“Scalable Training of Mixture-of-Experts Models with Megatron Core” This NVIDIA MoE report walks through the hard part of MoE training. The key is not to add more parameters, but keeping sparse models efficient when only a small part of the model runs for each token. For
@askalphaxiv
-
Paperscrolling: AI-Powered Research Discovery Tool Launches
By
–
If doomscrolling X is part of your research workflow, we built something for you.
— alphaXiv (@askalphaxiv) 12 mars 2026
Introducing Paperscrolling 🚀
Get the most trending research with key ideas, figures, and audio explanations from alphaXiv Briefs pic.twitter.com/8uGbaheOvbIf doomscrolling X is part of your research workflow, we built something for you. Introducing Paperscrolling Get the most trending research with key ideas, figures, and audio explanations from alphaXiv Briefs
-
On-Policy Self-Distillation Improves Reasoning LLM Efficiency
By
–
“On-Policy Self-Distillation for Reasoning Compression” This paper shows that the new bottleneck for reasoning LLMs isn’t “too little reasoning”, it’s that more reasoning tokens often increase mistakes. So they fixed it without RL length penalties or external verifiers, but all
-
Research Analysis: Massive Activations and Attention Sinks in Transformers
By
–
Another by Yann LeCun! “The Spike, the Sparse and the Sink” This paper shows that massive activations and attention sinks come from the same pre-norm Transformer pipeline rather than being separate anomalies. Early SwiGLU blocks act like directional quadratic amplifiers,
-
Research Analysis: Can LLM Chain-of-Thought Reasoning be Deceptive?
By
–
"Can Aha Moments Be Fake?" This paper shows that LLMs can generate CoT reasoning where the verbalized steps don't actually reflect the model's internal thinking process, like decorative thinking steps. The model can even verbalize an aha moment where it pretends to correct
-
Introducing arXivQA: Training retrieval agents for arXiv search
By
–
Introducing arXivQA: Training retrieval agents for arXiv search We curate a multi-hop arXiv dataset based on real queries and use rubric-as-rewards to train a Qwen model for production-grade retrieval Our latest blog details our experience training with both RLVR and rubrics
-
Introducing GPT 5.4 for Research Paper Analysis and Contextual Queries
By
–
Introducing GPT 5.4 for understanding research papers 🚀
— alphaXiv (@askalphaxiv) 9 mars 2026
Highlight any section of a paper to ask questions and “@” other papers for quick context, comparisons, and benchmark references pic.twitter.com/WmWJDoK5kfIntroducing GPT 5.4 for understanding research papers Highlight any section of a paper to ask questions and “@” other papers for quick context, comparisons, and benchmark references
-
Optimizing Claude Code usage for research with alphaXiv data
By
–
If you're using Claude Code for research: stop making it read directly from PDFs
— alphaXiv (@askalphaxiv) 9 mars 2026
We've introduced a SKILL.md that fetches structured, AI-friendly paper overviews from alphaXiv 👀 pic.twitter.com/AYg3n0hnB2If you're using Claude Code for research: stop making it read directly from PDFs We've introduced a SKILL.md that fetches structured, AI-friendly paper overviews from alphaXiv
-
ByteDance Seed Releases Helios: Real-Time Long Video Generation Model
By
–
"Helios: Real Real-Time Long Generation Model" More video gen techniques shared by ByteDance Seed! They show you can get minute-scale and temporally stable video generation in real time by training a 14B autoregressive diffusion model to expect and correct its own
-
New Paper Evaluates AI Coding Agents on Long-Horizon Codebase Maintenance
By
–
Even the SoTA models still struggle with long-horizon code maintenance! “SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration” Right now, most coding-agent evals are static, where you solve a single issue, and pass one-off tests. This paper