AI Research Highlights

Tanmay Gupta, Piper Wolters, Zixian Ma et al.

An open 4B and 8B visual web agent plus large mixed training set that beats comparable open agents and some larger closed systems, giving builders a reproducible browser-automation stack without HTML or accessibility-tree dependence.

Yuxuan Zhang, Yubo Wang, Yipeng Zhu et al.

A live-web benchmark across 144 production sites and everyday tasks, showing frontier agents still complete only a small slice of real user workflows and giving builders a far more realistic yardstick than sandboxed browser evals.

Runpeng Geng, Chenlong Yin, Yanting Wang et al.

A unified prompt-injection evaluation platform with adaptive attacks that exposes how brittle many current defenses remain across tasks, making it useful core infrastructure for teams shipping tool-using or retrieval-augmented agents.

Shilin Yan, Jintao Tong, Hongwei Xue et al.

Act Wisely separates task accuracy from tool-efficiency rewards so multimodal agents learn when not to call tools, cutting unnecessary invocations by orders of magnitude while improving accuracy, latency, and cost.

Ashima Suvarna, Kendrick Phan, Mehrab Beikzadeh et al.

SUPERNOVA turns natural-instruction datasets into RL-ready supervision for general reasoning, delivering large gains beyond math and code and giving post-training teams a practical recipe for broader reasoning improvement.

Boyang Zhang, Sebastián G. Acosta, Preston Carlson et al.

ParseBench is a 2,000-page enterprise document benchmark that scores tables, charts, formatting, faithfulness, and grounding the way agents actually need them, exposing why text-similarity metrics miss business-critical parsing failures.

Andrey Bocharnikov, Ivan Ermakov, Denis Kuznedelev et al.

This study shows popular KV-cache offloading schemes break on context-intensive workloads like structured extraction, then offers a simpler strategy that preserves far more accuracy for long-context production inference.

Tongbo Chen, Zhengxi Lu, Zhan Xu et al.

KnowU-Bench evaluates personalized mobile agents in live GUI environments, including when to ask, act, or stay silent, which is much closer to real assistant behavior than static preference benchmarks.

Khushal Sethi

TrACE spends extra rollouts only on uncertain agent steps, matching fixed self-consistency accuracy with far fewer model calls and offering an easy path to cheaper agent inference.

Wenbo Hu, Xin Chen, Yan Gao-Tian et al.

OpenVLThinkerV2 introduces a more stable RL objective and task-shaping recipe for open multimodal reasoning, helping a generalist model balance perception with multi-step thinking across 18 visual benchmarks.

Yunsong Zhou, Hangxu Liu, Xuekun Jiang et al.

SIM1 builds physics-aligned real-to-sim twins for deformable manipulation, letting purely synthetic training reach real-data parity at a fraction of collection cost and making sim-scaled robotics learning much more practical.

Jiayuan Ye, Vitaly Feldman, Kunal Talwar

Pruning and rebalancing pretraining data can improve factual memorization enough for a 110M model to match a 1.3B baseline on entity facts, highlighting data mix as a real scaling lever.

Addison J. Wu, Ryan Liu, Shuyue Stella Li et al.

This paper turns chatbot advertising into a concrete alignment problem, probing how model behavior shifts when user benefit and platform revenue diverge.

Sai Srinivas Kancheti, Aditya Kanade, Rohit Sinha et al.

Faithful GRPO adds consistency and grounding constraints to multimodal RL training, sharply reducing unfaithful visual reasoning traces while also improving final spatial reasoning accuracy.

Ziwei Zhou, Zeyuan Lai, Rui Wang et al.

AVGen-Bench finds that today's flashy text-to-audio-video systems are still semantically unreliable, especially for speech, text rendering, physical reasoning, and musical pitch control.