Simin Huo

Simin Huo

The ability to process ultra-long contexts is crucial for large language models (LLMs) to perform long-horizon tasks. While recent efforts have extended context windows to 1M and beyond, model performance degrades when sequence length exceeds the pre-trained range of positional encodings (e.g., RoPE), i.e., position exhaustion. This fundamental limitation must be overcome to achieve a truly infinite context. To address it, we propose Periodic RoPE (P-RoPE), a positional encoding mechanism designed to circumvent this exhaustion. It operates in conjunction with sliding window attention (SWA) to capture local dependencies and relative positions within each window. This local layer is then complemented by a global attention layer with No Positional Encoding (NoPE), enabling unbounded interaction across the entire sequence without positional constraints. By stacking these two types of layers, the model avoids the need for positional extrapolation to generalize longer and theoretically supports an infinite context window. Empirical results show that our model, MiniWin, outperforms MiniMInd with standard GPT architectures in long-context efficiency and stability. Our work provides a possible pathway toward LLMs with genuine infinite-context understanding. The code is available at \href{https://github.com/Cominder/miniwin}{https://github.com/Cominder/miniwin}.

Simin Huo, Ning Li

Token compression is crucial for mitigating the quadratic complexity of self-attention mechanisms in Vision Transformers (ViTs), which often involve numerous input tokens. Existing methods, such as ToMe, rely on GPU-inefficient operations (e.g., sorting, scattered writes), introducing overheads that limit their effectiveness. We introduce MaMe, a training-free, differentiable token merging method based entirely on matrix operations, which is GPU-friendly to accelerate ViTs. Additionally, we present MaRe, its inverse operation, for token restoration, forming a MaMe+MaRe pipeline for image synthesis. When applied to pre-trained models, MaMe doubles ViT-B throughput with a 2% accuracy drop. Notably, fine-tuning the last layer with MaMe boosts ViT-B accuracy by 1.0% at 1.1x speed. In SigLIP2-B@512 zero-shot classification, MaMe provides 1.3x acceleration with negligible performance degradation. In video tasks, MaMe accelerates VideoMAE-L by 48.5% on Kinetics-400 with only a 0.84% accuracy loss. Furthermore, MaMe achieves simultaneous improvements in both performance and speed on some tasks. In image synthesis, the MaMe+MaRe pipeline enhances quality while reducing Stable Diffusion v2.1 generation latency by 31%. Collectively, these results demonstrate MaMe's and MaRe's effectiveness in accelerating vision models. The code is available at https://github.com/cominder/mame}{https://github.com/cominder/mame.

Simin Huo, Ning LI

Video-language models (VLMs) face rapid inference costs as visual token counts scale with video length. For example, 32 frames at $448{\times}448$ resolution already yield >8,000 visual tokens in Qwen3-VL, making LLM prefill the dominant throughput bottleneck. Existing methods often rely on global similarity or attention-guided compression, incurring offsets to their gains. We propose \textbf{Temporal Token Fusion (TTF)}, a training-free, plug-and-play pre-LLM token compression framework that exploits structured temporal redundancy in video. TTF automatically selects an anchor frame, then for each subsequent frame, performs a local window similarity search (e.g.,$3\times 3$), fusing tokens that exceed a threshold. The compressed sequence maintains positional consistency across both prefill and decoding through coordinate realignment, enabling seamless integration with existing VLM pipelines. On Qwen3-VL-8B with threshold t=0.70, TTF removes about 67\% of visual tokens while retaining 99.5\% of the baseline accuracy and introducing only ${\approx}0.16$\,GFLOPs of matching overhead. Overall, TTF offers a practical, efficient solution for video understanding. The code is available at \href{https://github.com/Cominder/ttf}{https://github.com/Cominder/ttf}

Simin Huo, Ning Li

We introduce Iwin Transformer, a novel position-embedding-free hierarchical vision transformer, which can be fine-tuned directly from low to high resolution, through the collaboration of innovative interleaved window attention and depthwise separable convolution. This approach uses attention to connect distant tokens and applies convolution to link neighboring tokens, enabling global information exchange within a single module, overcoming Swin Transformer's limitation of requiring two consecutive blocks to approximate global attention. Extensive experiments on visual benchmarks demonstrate that Iwin Transformer exhibits strong competitiveness in tasks such as image classification (87.4 top-1 accuracy on ImageNet-1K), semantic segmentation and video action recognition. We also validate the effectiveness of the core component in Iwin as a standalone module that can seamlessly replace the self-attention module in class-conditional image generation. The concepts and methods introduced by the Iwin Transformer have the potential to inspire future research, like Iwin 3D Attention in video generation. The code and models are available at https://github.com/cominder/Iwin-Transformer.

Simin Huo

We propose \textbf{ULU}, a novel non-monotonic, piecewise activation function defined as $\{f(x;α_1),x<0; f(x;α_2),x>=0 \}$, where $f(x;α)=0.5x(tanh(αx)+1),α>0$. ULU treats positive and negative inputs differently. Extensive experiments demonstrate ULU significantly outperforms ReLU and Mish across image classification and object detection tasks. Its variant Adaptive ULU (\textbf{AULU}) is expressed as $\{f(x;β_1^2),x<0; f(x;β_2^2),x>=0 \}$, where $β_1$ and $β_2$ are learnable parameters, enabling it to adapt its response separately for positive and negative inputs. Additionally, we introduce the LIB (Like Inductive Bias) metric from AULU to quantitatively measure the inductive bias of the model.