Tag
The paper proposes Latent Sparse Attention (LSA), which hierarchically clusters Keys in latent space so that Queries only need to score against cluster centers. This achieves a 72.4x speedup in Prefill attention and a 42.5x speedup in Decode for a single attention layer of Qwen3.8-27B under 10M ultra-long context, reduces end-to-end first-token latency to about 1/9, and keeps accuracy essentially unchanged. The method includes two variants: distillable D-LSA and natively trained N-LSA, with code and distilled weights open-sourced.
Meta's DecisionRank paper introduces item-space on-policy distillation to transfer a generative LLM teacher's ranking knowledge into a single-token decision-model reranker, achieving state-of-the-art results on TREC DL, BEIR, and BRIGHT while running up to 24.5x faster than its generative counterpart and surpassing its 27B teacher at just 1.7B parameters.
The paper introduces two kernel-based interpolatory KV-cache compression methods, GAIM and Q-DAIM, along with linear-time surrogate variants and streaming attention extensions, achieving performance comparable to state-of-the-art on LongBench-E.
Glance integrates reranking and generation into a single 9B backbone using lightweight decision adapters and pointer heads that score passages from a shared prefill, reducing memory overhead while beating bge-reranker-v2-m3 and cross-encoders on MS MARCO benchmarks and HotpotQA answer F1.
GS-MoE 介绍了一种免训练的 MoE 压缩框架，通过分组子空间共享（低秩）与基于 Nyström 的神经元剪枝两条压缩轴同时削减专家参数；在 DeepSeek-MoE-16B-Base 上以 60% 参数压缩保留约 90% 的推理性能，优于 D2-MoE 的 73%。
GLOW-FM is a one-step latent flow-matching framework for efficient generative low-light image enhancement, performing conditional transport in a frozen Tiny-VAE latent space with Terminal Velocity Matching to achieve state-of-the-art cross-dataset generalization using a single function evaluation and 34 GFLOPs.
Mass General Brigham researchers present Clinical System One (CS1-4B), a 4B-parameter decision model that performs fast, typed judgments over clinical records, outperforming open decision models up to 27B parameters on 13 of 14 tasks while running on a single desktop GPU. Paired with a generalist LLM agent, it enables longitudinal chart review and clinical trial screening with 690-fold reduced token usage.
该论文提出一个 141M 参数的编码器式 PII 检测模型，使用来自 77 个公共数据集、按统一上下文检测标准重新标注的 112 万条样本训练，在 AI4Privacy 和 Nemotron-PII 基准上以更低的 GPU 内存和 2–10 倍吞吐量取得领先召回率。
A hands-on review of Nemotron 3.5 Bolt, noting its efficient KV cache, 2-bit/NVFP4 quantization, and Mamba2 hybrid attention that make it well-suited for small multi-user setups on limited hardware, despite only mediocre intelligence by current standards.
The paper introduces HybridAgents, an LLM agent architecture that dynamically switches between latent-space reasoning and text-based action via a lightweight HybridLink, combined with HybridSFT and HybridRL post-training, achieving 10.4% performance gains with large token, latency, and cost reductions across 11 benchmarks.
NVIDIA 发布 Nemotron Labs Dory1.0-4B-BF16，一款 4B 参数的循环式(latent-recurrent)大语言模型，采用 normalized-GPT 架构与滑窗/全注意力混合层，支持弹性推理计算与最后一次循环 KV 缓存，并原生支持 NVFP4 预训练，采用 OpenMDW-1.1 开放许可。
提出 Selection-Safe Conformal Early Exit 方法，通过终端路径分数与单一 split-conformal 分位数校准，在数据依赖的自适应深度选择下仍保持无分布覆盖保证，无需 α/L 修正，并在 CIFAR-100/ImageNet 等任务上验证了路径选择相比朴素逐层选择的覆盖优势。
这篇论文提出将注意力的指数核展开为 n 项截断 Taylor 级数，并将每项分解为张量幂内积，从而用固定数量的循环状态实现线性时间 Transformer，实测在长上下文下相比 softmax 注意力有显著加速和常数解码延迟。
AgriJev 将 Qwen3.5 系列（0.8B–9B）适配为农业领域的候选评分模型，通过单次前向传播直接输出候选分数与相对概率；AgriJev-9B 在 17 个外部测试视图上达到 75.44% 平均准确率，且评分延迟约为同规模答案生成的一半。
该论文提出了一种名为 Conformal Adaptation 的框架，通过共形风险控制（conformal risk control）在低分辨率下认证预测信号，再对已认证区域进行精细化计算，从而在 3D 医学图像分割、星系流量预测和钙信号预测等任务上实现 2.5–72 倍的墙钟加速，同时保留不确定性估计。
MJev-Spatial 提出一种 System 1 多模态决策模型，将空间推理任务统一为单次前向的分类框架，利用预训练的 query/key 投影为候选选项打分，在 12 项基准中以 65.64% 取得最高总分，并比自回归基线快 2.5 至 5.6 倍。
ReaDy introduces a speculative decoding method that uses hybrid linear attention recurrent states as dynamic context for drafting, achieving a 3.16x decoding speedup over autoregressive generation with Qwen3.8-27B on an NVIDIA B200 while improving throughput and acceptance length over DSpark across seven benchmarks.
Liquid AI 发布 LFM2.5-24B-A2B，一个 24B 总参数、仅 2B 激活参数的推理型 MoE 模型，专为在本地设备上运行快速、私密的智能体而设计，在工具使用和智能体任务上可与参数量多约 50% 的模型竞争。
介绍 Looped Diffusion Transformer 论文：通过在每个去噪步骤内重复运行共享的 transformer 块，并结合深度监督和自调制注意力，使 260M 模型以 4.9 倍更少的推理计算超越 6.5 倍大的模型，且循环深度可在采样时调节；作者还用 MIT 开源权重在 M5 Pro 上实测了不同深度的效果。
The paper proposes Predictive-State Flows (PSF), a continuous diffusion language model design where a rectified-flow prior generates block-level predictive states and a block-causal decoder generates tokens in parallel, significantly reducing diffusion states and sequential depth while improving efficiency and generative quality over models like ELF-B.