adapters

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.

IBM releases Granite Switch 4.1.1, a backwards-compatible patch update to its modular Switch architecture that adds an updated multilingual Granite Speech ASR model and an improved Decision adapter without retraining the base model.

This paper evaluates the 305M-parameter Gemma 4/EmbeddingGemma 2 speech encoder on Saudi and Gulf Arabic ASR without modifying its weights, comparing a frozen CTC reader and an adapter-augmented system against Cohere Transcribe Arabic and other baselines, showing competitive CER and strong efficiency gains. Code and results are publicly released.

IBM 发布 Granite Switch 4.1 30B/8B，这是一个模块化 LLM，将基座模型、多个 aLoRA 适配器（决策、guardian）和语音编码器组合在一起，通过单个端点即可在聊天、通用/专业决策和语音转文本之间切换，采用 Apache 2.0 许可。

IBM 推出 Granite Switch 4.1（8B 与 30B），通过 aLoRA 适配器在同一端点上组合聊天、决策与语音转写能力，几乎不额外占用 VRAM。新增 Decision 与 Guardian aLoRA，并嵌入 OpenASR 榜单最快的 Granite-Speech-5.0-Turbo-CTC 模型。

CAFE提出将CKA特征蒸馏的梯度限制在英语旁路的语言路由适配器中，解决了多语言紧凑ASR中英语与非英语WER的权衡问题，在Open ASR Leaderboard上将平均WER从8.55%降至8.02%。

论文提出 MaRK（Markov-adapted Recurrent Kernels），一种将上下文向量直接映射为冻结 SSM 递归参数有界调制的动态算子条件化框架，用于将预训练 SSM 适配到迭代扩散生成任务，在 111M Hydra 骨干上仅需 6.3–11M 可训练参数，其中 Chebyshev 变体表现最佳。

The author is discussing which codebase design techniques from the pre-AI era are still useful for AI agents, as part of writing a course on codebase design.

ACE introduces a method to consolidate redundant adapters across experts in MoE large language models for more efficient parameter-efficient fine-tuning, achieving up to 1.48× training speedup without increasing peak memory.

LatentPress introduces a method to compress conversational and document context into continuous memory tokens, enabling frozen decoders to read directly without text reconstruction, achieving higher compression ratios and improved performance on long-context tasks.

The paper describes a submission to the WMT 2026 MIST shared task, using the Tiny Aya Global model with task-specialized QLoRA adapters for multilingual summarization and question answering.

The author built two open-source projects, A2A Adapter and Hybro, to enable interoperability between different AI agents, allowing them to collaborate without replacing each other.

Tencent researchers propose YOLO-PEFT, a structure-aware framework that treats adapter placement as constraint planning, outperforming full fine-tuning on YOLO detectors while using significantly fewer trainable parameters.

AuroSFT introduces a parameter-efficient adapter-based framework for multi-task supervised fine-tuning that rolls back adapter checkpoints at task-wise peaks instead of full-model checkpoints, achieving higher average accuracy than mSFT.

SemiAdapt-Instruct proposes a modular framework that discovers latent instruction domains, trains per-domain LoRA adapters in parallel, and routes among them without extra parameters, enabling extensible instruction tuning where new domains require only single-adapter updates instead of full retraining.

LARA is a method for efficient adaptation that adds low-rank corrections to a frozen model's residual stream instead of modifying weights, matching LoRA's performance while enabling composable behaviors and inference-time steering.

SafeGene proposes a reusable safety-adapter module that decouples safety capabilities from task-specific updates, enabling efficient restoration of safety alignment in open-weight LLMs after downstream fine-tuning through few-shot recalibration.

This paper introduces Code2LoRA, a hypernetwork-based method to generate adapters for code language models, addressing challenges under software evolution.

This paper explores using parameter-efficient fine-tuning (PEFT) as a compact substrate for persistent personal models, studying scaling up, down, and out, and introduces MinT for managing adapters.

The author benchmarks serving 1,000 LoRA adapters on one GPU using vLLM, finding that active adapter count and traffic shape are the real bottlenecks, and provides recommendations for tuning max_loras.