Transformers are great at NLP and computer vision tasks, but I was surprised to learn they still lag behind simple linear models at time series forecasting. The issue is how most Transformer architectures treat each timestamp as a token and fuse all th…

Multimodal sentiment analysis combines text, audio and video to understand human emotions. But extra inputs can add irrelevant or conflicting signals. So filtering matters. Researchers made a "Adaptive Language-guided Multimodal Transformer" …

I like Cities: Skylines, but struggle at building roundabouts. Turns out, despite being safer than intersections, they're also tricky to design in real life – small tweaks can ruin traffic flow. They're designed iteratively. This is a pain for …

Researchers from Brown University presented a new study supporting that translating unsafe prompts into `low-resource languages` allows them to easily bypass safety measures in LLMs. By converting English inputs like "how to steal without getting …

Can LLMs actually improve their own reasoning by self-correcting mistakes? A new paper from DeepMind and the University of Illinois looks to answer this quantitatively. The results show that unaided, LLMs struggle at self-correction for reasoning tasks…

LLMs keep getting more capable at generating natural language. But there's always room for improving the quality and alignment of their responses. Typically this requires lots of human effort to collect more training data. So researchers are explor…

LLMs like GPT-3 struggle in streaming uses like chatbots because their performance tanks on long texts exceeding their training length. I checked out a new paper investigating why windowed attention fails for this. By visualizing the attention maps, th…

When trying to get language models to solve complex math problems, researchers kept running into limits. Models like GPT-3 and ChatGPT still struggle with advanced algebra, calculus, and geometry questions. The math is just too abstract and symbol-heav…

When visualizing the inner workings of vision transformers (ViTs), researchers noticed weird spikes of attention on random background patches. This didn't make sense since the models should focus on foreground objects. By analyzing the output embed…

A new paper proposes Show-1, a hybrid model that combines pixel and latent diffusion for efficient high-quality text-to-video generation. Both of these approaches have tradeoffs, so researchers at the National University of Singapore tried a hybrid app…