Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Speech-LLMs often exhibit prompt overfitting, where models solely trained on automatic speech recognition (ASR) instruction fail to generalize to new instructions such as speech translation and contin...
LLM tool-use agents operate in dynamic environments where many actions carry operational risk. However, most safety mechanisms react only after errors manifest. Existing pre-emptive approaches either...
Neural audio synthesis models like the Realtime Audio Variational autoEncoder (RAVE) achieve impressive genera tion quality, yet how their internal representations encode musical features remains poor...
As LLMs increasingly assist in moral reasoning, omission bias, the tendency to prefer inaction even when equivalent framings reverse substantive outcomes, poses a significant risk of skewed decision-m...
Prior Fitted Networks (PFNs) such as TabPFN now rival established statistical procedures across prediction and estimation tasks. A natural explanation is that PFNs have the property of statistical ada...
Recommender systems learn from observed user-item interactions, but explicit negative feedback is often unavailable. Since deep learning models require negative signals for training, negative sampling...
In this paper, we show that standard evaluations of high-resolution Model Inversion Attacks (MIAs) significantly underestimate training-data privacy leakage. State-of-the-art privacy defenses, standar...
While AI weather models now rival operational forecasts, how they represent the atmosphere internally remains an open question: feature attribution reveals which input patterns matter, not what the mo...
The article explains the transition from the original Cowork system, which ran model inference and tool calls in a local Anthropic VM, to a new cloud‑based version that hosts both inference and the VM. The new design isolates each session in its own sandbox, eliminates local resource drain, and allows file access through the desktop app, addressing concerns about battery usage, performance, and continuity when a laptop is closed.
Joint Models simultaneously model longitudinal and survival outcomes, leveraging patterns in patients' longitudinal trajectory to improve the prediction of survival outcomes. The classical parametric...
SPDAlign is an interpretable framework that aligns domain-specific EEG data by applying linear transformations on the Symmetric Positive Definite manifold, correcting mean differences and global rotations across domains using Wasserstein Procrustes. It leverages a generative model of EEG to recover distribution shifts caused by factors like head geometry, and demonstrates competitive performance on extensive public EEG datasets. The method is globally linear, enabling identification of frequency ranges, spatial patterns, and cross-subject variability without requiring labeled calibration data.
MoCAR (Motion-code Coordinate-aware AutoRegression) is a decoder‑only framework that treats continuous trajectory forecasting as next‑code prediction in a coordinate‑aware latent space. It learns a continuous motion‑code space from endpoint‑normalized trajectory segments, using historical codes as a teacher‑forced prefix and generating future codes autoregressively while updating local scene context. The method achieves top‑tier performance on Argoverse benchmarks, transfers well from AV2 to AV1 in zero‑shot evaluation, and improves on turn‑heavy scenarios without requiring trajectory‑space re‑tokenization or proposal‑and‑refinement pipelines.
The paper evaluates CLIP’s zero‑shot gender estimation on full‑face and periocular images from the Adience dataset. Using image‑text similarity with male/female prompts, CLIP achieves 95.54% accuracy on full faces without task‑specific training. Periocular predictions are initially biased toward males, but threshold alignment improves accuracy to 85.29%, and a linear SVM on CLIP features yields a modest 86.17% accuracy, slightly better than prior Adience results.
The paper introduces TrustMI, a method to causally control how large language model assistants decide to trust their users. By creating 2,000 contrastive conversations that vary in ability, benevolence, and integrity, the authors learn steering matrices that adjust trust decisions along linear directions in model activations while keeping the model parameters frozen. Experiments across six instruction‑tuned models show that these steering changes reliably alter trust decisions and affect safety‑related behaviors such as compliance with harmful requests, prompt injections, and insider threats.
The paper examines the reliability of deepfake detectors, noting a decline in accuracy from 99.5% to 76% over four years and a drastic drop below 2% when adversarial perturbations are applied. It argues that content alone cannot determine provenance because generators can perfectly reproduce authentic images, so a new detection approach is proposed that certifies authenticity only if a generator cannot faithfully reconstruct the content. The authors demonstrate that their calibrated detector can limit false certifications to 1% while maintaining robustness against bounded‑perturbation attacks, though it struggles with arbitrary transformations.
The paper investigates how poisoning attacks—specifically label flip and noise injection—affect online AutoML learners in IoT networks and evaluates the effectiveness of adversarial training (AT) as a defense. Experiments on streaming-capable models (Hoeffding Tree, Leveraging Bagging, Adaptive Random Forest, Hoeffding Adaptive Tree, and Streaming Random Patches) show that AT-SRP achieves the highest F1-score (0.904) against label flip attacks, while AT-LB achieves the highest F1-score (0.933) against noise injection. Drift detection methods were also employed for rolling accuracy and prequential evaluation.
The paper introduces U-LoRA, an adaptive version of Low‑Rank Adaptation that uses conditioned gating to generate token‑specific utilization coefficients for the low‑rank subspace. By coordinating these coefficients with sequence‑level context and applying a bias‑corrected EMA prior, U‑LoRA achieves more consistent adaptive patterns and reduces noise from batch fluctuations. Experiments on mathematical reasoning and natural language understanding tasks show that U‑LoRA matches or surpasses strong LoRA baselines while keeping the same parameter budget.
The paper introduces a geometric measure of coarticulation for speech‑driven 3D facial animation, comparing lip‑path length to the shortest route through vowel, consonant, and vowel positions. Using only forced alignment, the measure evaluates four state‑of‑the‑art animation methods, revealing that all produce flatter lip trajectories than captured speech and that some methods lose 15–60% of the fast articulatory component. A pre‑registered viewer study confirms that damping real motion lowers perceived quality while exaggeration is not penalized, and viewers prefer real speech in 73.4% of sentence comparisons.
The paper investigates how visual and textual information are fused in Multimodal Large Language Models (MLLMs). By analyzing concatenation and native multimodal architectures through alignment decoupling, attention routing, entropy, intrinsic dimensionality, and causal interventions, the authors uncover two distinct fusion pathways: concatenation models use a text‑first, vision‑later strategy, while native models integrate vision and text earlier and reorganize feature spaces. The study also employs visual CKA to test the Platonic Representation Hypothesis, offering a mechanistic view of multimodal fusion and informing architecture‑aware diagnostics.
The study explores how the language used for reasoning affects retrieval‑augmented generation (RAG) in a monolingual German setting. Using a German RAG question‑answering testbed based on the tabletop game The Dark Eye, the authors show that aligning the reasoning language with the query and retrieved documents improves performance, with German reasoning outperforming French reasoning. However, German reasoning still does not surpass the model’s native English reasoning, indicating that native multilingual reasoning is necessary for optimal results.