardyce and Ph.D. candidate Mr. Martin Zaki from Deakin’s Institute for Frontier Materials (IFM) have delivered a world first in next-generation materials research. Silkworm silk is a protein-based fiber with mechanical properties rivaling petroleum-derived synthetic fibers, yet spun using a fraction of the energy. Despite decades of research, aspects of natural silkworm spinning remain a mystery.

Published in Advanced Materials, the IFM discovery takes researchers one step closer to solving this mystery by wet-spinning a new class of silk that produces fibers that outperform natural silk.

This research, led by Dr. Allardyce and Mr. Zaki, with expert input from Sheffield University’s Professor Chris Holland, involves sidestepping degumming—a commonplace industrial process—and experimenting with dissolving whole silk fibers.

In Europe alone, approximately 2 million people live with chronic inflammatory bowel diseases (IBD), and their incidence has been rising steadily in recent decades. However, a small proportion of the European population carries a genetic variant that provides natural protection against IBD.

A newly published study in the journal eBioMedicine explores how this protective variant can be leveraged to develop modern therapies, demonstrating the potential of evolutionary medicine in addressing chronic diseases of the modern era.

The study, led by the Institute of Clinical Molecular Biology (IKMB) at Kiel University, brought together researchers from genetics, medicine, and archaeology.