Seungjun Shin

Seungjun Shin, Jaehoon Oh, Dokwan Oh

Attention mechanisms are central to the success of large language models (LLMs), enabling them to capture intricate token dependencies and implicitly assign importance to each token. Recent studies have revealed the sink token, which receives disproportionately high attention despite their limited semantic role. In this paper, we first expand the relationship between the sink token and other tokens, moving beyond attention to explore their similarity in hidden states, considering the layer depth. We observe that as the layers get deeper, the cosine similarity between the normalized hidden states of the sink token and those of other tokens increases, and that the normalized hidden states of the sink token exhibit negligible changes. These imply that other tokens consistently are directed toward the sink token throughout the layers. Next, we propose a dynamic token selection method, called OrthoRank, using these findings to select important tokens. Specifically, in a certain layer, we define token importance by the speed at which the token moves toward the sink token. This is converted into orthogonality with the sink token, meaning that tokens that are more orthogonal to the sink token are assigned greater importance. Finally, through extensive experiments, we demonstrated that our method results in lower perplexity and higher zero-shot accuracy compared to layer pruning methods at the same sparsity ratio with comparable throughput, while also achieving superior performance on LongBench.

Jegwang Ryu, Minkyu Kim, Seungjun Shin et al.

The efficient storage and transmission of language model weights is becoming increasingly important, as their scale and adoption continue to grow. However, as our understanding of this new data modality is limited, designing a good compression algorithm for language model weights heavily relies on manual, trial-and-error approaches. In this paper, we propose a learned compression framework that trains neural codecs directly from pretrained language model weights. Unlike conventional data (e.g., images), language model weights pose unique challenges: the sizes and shapes of weight tensors vary significantly, and the reconstruction quality must be judged by downstream model predictions rather than naïve MSE loss. To address this, we introduce Neural Weight Compression (NWC), a novel autoencoder-based neural codec tailored to model weight compression. The proposed method inherits the advantages of autoencoder-based codecs while incorporating three technical components: (1) column-wise tensor chunking and normalization; (2) an importance-aware training loss; (3) an inference-time error compensation mechanism guided by model outputs. Experiments on open-weight language models show that NWC achieves competitive or state-of-the-art accuracy-compression tradeoffs, with particularly strong results at 4-6 bit precisions where accuracy remains nearly on par with FP16 models.

Seungjun Shin, Suji Kim, Dokwan Oh

Implicit neural representations (INR) has found successful applications across diverse domains. To employ INR in real-life, it is important to speed up training. In the field of INR for video applications, the state-of-the-art approach employs grid-type parametric encoding and successfully achieves a faster encoding speed in comparison to its predecessors. However, the grid usage, which does not consider the video's dynamic nature, leads to redundant use of trainable parameters. As a result, it has significantly lower parameter efficiency and higher bitrate compared to NeRV-style methods that do not use a parametric encoding. To address the problem, we propose Neural Video representation with Temporally coherent Modulation (NVTM), a novel framework that can capture dynamic characteristics of video. By decomposing the spatio-temporal 3D video data into a set of 2D grids with flow information, NVTM enables learning video representation rapidly and uses parameter efficiently. Our framework enables to process temporally corresponding pixels at once, resulting in the fastest encoding speed for a reasonable video quality, especially when compared to the NeRV-style method, with a speed increase of over 3 times. Also, it remarks an average of 1.54dB/0.019 improvements in PSNR/LPIPS on UVG (Dynamic) (even with 10% fewer parameters) and an average of 1.84dB/0.013 improvements in PSNR/LPIPS on MCL-JCV (Dynamic), compared to previous grid-type works. By expanding this to compression tasks, we demonstrate comparable performance to video compression standards (H.264, HEVC) and recent INR approaches for video compression. Additionally, we perform extensive experiments demonstrating the superior performance of our algorithm across diverse tasks, encompassing super resolution, frame interpolation and video inpainting. Project page is https://sujiikim.github.io/NVTM/.