Adyasha Maharana

Can LLMs keep track of very long conversations? We evaluate 'conversational memory' of LLMs via 3 tasks on our dataset of multi-session multimodal dialogs --> LLMs struggle to remember, reason over history, draw long-range temporal/causal connections arxiv.org/abs/2402.17753 🧵

Will be at #NeurIPS2025 from 2nd to 6th Dec. Excited to chat about async RL, Environment Exploration, Agents/Tool use, User Simulator, Synthetic Data Generation or any other topic!! You can find me at the @databricks booth @ Tue 12 - 4pm

🚨 New #CVPR2025 Paper 🚨 🏀BASKET: A Large-Scale Dataset for Fine-Grained Basketball Skill Estimation🎥 4,477 hours of videos⏱️ | 32,232 players⛹️ | 20 fine-grained skills🎯 We present a new video dataset for skill estimation with unprecedented scale and diversity! A thread👇

🚨 Introducing Adapt-♾: Scalable Lifelong Multimodal Instruction Tuning via Dynamic Data Selection! ▶️ New multimodal instruction tuning datasets are continuously released, often containing redundant/similar content or targeting highly varied skills (i.e., tasks). ▶️ How can we enable scalable, lifelong instruction tuning for MLLMs, where a temporal stream of multi-task, multimodal instruction-tuning datasets are continually added to the existing training pool? ▶️ We present Adapt-♾, a scalable and adaptive data selection strategy that facilitates the effective learning of MLLM on new skills while reinforcing previously learned skills over time. 📖: arxiv.org/abs/2410.10636 Thread 🧵👇

[1/7] New paper alert! Heard about the BitNet hype or that Llama-3 is harder to quantize? Our new work studies both! We formulate scaling laws for precision, across both pre and post-training arxiv.org/pdf/2411.04330. TLDR; - Models become harder to post-train quantize as they are overtrained on lots of data, so that eventually more pretraining data can be actively harmful if quantizing post-training! - The effects of putting weights, activations, or attention in varying precisions during pretraining are consistent and predictable, and fitting a scaling law suggests that pretraining at high (BF16) and next-generation (FP4) precisions may both be suboptimal design choices! Joint work with @ZackAnkner @bfspector @blake__bordelon @Muennighoff @mansiege @CPehlevan @HazyResearch @AdtRaghunathan.