The paper presents a training‑free weight‑editing technique that reduces language model compliance with harmful instructions. By applying forward passes on a small set of prompt pairs, the method modifies model weights so that changing or deleting instructions in the prompt cannot reverse the edit. Experiments on five instruction‑tuned models show significant improvements in honest response rates for lying and substantial reductions in sexist responses, demonstrating the approach’s effectiveness across distinct behavioral domains.
By Ali Ebrahimpour-Boroojeny, Xiaocong Yang, Samuel Schapiro, Zahra Borhani, Francisco R. Ortega, Yujia Zheng, ChengXiang Zhai
The paper reports two data‑only poisoning attacks on protein design systems. In the first, poisoned sequences were added to the pre‑training corpus of large autoregressive language models, causing the models to reliably generate sequences triggered by specific patterns after as few as 200 contaminated examples. The second attack inserted malicious records into a retrieval corpus, enabling a single query‑aware record to dominate retrieval results and, in many cases, alter the generated protein sequence or transfer a new domain via homology.
By Charalampos Koilakos, Kimon-Antonios Provatas, Michail Patsakis, Candace SY Chan, Georgios A. Pavlopoulos, Ilias Georgakopoulos-Soares
The paper introduces Update‑Matched Unlearning (UMU), a method that applies the same simulated relearning update to both a defended model and an undefended reference to better preserve plasticity while resisting relearning of deleted knowledge. Experiments on Llama and Qwen models with TOFU and FaithUn datasets show that UMU reduces harmful and benign relearning by 30.4 and 9.9 points on average and increases new‑task accuracy by 3.5 points, supporting the reference‑timing hypothesis. The study demonstrates that robust unlearning should selectively prevent recovery of deleted knowledge without suppressing the model’s ability to learn new tasks.
By Jie Li, Youming Tao, Bowei He, Meng Ding
This paper studies the statistical properties of Stochastic Gradient Descent (SGD) with a constant step size using the framework of iterated random functions. It demonstrates convergence in Euclidean distance via Geometric Moment Contraction, establishes quenched central limit theorems, invariance principles, and α-stable convergence for averaged SGD regardless of initialization, and explores the effect of step sizes through derivative processes. The authors further develop a Richardson-Romberg extrapolation with multiple step sizes, propose a bias-reduced variant, and introduce an online inference method that yields empirical confidence intervals with coverage approaching nominal levels as sample size increases.
By Jiaqi Li, Zhipeng Lou, Stefan Richter, Wei Biao Wu
The paper investigates Kolmogorov‑Arnold networks (KANs) and their approximation properties. It shows that while KANs can achieve dimension‑independent approximation exponents for targets with smooth univariate edge functions, this advantage is tied to the target structure rather than the architecture itself. Empirical tests reveal that KANs outperform MLPs only when each nonlinear component depends on a single input coordinate; otherwise, MLPs can be more efficient, especially under tight parameter budgets or for interacting targets.
By Yunhao Ni, Yuxin Guo, Jie Luo, Wenjun Wu, Lei Huang
The paper investigates privacy and integrity risks in federated fine‑tuning of a pretrained MobileNetV4 model on Quick, Draw! data. It shows that a single client update can reveal private labels, that label‑count recovery degrades with more clients, and that backdoor and label‑flipping attacks can succeed under FedAvg but are mitigated by FLTrust. The study also demonstrates a live poisoning session using Indigma’s FedMaestro orchestrator.
By Vasileios Perifanis, Nikolaos Pavlidis
The paper investigates how regularized importance sampling in policy gradient estimation breaks reward invariance and proposes using the covariance between reward and importance-weighted score to restore it. It compares two centering techniques—group-mean baseline and score centering—showing they differ in the mean subtracted and yield distinct unbiased covariance estimators with different variances. The study finds that for sequence-level importance sampling, score centering offers no added benefit beyond the group-mean baseline, whereas for token-level sampling, omitting score centering can introduce bias when the importance ratio is regularized.
By Shuo Han, Ziqing Hu
The study examines how ensemble spread from Long Short-Term Memory (LSTM) models, used for streamflow forecasting, behaves under increasingly extreme flow conditions due to climate change. It finds that prediction error grows relative to the corrected spread—up to four times larger at the highest flows—and that under-dispersion is most pronounced for extreme flows beyond the training record, reaching an 18‑fold discrepancy. While the spread still ranks errors well (near‑oracle ranking), it systematically understates risk, especially for the most critical extreme events.
By Justin Gallant, Tyler Wilson
MIDI-REPA is a framework that aligns symbolic music representations with intermediate layers of a flow-matching DiT model to improve text-to-music generation. By deriving MIDI supervision from automatic music transcription and aligning a MuScriptor-derived piano roll with an early DiT layer throughout the flow trajectory, the method achieves competitive performance on the ATTM objective metrics without extra training data or techniques. The approach is trainable on a single RTX 3090 in about 31 hours and is supported by released transcription pipeline, training code, and audio samples.
By Song-Ze Yu, Fang-Duo Tsai, Yi-Hsuan Yang
The paper introduces the pooled selection-rate standard deviation (pooled SD) as a way to quantify how unevenly multiple-choice option letters are chosen across different orderings. It shows that pooled SD decomposes into two label‑free factors: the flip rate, which measures how often a model’s chosen option changes with ordering, and the positional skew, which captures the variability in option‑letter selection rates. Experiments on 41 checkpoints and 12 benchmarks reveal that prior correction mainly reduces positional skew, while aggregating each ordering with its reversal mainly reduces flip rate; applying both mitigations together further narrows the accuracy range across answer letters.
By Hyunjong Ok, Minjae Park, Jaeho Lee
The paper presents a theoretical analysis of score mismatch in discrete diffusion models, deriving a KL divergence upper bound that quantifies the bias introduced when the training and target distributions differ. It examines partial-observation constraints such as text infilling and sentence completion, showing that the Conditional SEDD sampler can incur a bias that grows with dimension, vocabulary size, and compatibility. To mitigate this, the authors propose a Randomized Hard Clamping (RHC) sampler for coordinatewise set constraints, demonstrating through experiments that it effectively reduces bias and improves sampling performance.
By Renxiang Huang, Yuchen Liang, Yingbin Liang, Ness Shroff, Lifeng Lai
The paper investigates Classifier-Free Guidance (CFG) in Text-and-Image-to-Image (TI2I) flow matching for controllable image editing. It finds that CFG can be implicitly harmful during the noisy phase of flow modeling, reducing instruction following and controllability. The authors propose a simple early-flow correction that improves the alignment‑preservation tradeoff without additional cost, and validate the method on both pixel‑space and latent TI2I flows.
By Andreas Floros, Elena Koritska, Teofilo de Campos, Christoph Vogel, Chu-Tak Li, Anuj Sharma, Alexandros Neophytou
MARS is a Momentum‑Aware Adaptive Rating System that models a question bank as a directed weighted knowledge graph and updates learner proficiency using dynamic rating sensitivity, response quality, and momentum‑aware learner velocity. It introduces a time‑budgeted mark‑maximization objective to select questions within a fixed test duration under negative marking. Experiments on ASSISTments 2009 and MARS‑SimExam show that MARS delivers smoother rating updates than Fixed Elo and achieves the highest score under time‑budgeted negative‑marking routing, improving total score by 8.09 over the strongest baseline and reducing overtime violations from 50% to 5.6%.
By Vikrant Sahu, Khethavath Sunil Naik, Gagan Raj Gupta
The paper proposes an AI-native, robust network for sixth-generation (6G) communication systems, where multiple base stations serve extended reality (XR) devices. It introduces a distributed deep learning approach that jointly manages transmit power, computing power, and computation allocation while accounting for uncertainties in channel and computing states. The method uses a deep neural network trained centrally or via federated learning, injecting uncertainty samples after the DNN output to minimize worst-case delay under power and computation constraints, achieving near-centralized performance with improved privacy and scalability.
By Robert-Jeron Reifert, Hayssam Dahrouj, Alaa Alameer Ahmad, Haris Gacanin, Aydin Sezgin
The paper introduces Where to Steer (W2S), a framework that adaptively selects which layer of a large language model to intervene on based on the input. It argues that optimal steering layers vary across inputs and demonstrates that W2S outperforms fixed-layer steering methods on several LLMs and alignment tasks, both in‑distribution and out‑of‑distribution. The study highlights input‑dependent layer selection as a complementary design dimension to steering vector extraction.
By Soham Gadgil, Chris Lin, Su-In Lee
The paper introduces IBIS, a framework for designing canaries that are efficient for one‑run privacy auditing. IBIS optimizes each canary’s membership inference attack score while minimizing interference among canaries by encouraging orthogonality in embedding space. The method uses influence‑based greedy initialization and a bilevel optimization algorithm that updates a single model incrementally, achieving better performance‑cost trade‑offs and reducing computational cost by up to two orders of magnitude compared to existing approaches.
By Mathieu Dagr\'eou, Valentin Dorseuil, Mohammad Yaghini, Nicolas Papernot, Aur\'elien Bellet
The paper presents the first adversarial linear bandit algorithm based on Follow-the-Perturbed-Leader (FTPL) that achieves √T regret over general convex action sets. It achieves a regret of ≥O(d^{5/4}√T) using only linear optimization oracles, and a near‑optimal ≥O(d√T) regret when an optimal‑design exploration distribution is available. The work improves on previous FTPL algorithms and generalizes earlier results for special action sets, offering a simpler and more computationally efficient approach that relies on adaptive perturbation distributions and a new analysis framework based on Fisher information.
By Chen-Yu Wei
The paper introduces a multimodal emotion recognition system that integrates facial expressions, speech, text, and body movement to aid clinicians in psychological evaluation. While individual modalities perform well on benchmarks, their real‑world accuracy drops to 82.97%–87.64%, whereas the fused multimodal system reaches 96.43% accuracy with no state below 92.31%. These findings demonstrate that multimodal fusion enhances robustness in realistic settings, supporting more reliable evaluator assistance.
By Kris Kraack
The paper constructs coordinate‑Bethe‑Ansatz joint laws for parallel token proposals using the periodic XXZ chain and compares them to maximum‑entropy pairwise projections that match the same one‑ and two‑site moments. In an eight‑site, three‑excitation sector, symmetry reduces the difference between the full probability law and its pairwise projection to a single contiguous three‑site moment δ, which changes sign and vanishes at Δ=(3+√3)/2, making the eigenstate law pairwise. Nonlinear two‑token and three‑token readouts preserve token marginals while revealing distances proportional to |δ|, with explicit deterministic binary distances computed at Δ=1.5; the authors also provide an exact target‑alignment formula and prove a strict speculative accepted‑prefix advantage in an unrestricted open neighborhood of the constructed target.
By Abraham Chachamovits
PRISM is a framework that evaluates how activation interventions transfer across precision, languages, and model scales to mitigate unsafe tool‑calling behaviors in language models. By leveraging lower‑resource settings such as quantized NF4 labels and cross‑lingual transfer, PRISM demonstrates that alignment improvements can be achieved with minimal scenarios and that these gains transfer to held‑out languages and larger models. The study shows that auxiliary supervision from compressed models and multilingual data can effectively uncover and reduce hazardous tool‑calling behaviors across different model families.
By Giovana Kerche Bon\'as, Thales Sales Almeida, Marcelo da Silva Reis, Rodrigo Nogueira, Marcos Meideiros Raimundo