lora

Recommends the locally deployable face swap LoRA "BFS-Best-Face-Swap," supporting face, head, and body replacement while preserving the original hair, skin texture, and lighting. Provides ComfyUI workflows, Draw Things presets, and Python Diffusers deployment options.

The paper presents Saidzoda-Tajik-LFM-2.6B, a Tajik adaptation of Liquid AI's LFM2.5-2.6B-Base for tool calling on modest hardware, using LoRA continued pretraining followed by supervised fine-tuning. Tajik tool-call accuracy jumps from 20.69% to 64.14%, but the authors show CPT wipes out native tool-call formats and harms English MMLU (60.8 to 52.5).

BYOD presents a LoRA-only recipe for efficiently converting autoregressive language models into diffusion language models, achieving conversion of Gemma 2 9B, Llama 3.1 8B, Qwen2.5 7B, and Ministral 8B in about 3-4 hours on a single Colab GPU while keeping all base parameters frozen.

The paper evaluates multiple LLM-based strategies for knowledge entity relation extraction from academic abstracts and full texts, finding that a LoRA-adapted T5 model achieves the best macro F1 scores while hybrid SLM/LLM strategies further improve performance.

This paper introduces bidirectional KV-cache patching as a method to trace how fine-tuned behaviors are instantiated in language models, showing that behaviors like responding in German or emergent misalignment can be transferred between a LoRA fine-tuned model and its base model via KV cache reuse, with effects concentrating on pivotal tokens such as system-prompt and chat-template positions.

HyperRepair is a framework that recovers the general capabilities of fine-tuned language models while retaining domain specialization at inference cost only, using a hypernetwork that generates a Repairing Side Net from activation fingerprints. It is evaluated across two model families, 12 LoRA adaptation datasets, and 15 general benchmarks, achieving 93.4% of base models' general score while keeping 99.5% of fine-tuned domain performance.

提出 CLIK，一个上下文条件化的参数生成器，通过单次前向传播将上下文映射为 LoRA 参数，把知识内化到模型权重中，无需测试时梯度更新，在文档问答和多轮对话上优于现有超网络方法。

Liquid AI released LoRA adapters trained on LFM2.5-24B for use with Pi and Claude Code, enabling more efficient agentic development with up to 40% fewer tool calls on Pi.

Princeton 与 Meta 的新论文提出 Self-Direction 方法：让失败的 AI 智能体反思并将教训训练成约 0.02% 大小的 LoRA 适配器，由路由器在遇到相似任务时自动加载，从而显著减少重复错误。在 Qwen3.6-27B 上，GAIA 任务成功率从 38.8% 提升至 55.8%，且跨 185 个早期错误的修复持久性远超普通重试。

提出一种层级低秩重参数化方法，将PEFT方法的可训练参数重组为近似方阵并递归进行低秩分解，可在压缩约一个数量级参数的同时保持甚至提升LoRA/DoRA在常识推理与数学推理任务上的性能。

A new arXiv paper introduces Functional Rank Allocation (FRA), an unlabeled method for compressing large fleets of LoRA adapters via functional rank-distortion frontiers and fleet-level minimax allocation under a shared rank budget, improving worst-case and tail utility over uniform spectral compression across 78 adapters from three public fleets.

A plug-and-play repo implementing self-distillation with an EMA teacher that scales to large models by combining LoRA adapters with DeepSpeed ZeRO-3, avoiding the memory and latency overhead of holding separate full teacher and student models. Tested on Qwen3.5 and ToolAlpaca, it includes adapter switching, EMA checkpointing, and memory-optimized DeepSpeed configs.

PrunaAI released 6-step LoRA adapters for Qwen-Image-2.1 that cut image generation from 40 steps to 6, with one variant removing the prompt-upsampling step entirely for ~5x faster end-to-end generation.

该论文提出 Parametric Session State (PSS)，一种将长对话交互编译为固定形状的 LoRA 会话记忆层，使 RAG 系统在无需重放历史的情况下恢复未明确表述的查询中的检索意图，并在受控与公开评测中验证其跨模型与检索后端的有效性。

A paper analyzing LoRA's sensitivity to learning-rate choices and proposing Init[A²], a new initialization scheme with variance scaling that suppresses initialization-induced noise, stabilizing asymmetric LoRA+ training and widening the stable learning-rate region for LLM fine-tuning.

Venus-DefakerOne2 is a unified anti-forgery vision foundation model that broadens detection across forgery forms (digital manipulation, physical spoofs) and media types (images, videos) using Tide-Style-Tuning and entropy-based test adaptation, achieving state-of-the-art results on multiple forensic benchmarks.

The paper proposes DIAL-S2ST, a post-training method that learns inference-time latency control in simultaneous speech-to-speech translation models via self-distillation from lookahead predictions and a jointly trained LoRA adapter, demonstrating flexible latency control on French-to-English translation using Hibiki with only 10 hours of source speech.

Rollback 提出了一种利用世界模型回溯到过去的决策点、探索反事实动作并生成细粒度动作级监督的策略改进范式，同时通过真实机器人反馈持续优化世界模型动力学，在三项真实任务上显著提升了机器人策略成功率。

Licon VBVR 是一个针对 MiniMax-H3 的图像转视频 LoRA，基于重建的 VBVR 数据训练，可减少物体复制、裁剪与穿插问题，并提升角色动作逻辑与空间一致性。

这篇论文介绍了一种"typed decision"System 1 接口：通过一次前向传播直接从模型 logits 或小分类头输出固定选项的概率，用于替代昂贵的 frontier LLM 判官。作者部署了一个自托管的 rank-16 LoRA 判官来分流 14.8 万条 agent 编码会话记录，成本仅为前沿模型判官的约 1%；并在 JevBench 上用 35B 开源模型达到 79.3%，与商业参考系统（73.0%）无显著差异。