Ramon Astudillo
@RamonAstudill12
Followers
559
Following
3K
Media
18
Statuses
2K
Principal RS at IBM Research AI. Speech, Formal/Natural Language Processing. Currently LLM post-training, structured SDG/RL. Opinions my own and non stationary
Manhattan, NY
Joined April 2019
All this "perplexity went down but benchmark did not go up" as if it was fully unexpected. It's "transfer learning" right? there should be a limit to the transfer, i.e. objectives are not the same.
0
0
0
RT @yacineMTB: @lexfridman you know what i learned? 1m qps is actually easier than 100k qps. at my last big tech wagie job, my last eng lea….
0
45
0
RT @RulinShao: 🎉Our Spurious Rewards is available on ArXiv! We added experiments on.- More prompts/steps/models/analysis. - Spurious Prom….
0
40
0
The 15th edition of the Lisbon Machine Learnings School (LxMLS 2025) is looking for its monitor team. As always alumni are especially welcome. Apply before the month ends!.
0
2
3
RT @srush_nlp: Really enjoyed this paper. Would love to see this kind of work in other domains.
0
4
0
RT @natolambert: I hear people are pretty into GRPO and RL these days, so I wrote up a pretty comprehensive research survey of recent paper….
0
95
0
RT @LChoshen: AI doesn’t get your culture?❌ butchers your language? 😤.With FeeL – you can fix that🛠️🌍. 💬 Talk to AI in your language. ✏️ Co….
0
9
0
RT @AsafYehudai: Survey on Evaluation of LLM-based Agents 🤖. Our paper is the first to provide a comprehensive overview of LLM-based agent….
0
83
0
Another way to see O-models is that the best way to select your pre-training data is a good initial selection using human categories pre-trained into a model with enough bits, and then Reinforcement Learning at post-training.
0
0
0
"The future immigrant lodging house" Judge magazine 1890
0
0
0
Feels very timely. Markets do not like feudalism for obvious reasons.
I don't know why people say the world runs on money, it clearly runs on certainty. Certainty enables all from personal relations to work, to funding debt of the biggest corporations and countries.
0
0
0
RT @DimitrisPapail: We should be seriously asking, how a 1.5B model that can't answer basic questions can also be that good at competition….
0
91
0
What if Chinese translations of mathematical problems present in English test sets (e.g. MATH) were not filtered from the pre-training corpora of Qwen/DeepSeek? this means the knowledge is there, just translated. This would also explain language switching when RL-ing CoT 👇.
1
0
0
We either are about to discover something amazing with a brotastic name like "abstraction-hyper-grokking" or we have serious test poisoning, maybe two fold. Here are LIMO/s1 results side to side. Same base model, 800/1K SFT on human/LongCoT-Machine highly curated data.
1
0
1
RT @andre_t_martins: Good to see @EU_Commission promoting OS LLMs in Europe. However (1) "OpenEuroLLM" is appropriating a name (#EuroLLM) w….
0
13
0
RT @Yikang_Shen: It's good to see Deepseek v3 draw everyone's attention to reducing the training cost of LLM. Over the last two years, we….
0
52
0