Efficient Adaptation of Large Language Models

The field of large language models (LLMs) is moving towards more efficient and robust adaptation methods. Recent developments focus on reducing the effects of overfitting and degeneralization, especially in low-data settings. Techniques such as corrective self-distillation, prompt-conditioned parameter generation, and sparse fine-tuning are being explored to improve the performance of LLMs on specific tasks while maintaining their general-purpose capabilities. Notably, some studies have demonstrated the importance of mitigating forgetting during domain-specific continued pre-training and the benefits of using novel training recipes to strengthen performance on both translation and general-purpose tasks. Noteworthy papers include: Minifinetuning, which introduces a method for language model domain adaptation that reduces the effects of overfitting-induced degeneralization. EvoLM, which presents a model suite that enables systematic and transparent analysis of LMs' training dynamics across multiple stages. Drag-and-Drop LLMs, which proposes a prompt-conditioned parameter generator that eliminates per-task training by mapping a handful of unlabeled task prompts directly to LoRA weight updates.

Sources

Minifinetuning: Low-Data Generation Domain Adaptation through Corrective Self-Distillation

EvoLM: In Search of Lost Language Model Training Dynamics

Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights

SparseLoRA: Accelerating LLM Fine-Tuning with Contextual Sparsity

Prmpt2Adpt: Prompt-Based Zero-Shot Domain Adaptation for Resource-Constrained Environments

Tower+: Bridging Generality and Translation Specialization in Multilingual LLMs

KnowMap: Efficient Knowledge-Driven Task Adaptation for LLMs

Breaking Barriers: Do Reinforcement Post Training Gains Transfer To Unseen Domains?

Optimising Language Models for Downstream Tasks: A Post-Training Perspective

SharpZO: Hybrid Sharpness-Aware Vision Language Model Prompt Tuning via Forward-Only Passes

Where to find Grokking in LLM Pretraining? Monitor Memorization-to-Generalization without Test

Built with on top of