Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Multi subject video editing modifies designated subjects while preserving non target content, but faces cross subject attribute leakage, and occlusion ambiguity. Existing approaches rely on masks and...
Privileged on-policy distillation improves multimodal reasoning by allowing a teacher to evaluate student trajectories using rich, training-only visual evidence. Both models score these trajectories w...
Classification and regression trees are typically constructed using a greedy splitting rule that maximizes the immediate reduction in prediction error at each node. Although this strategy is computati...
We introduce ParsHate, a manually annotated dataset of 10,000 Persian tweets spanning 2013-2022, representing the first decade-long benchmark for hate speech detection in Persian. The dataset contains...
Probabilistic certification of Bayesian neural networks lower-bounds the posterior probability that a model satisfies a verifier-defined safety property. In one-shot federated Bayesian learning, howev...
Large language model (LLM) agents increasingly operate over long-horizon interactions involving tool use, persistent state, evolving authorization, and external environment feedback. In such settings,...
Many operational cases are documented more than once, at different workflow stages and for different purposes, yet model evaluations normally select one of these records before model comparison begins...
Long-horizon scientific discovery requires agents to alternate between exploration, disciplined execution, and critical reassessment as evidence changes. Current language models are trained primarily...
Record-level differential privacy exposes a structural misalignment in personalized federated learning when client-specific variation is low-dimensional while training repeatedly releases high-dimensi...
Multimodal image registration is a key component of many clinical workflows, yet it remains challenging because corresponding anatomical structures often exhibit substantially different image intensit...
Multilingual Large Language Models (LLMs) traditionally rely on a single vocabulary shared by all supported languages, which can lead to uneven compression across them. Moreover, their large embedding...
The paper introduces a three‑stage training pipeline that builds compact, efficient dense retrievers without requiring ground‑truth relevance labels. Using cross‑lingual alignment, relational knowledge distillation, and contrastive fine‑tuning, the authors develop PolDense (six Polish models ranging from 17 M to 1 B parameters) and EuroDense (a 435 M‑parameter model covering nine European languages). Extensive evaluation on 41 Polish and 150 multilingual tasks shows that PolDense‑1B outperforms larger retrievers up to 9 B parameters, while EuroDense leads in task‑averaged and language‑averaged performance among models below 1 B parameters.
The paper investigates on-device language models (ODLMs) for predicting stress in a mobile health context, focusing on privacy-preserving, cloud-independent inference. Using zero‑shot prompting, the authors evaluate ODLMs across multimodal data—objective sensor features and subjective self‑reports—measuring predictive accuracy, latency, and throughput. Results indicate that sensor features slightly outperform self‑reports, and that lightweight sub‑2B models deliver low latency with predictable resource usage, underscoring both the potential and practical limits of ODLMs for mobile mental health.
The paper investigates how large language models (LLMs) used as judges in absolute scoring tasks exhibit systematic biases that compromise reliability. It shows that a judge’s task accuracy strongly predicts both its judging accuracy and its directional bias, yet more capable examinee models consistently receive more lenient judgments. To mitigate these biases, the authors propose a calibrated weighted majority voting (WMV) ensemble that estimates judges’ error rates from inter-judge agreement patterns, achieving near-oracle performance without labeled data and improving both accuracy and fairness.
The paper introduces an adversarial reinforcement learning framework that learns the sparsest Denial-of-Service (DoS) attack schedule capable of destabilizing self‑triggered reinforcement learning controllers (RL‑STC). It proves a lower bound on the minimum number of jamming actions needed to force a crash and demonstrates that the learned adversary consistently defeats four different defenders—one LQR and three RL‑STC—across Pendulum, CartPole, and Quadrotor2D environments, outperforming greedy and periodic baselines in jam‑time‑per‑failure. The study also shows that the adversary remains effective under Gaussian observation noise and limited state information.
The paper introduces a Differentially Private Federated Proximal (DP‑FedProx) framework for predicting customer churn in telecom networks. It compares DP‑FedProx with FedAvg, DP‑FedAvg, FedProx, and several centralized/local models on two public churn datasets, showing that DP‑FedProx consistently outperforms FedAvg variants and matches the best centralized model with only a slight accuracy loss while offering privacy guarantees. SHAP analysis reveals that DP‑FedProx prioritizes revenue‑group features, highlighting its practical balance between performance and privacy.
The paper proposes a method for routing requests to a fixed pool of quantized Mixture-of-Experts (MoE) instances, aiming to maximize throughput while respecting a quality‑degradation budget. It introduces Fragility‑Weighted Perplexity (FWP) as a request‑specific risk metric derived from prompt tokens, and uses a window‑level linear program to compute a reduced‑reward score that aligns with the LP optimum. Experiments on Qwen prompts show that FWP‑based allocation improves throughput by 2.5% over request‑agnostic mixing and static configurations.
arXiv:2609. 12591v1 Announce Type: new Abstract: Foundation models are increasingly adapted through fine-tuning, model editing, and alignment procedures while retaining previously acquired capabilities.
arXiv:2609. 12594v1 Announce Type: new Abstract: We study the classical Moreau--Yosida unadjusted Langevin algorithm (MYULA) for $\pi(\,\mathrm{d} x)\propto e^{-f(x)-g(x)}\,\mathrm{d} x$, where $f\in C^2(\mathbb{R}^d)$ is $m$-strongly convex with $L_f$-Lipschitz gradient and $g:\mathbb{R}^d\to\mathbb{R}$ is convex and globally $G$-Lipschitz.
The paper revisits the distortion problem in Reinforcement Learning from Human Feedback (RLHF) for aligning large language models. It shows that exponential degradation in user utility, previously attributed to RLHF, actually stems from a mismatch between the preference data distribution and the KL reference policy. By deriving tight bounds across different KL regularization regimes, the authors demonstrate that when the two distributions match, RLHF achieves near‑optimal distortion, and they recommend using on‑policy data or pre‑fine‑tuning on data close to the true preference distribution.