Steering the Noise: Turning Random Perturbations into Effective Descent for Memory-Efficient LLM Fine-Tuning
School of Intelligence Science and Technology（智能科学与技术学院）；Institute for Artificial Intelligence（人工智能研究院）；Peking University（北京大学）；State Key Laboratory of General Artificial Intelligence（通用人工智能国家重点实验室）
Fine-tuning large language models (LLMs) achieves strong performance but is often limited by the memory overhead of backpropagation. Zeroth-order (ZO) optimization avoids this overhead by estimating gradients through forward passes alone, yet it typically converges slowly because random Gaussian perturbations yield high-variance gradient estimates in high-dimensional parameter spaces. In this paper, we propose a plug-and-play framework that turns random perturbations into more effective descent directions. The key idea is to draw a small pool of candidate perturbations, evaluate their loss values, and then select or combine those that are best aligned with the optimization objective. We develop two instantiations of this idea: MeZO-GV, which forms a guiding vector from the contrast between low-loss and high-loss perturbation groups, and MeZO-Greedy, which keeps the single best perturbation within a fixed evaluation budget. We theoretically show that both strategies yield a larger per-step reduction in the objective than standard ZO estimation, leading to improved convergence rates. Experiments on LLMs of different scales and architectures confirm that the proposed methods integrate naturally with existing ZO optimizers and consistently improve convergence speed and task accuracy. On OPT-13B, our approach outperforms all ZO baselines across 11 benchmarks and exceeds gradient-based methods on 9 of them, while retaining the memory efficiency of forward-only optimization.
Subspace Control: Turning Constrained Model Steering into Controllable Spectral Optimization
OPTML, Michigan State University（OPTML，密歇根州立大学）；MIT-IBM Watson AI Lab, IBM Research（MIT-IBM Watson AI Lab，IBM研究院）；University of Minnesota（明尼苏达大学）；University of Michigan（密歇根大学）；University of Colorado Boulder（科罗拉多大学博尔德分校）
In Iterative Learning Control (ILC), a sequence of feedforward control actions is generated at each iteration on the basis of partial model knowledge and past measurements with the goal of steering the system toward a desired reference trajectory. This is framed here as an online learning task, where the decision-maker takes sequential decisions by solving a sequence of optimization problems having only partial knowledge of the cost functions. Having established this connection, the performance of an online gradient-descent based scheme using inexact gradient information is analyzed in the set…
Turning Logic Against Itself : Probing Model Defenses Through Contrastive Questions
Ubiquitous Knowledge Processing Lab (UKP Lab), Department of Computer Science and Hessian Center for AI (hessian.AI), Technical University of Darmstadt（德累斯顿技术大学计算机科学系、海斯堡人工智能中心（hessian.AI）、通用知识处理实验室（UKP Lab））
针对推理驱动的隐蔽越狱漏洞，本文提出 POATE 对比推理攻击，并设计 Intent-Aware CoT 与 Reverse Thinking CoT 以检测恶意意图、拒绝有害输出。
Bayesian Deep Learning for Remaining Useful Life Estimation via Stein Variational Gradient Descent
Concordia University（康考迪亚大学）；Gina Cody School of Engineering and Computer Science（吉娜·科迪工程与计算机科学学院）；Università degli Studi di Padova（帕多瓦大学）；Mila-Quebec AI Institute（米拉-魁北克人工智能研究所）
Percolation Dynamics in Optimization: Variance Cascades and Nested Symmetry
School of Computation, Information and Technology, Technical University of Munich（慕尼黑工业大学计算、信息与技术学院）；Munich Center for Machine Learning (MCML)（慕尼黑机器学习中心）
Combinatorial optimization (CO) layers in machine learning (ML) pipelines are a powerful tool to tackle data-driven decision tasks, but they come with two main challenges. First, the solution of a CO problem often behaves as a piecewise constant function of its objective parameters. Given that ML pipelines are typically trained using stochastic gradient descent, the absence of slope information is very detrimental. Second, standard ML losses do not work well in combinatorial settings. A growing body of research addresses these challenges through diverse methods. Unfortunately, the lack of well…
Ensembling Large Language Models to Characterize Affective Dynamics in Student-AI Tutor Dialogues
Harvard Graduate School of Education（哈佛教育研究生院）；Harvard University（哈佛大学）；Personal Robots Group, Media Lab（个人机器人小组、媒体实验室）；Massachusetts Institute of Technology（麻省理工学院）
