The Who's Who of AI

“2026 - agents should help users construct preferences, not just elicit them - Irena Saracay, Ludwig Schmidt, Carlos Guestrin 4/n”

“Beyond expert users: agents should help users construct preferences, not just elicit them Irena Saracay, Ludwig Schmidt, Carlos Guestrin https://t.co/C6dYa9Tgwh [𝚌𝚜.𝙰𝙸] https://t.co/LTMprro9iE”

“TL;DR: improved training-inference trade-off of drifting models Faster training & comparable FID, costing increased memory usage First author Ali Falahati, co-supervised w/ @elliot-creager.bsky.social & Shubhankar Mohapatra Paper: arxiv.org/abs/2605.12183 C…”

“DriftXpress transforms one-step generative modeling by cutting training time while ensuring high-quality outputs. By applying projected RKHS fields, it sustains drifting models' effectiveness and enhances inference, promising a new era in efficient image ge…”

“D-OPSD transforms training for step-distilled diffusion models, enabling on-policy self-distillation to learn new concepts without sacrificing efficient few-step inference. This enhances image quality and response speed for AI-generated content. https://arx…”

“absolutely goated paper for anyone working with distilled diffusion models lately: arxiv.org/abs/2605.05204”

“Researchers unveiled VLM-IE3D, a model merging implicit and explicit 3D geometries from RGB videos, enhancing spatial awareness for 3D tasks. This method improves performance in 3D understanding, closing the gap between 2D data and complex spatial capabilit…”

“Soumya Dutta, Tejas Indulal Dhamecha, Pannaga Shivaswamy On Improving Faithfulness of Podcasts from Documents https://arxiv.org/abs/2607.21961”

“Research presents 'catch-n-repair,' a framework that improves faithfulness in AI-generated podcasts by identifying and correcting ungrounded turns. This method consistently enhances performance across various Language Models, resulting in more reliable audi…”

“Glob3R: Global Structure-from-Motion with 3D Foundation Models Junyuan Deng et 9 al. tl;dr: finetune Pi3 matching head in RoMav2 style, use SALAD to sort seqs, do Glomap-like global SfM arxiv.org/abs/2607.09225”

“Introducing Glob3R, a novel framework that enhances 3D reconstruction by optimizing foundation model predictions, enabling accurate global motion averaging and efficient scene recovery. This method outperforms traditional techniques, providing high-fidelity…”

“A study reveals vision language models misinterpret simple images as illusions, showcasing processing issues. These "illusion-illusions" expose AI's perceptual gaps, questioning the sophistication of machine recognition against human perception. https://arx…”

“Forecasting Scientific Progress with Artificial Intelligence arxiv.org/abs/2605.22681 #ai #science #research”

“A study shows frontier AI models excel in reasoning but struggle to predict future advancements, often underestimating timelines. This highlights limits in AI's predictive capabilities, calling for improved evaluation metrics in research prioritization. htt…”

“Renuka Oladri, Niveda Jawahar, Abdirisak Mohamed Token Budget Saturation and Mechanistic Early Detection of Reasoning Non-Convergence in Chain-of-Thought Models https://arxiv.org/abs/2607.21433”

“Research shows that models like DeepSeek-R1-Distill-Qwen-7B can achieve high math accuracy with only 256 thinking tokens, underscoring that longer reasoning is not always better. Convergence signals in internal states indicate routes for more efficient infe…”

“A study shows blind and low-vision individuals can create bespoke assistive technologies with ProgramAT. In two months, participants developed 37 tools, addressing unmet needs and showcasing AI's role in empowering users to craft personalized tech. https://…”

“A study reveals OTTER, a universal multilingual NER model, surpassing benchmarks and enhancing performance across 100+ languages while remaining efficient. Key innovations include cross-encoders and systematic evaluation of choices, setting a new standard f…”

“Google DeepMind has unveiled Test-Time Scaling via Error Localization (TTEL), a method that improves large language models' performance. TTEL identifies token-level errors, enhancing efficiency and achieving 71% accuracy on complex tasks with fewer tokens. …”

“Research on pluralistic value alignment in AI highlights a key gap: existing models lack empirical support and guidelines. The authors emphasize a need for practical frameworks to guarantee diverse perspectives in AI systems. https://arxiv.org/abs/2607.22305”

“Research indicates language models may fail not from lack of ability, but rather from "context anxiety"—self-doubt in estimating token needs. Fine-tuning these models to reduce anxiety boosts accuracy and efficiency, revealing a pathway to enhance AI reason…”

“Charles Lu, Olivia Burke, Debby Cheng, Adam Kashlan, Caitlyn Duffy, Zeyun Lu, Lirit Fuksman, Jin Ning Tian, Andrew Sedlack, Priya Katyal, ... Human-in-the-Loop Large Language Model Framework for Identification of Cutaneous Immune-Related Adverse Events http…”

“A study shows a multi-agent language model framework improves cutaneous immune-related event identification in clinical notes, increasing accuracy and halving review time. This method combines AI with expert oversight for efficient immunotherapy toxicity mo…”

“Zixin Chen, Peng Liu, Haobo Li, Rui Sheng, Jianhong Tu, Xiaodong Deng, Fei Huang, Kashun Shum, Dayiheng Liu, Huamin Qu WILDTRACE: Benchmarking Natural Evidence Trails in Long-Context Reasoning https://arxiv.org/abs/2607.09328”

“WILDTRACE benchmarks long-context reasoning with 481 tasks from natural sources that require true evidence integration. By emphasizing source-internal trails, it highlights gaps in AI systems' reasoning, paving the way for advancements in document processin…”

“Phuong Huu Vu Tran, Long Minh Vo, Son Nguyen Minh Le, Hoang Van LLM-INSTRUCT at UZH Shared Task 2026: Constraint-Aware Retrieval and Selective Debate for Paragraph-Level Argument Mining https://arxiv.org/abs/2607.20430”

“LLM-INSTRUCT took first place in the UZH Shared Task 2026, innovating paragraph-level argument mining with metadata-aware retrieval, constrained decoding, and selective debate strategies to enhance accuracy and reliability in processing institutional texts.…”

“This study introduces the Entropy-Scaled Trust Region (ESTR) to enhance asynchronous reinforcement learning. It stabilizes training by filtering noise, achieves 2.6× speedup without losing accuracy, and sets benchmarks for long-horizon reasoning in language…”