Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2608. 28063v1 Announce Type: new Abstract: Multi-organ ultrasound classifiers increasingly combine attention, mixture-of-experts routing, uncertainty gating, and evidential deep learning (EDL) objectives to address heterogeneous anatomy and acquisition.
DiffAnon is a diffusion‑based voice anonymization method that uses classifier‑free guidance to give users continuous, inference‑time control over how much prosody is preserved. By refining acoustic detail over semantic embeddings from an RVQ codec, the model allows smooth interpolation between strong anonymization and high prosodic fidelity within a single architecture. This is the first framework to provide structured, interpolatable prosody control while maintaining competitive privacy and utility across different operating points.
arXiv:2608.27856v1 Announce Type: new Abstract: Recent advances in large language models are enabling autonomous clinical agents to perform increasingly complex electronic health record (EHR) modelin...
arXiv:2608.27782v1 Announce Type: cross Abstract: Memorization in large language models is measured through a zoo of definitions whose formal relations are unknown, and differential privacy (DP) is t...
arXiv:2608.27658v1 Announce Type: new Abstract: Subword tokenization hinders low-resource language processing by imposing frequency patterns from dominant languages onto script-sharing variants. Byte...
The paper introduces Dynamic Alignment Compensation (DAC), a training‑free inference‑time technique designed to reduce hallucinations in Large Vision‑Language Models (LVLMs). DAC monitors cross‑modal representation drift across decoder layers and generation steps, applying lightweight residual compensation through Layer‑wise Semantic Compensation and Sequential Semantic Correction. Experiments on nine multimodal benchmarks across various LVLM backbones demonstrate that DAC consistently lowers hallucination rates while preserving overall performance.
Large reasoning models generate chain-of-thought (CoT) before answering but struggle with safety alignment and can be misled by flawed premises. The paper introduces RECAP, a reinforcement learning approach that trains models to override flawed reasoning paths and produce safe, helpful responses without extra training cost. RECAP improves safety, jailbreak robustness, and reduces overrefusal while preserving core reasoning abilities and inference token budget.
The UIC-AIHealth4All system was presented for the ArchEHR-QA 2026 shared task on grounded question answering from electronic health records. It participated in evidence identification, answer generation, and answer‑evidence alignment, using an answer‑first pipeline that generates candidate answers with cited note sentences before classifying the full evidence set. The system ranked third in evidence identification, ninth in answer generation, and fifth in answer‑evidence alignment, and a linguistic analysis showed its outputs were harder to read than clinician‑authored references, highlighting the need for readability optimization in clinical NLP.
Cyc3D is a new benchmark for image‑to‑3D generation that evaluates both Cross‑View Object Consistency and Representation Quality. It introduces a closed‑loop View‑Cycle Structural Consistency protocol to measure geometric, perceptual, and semantic drift across repeated render‑regenerate cycles, and also assesses asset usability through geometric structure, fidelity, mesh discretization, and UV quality. Experiments show that closed‑source models outperform open‑source baselines yet still score below 48 on cycle stability, highlighting a gap between visual plausibility and robust 3D understanding.
iOSPointMapper is a mobile app that performs real‑time, privacy‑conscious sidewalk mapping using on‑device semantic segmentation, LiDAR depth estimation, and fused GPS/IMU data on recent iPhones and iPads. It detects and localizes sidewalk‑relevant features such as traffic signs, traffic lights, and poles, and includes a user‑guided annotation interface for validating outputs before submission. The anonymized data is transmitted to the Transportation Data Exchange Initiative (TDEI), where it integrates with broader multimodal transportation datasets, and evaluations show the app’s potential for enhanced pedestrian mapping.
EvoHarmBench is a dynamic adversarial evaluation framework that simulates how users iteratively modify harmful content to evade moderation. It uses an optimization loop that evolves evasion strategies at the semantic-cluster level while maintaining human readability, and tests 229 semantic sub-clusters across five violation categories derived from 5,002 real-world adversarial samples. The study shows that even state‑of‑the‑art LLM‑based moderators can be bypassed with an 80.3% success rate after twelve iterations, highlighting significant vulnerabilities in current systems.
GeoFF3D is a new feed‑forward 3D reconstruction method designed for large‑scale UAV mapping. It uses a coordinate‑anchored model that predicts camera poses and dense point maps directly in a gravity‑aligned Z‑up metric frame, while a spatial large‑scale reconstruction framework (SLRF) partitions images into overlapping chunks, propagates shared‑view priors, and aggregates local reconstructions hierarchically. Across nine aerial mapping blocks, GeoFF3D achieves the best average reconstruction quality, improving F@5 from 0.829 to 0.877, and can reconstruct 2,000 images in about five minutes.
The paper shows that post‑training quantization can introduce backdoors in large language models that are not detected by source‑precision checks. By formalizing the validation‑deployment gap with Quantization Behavioral Equivalence Classes (QBECs), the authors demonstrate that models can pass full‑precision tests yet exhibit malicious behavior after INT8 or 4‑bit compression. Experiments on machine translation and political stance classification reveal significant corruption and ideological shifts, and cross‑quantizer analysis indicates that attack persistence depends on the quantization scheme and architecture rather than just bit‑width.
The paper introduces Visual Metaphor Transfer (VMT), a task that requires models to extract the abstract ‘creative essence’ from a reference image and apply it to a new target subject. It proposes a multi‑agent framework based on Conceptual Blending Theory, using a Schema Grammar to separate relational invariants from visual entities. The system includes perception, transfer, generation, and diagnostic agents, and experimental results show it outperforms state‑of‑the‑art baselines in metaphor consistency, analogy appropriateness, and visual creativity.
Cut‑ViT introduces a task‑specific pruning pipeline for visual foundation models that uses gram anchoring matrices and subspace decomposition to align feature representations between native and pruned DINOv3 models. The method incorporates basis‑agnostic and residual constraints to preserve robustness across spatial and channel dimensions, and employs spectral entropy adaptation to tailor the pruning objective to downstream tasks. Experiments demonstrate that Cut‑ViT achieves state‑of‑the‑art performance on six tasks across nine datasets while reducing pruning time to about one minute on a single A100 GPU, using only 20.9% of the time and 45.5% of the GPU memory compared to prior methods.
The paper introduces Relax Forcing, a training‑free memory mechanism for autoregressive video diffusion that structures temporal context into Sink, Tail, and History frames. By selecting History frames with a relaxation criterion, the method reduces error accumulation and attention overhead while preserving motion dynamics. Experiments on VBench‑Long demonstrate that this structured memory improves long‑video generation quality over existing baselines.
The paper introduces QK Product Steering, a data‑free, training‑free method that edits the query‑key product in vision‑language models to reduce object hallucination. By suppressing a few dominant singular modes in selected middle layers and mapping the edited product back to query weights, the approach lowers hallucination rates without affecting inference cost. Experiments on three GQA‑based VLMs show a 4.0% average reduction in CHAIR$_s$, with the effect localized to symmetric mutual‑attention channels.
SUFLECA is a weakly supervised framework that improves zero‑shot CAD‑to‑image alignment by scaling geometry‑grounded feature learning using Normalized Object Coordinates across up to 12 real and synthetic datasets. It introduces a geometrically consistent matching algorithm that reliably establishes CAD‑to‑image correspondences, enabling accurate, sub‑second alignment without iterative pose refinement. On the ScanNet25k benchmark, SUFLECA achieves 32.8%/42.6% category/instance accuracy, outperforming the strongest zero‑shot baseline by 9.7/12.5 percentage points and surpassing existing pose‑supervised methods for the first time.
The paper introduces PMFRec, a federated cold-start recommendation framework that addresses personalization, compositionality, and communication inefficiencies. PMFRec generates user-specific item representations from attribute features, employs a global multi-view encoder with adaptive gating and orthogonality to capture complementary semantics, and fuses collaborative and attribute knowledge into a single exchanged representation. Experiments on real-world datasets demonstrate that PMFRec outperforms strong baselines in cold-item recommendation while improving user-level fairness, warm-scenario adaptability, and robustness under Local Differential Privacy.
The paper introduces Concept-Targeted Attribution (CTA), a method that trains attribution graphs to explain the emergence of internal concept representations in language models, rather than just the final token prediction. CTA produces probe-specific circuits that reveal which internal computations drive a linear probe’s accuracy, and cross-layer transcoders demonstrate that these graphs contain predictive structure across multiple concept categories. Causal ablations show that probe-targeted and logit-targeted graphs capture distinct mechanisms, with probe-relevant features affecting internal concept scores and logit-relevant features altering generated tokens.