Ting-En Lin

Research scientist at Alibaba Group, Qwen-Character.

[Ting-En Lin @tnlin_tw]

🚀 Excited to introduce MMEvol, which improves MLLM through complex and diverse instruction with perceptual, cognitive, and interactive evolution.
🌟 Achieving a 3.1% accuracy boost across 13 VL tasks.
Code and data will be released soon; stay tuned!
📄 https://t.co/26KJwirI85

[Fuli Luo @_LuoFuli]

Nearly half a year of silence. We spent it studying one problem: how far RL can scale.

MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.

Streaming the run: https://t.co/ZSxahzJRju

[DAN KOE @thedankoe]

https://t.co/7l7Jef99QZ

[Percy Liang @percyliang]

🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung scaling ladder from 1.6B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.

[Ting-En Lin @tnlin_tw]

@JustinLin610 ❤️🥃

[Oliver Henry @oliverhenry]

https://t.co/Mg6IwpDnGK

[Ting-En Lin @tnlin_tw]

Feeling overwhelmed. Rebuttals usually help raise reviewer scores, and without that, acceptance chances are low. With ACL 2026 overlapping with ICLR, some are thinking about withdrawing and resubmitting to ACL to avoid the uncertainty as LLM research moves so quickly. @aclmeeting

[Ting-En Lin @tnlin_tw]

@iclr_conf Why are rebuttal discussions being reverted? This is unfair to both authors and reviewers who worked hard to improve the paper.

New ACs should be able to see the rebuttal history, otherwise what’s the point of the process?

[Ting-En Lin @tnlin_tw]

should we just forward the previous discussion and the reviewers’ promises to raise their scores once issues are addressed?

[Susan Zhang @suchenzang]

[episode 126 of frontier lab gossip: OH in the Louvre]

>> yo! did you hear? the ***** team trained a better *** than the [[ TBD ]] folks!

> what?? how did that happen...

>> oh man it's been a mess
>> first guy that came over got backstabbed by second guy who came over
>> so they put an OAI guy in charge of mid/post-train, aka distill-from-gpt-oss
>> they're running pretrain like Gemini does - same methodology
>> and overall arch like deepseek
>> smallest model was ok, next size up was concerning, and the one after that was just bad
>> so they delay scaling further, tried again, still bad
>> now [[ TBD ]] wants help from regular plebs
>> plebs clean their own data, simplify a bunch of stuff, get [[ N ]]x higher MFU
>> and immediately show a better scaling trend than the [[ TBD ]] setup
>> so [[ TBD ]] retrains again, still bad
>> OAI guy blames pretraining, pretraining guy blames data, so data guy gets fired

> what's the 🍎 guy doing?

>> building his own infra and waiting it out
>> there's at least 2 codebases already, so why not make it 3

> is that useful?

>> who knows... nearly no progress since beginning of year
>> [[ old_giant_model ]] is still better than all [[ TBD ]] models so far
>> and now they're too busy pointing fingers, alienating actual talent, and fighting amongst each other
>> to make anything better

> hm... does [[ big_man_up_top ]] know about all this?

>> [[ new_guy_in_charge ]] controls all info to him right now, so he probably doesn't know
>> that he's always had the talent
>> the only difference now is who gets to repackage whose work as their own
>> in order to keep up the facade of being a worthwhile investment

-----

[[ ... ]] content not to be taken literally; replace at your own discretion

[Jason Wei @_jasonwei]

There are traditionally two types of research: problem-driven research and method-driven research. As we’ve seen with large language models and now AlphaEvolve, it should be very clear now that total method-driven research is a huge opportunity.

Problem-driven research is nice because you have a consistent and specific goal. The goal is usually virtuous, so it feels good to have a mission and identity. However, it just doesn’t work due to The Bitter Lesson. Basically everything in classical NLP (machine translation, summarization, chatbots) lost to simple scaling. ChatGPT is a prime example—it used nothing from chatbot research and certainly wasn’t the intended end goal of OpenAI’s 2022 research program, but was a huge hit because someone (John Schulman et al) figured out the right way to package large language models as a product.

Method-driven research feels less stable because you’re constantly searching for problems and you have to be opportunistic. But I believe AI will allow method-driven research to dominate progress in most fields of science, one-by-one. The latest method (or “hammer”), as we’ve seen in AlphaEvolve, is ruthless search and optimization against a reward function (whether this requires RL or not is a separate discussion). Things that problem-driven researchers have been trying to solve for a long time like the kissing number problem will become nails hit by the hammer. Eventually the hammer will become bigger, stronger, and more general and will hit more and more nails.

So a very important meta-skill for the next decade will be knowing how to create the right environments to use The Hammer. Ironically, the problem-driven researchers, who by definition are experts in a specific problem, are well-positioned to create these environments. If, that is, they can put down their egos and pick up the hammer.

[Haibin @eric_haibin_lin]

Recent updates on @verl_project (RL lib for LLMs):

Engine:
- Megatron qwen & GRPO support, v0.11 upgrade
- vllm v0.7 integration with v1 mode
- experimental sglang integration

Algorithm & recipes:
- vision language reasoning with qwen2.5-vl
- PRIME, RLOO, remax, math-verify rewards, etc

Docs:
- tutorial for distributed training setup and debugging
- programming model tutorial

Hardware:
- experimental AMD support

And many awesome community projects such as code-R1, Easy-R1, Search-R1, RAGEN, etc. Big thank you to the community!

Working on multi-turn & environment/tool supports. Stay tuned...

[Jason Wei @_jasonwei]

Nice paper from Deepmind takes a fresh angle on factuality: https://t.co/Se4aSpDeWL

While most existing factuality datasets focus on public world knowledge, this paper evaluates whether responses are consistent with a provided document as context. This is an elegant and tractable problem formulation because the surface area is well-specified and there are no disputes about what the ground truth should be.

This type of factuality is hugely important for personalizing LLMs, which is a direction that the field has not been focusing on so far. Current LLM products (for the most part) are used in a context-agnostic way, for example, answering facts about the world or debugging a small piece of code. But a huge piece of intelligence is the ability to understand and follow the context for a very specific use case.

Staying true to context is required not just for tasks like software engineering but even for tasks that don’t require lots of specialized training, such as taking someone’s fast food order. A language model that is able to totally not hallucinate given a long piece of context seems like a goal that we should be able to get to soon that will be transformative for these types of use cases

[Ting-En Lin @tnlin_tw]

@huybery Glad to help😏

[OpenAI @OpenAI]

12 days.
12 livestreams.
A bunch of new things, big and small.

12 Days of OpenAI starts tomorrow.

[Ting-En Lin @tnlin_tw]

@brianckwu @zraytam I Identify as an Apache Attack Helicopter 🚁

[OpenAI @OpenAI]

GPT-4o got an update 🎉

The model’s creative writing ability has leveled up–more natural, engaging, and tailored writing to improve relevance & readability.

It’s also better at working with uploaded files, providing deeper insights & more thorough responses.

[Ting-En Lin @tnlin_tw]

Exciting news! 🚀 MMEvol's dataset and code are now officially available! If you find it useful, please give us a star or cite our paper 🌟
https://t.co/XioVDLxa1s

[Ting-En Lin @tnlin_tw]

🚀 Excited to introduce MMEvol, which improves MLLM through complex and diverse instruction with perceptual, cognitive, and interactive evolution.
🌟 Achieving a 3.1% accuracy boost across 13 VL tasks.
Code and data will be released soon; stay tuned!
📄 https://t.co/26KJwirI85

[Ting-En Lin @tnlin_tw]

@zraytam "Posters are often better than orals."
That is true. I also prefer posters to meet new friends.

[Ting-En Lin @tnlin_tw]

@ym_tseng @customnlp4u Great work! congrat

[Ting-En Lin @tnlin_tw]

@zraytam Thanks for the support 😄😄