The paper investigates how attention dynamics evolve across recurrent depth in language models, finding that attention support stabilizes early while hidden states and outputs take longer. It proposes WISE, a training‑free method that uses full attention in early steps and then reuses the discovered sparse working set for later steps, preserving performance on multi‑hop QA tasks. Experiments show that WISE maintains quality up to 2K context, offers measurable speedups, and highlights the importance of recurrent discovery of attention support.
By Ke Wan, Chen Chen
arXiv:2604. 07822v2 Announce Type: replace-cross Abstract: We study implicit reasoning, i.
By Harsh Kohli, Srinivasan Parthasarathy, Huan Sun, Yuekun Yao
arXiv:2610.25076v2 Announce Type: replace
Abstract: Mathematical problem solving and autonomous agents are increasingly important applications of language models, demanding stronger reasoning capabil...
By Yiqin Hao, Boyi Zeng, Zitong Wang, He Li, Shixiang Song, Feichen Song, Yifan Liu, Ran Tao, Bryan Dai, Zhouhan Lin
arXiv:2603.21676v2 Announce Type: replace-cross
Abstract: Standard Transformers have a fixed computational depth, limiting their ability to generalize to tasks that require variable-depth reasoning....
By Hung-Hsuan Chen
arXiv:2608. 15062v1 Announce Type: cross Abstract: Scaling transformer language models creates an inherent tension between expressivity and memory efficiency.
By Amr Hegazy, Amr Alanwar, Mostafa Elhoushi
The paper introduces Gated Recurrent Transformers, a depth‑sharing architecture that brackets a single shared core with fixed prelude and coda blocks and uses a lightweight projection and element‑wise update gate to modulate recurrent updates. This design allows functional specialization across recurrences while reducing memory footprint. Experiments show that, under equal FLOPs or parameter budgets, the recurrent model matches or surpasses deeper GPT‑2 Small baselines, achieving similar or better accuracy with fewer parameters and lower peak decoding memory.
By Amr Hegazy, Amr Alanwar, Mostafa Elhoushi
Prism Attention introduces a proposal‑refined sparse attention framework that replaces static cross‑layer index reuse with a coarse‑to‑fine strategy, allowing layer‑wise adaptability while maintaining efficiency. The method leverages the observation that high‑attention tokens remain largely stable across layers, with changes concentrated near the boundary of previously selected regions. Experiments demonstrate state‑of‑the‑art performance on reasoning and long‑context benchmarks, achieving a 3.4× decoding speedup over FlashAttention at 128K context length.
By Zhenxu Tian, Kebin Liu, Zhengwu Yang, Yi Su, Qingqing Dang, Yanlin Sha, Kaipeng Deng, Dianhai Yu, Yanjun Ma, Juntao Li, Min Zhang
arXiv:2608.15062v3 Announce Type: replace-cross
Abstract: Scaling transformer language models creates an inherent tension between expressivity and memory efficiency. While unique weights across layer...
By Amr Hegazy, Amr Alanwar, Mostafa Elhoushi
arXiv:2607. 21291v1 Announce Type: cross Abstract: Large language models (LLMs) achieve strong generation and reasoning performance, but the Transformer architecture incurs high inference cost.
By Yidu Wu, Xiang Wang, Kejie Zhao, Zhangchi Wang, Qinghai Guo, Xiaoying Tang
The Asymmetric Looped Transformer (AsymLoop) separates context encoding from state refinement by using a single-pass causal encoder to produce fixed global key-value caches and a weight-tied recurrent decoder to generate local key-values from its evolving state. Joint attention over both sources allows each update to combine persistent contextual information with the current intermediate state. Experiments on FineWeb-Edu across models from 60M to 1B show lower validation perplexity and higher downstream accuracy than full-loop models, with a 1B model improving zero-shot commonsense reasoning accuracy from 49.58% to 50.85%.
By Cheng Liang, Tengfei Zhang, Wenjie DU, Jianyu Wu, Changlian Ma, Shuchen Pu, Ya Zhang, Yanfeng Wang, Zhouhan Lin
arXiv:2607. 19358v1 Announce Type: new Abstract: Recent advances in long chain-of-thought reasoning models such as DeepSeek-R1 have led to increasingly longer inference context lengths under the test-time scaling paradigm.
By Yu Zhao, Zekun Zhang, Fan Jiang, Bo Zeng, Linlong Xu, Shimin Shan, Yu Liu, Longyue Wang, Weihua Luo
The paper investigates whether extra computation in recurrent Transformers should be allocated to more temporal steps or greater physical depth. Using Latent Recurrent Transformers (LRTs), the authors insert a latent thought token between vocabulary tokens, allowing each token to pass through the same $L$ layers twice while sharing parameters. Experiments on 16‑ and 20‑layer mixture‑of‑experts NanoChat backbones show that a single thought token brings a shallower model within 0.006–0.004 bits per byte of a double‑depth counterpart, recovering 67–81% of the improvement with roughly 48% fewer parameters.
By Zeyi Huang, Xuehai He, Yong Jae Lee, Yelong Shen