NVIDIA TensorRT Edge-LLM 在 Jetson AGX Thor 上以 6.4 倍速度完成 MLPerf Edge Agentic 基准测试

DFlash 由 UC San Diego 团队提出，是一种用于投机解码的轻量块扩散草稿模型。它一次前向推理生成整块 token，再由目标模型并行验证，保证输出无损。相比 EAGLE-3，DFlash 实现最高 2.5 倍加速，在 Qwen3-8B 等多种模型上平均无损加速超过 6 倍（MATH-500 达 6.08×）。在 NVIDIA Blackwell 上（TensorRT-LLM），gpt-oss-120b 模型吞吐量提升最高 15 倍，约为 EAGLE-3 的 1…

Qwen3.5在TokenSpeed推理引擎上，针对智能体工作负载达到了创纪录的580 tokens per second （tps）速度。这一成果由通义千问推理团队、lightseekorg Foundation TokenSpeed团队、NVIDIA及Mooncake团队共同实现，并采用了tri_dao的FlashAttention-4 （FA4） 优化。此里程碑标志着开源大语言模型推理性能的边界得到了推动，相关详情可查阅PyTorch社区博客。

Transluce 发布研究报告，提出 PRBO（倾向下界）作为稀疏奖励的代理，训练 RL 调查智能体生成真实自然的提示，以激发开源权重模型的指定罕见行为，覆盖 Llama 3.1/4、Qwen 2.5 和 DeepSeek-V3。

Large language models can easily generate explanations for their own outputs, but such self-explanations are not necessarily faithful to the model's behavior. We study this issue through counterfactual self-explanations, where a model mini…

Grounded language-model systems are often evaluated by final answer accuracy, yet a correct answer can be unsupported, drawn from the wrong source, or produced when evidence is insufficient or contradictory. We introduce EviScope, a paired…

As language models become more capable, long-term collaboration in learning, reasoning, and decision-making calls for a deeper understanding of the people they serve. Yet training such human-aware language models faces a fundamental superv…

研究者提出 Mind2Dialogue 框架，通过心理学引导的模拟器生成用户心理状态，并将其转化为特权监督信号，训练能理解用户未言明信念与目标的语言模型。在完整语料上训练后，Qwen、Llama 和 OLMo 指令微调基线在所有已报告的个性化指标上均有提升，偏好遵循生成任务提升 26.6 至 40.9 个百分点，Qwen 和 Llama 的信念与行动推理能力也同步增强。

SQS 提出一种统一框架，通过贝叶斯变分学习同时进行剪枝与低比特量化，采用 spike-and-slab 先验结合高斯混合模型（GMM）建模量化权重，在 ResNet、BERT-base、Llama3.2 和 Qwen2.5 上实现比现有方法更高的压缩率，且性能下降相当。在相同位宽下所需参数更少，在 2-bit 和 4-bit 精度下表现尤为突出。

Continual knowledge-updating methods are often declared superior from one final checkpoint and one conventional adapter rank. We show that this can be insufficient to identify the better operating point. Holding a periodic hierarchy fixed,…

How do the methods used to train language models to refuse harmful requests shape how that refusal actually works inside the model? We compare three post-training methods - supervised fine-tuning, reasoning-augmented fine-tuning (training …