HackerTrans
TopNewTrendsCommentsPastAskShowJobs

FranckDernoncou

no profile record

Submissions

Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution

github.com
246 points·by FranckDernoncou·2 mesi fa·45 comments

comments

FranckDernoncou
·2 mesi fa·discuss
Paper: https://arxiv.org/abs/2605.12825 ; Code+models: https://github.com/chiennv2000/orthrus ; Disclosure: co-author.

Idea: Inject a trainable diffusion attention module into each layer of a frozen AR Transformer. Both heads share one KV cache. Diffusion head projects K=32 tokens in parallel; AR head verifies in a second pass and accepts the longest matching prefix. Output distribution is provably identical to the base model.

Results:

- Up to 7.8x TPF, ~6x wall-clock on MATH-500.

- 16% of params trained, <1B tokens, 24h on 8xH200.

- vs. diffusion LMs (Dream, Fast-dLLM-v2, SDAR, Mercury, Gemini Diffusion): they modify base weights and lose accuracy (Fast-dLLM-v2: -11 pts on MATH-500). Orthrus freezes the backbone; accuracy matches Qwen3-8B exactly.

- vs. Speculative Decoding (EAGLE-3, DFlash): no external drafter, no separate cache, zero TTFT penalty (no drafter to init/sync). KV overhead is O(1) (~4.5 MiB flat). Acceptance length on MATH-500: 11.7 vs. 7.9 (DFlash) vs. 3.5 (EAGLE-3).

- Single-step denoising beats multi-step (6.35 vs. 3.53 TPF). KL distillation beats CE on acceptance rate.

Limitations: strictly bounded by the frozen base model (inherits its biases, hallucinations, knowledge gaps); Qwen3-only evaluation; greedy + rejection sampling only.
FranckDernoncou
·7 mesi fa·discuss
But harder questions will typically get no replies and/or get closed for being too niche and/or get downvoted because someone got upset they didn't understand the question. That's my experience.
FranckDernoncou
·7 mesi fa·discuss
> LLMs are terrible, but stackoverflow is definitely worse.

Exactly. Some stackoverflow users keep complaining about AI slop but they fail to realize that human slop is much more rampant.
FranckDernoncou
·7 mesi fa·discuss
> In fact, SO has a great opportunity to specialize in answering questions that LLMs cannot answer (bleeding edge technology, complex debugging problems, emerging issues, you name it.)

But they are missing that opportunity because such questions typically get no answer and no votes (or sometimes even downvoted as some users tend to downvote questions they don't understand), which caused them to be automatically deleted later. And some of these questions will be closed e.g. as too niche (complex debugging problems). I've posted over 7000 questions across Stack Exchange and that's my experience.

> quick and confident-sounding answers to whatever problem you throw at them.

Just like an answer on Stack Overflow.
FranckDernoncou
·7 mesi fa·discuss
The entire Stack Exchange network is collapsing: [The number of questions has decreased 10-fold: where did users go and what can be done to prevent that?](https://meta.stackexchange.com/q/408728/178179)

As for Stack Overflow, the number of questions has decreased 25-fold (!) since 2018. A single person could now answer all new questions if they had nothing else to do.