/3 DFlash replaces the autoregressive draft head with a block diffusion model. The drafter gets the target model's hidden states injected into the KV cache of every draft layer (not just the first like EAGLE-3). One denoising step. Entire block of 16 tokens produced in a single
RESEARCH
-
DFlash 6x lossless acceleration on Qwen3-8B
By
–
Diffusion LLMs have been losing to autoregressive models for two years. UC San Diego just found the one thing they're actually better at: not generating, drafting. DFlash hits 6x lossless acceleration on Qwen3-8B. 2.5x faster than EAGLE-3. Check out the detailed breakdown
-
GLM 5.2 enables trusted open model for automated research tasks
By
–
1/2 This is obviously not a comprehensive benchmark, but it’s clear that we finally have an open model that can be trusted and depended upon on difficult research tasks. You can easily run autoresearch yourself with GLM 5.2 by changing ‘arxiv’ to ‘autoarxiv’ for any arXiv URL:
-
Comparison of GLM 5.2 and Opus 4.8 on SDPO paper reproduction
By
–
Here’s a fun comparison between GLM 5.2 and Opus 4.8 on a one-shot reproduction of the SDPO paper This is a hard task: the model must resolve messy verl issues and then run ablations to completion and confirm the paper’s claims. – GLM 5.2 costs $6.21 while Opus 4.8 cost us
-
Training step with refined scaffold and reward hacking guards
By
–
Each training step has the model propose a refined scaffold for a task, which it then uses to generate a solution, with reward flowing back to both stages. There are three layers of guard against reward hacking. As DeepReinforce states, the 9B variant achieves a score of 43.1 on
-
Trying to replace Fable with an ensemble of smaller models
By
–
Me trying to replace Fable with an ensemble of smaller models pic.twitter.com/0dATTnoMbn
— Peter Gostev (@petergostev) 25 juin 2026Me trying to replace Fable with an ensemble of smaller models
-
Anthropic’s distillation continues unabated despite update
By
–
This is an interesting update that Anthropic published in their official letter. But in short: they apparently haven't managed to stop the distillation. It's continuing almost seamlessly, just like before.
-
Class of 2026 faces AI hiring bias and systemic rejection
By
–
The Class of 2026 is entering one of the toughest labor markets in years. With 90% of U.S. employers using AI screening tools, what is the impact on job seekers? This study of hiring algorithms finds concerning patterns of systemic rejection: https://
hai.stanford.edu/news/ai-hiring
-tools-can-yield-racial-bias-and-systemic-rejection
… -
GuidedVLA Helps Robots Focus by Splitting Action Decoder
By
–
Can robots stop getting distracted by visual clutter and actually focus on what matters for the task? Researchers from Fudan, SJTU, and HKU present GuidedVLA. Instead of treating the action decoder as a single black box, they split it into specialized “attention heads”—each
-
Hermes Agent learns by failing and recording lessons
By
–
Hermes /learn explained. agents usually learn the hard way. they struggle through a task live, fail a few times, find the path that works, and only then write down what they figured out. the lesson costs you a painful session before it becomes reusable. Hermes Agent by Nous