[1/9] VLMs caption well, but no simple RL trains them in multi-step sims and shows gains. VL-DAC: a lightweight RL algorithm on top of VLM + cheap sims yields agents that finish long quests and transfer to skill specific benchmarks with no tuning. HF link: https://t.co/HIGDN7NisA
[1/10] ๐ Proud to announce our paper โThe Differences Between Direct Alignment Algorithms are a Blurโ is now #1 on Hugging Face's Daily Papers list @_akhaliq! We explore DAAs for language model alignment, comparing methods and uncovering insights.
https://t.co/aIPbKqB1S1
๐ Exciting News! Our research, "EXACT: How to train your accuracy", will be published in Pattern Recognition Letters (Q1)! ๐
A huge thanks to my co-authors, @stdereka and @Scitator!
1/7 Deep ensembles deliver exceptional results in uncertainty estimation (@balajiln) and OOD detection (@SharonYixuanLi). But how can we push their performance even further? Our latest paper, accepted at @ieeeICIP#ICIP2024, uses saliency maps to enhance ensemble diversity. ๐งต๐
6/7 All our comparisons were conducted using the OpenOOD benchmark. Huge thanks to @SharonYixuanLi, @JingkangY, and @liuziwei7 for their invaluable benchmark!