Can we finally trust what LLMs are really thinking? Yutong Gao and researchers from Peking University, Purdue, and Nanjing University present a new survey on intrinsic interpretability — building transparency directly into model architecture rather than relying on post-hoc
@jiqizhixin
-
PAT3D: AI Model for Physics-Aware 3D Scene Generation
By
–
What if your text prompts could generate 3D scenes that actually obey gravity and don’t clip through each other? Carnegie Mellon, HKU, HKUST, and Genesis AI present PAT3D. It combines vision-language models with a physics simulator to arrange objects into stable,
-
CogAlpha: Using LLMs as Adaptive Agents for Financial Market Analysis
By
–
Can AI find hidden patterns in stock markets that traditional methods miss? Researchers from China Mobile and the University of Hong Kong present CogAlpha — a framework that treats large language models as adaptive cognitive agents. Instead of relying on opaque neural nets or
-
Doc-V* Agent Introduces OCR-Free Efficient Document Retrieval
By
–
Struggling with AI that can't handle long, multi-page documents? Researchers from HUST and Xiaomi introduce Doc-V*, an OCR-free agent that starts with a thumbnail overview, then actively retrieves and reads only the relevant pages, storing evidence in a working memory. It
-
The State of Reinforcement Learning for Reasoning LLMs
By
–
Awesome blog! State of RL for reasoning LLMs https://
aweers.de/blog/2026/rl-f
or-llms/
… -
TokenSpeed: A High-Performance LLM Inference Engine for Agentic Workloads
By
–
Nice, a new and super fast LLM inference engine! TokenSpeed is a speed-of-light LLM inference engine designed for agentic workloads, with TensorRT-LLM-level performance and vLLM-level usability. Project: https://
github.com/lightseekorg/t
okenspeed
… -
Automated Data Engine Trains AI for 3D Scene Understanding from Video
By
–
What if unlabeled YouTube videos could teach AI to understand 3D scenes? BIGAI & collaborators built an automated data engine that extracts 3D training data from raw internet videos — no manual labeling needed. Their model achieves strong zero-shot results on 3D detection,
-
Comprehensive Survey of World Models for Robot Learning
By
–
World Model for Robot Learning: A Comprehensive Survey Paper: https://
arxiv.org/abs/2605.00080
v1
…
Repo: https://
github.com/NTUMARS/Awesom
e-World-Model-for-Robotics-Policy
…
Page: https://
ntumars.github.io/wm-robot-surve
y/
… -
Comprehensive Survey on World Models for Robot Learning
By
–
Nice survey! Researchers from NTU, UC Berkeley, Stanford, Oxford, Microsoft, ETH Zurich, Princeton, and Harvard present a comprehensive survey on world models for robot learning. World models are predictive “internal simulators” that let robots imagine how environments evolve
-
New Research on Adaptive Optimization Beyond SGD and Adam
By
–
Anon: Extrapolating Adaptivity Beyond SGD and Adam Paper: https://
arxiv.org/abs/2605.02317
v2
…
Code: https://
anonymous.4open.science/r/Anon-6511/