“Qwen3.5-Omni Technical Report” Qwen3.5-Omni is a fully omnimodal model that understands text, images, video, and audio, then responds with text or real-time speech. The architecture they use is a Thinker-Talker split, where the Thinker performs multimodal reasoning, while the
@askalphaxiv
-
AlphaXiv Launches Markdown Papers and Structured Skill Overview Tool
By
–
Yup we have every paper in markdown! Example: https://t.co/oEkZ3sJE8s. We also introduced a SKILL.md that fetches structured, AI-friendly paper overviews from alphaXiv https://t.co/dGzA5Busw4
— alphaXiv (@askalphaxiv) 20 avril 2026Yup we have every paper in markdown! Example: https://
alphaxiv.org/abs/1912.01214
.md
…. We also introduced a SKILL.md that fetches structured, AI-friendly paper overviews from alphaXiv -
1D Ordered Tokens Improve AR Image Model Test-Time Search
By
–
“(1D) Ordered Tokens Enable Efficient Test-Time Search” Most AR image models generate 2D grid tokens patch-by-patch, so partial generations are hard to verify. This paper shows that 1D coarse-to-fine tokens make search much easier, with early tokens encode global semantics,
-
Think-Anywhere: Reasoning in Code Generation for Edge Cases
By
–
“Think Anywhere in Code Generation” Most reasoning LLMs think before writing code. But coding often gets hard because the tricky parts only gets revealed mid-implementation when the edge cases or final return logic appear. So this paper introduces Think-Anywhere, where models
-
Parcae Scaling Laws: Recurrence Approach for Stable Language Models
By
–
"Parcae Scaling Laws For Stable Looped Language Models" LLMs usually scale with more parameters or more data. But this paper scales a new axis for transformers through recurrence. So instead of making the model bigger, it reuses the same Transformer layers multiple times.
-
Prefill-as-Service Enables Cross-Datacenter LLM Serving
By
–
new paper from Kimi! "Prefill-as-a-Service makes long-context LLM serving cross-datacenter" Main idea: smaller KV Cache turns prefill into a scalable remote service. Prefill is compute-bound, decode is memory-bandwidth-bound, but KV cache transfer keeps them trapped in the
-
Google’s Reasoning-Driven Synthetic Data Generation Approach
By
–
“Reasoning-Driven Synthetic Data Generation and Evaluation” This paper from Google gave us some insights on how frontier labs currently generate synthetic data. The key is that they treat synthetic data generation as distribution design rather than prompt design, as this builds
-
ELT: Elastic Looped Transformers for Visual Generation
By
–
“ELT: Elastic Looped Transformers for Visual Generation” ELT makes visual generation recurrent by looping a shared transformer block instead of stacking unique layers. Its core idea, Intra-Loop Self-Distillation, trains intermediate loop states to match the full-depth
-
Single Global Merge Improves Decentralized Learning Performance
By
–
"On the Surprising Effectiveness of a Single Global Merging in Decentralized Learning” This paper shows that in non-IID decentralized SGD, a single final global merge can substantially improve global test performance under severe communication constraints, even recovering much