Aswin RRV

@sheriyuo Intentionally*

I think, crowd-review system (like whats happening in twitter/X) is better than these conferences. Some issues I have seen: Reviewer Sabotage (Like internationally asking for orthogonal experiments in the last day of the rebuttal) Unresponsive ones Unjustified Rejections and so on and on

I’ve already seen dozens of papers just like 544 and 5444 get rejected, and I’m honestly confused. Feels like something’s off with the review process lately. 😕

@rajammanabrolu Which one is the reasoning during midtraining one?

My lab and collaborators had 4 papers on everything from multi objective alignment, reasoning during mid training, multimodal synthetic data, and generating RL tasks accepted to #ICML2026! Come hang out with us in Seoul and we can talk about the exciting follow-ups!

Anyone got reviewer sabotaged??? 🙂 #icml_2026 #peer_review

@wjwang2003 @AzmineWasi @icmlconf Yeah, if they ask this question. I think its accept. If they ask Opt-in or Veto, then it reject.

@aswinrrv @AzmineWasi @icmlconf someone say they have another type

@icmlconf ICML Position Paper decisions seems out, indirectly 👀 Public-release or In-person presentation...?

@giffmana I think this is not something surprising right? Well, consider a mode collapsed RL model trained on math and you evaluate it on say Code task, you can see some style transfer happening.

In other words: their RL transfers/generalizes.

🫡

@FazlBarez @icmlconf Postponed or preponed?

Interesting way of announcing decisions won’t be on time @icmlconf 🙈

@sheriyuo What about Prime-RL, isn't it stable?

When you’re running RL experiments with verl Me: Damn, I can run GRPO and GSPO but not DAPO > Spent a day or two debugging with the infra guy, then realized upgrading vllm to 0.18.0 fixes it, but now my verl needs to be updated Me: Let the Code Agent handle the migration Agent: I spent an hour writing a bunch of garbage code for you Me: wtf, so I rewrote everything myself in an hour and finally got DAPO running > One day later Me: wtf, why does it only work for dense models, why is MoE broken again > xxxxx tons of error logs (random order from Ray) Me: feed it to DeepSeek -> DeepSeek edits -> ... -> ten turns later DeepSeek: you should downgrade vllm to 0.17.0 Me: f**k verl

Training giant Mixture-of-Experts (MoE) models from scratch is incredibly expensive. What if we could grow their capacity mid-training without increasing inference costs? Introducing Expert Upcycling! A new compute-efficient recipe for scaling MoEs that saves ~32% in GPU hours. 🚀👇 Full paper: huggingface.co/papers/2604.19… code here: github.com/amazon-science…

@novasarc01 😂😂😂

i am seeing that similar to mood swings researchers have policy swings (jumping from on-policy to off-policy and vice-versa). when training is unstable everyone becomes deeply on-policy. when rollouts get expensive everyone rediscovers off-policy like it is a forgotten religion/ancient sacred thing.

A novel argument to do a PhD in 2026 is to expand the training set of AI models by a unique 100-pager. Before, only a few experts in the world would read it — now, every single model can do so & benefit from reusing it in unexpected contexts. It all happened so fast…

First in the country to establish the Ministry of AI! Super happy about this. @TVKVijayHQ #TVKVijay‌ #TVKVijay‌HQ

ஜனநாயகன் திரைப்படம் இணையத்தில் யாராலோ வெளியிடப்பட்டிருப்பது அதிர்ச்சியையும்,வேதனையையும் அளிக்கிறது. திரை அமைப்புகள் இதற்கு எதிராகக் குரல் எழுப்பி, அரசு இதைச் செய்தவர்களைக் கண்டுபிடித்து கடுமையான தண்டனை அளிக்க வேண்டும். இது போன்ற குற்றம் இனியும் தொடரக்கூடாது.

The leak of #Jananayagan is not an accident - it is the result of systemic failure. Had due process been timely, we would not be here. Inordinate delays in certification created fertile ground for piracy. When legal access is stalled, illegitimate channels take over. Piracy is beyond politics; it is an attack on the art and artist itself. It endangers the work of hundreds of artists and technicians, and the investments of honest tax paying producers, exhibitors and theatre owners, all who sustain the cinema we love. Who protects the creator when the system fails? We need accountability, swift certification, strict enforcement, and real-time takedowns. I trust true lovers of cinema will unite and give a befitting response by watching the film legally in theatres, as you stood with me in the past.

@langer_han @DHolzmueller @icmlconf Final justification is visible for everyone?

@DHolzmueller @icmlconf There is a final justification which is shared with everyone which I am guessing I only got one so far and 3 have ghosted. My concern is the ghosting nature of reviews.

Really bummed that 2 @icmlconf reviewers just decided to ghosts us after putting their questions in rebuttal acknowledgement and not looking at our responses. Come on folks we can do better than this.

cs grad:

Okay, I am gonna stay out of x today! :)