RestoreKV: Recovering Full-Cache Behavior Under Aggressive Query-Agnostic KV Cache Eviction

Query-agnostic KV cache eviction compresses a context once and reuses the resulting cache for arbitrary future queries, but performance can collapse under tight budgets. Existing methods primarily improve which original KV pairs are retained. We introduce RestoreKV, which complements this selection-based formulation with learned restoration under the same total KV budget. Our key insight is that, although the information lost through eviction is context-specific, the mechanism for generating its compact complement can be shared across contexts. After context prefill, a few restore tokens attend to the full KV cache in a single LoRA-adapted pass, generating a compact, context-conditioned restore cache. The base importance scorer and eviction rule remain unchanged, and the adapters are disabled for all subsequent queries and decoding. RestoreKV is trained through parameter-efficient self-distillation from the frozen full-cache model, optimizing only $0.4\%$ of the parameters and requiring no task-specific tuning. Across four backbones and four long-context benchmarks, RestoreKV substantially reduces compression-induced degradation. On Qwen3-4B, it improves 59 of 60 paired, budget-matched settings across five base eviction methods; at a $5\%$ budget, it raises KVzip from $38.2$ to $73.2$ on RULER-4K. Applied to KVzip+, RestoreKV reaches $86.4$ RULER accuracy at $16\times$ compression on the KVPress Benchmark, while adding less than $0.5\%$ one-time cache-construction overhead in a 32K-context evaluation. Our project page is available at https://paper.pnu-cvsp.com/RestoreKV/

Efficient Neural Video Representation with Temporally Coherent Modulation

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/.

House of Cards: Massive Weights in LLMs

Massive activations, which manifest in specific feature dimensions of hidden states, introduce a significant bias in large language models (LLMs), leading to an overemphasis on the corresponding token. In this paper, we identify that massive activations originate not from the hidden state but from the intermediate state of a feed-forward network module in an early layer. Expanding on the previous observation that massive activations occur only in specific feature dimensions, we dive deep into the weights that cause massive activations. Specifically, we define top-$k$ massive weights as the weights that contribute to the dimensions with the top-$k$ magnitudes in the intermediate state. When these massive weights are set to zero, the functionality of LLMs is entirely disrupted. However, when all weights except for massive weights are set to zero, it results in a relatively minor performance drop, even though a much larger number of weights are set to zero. This implies that during the pre-training process, learning is dominantly focused on massive weights. Building on this observation, we propose a simple plug-and-play method called MacDrop (massive weights curriculum dropout), to rely less on massive weights during parameter-efficient fine-tuning. This method applies dropout to the pre-trained massive weights, starting with a high dropout probability and gradually decreasing it as fine-tuning progresses. Through experiments, we demonstrate that MacDrop generally improves performance across zero-shot downstream tasks and generation tasks.