Tag
The team shares Latent Sparse Attention (LSA), an open research release enabling subquadratic attention that matches full-attention accuracy at ultra-long context, making sparse attention up to 22.7× faster and full attention up to 72.4× faster.
The paper investigates three calibration approaches—HiGPTQ, HiAutoRound, and HiFlatQuant—for quantizing LLMs with the HiFloat4 block floating-point format, reducing the accuracy gap to BF16 below 1% and outperforming NVFP4 across models and tasks.
The paper introduces Agentic Soft-Token Compression (ASTC), a training-free method that uses an LLM's native recurrent dynamics (GDN-attention) to compress long-horizon agent context into soft tokens, matching or exceeding uncompressed accuracy with up to 50.68% fewer Attention Pairs.
TraceSpec is a framework linking LLM agent trajectory evaluation with adaptive draft training for speculative decoding, introducing a TraceSpec-Traj corpus and improving replay throughput by 12.5% to 22.6% over its DSpark base across multiple model sizes.
EDU-Jev introduces a family of single-pass 'System One' decision models (0.8B/2B/4B) that return probability distributions over fixed options for educational assessment tasks, achieving 67.6% average accuracy across twelve benchmarks and outperforming leading commercial decision APIs by 9.4 points.
ThinkJev is a training-free method that builds fast (System 1) and slow (System 2) answering paths from a single hybrid-thinking open model, reading calibrated option probabilities directly or after variable-length thought, so it only pays for thinking where it statistically guarantees a target error rate.
论文提出 conditional interaction 框架，通过 System-1 轻量预测器筛选高价值交互、仅让 System-2 执行被接纳的候选，从而以 1/ρ_u 因子降低执行计算成本，并主张以密度可测的 ρ_u 和 informative compute 更好地描述扩展效率。
该论文提出 Packed-SS，一种自投机解码调度方案，通过让扩散起草器在验证当前草稿时提前一个调用构建下一草稿，实现单次模型调用完成验证与起草，输入规模随投机宽度线性增长，并在 Nemotron 3B/8B 与 SetDLM 1.7B 上取得更高吞吐量。
论文提出 QViD，一种免训练的查询感知视觉token剪枝方法，通过发现查询-视觉交互矩阵的低秩结构并分解出主导潜在交互模式，在图像与视频理解基准上以相同token预算持续提升性能，尤其在高压缩率下增益显著。
提出了 RAQ-MLA，一种针对 Multi-Head Latent Attention 缓存的角色感知量化方法，无需训练即可将 MLA 缓存压缩至约 2.5 比特，在四个 MLA 模型上推理与编码任务几乎无损。
TraceQuery 是一个轻量级稀疏解码框架，通过全去噪轨迹监督预测需要重新计算的少量位置，为扩散大语言模型（dLLM）推理提供最优的质量-计算权衡。
论文提出 S-SALAAD 框架，在训练中联合学习模型权重及其低秩与结构化稀疏分量，配合预算感知压缩和专用 CPU/GPU kernel，将单个检查点的弹性压缩性转化为更低的内存占用与推理延迟，适用于边缘设备部署。
该论文研究全模态大模型（Omni-LLM）的推理效率问题，提出按查询自适应选择模态的方法，将问题形式化为准确率-成本优化，并通过成本敏感学习实现在匹配或超越全模态准确率的同时大幅降低 token、延迟与金钱成本。
LOWN introduces certificate-guided visual routing for diffusion vision-language models, skipping image processing when text-only predictions agree with image-conditioned ones, retaining 99.6% of LLaDA-V's score across ten benchmarks at 35.3% of decoding FLOPs.
ThinkBit 是一个严格标量的两比特量化框架，通过分层比特分配、Gumbel-softmax 搜索码字、知识蒸馏重校准和量化感知训练来保留推理能力，在 Qwen3-1.7B 和 14B 模型上以远小于 bfloat16 的体积取得接近原模型的 MATH-500 成绩。
该论文研究了推测解码中目标模型概率的复用问题，提出在树分配前利用目标概率进行重排序和首词采样，可在小树预算下提升期望接受率并加快解码速度。
This paper shows that regenerating target model responses shrinks the gains that draft-training objectives like knowledge distillation provide in speculative decoding, replicating findings across vision-language (Qwen3-VL-8B) and text (Qwen3-4B) settings with multiple controls.
PACE 是一种免微调的后训练四比特量化器，用于线性注意力循环状态，通过降低量化频率（每隔几个 token 才舍入一次）并采用列对齐的量化尺度来减少误差，从而在 4B-35B 参数混合模型上以约三分之二的内存达到与八比特 DAMP 相当的性能。
The paper presents Lafa, an 8B-parameter blockwise continuous flow language model trained with under 30B tokens that matches or surpasses autoregressive quality while achieving 2.54x inference throughput via self-speculative decoding and blockwise parallel decoding.
MiniMax 发布 MiniMax M3.1 Flash，作为新一代 Flash 模型，在 agentic 任务、编码、Computer Use 与网络安全方面实现全面跃升，在线推理速度提升近三倍，使用成本降低 70%，输入/输出定价为每百万 token $0.15 / $0.60。