Reasoning with Exploration: An Entropy Perspective
Adding a single entropy term to the RL advantage function lifts PassK scores dramatically, especially for very large K.
Daixuan Cheng, Shaohan Huang2026-03-14261 cites
ReadPapers
UI-R1: Enhancing Efficient Action Prediction of GUI Agents by
A rule‑based RL framework lifts multimodal LLM accuracy on GUI tasks by up to 18 percent.
Zhengxi Lu, Yuxiang Chai2026-03-14210 cites
ReadSpikingformer: A Key Foundation Model for Spiking Neural Networks
A new spiking transformer backbone removes integer‑float multiplications from residual paths while keeping global attention.
Chenlin Zhou, Liutao Yu2026-03-14165 cites
ReadVisionReward: Fine-Grained Multi-Dimensional Human
A new reward framework improves preference prediction by 17.2% and raises text‑to‑video win rates by over 30%.
Jiazheng Xu, yu huang2026-03-14155 cites
ReadTOUCAN: Synthesizing 1.5M Tool-Agentic Data from
A new dataset of 1.5 M tool‑agentic trajectories, built on authentic Model Context Protocols, pushes open‑source LLM agents toward real‑world complexity.
Zhangchen Xu, Adriana Meza Soria2026-01-0193 cites
ReadWhen Truth Is Overridden
Large Language Models (LLMs) often exhibit sycophantic behavior, agreeing with user-stated opinions even when those contradict factual knowledge. While prior work has documented this tendency, the internal mechanisms that enable such behavior remain poorly understood.
Keyu Wang, Jin Li2026-03-1484 cites
ReadMoLe-VLA: Dynamic Layer-skipping Vision Language
MoMo‑VLA cuts inference time by more than a third while raising success rates on ten manipulation tasks.
Rongyu Zhang, Menghang Dong2026-03-1478 cites
ReadMMAU-Pro: A Challenging and Comprehensive Benchmark
The new MMAU-Pro suite tests AI on 5,305 audio clips across 49 distinct skills.
Sonal Kumar, Šimon Sedláček2026-03-1476 cites
ReadPRISM: Self-Pruning Intrinsic Selection Method for
PRISM slashes GPU time by roughly 70% and pushes accuracy to 101.7% of the full‑data baseline without any extra training.
Jinhe Bi, Aniri2026-01-0176 cites
Read