Reinforcement learning for LLMs—fully open-sourced DAPO trains Qwen2.5-32B with RL, hitting 50 points on AIME 2024, outperforming DeepSeek-R1-Zero after just 50% of the training steps. Open-source code & dataset
Improved training stability Trending #1 on alphaXiv
@askalphaxiv
-
DAPO: Open-Source RL Training for Qwen2.5-32B LLMs
By
–
-
ArXiv Paper Curation Platform Powered by Community Engagement
By
–
We’re curating the most viewed and liked papers on arXiv every day 👀
— alphaXiv (@askalphaxiv) 20 mars 2025
You can now browse through AI papers based on what our community of hundreds of thousands of researchers enjoy most
Crowdsourced from alphaXiv and our Chrome extension
Goodreads for arXiv🚀 pic.twitter.com/yEqKLZG6CiWe’re curating the most viewed and liked papers on arXiv every day You can now browse through AI papers based on what our community of hundreds of thousands of researchers enjoy most Crowdsourced from alphaXiv and our Chrome extension Goodreads for arXiv
-
Transformers Excel at 3D Vision with Multi-Image Processing
By
–
New paper from Oxford and Meta AI demonstrates how powerful transformers are becoming for 3D vision tasks. This single model can: • Process 1-200+ images simultaneously • Predict complete 3D scene geometry • Operate in <1 second Trending on alphaXiv
-
Technical Advances in AI Safety and Model Fine-tuning Methods
By
–
Technical Report http://
intology.ai/blog/zochi-tec
h-report
… Siege Multi Turn Jailbreak http://
alphaxiv.org/abs/2503.10619 Compositional Subspace Representation Fine tuning http://
alphaxiv.org/abs/2503.10617 -
Two papers advance LLM jailbreaking and multi-domain adaptation
By
–
Paper 1: Siege sets the state-of-the-art on jailbreaking, formalizing multi-turn attacks as a tree search, achieving an 100% attack success rate on leading LLMs. Paper 2: CS-ReFT adapts LLMs to multiple new domains simultaneously by editing model subspaces, enabling
-
Zochi AI Scientist Publishes Accepted Papers at ICLR 2025
By
–
@IntologyAI releases Zochi, an Artificial Scientist that produced multiple accepted papers at ICLR 2025 workshops, presenting state-of-the-art results! Zochi discovers novel methods in diverse domains, from parameter-efficient tuning & jailbreaking to computational biology.
-
InAs-Al Hybrid Devices Achieve Topological Gap Protocol
By
–
Comment on “InAs-Al hybrid devices passing the topological gap protocol”,
Microsoft Quantum, Phys. Rev. B 107, 245423 (2023) Read and discuss the paper here: http://
alphaxiv.org/abs/2502.19560 -
Microsoft Quantum Results Face Significant Reproducibility Critique
By
–
Significant critique of Microsoft Quantum's results just released on arXiv: – Protocol definition differs between published paper and code – Results depend on unexplained measurement choices, not device properties Trending on alphaXiv
-
Looped Transformers Enhance Reasoning Through Depth Over Parameters
By
–
Reasoning with Latent Thoughts: On the Power of Looped Transformers This paper explores the potential of looped transformers—models that reuse the same layers multiple times—for reasoning tasks. It argues that depth, rather than parameter count, is the key driver of reasoning
-
ESPnet-SpeechLM: Open Source Speech Language Model Toolkit
By
–
ESPnet-SpeechLM: An Open Speech Language Model Toolkit ESPnet-SpeechLM is an open-source toolkit for developing Speech Language Models (SpeechLMs) and voice-driven applications. It simplifies speech processing by framing tasks as sequential modeling problems, automating data