Tag
提出 Squelch——一种无路由、无新增可训练参数的 rank-one LoRA 专家混合方法，通过噪声底限软阈值和 Stein 无偏风险估计为每个 token 自适应选择激活秩，并在常识推理、知识、数学、代码和 GLUE 基准上显著优于 LoRA 与最强混合基线。
The paper introduces Matryoshka Attribution (MAttr), a mask-learning method for attributing language model outputs to internal components, achieving top performance on the Mechanistic Interpretability Benchmark and demonstrating practical use in modifying LLM behaviors by adjusting weights.
Introduces SP3O, a novel reward-model-free, critic-free, gradient-based preference-based RL algorithm that leverages segment-level preferences, demonstrating improved performance in robotic control and LLM fine-tuning, especially for long-horizon tasks.
This paper introduces KGPS, a Kalman-guided prompt selection method for adaptive RL finetuning of LLMs, which models prompt difficulty as a dynamic state to improve accuracy and rollout efficiency.
FoRA introduces a parameter-efficient fine-tuning method that selects task-informative layers via Fisher scores and trains LoRA down-projections on the Stiefel manifold, reducing parameters while preserving accuracy.
NVIDIA and Unsloth have published a technical guide detailing three low-level optimizations that can accelerate LLM fine-tuning by up to 25%, including packed-sequence caching, double-buffered checkpointing, and optimized MoE routing. The guide provides deep systems-level explanations and benchmarks aimed at ML engineers and developers.
Researchers from Bangladesh University of Engineering and Technology present CBRS, a multi-platform framework that filters and parses blood donation requests from social media using a dual-layer architecture and a novel 11K bilingual dataset in Bengali and English. Their LoRA fine-tuned Llama-3.2-3B model achieves 99% filtering accuracy and 92% zero-shot parsing accuracy, outperforming GPT-4o-mini and other LLMs with 35× reduced token usage.
Value Gradient Flow (VGF) presents a scalable approach to behavior-regularized reinforcement learning by formulating it as an optimal transport problem solved through discrete gradient flow, achieving state-of-the-art results on offline RL and LLM RL benchmarks. The method eliminates explicit policy parameterization while enabling adaptive test-time scaling by controlling transport budget.