Iterative reflections for LLMs can outperform heavy RL? This paper shows that having the LLM reflects on its own trajectories, rewrite its own prompts, and evolve a diverse pool of candidates beats RL w/ GRPO so far on four reasoning tasks . 10% improv with 35x fewer rollouts!
MACHINE LEARNING
-
COLT 2025 Conference: Learning Theory Authors Challenge Perspectives
By
–
COLT authors beg to differ: https://
learningtheory.org/colt2025/ -
Training weights in machine learning models explained
By
–
sorry, by train i mean actually updating weights, but i appreciate the comment!
-
Training and Fine-tuning Models: Community Discussion Opportunity
By
–
If you train/fine-tune models, I'd love to chat with you. Please comment or DM and I'll reach out!
-
Why AI Is Ending The Era Of The End User
By
–
Why AI Is Ending The Era Of The End User
#AI #AIio #AIInnovation #ML #DataScience #Futureofwork @demishassabis @Ronald_vanLoon @TamaraMcCleary @geoffreyhinton @goodfellow_ian @jeffdean @erikbryn http://
ow.ly/HKSL30sOj63 -
Attention is All You Need paper implementation with visualizations
By
–
Primera prueba, usando el paper de Attention is All You Need y muy sorprendido! Las visualizaciones acompañan muy bien al resultado final y lo hace más digerible (y sólo puedo pensar todo lo que puede mejorar esto!).
— Carlos Santana (@DotCSV) 29 juillet 2025
Pero es que además la narración, más expositiva que la de los… pic.twitter.com/MlCRQ2F1V2Primera prueba, usando el paper de Attention is All You Need y muy sorprendido! Las visualizaciones acompañan muy bien al resultado final y lo hace más digerible (y sólo puedo pensar todo lo que puede mejorar esto!). Pero es que además la narración, más expositiva que la de los
-
BAIR Faculty Awarded ACL Doctoral Dissertation Prize
By
–
Congratulations to BAIR Faculty @sewon__min who has been awarded the 1st ACL Computational Linguistics Doctoral Dissertation Award for the dissertation: Rethinking Data Use in Large Language Models!
-
GPT-2 Negation Understanding: From 2019 Struggles to Modern Progress
By
–
i fondly recall arguing in 2019 with another researcher about how whether models understand negation. this really was a major sticking point at the time, GPT-2 couldn't reliably distinguish between "I did like that" and "I didn't like that". at some point we fully blew by the
-
Trackio: Hugging Face Lightweight Experiment Tracking Library
By
–
Introducing Trackio A Lightweight Experiment Tracking Library from Hugging Face uv pip install trackio
-
Zero-Cost Model Serving: SFT vs RLHF Strategy
By
–
hypothetical situation – i am an AI company that's reduced the cost of transferring and storing models to zero. i can serve each user their own model with no overhead what do i do? directly SFT user-specific models on their data? or RLHF on the chat ratings? something else?