arXiv:2605.13030v2 Announce Type: replace Abstract: Model merging combines task experts into one model and avoids joint training, retraining, or deploying many expert models, but the merged model often still underperforms task experts. We study this performance gap through feature drift, the difference between features produced by the merged model and by the expert on the same input. Our theory decomposes this drift into upstream propagation and local mismatch, tracks how it propagates and combines through later layers in forward order, and links final feature drift to output drift. This view motivates FeatCal, which uses a small calibration set to calibrate the merged model weights layer by layer in forward order, reducing feature drift while staying close to merged weights and preserving the benefits of model merging. FeatCal uses an efficient closed-form solution to update model weights, with no gradient descent, iterative optimization, or extra modules. On the main CLIP and GLUE benchmarks, FeatCal beats Surgery and ProbSurgery, the closest post-merging calibration baselines: 85.5% vs. 77.0%/78.8% on CLIP-ViT-B/32 Task Arithmetic (TA) and 85.2% vs. 83.7%/82.2% on FLAN-T5-base GLUE. On CLIP-ViT-B/32, 8 examples per task reach 82.9%, and 256 examples per task take 53 seconds, about 4x faster than both baselines, showing better sample efficiency and lower calibration cost.
arXiv:2605.12765v4 Announce Type: replace Abstract: Large Language Models (LLMs) memorize vast amounts of training data, raising concerns regarding privacy, copyright infringement, and safety. Machine unlearning seeks to remove the influence of a targeted forget set while preserving model performance, ideally approximating a model retrained from scratch without it. Once an LLM is in use, every new request to make it forget specific content demands updating its weights. However, unlearning through parameter updates is expensive, hard to audit, and can be undone by quantization. We show that unlearning can be enforced entirely at inference time, without training, gradients, or weight changes. We introduce Inference-Time Unlearning via Gated Activation Redirection (GUARD-IT), a training- and gradient-free method that unlearns via input-dependent activation steering at inference time. GUARD-IT stores the content to be forgotten as a small library of activation directions, and during inference, it routes each query through a similarity gate that activates only for relevant directions and applies them as a norm-preserving rotation of the residual stream, while unrelated queries pass through the unmodified model. The same design carries across three model families and nine checkpoints from 0.8B to 8B parameters, and new forget requests are absorbed by one offline pass of forward passes. On TOFU, against 16 gradient-based and inference-time baselines, and on MUSE and WMDP, GUARD-IT forgets without breaking the model, and on TOFU it is the only method that suppresses memorization in every Llama configuration without collapsing: it keeps utility and fluent generation in every configuration, moves the model's output distribution closest to a model that never saw the forgotten data on the forget01 split, survives 4- and 8-bit quantization and ten sequential forget requests, and holds under jailbreak attacks.
arXiv:2605.12174v2 发布类型：替换 摘要：在大规模学习中，针对随机小批次求解最优传输（OT）是精确 OT 的常见替代方案。在流匹配（FM）中，这种替代方案用于获得类似 OT 的耦合，从而使概率路径变直并降低数值积分成本。然而，重复执行小批次 OT 所诱导的总体层面耦合仍未得到充分理解。我们将这一耦合形式化为期望批次 OT 传输计划 $\overline{\pi}_{k}$，其通过对独立的、大小为 $k$ 的小批次上的经验 OT 传输计划取平均得到。随后，我们证明了其在大批次下的一致性，并在与生成建模相关的半离散情形中，推导出传输成本偏差以及 $\overline{\pi}_{k}$ 收敛至 OT 传输计划的速率。对于 FM，这得到了一种总体耦合，其诱导的速度场具有足够的正则性，能够定义从源分布到离散目标分布的唯一流。最后，我们在一个易于处理的双原子模型以及合成数据和图像实验中，量化了 OT 批次大小与数值积分之间的相互作用。
arXiv:2604.13966v2 Announce Type: replace Abstract: We study value adaptation in offline-to-online reinforcement learning under general function approximation. Starting from an imperfect offline pretrained $Q$-function, the learner aims to adapt it to the target environment using only a limited amount of online interaction. We first characterize the difficulty of this setting by establishing a minimax lower bound, showing that even when the pretrained $Q$-function is close to optimal $Q^\star$, online adaptation can be no more efficient than pure online RL on certain hard instances. On the positive side, under a novel structural condition on the offline-pretrained value functions, we propose O2O-LSVI, an adaptation algorithm with problem-dependent sample complexity that provably improves over pure online RL. Finally, we complement our theory with neural-network experiments that demonstrate the practical effectiveness of the proposed method.
arXiv:2604.10166v2 Announce Type: replace Abstract: Intelligent operation of thermal energy networks aims to improve energy efficiency, reliability, and operational flexibility through data-driven control, predictive optimization, and early fault detection. Achieving these goals relies on sufficient observability, requiring continuous and well-distributed monitoring of thermal and hydraulic states. However, district heating systems are typically sparsely instrumented and frequently affected by sensor faults, limiting monitoring. Virtual sensing offers a cost-effective means to enhance observability, yet its development and validation remain limited in practice. Existing data-driven methods generally assume dense synchronized data, while analytical models rely on simplified hydraulic and thermal assumptions that may not adequately capture the behavior of heterogeneous network topologies. Consequently, modeling the coupled nonlinear dependencies between pressure, flow, and temperature under realistic operating conditions remains challenging. In addition, the lack of publicly available benchmark datasets hinders systematic comparison of virtual sensing approaches. To address these challenges, we propose a heterogeneous spatial-temporal graph neural network (HSTGNN) for constructing virtual smart heat meters. The model incorporates the functional relationships inherent in district heating networks and employs dedicated branches to learn graph structures and temporal dynamics for flow, temperature, and pressure measurements, thereby enabling the joint modeling of cross-variable and spatial correlations. To support further research, we introduce a controlled laboratory dataset collected at the Aalborg Smart Water Infrastructure Laboratory, providing synchronized high-resolution measurements representative of real operating conditions. Extensive experiments demonstrate that the proposed approach significantly outperforms existing baselines.
arXiv:2604.04611v3 Announce Type: replace Abstract: Federated learning (FL) enables multiple clients to collaboratively train a global model by aggregating local updates without sharing private data. In this work, we focus on cross-silo FL, where each client typically represents an independent organization. However, cross-silo FL can face the challenge of free-riders, clients who submit fake model parameters without performing actual training to obtain the global model without contributing. Chen et al. proposed a free-rider detection method based on the weight evolving frequency (WEF) of model parameters. This detection approach is practical because it requires neither a proxy dataset nor pre-training. Nevertheless, it struggles to detect ``dynamic'' free-riders who behave honestly in early rounds and later switch to free-riding, particularly under global-model-mimicking attacks such as the delta weight attack and our newly proposed adaptive WEF-camouflage attack. In this paper, we propose a novel detection method S2-WEF that simulates the WEF patterns of potential global-model-mimicking attacks on the server side using previously broadcast global models, and identifies clients whose submitted WEF patterns resemble the simulated ones. To handle a variety of free-rider attack strategies, S2-WEF further combines this simulation-based similarity score with a deviation score computed from mutual comparisons among submitted WEFs, and separates benign and free-rider clients by two-dimensional clustering and per-score classification. This method enables dynamic detection of clients that transition into free-riders during training without proxy datasets or pre-training. We conduct extensive experiments across four datasets and five attack types, demonstrating that S2-WEF provides robust dynamic free-rider detection across diverse settings.
arXiv:2603.26024v4 Announce Type: replace Abstract: Identification of causal directionality in bivariate numerical data is a fundamental research problem with important practical implications. This paper presents two alternative methods to identify direction of causation by considering conditional distributions: (1) Anticipated Asymmetric Geometries (AAG) and (2) Monotonicity Index (MI). The AAG method compares the actual conditional distributions to anticipated ones along two variables. Different comparison metrics, such as Pearson correlation, cosine distance, Hellinger distance, Jaccard index, Jeffreys divergence, K-L divergence, K-S distance, MAE, MSE, mutual information, and Wasserstein distance have been evaluated. Anticipated distributions have been projected as normal based on dual response statistics: mean and standard deviation. The MI method compares the calculated monotonicity indexes of the gradients of conditional distributions along two axes and exhibits count of gradient sign changes. Both methods assume stochastic properties of the bivariate data and exploit anticipated unimodality of conditional distributions of the effect. The proposed methods are straightforward and include only a limited number of hyperparameters that affect the accuracy of the identification. For a given set of hyperparameters, both the AAG and MI methods provide a unique, deterministic solution. To address sensitivity to hyperparameters, tuning has been done by utilizing a full factorial Design of Experiment. It turns out that the AAG method outperforms MI, achieving top weighted accuracies of 82.7% with simple tuning and 84.4% with size-adaptive tuning, compared with 81.6% for GRCI or 82.0% for CAREFL-H on the 99 pairs of the Tubingen real-world cause-effect examples. To evaluate the decisiveness of the identification method, a decision tree was fitted on the input data's symmetrical bivariate statistics to detect misclassified cases.
arXiv:2603.07233v2 Announce Type: replace Abstract: Predicting transcriptional responses to genetic perturbations is fundamental to functional genomics and therapeutic discovery. Recent deep learning models have shown promise in single-cell perturbation response prediction, but they typically generate each response in isolation, without explicitly leveraging related perturbations. We introduce PT-RAG (Perturbation-aware Two-stage Retrieval-Augmented Generation), a plug-in retrieval-and-conditioning module for generative cellular perturbation response. PT-RAG augments an existing perturbation-response backbone with learned access to related perturbation contexts. The key challenge is that relevance is not fixed in this setting: functionally related genes may elicit different effects across cell types. PT-RAG addresses this with a two-stage retrieval mechanism: GenePT-based semantic retrieval first identifies K candidate perturbations, after which a differentiable Gumbel-Softmax selector adaptively selects retrieved contexts conditioned on the control cell state, the query perturbation, and each candidate perturbation. We evaluate PT-RAG on two backbones, a STATE-style generator used as a frozen random reservoir and a fully trained scGPT, across cross-cell-type and cross-perturbation generalization tasks. PT-RAG consistently improves distributional similarity and often overall predictive quality; for example, on scGPT cross-cell-type results, the 2-Wasserstein distance drops by 5.9% relative to scGPT alone. The code to reproduce our experiments is available at https://github.com/difra100/PT-RAG_NIPS.
arXiv:2602.19661v3 Announce Type: replace Abstract: Temporal information in structured electronic health records (EHRs) is often lost in sparse one-hot or count-based representations, while sequence models can be costly and data-hungry. We propose PaReGTA, an LLM-based encoding framework that (i) converts longitudinal EHR events into visit-level templated text with explicit temporal cues, (ii) learns domain-adapted visit embeddings via lightweight contrastive fine-tuning of a sentence-embedding model, and (iii) aggregates visit embeddings into a fixed-dimensional patient representation using hybrid temporal pooling that captures both recency and globally informative visits. The resulting fixed-dimensional patient representations can be used with conventional downstream machine-learning models. To examine factor-level sensitivity, we use PaReGTA-RSS (Representation Shift Score), a prespecified factor-removal analysis that recomputes patient representations after removing clinically defined factor groups and quantifies the resulting change in the fitted logit of a fixed logistic-regression model. We evaluated PaReGTA in a cohort of 39,088 patients with migraine from the All of Us Research Program (AoU) on a retrospective patient-level classification task with an EHR-derived target. In an exploratory comparison on the fixed held-out test cohort of 7,818 patients, PaReGTA-Gap + LightGBM, used as a post hoc analytical reference, had the highest reported AUC, accuracy, and F1 values among the evaluated sparse, BERT-based EHR, and recurrent approaches in this cohort.
arXiv:2510.12843v2 Announce Type: replace Abstract: Spiking neural networks (SNNs) promise energy-efficient artificial intelligence on neuromorphic hardware but struggle with tasks requiring both fast adaptation and long-term memory, especially in continual learning. We propose Local Timescale Gating (LT-Gate), a neuron model that combines dual time-constant dynamics with an adaptive gating mechanism. Each spiking neuron tracks information on a fast and a slow timescale in parallel, and a learned gate locally adjusts their influence. This design enables individual neurons to preserve slow contextual information while responding to fast signals, addressing the stability-plasticity dilemma. We further introduce a variance-tracking regularization that stabilizes firing activity, inspired by biological homeostasis. Empirically, LT-Gate yields significantly improved accuracy and retention in sequential learning tasks: on a challenging temporal classification benchmark it achieves about 51 percent final accuracy, compared to about 46 percent for a recent Hebbian continual-learning baseline and lower for prior SNN methods. Unlike approaches that require external replay or expensive orthogonalizations, LT-Gate operates with local updates and is fully compatible with neuromorphic hardware. In particular, it leverages features of Intel's Loihi chip (multiple synaptic traces with different decay rates) for on-chip learning. Our results demonstrate that multi-timescale gating can substantially enhance continual learning in SNNs, narrowing the gap between spiking and conventional deep networks on lifelong-learning tasks.