A look at why AI models, even after learning to reason step-by-step on certain problems, often revert to inefficient guesswork when faced with slightly harder or unfamiliar versions. An intervention, co-training them on seemingly unrelated tasks, can restore their confidence in using the correct reasoning strategy under new conditions.
Not All Reasoning Generalizes: A Case Study on Learning Generalizable Reasoning Strategies in RLVR · Ziqi Pang, Yu-Xiong Wang · arXiv:2611.03627
AI coding assistants can complete requests while causing unintended harm, such as deleting unrelated files, and top models often succeed without staying safe. As these AI agents gain more control over systems, the benchmark helps show hidden dangers when success is not safe.
VeriHarm-Bench: Measuring Harmful Success in LLM Coding Agents Verifiably · Wei-Lin Chen, Zhepei Wei, Xinyu Zhu et al. · arXiv:2611.03575
A benchmark to see if large AI models (LLMs) can predict future research directions by comparing their generated ideas for 2026 with actual papers published that year, based on a 2025 knowledge snapshot.
IdeaTaste: Do LLMs Share the Research Taste of Human Researchers? · Shanshan Zhong, Chen Xu, Jingjie Ning et al. · arXiv:2611.03424
Large language models exposed to more monofacts, pieces of information seen only once in training, consistently made more scientific discoveries. This offers a concrete way to understand and potentially improve how much scientific discovery large language models can do.
Tracing the Scientific Discovery Capacity of LLMs to Their Training Data · Arda Uzunoglu, Mahler Revsine, Michael Schatz et al. · arXiv:2611.03617
Reviso is an AI system that makes better video captions by learning which parts of a video are most important to observe and how to gather relevant details. More accurate captions help AI understand complex video content and support tasks like answering questions about it.
Reviso: Not All Visual Evidence Deserves Equal Attention in Video Captioning · Tengfei Liu, Yang Shi, Yuran Wang et al. · arXiv:2611.03602
SpecSlot lets an AI focus global attention on an image's overall structure and handle fine details locally. It greatly cuts the computing cost of sharp images, making advanced image AI more accessible and scalable.
Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers · Jiongfan Yang, Weichen Fan, Haiwen Diao et al. · arXiv:2610.40647
Library of Layers lets an AI model dynamically choose the best order to process information instead of following a fixed sequence. It lets an AI adapt its internal thinking for different problems or languages to make it more efficient and interpretable.
Library of Layers (LoL): Adaptively Ordering Transformer Layers at Runtime · Abdellah Rahmani, Sean McLeish, Ang Li et al. · arXiv:2611.03526
Advanced AI language models did okay with familiar molecules but struggled significantly with truly novel ones unless given access to specialized chemistry software. For complex scientific problems, their strength may lie in intelligently using existing expert tools.
LLMs Depend on Specialist Tools for Out-of-Distribution Generalization in MS/MS Structural Elucidation · Gabriel Asher, Cailum M. K. Stienstra, Tomo Oga et al. · arXiv:2611.03308
Strategist is an AI that learns and evolves step-by-step reasoning procedures, building a library of reusable components as it goes. It can let AI systems automatically improve how they think, making them more adaptable and efficient without constant human intervention.
Strategist: Evolving Reasoning with Agents · Taiming Lu, Qi Xu, Zhigang Hua et al. · arXiv:2611.03618
FADE separates an AI's evolving skills from each task's inherent difficulty, so training focuses on problems that are not too easy or too hard. It may make training advanced AI models more effective, yielding better results with the same effort.
FADE: Factored Ability--Difficulty Filtering for Online RLVR Scheduling · Tomas Duncun, Jonah Ellison, Mateo Ruiz et al. · arXiv:2611.03625