Tag
This article explains five parameter-efficient fine-tuning techniques for large language models, such as LoRA and VeRA, detailing how each method adapts model weights with minimal updates.
The paper investigates adapting CLIP for regional geolocalization, finding that encoder adaptation improves accuracy by leveraging intact scene configuration rather than coarse geographic cues.
This article explains test-time training, where AI models adapt during inference to improve personalization and reduce memory usage, but at the cost of increased per-user compute. It discusses implications for serving models at scale, balancing long context and user concurrency.
This paper introduces KV-Skill, a design space of external factorized operators that frozen language models read through a lightweight interface, enabling task knowledge to be acquired from text or rewards and deployed independently. Experiments across ten benchmarks show consistent improvements over text skills, prefix tuning, and LoRA, with composable loadable skills.
This paper introduces NOMADD, a post-hoc method to reduce concept drift in tabular models by fitting base models on labeled periods and extrapolating compressed parameter changes. It achieves competitive performance with Drift-Resilient TabPFN at a fraction of training and inference cost.
Proposes ROAM, a framework that uses LLM world knowledge and reasoning to adapt frozen specialist models to unseen scenarios without retraining, achieving over 20% MAE reduction with minimal overhead.
This research explores the concept of constraining a model's learning to only what trusted LoRA adapters can express, aiming to improve safety and reliability in fine-tuning.
MiCA (Minor Component Adaptation), a new fine-tuning method that initializes adapters in the minor singular subspace for better knowledge uptake and less forgetting, has been merged into the Hugging Face PEFT library. It is available via the PEFT main branch and integrates through the existing LoRA interface with init_lora_weights='mica'.
Explains five parameter-efficient fine-tuning techniques: LoRA, LoRA-FA, VeRA, Delta-LoRA, and LoRA+, detailing how each modifies model weights during adaptation.
This paper presents EMA, a model adaptation system for learning-based systems that reduces training and labeling costs while improving system performance in evolving environments.
Google introduces T5Gemma, a new collection of encoder-decoder models adapted from the Gemma 2 decoder-only architecture, offering improved quality-efficiency trade-offs for tasks like summarization and translation.
OpenAI has released fine-tuning capabilities for GPT-3.5 Turbo, allowing developers to customize models for specific use cases with improved performance, steerability, and output formatting. The update enables fine-tuned GPT-3.5 Turbo to match GPT-4 performance on certain tasks while reducing prompt sizes by up to 90%.