Tag
TempQuant 提出自适应学习量化语言模型的温度与 top-p 设置,通过校准使量化模型的下一词元概率分布更接近全精度模型,在 3-bit Qwen3-4B 上将 MATH 准确率从 34% 提升至 51%。
The paper shows that long-context generation quality in open-weight LMs is primarily a sampling problem rather than a training problem: samplers like top-nsigma that keep a small, steady candidate-token pool preserve accuracy as temperature rises, while top-p/top-k collapse (e.g., 0 vs 46-62 of 100 SWE-bench issues resolved at temperature 2).
The Verge shares exclusive details about Akai's upcoming MPC Sample Key, a 37-key instrument focused on pitched sample playback with chord, arpeggiator, looper features and open-source software for the modding community.
论文揭示连续扩散语言模型(如ELF)因生成过程缺少逐步离散采样步骤而导致长尾词汇严重缺失(稀有词生成率仅为人类文本的0.30×),并提出训练无关的GRAFT方法,通过在模型未确定时提交稀有词再重新生成序列,将稀有词比例恢复至人类水平的0.82–1.10×。
The paper introduces Conditional Poisson Sampling as a fixed-size drop-in replacement for Poisson Sampling in active statistical inference, showing it achieves no worse and often better asymptotic variance while eliminating random sample size issues, along with a released Python package.
作者分享在Strata上从IQ2_XS迁移到Unsloth UD-IQ4_XS时的经验:0.1.45之前缺少sampling配置会默认贪心解码导致模型循环,以及默认STRATA_ROUTE_TAIL_SKIP=7会牺牲4-bit量化质量,并给出配置与校准建议。
论文提出 Mine-Sample-Render (MSR) 测试时方法,从真实语料中挖掘可解释的类别结构并据此采样属性配置,再将生成模型用作条件渲染器,以提升天气报告、用户模拟、网页设计、文生图及3D资产等多种生成任务的输出多样性与语料分布对齐度。
提出训练无关的解码算法 One Block Ahead (OBA),在块扩散语言模型中并发去噪当前块与下一块并传递采样器状态,在 DiffusionGemma 上将每次前向传递的 token 数提升 42.3%、单请求吞吐提升 16.3%。
提出 Weighted Ensemble (WE),一个无需训练的框架,通过在生成过程中自适应重新分配采样预算来提升语言模型中稀有事件概率估计的效率与可靠性,并在安全相关任务上验证了其有效性。
该论文引入了一种以神经网络复杂度(Barron能量)为核心的最优传输框架,替代传统L²动能,从而在概率测度空间上定义新的几何结构,并证明了扩散生成模型在Barron几何下的次优性(超多项式score逼近下界)。
论文提出利用语言模型采样时附带的下一 token 条件概率,通过势函数(potentials)构造方差更小的期望估计器,在 KL 散度、AI 安全、蒸馏、水印等多种应用场景中实现了显著的方差降低。
Meta Superintelligence Labs finds that RL post-training narrows solution coverage of LLM agents (the 'Sharpening Tax'): base models with a light harness can solve more distinct agentic tasks at large K, while post-trained models win at pass@1. They propose PTGS, an adaptive-temperature RL method that reduces this tax while improving pass@1.
Introduces Gumbel Straight Flow, a method for distilling autoregressive language models into one-step flow maps that enable parallel sampling.
This paper introduces 'Sharpening Tax,' a metric quantifying how RL post-training improves pass@1 but reduces solution coverage (pass@K) in LLMs, and proposes posterior-tempered group sampling (PTGS), a plug-and-play Bayesian sampler that adapts temperature per prompt to pay a smaller tax while improving single-shot accuracy.
FluxLite introduces a training-free, inference-time proposal-control framework for discrete diffusion models that compensates jump-rate perturbations via a graph-divergence term in the Feynman-Kac potential, yielding two samplers (HEU and D-VCG) that substantially reduce reweighting variance and sampling error over standard SMC baselines.
This paper introduces a framework for jailbreaking black-box LLMs through text-only continuation interfaces, combining sample-based distribution reconstruction, risk-gated residual control, and speculative multi-token execution to bypass safety alignment without access to model weights or token probabilities.
This GitHub project uses the Jev model to create a lousy chatbot with various sampling strategies, developed with Claude's help as a fun experiment.
This paper introduces a method for vibe design agents to explore diverse UI alternatives by separating exploration from implementation through structured design specifications, as evaluated on 168 prompts and a large online experiment with over 300,000 tasks.
This paper introduces Newton Matching, a unified framework for fine-tuning and sampling in generative models, which addresses limitations of existing methods by treating learning as an iterative optimization process and leveraging conditional-matching structure.
The article discusses Poisson disk sampling, a technique for randomly placing points with a minimum distance, and details Bridson's efficient algorithm from a highly-cited 2007 paper, along with improvements for computer graphics and simulations.