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
🚨 Your RL only improves 𝗽𝗮𝘀𝘀@𝟭, not 𝗽𝗮𝘀𝘀@𝗸? 🚨
That’s not a bug — it’s a 𝗳𝗲𝗮𝘁𝘂𝗿𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗼𝗯𝗷𝗲𝗰𝘁𝗶𝘃𝗲 you’re optimizing.
You get what you optimize for. If you want better pass@k, you need to optimize for pass@k at training time.
🧵 How?
Nvidia just opensourced Describe Anything!
It can generate detailed descriptions for user-specified regions in images and videos, marked by points, boxes, scribbles, or masks
SIFT-50M: A Large-Scale Multilingual Dataset for Speech Instruction Fine-Tuning
- 50M examples across 14K hours of speech
- covers 5 languages with diverse speech understanding + generation tasks
- built using LLMs + expert models on open data
Trained on SIFT-50M, SIFT-LLM:
- outperforms prior speech-text LLMs on instruction following
- matches SoTA on core speech tasks
``SIFT-50M: A Large-Scale Multilingual Dataset for Speech Instruction Fine-Tuning,'' Prabhat Pandey, Rupak Vignesh Swaminathan, K V Vijay Girish, Arunasish Sen, Jian Xie, Grant P. Strimel, Andreas Schwarz, https://t.co/OeurkarULx
It would be a useful exercise to imagine where the AI industry would be today if AI research labs in industry had remained closed, had not open sourced their code, had patented everything and enforced their patents.
Imagine a world without PyTorch, and with enforced patents on Transformers, ResNet, Mask-RCNN, FPN, SAM, DINO, seq2seq, wav2vec, memory-augmented networks, Tansformers, BatchNorm, LayerNorm, Adam, denoising auto-encoder, joint-embedding architectures, and tons of SSL methods....
@UPPCLLKO@PuVVNLHQ@EMofficeUP Also visited SDO, EUDD IV Varanasi who is not accepting my application to change back to postpaid. This is the state of government services. Despite trying so hard to pay my bills and help gvt with revenue generation, not able to do so :(
@UPPCLLKO@PuVVNLHQ Not able to pay bills online as connection changed to prepaid to postpaid w/o my knowledge. Bills come like postpaid but portal not accepting payment with error that service not available for prepaid customers. 1912 is of no help here.
@MoHFW_INDIA@OfficeOf_MM Took my first two doses of Covid19 vaccine outside India, I'm not able to get booster dose. Hospitals are not able to do onsite registration, contrary to what is mentioned on CoWIN FAQ: https://t.co/W2eYpbBXvB
Multiple interpretations of ambiguous percept must be associated with multiple values of an explanatory latent variable.
By *latent*, I mean that they are not outputs but internal inputs.
What are the mechanisms in the brain for exploring the set of plausible values?
Still trying to figure out where most of the real signal is in typical 5-star rating distributions. It's definitely not average rating. Simply the number of ratings is usually quite good. The ratio of 5-star to 4-star is for some reason quite good too.
Memorization Without Overfitting: Analyzing the Training Dynamics of Large Language Models
abs: https://t.co/0mnkK6SPZu
show that larger models can memorize a larger portion of the data before over-fitting and tend to forget less throughout the training process