lora

该论文提出 Fiber Regulation Theory (FRT) 与 Fiber Acquisition Theory (FAT) 两个理论框架，刻画受保护的功能修复与可训练结构的形成/续期动力学，并在 Qwen2.5-32B 上通过受保护头修复与固定容量 LoRA 续期验证联合策略优于单独策略。

A developer shares custom general and fused LoRA kernels for llama.cpp that restore prompt-processing throughput on Strix Halo from ~300 t/s back up to ~1400 t/s when running LoRA adapters, and invites others to help upstream the changes.

该论文提出一个多粒度评估框架，结合字符错误率、音节错误率和分词错误率，评估藏语语音识别在多方言与不同适配策略下的表现，发现低字符错误率并不总带来更优的高层级指标，且 LoRA 适配在测试范围内仍逊于全量微调。

该论文提出 MoPE-LoRA，一种基于音素类别（发音方式）而非口音身份组织低秩专家的参数高效适配框架，用于带口音的英语语音识别；在 L2-ARCTIC 上零样本口音评估中相对单 LoRA 提升 12.3%。

论文提出 LSA-PAST，一种面向组合式增量学习（composition-IL）的参数高效方法，通过在冻结 ViT 中注入深度特化的 LoRA 适配器并用无复习的仿射变换将旧原型映射到新适配器空间，实现了精度-遗忘权衡的 state-of-the-art 表现。

SolarRenderer 将任意引擎的无纹理结构通道结合参考首帧，生成风格化完整视频，基于 MiniMax-H3 构建，通过控制 LoRA 与少步蒸馏 LoRA 实现实时逐帧结构一致的渲染，无需按游戏训练。

用户分享了使用 Jev（Qwen 3.8 27B LoRA 决策模型）驱动的贪吃蛇游戏，模型在每步操作时根据棋盘状态文本描述进行决策，延迟约 600ms，运行于 llama.cpp 环境中。

A small open model, qwen3-4b-deslop, is a Qwen3-4B-Instruct fine-tune trained on a MacBook Air that removes common AI writing tics (em dashes, 'not X, but Y', lists of three), cutting the slop score from 121 to 12.5 per 1000 words while keeping factual accuracy, and it runs locally via Ollama.

An educational post breaking down 5 parameter-efficient fine-tuning (PEFT) techniques for LLMs — LoRA, LoRA-FA, VeRA, Delta-LoRA, and LoRA+ — with visual explanations of how each adapts low-rank matrices to reduce memory and training costs.

MOSS-Image Turbo LoRA (4-step) v2.0 has been released on Hugging Face, now Euler-native so it runs on the stock flow-matching sampler in any standard LoRA loader with no CFG or custom pipeline files, delivering fast high-fidelity bilingual text rendering.

Kromoss is a full fine-tune of MOSS-Image-Base at 1024px, shipping with two 8-step turbo LoRAs (DMD and DMD2-distilled) and drop-in support for diffusers and ComfyUI, including an int8 quantized variant.

该论文提出使用参数高效微调（PEFT/低秩适配器）来适配超大多模态骨干网络进行脑编码，Brain LoRA 在 Algonauts 2025 挑战赛中取得最佳公开成绩，并通过 multi-LoRA kernel 支持 200+ 模型集成。

asmsense is a dataset of 228k stripped x86-64 functions with source-aware labels plus a Qwen3-8B+LoRA model for naming and describing functions in stripped binaries, along with a detailed roadmap of missing data such as ARM64/RISC-V support and GCC/ELF builds.

A discussion about using LoRA adapters as a cheaper, more flexible alternative to tools like Heretic for removing refusals/censorship from LLMs, referencing the weightless GitHub project and similar LoRAs for image model text encoders.

A 0.6B parameter d3-edge model, fine-tuned with SFT + LoRA on a few thousand examples, outperformed the 27B d3 decision model on typed decisions and nearly matched it on Banking77, highlighting the value of fine-tuning small specialized models over large general-purpose ones.

Hugging Face 上发布 lodestones/MOSS-Image-SFT-8-steps-dmd，一个将 MOSS-Image-SFT（50 步、CFG 3）蒸馏为 8 步、无 CFG 的工作进展检查点，采用 data-space DMD 方法，可作为 LoRA 用于 ComfyUI。

This paper shows that naive few-shot LoRA personalization on-device degrades out-of-scope (OOS) rejection AUROC across five intent-classification benchmarks while preserving in-scope accuracy, and proposes a training-free fix using Mahalanobis-distance-based rejection scoring that recovers most of the lost quality on 3 of 5 datasets.

阿里巴巴 TaoLive AIGC 团队发布了 TaoMate-H3-FL2VA，据称是首个实时流式首尾帧控制的音频-视频生成（FL2VA）模型，基于 MiniMax H3 的三步 LoRA，支持低延迟分块生成与 480p/768p/1080p 分辨率。

AcrossVAM2.0 提出一种冲突感知的点对应监督方法，用于参数高效的视频扩散模型适配，通过置信度加权跟踪头、零初始化残差反馈和非对称 PCGrad，在保持预训练外观先验的同时提升对应一致性，但并不均匀改善视频质量。

提出「在前沿LLM上迭代开发行为，再蒸馏到小模型（SLM）上线」的两阶段方法：用LoRA微调的Qwen3-8B在家庭护理排班智能体的生产决策上达到0.913一致率，优于教师模型GPT-5.4-nano自身重跑时72.7%的自我一致性，且延迟降低约2.5倍、成本大幅下降。论文已被NeurIPS 2026第二届企业任务智能体基准与应用研讨会（AABA4ET）接收。