Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2610.27719v1 Announce Type: new Abstract: Rapid advances in diffusion models have raised concerns about privacy, harmful content, and copyright violations. Concept erasure methods aim to addres...
arXiv:2610.27947v1 Announce Type: new Abstract: Human-centric audio-visual editing requires precise alignment of lip movements, gestures, and expressions with speech. Existing methods process inputs...
arXiv:2610.28289v1 Announce Type: new Abstract: Radiologists interpret follow-up chest X-rays by comparing current and prior examinations to identify new, resolved or progressing findings. Most repor...
arXiv:2610.28511v1 Announce Type: new Abstract: Tactile-visual-language models combine vision and touch through a shared language space to recognize objects and materials. In real-world deployment, c...
arXiv:2610.28530v1 Announce Type: new Abstract: Collaborative perception enables agents to retrieve the scene semantics and geometry beyond the field of view of any individual agent. This broader sce...
arXiv:2610.28571v1 Announce Type: new Abstract: ScreenCalib enables light detection and ranging (LiDAR)--camera extrinsic calibration using an ordinary programmable display, providing known visual ma...
arXiv:2610.28615v1 Announce Type: new Abstract: Deploying a trained 3D Gaussian Splatting scene under shrinking compute budgets requires one primitive set to remain effective when rendered at smaller...
arXiv:2604.27529v2 Announce Type: replace Abstract: How do CNN channels combine scene information into a class-specific response? We develop an input-conditioned adjoint inversion framework that rend...
arXiv:2610.27425v1 Announce Type: cross Abstract: We develop a Bayesian machine learning approach for estimating direct and spillover effects in cluster-randomized trials with treatment noncompliance...
arXiv:2610.26069v2 Announce Type: replace Abstract: Despite achieving high visual fidelity, video models still suffer from physical hallucinations, limiting their reliability as world simulators. Exi...
RealControlBench is a new benchmark that evaluates fine‑grained steering of large language models and agents across diverse behavioral targets and interaction settings. It extends beyond isolated generation tasks to include complex, execution‑grounded, multi‑turn agent interactions, covering instruction‑following control, tool‑use decision control, reasoning performance, reasoning length, and safety awareness. The benchmark jointly measures target controllability and capability preservation, testing robustness across prompts, domains, and model architectures, and finds that current steering methods are often unreliable in realistic scenarios.
The paper investigates whether the internal hidden states of large language models can improve the detection of hallucinations in model responses. By integrating hidden states with an output‑side baseline and evaluating on both open‑ended generation and multiple‑choice tasks, the authors find that adding hidden states yields modest AUROC gains of about 0.004 and 0.039. These results suggest that while hidden states contain additional predictive information, the overall improvement over strong standalone detection methods remains limited.
The paper introduces DeCert, a framework that evaluates the validity and verdict consistency of rationales generated by generative reward models (GenRMs) without needing reference explanations. DeCert translates these evaluations into preference supervision that considers both verdict correctness and justification quality, enabling the construction of rationale preference pairs that differentiate stronger from weaker justifications. Experiments across multiple model families and benchmarks show that DeCert outperforms strong generative judge baselines, improves rationale quality, and mitigates positional bias, raising position-consistent accuracy by up to 14.8 points.
The paper introduces a protocol for generating and evaluating codebooks for measuring Computational Discourse Quality Index (DQI) using Large Language Models (LLMs). It compares top‑down, bottom‑up, and hybrid approaches, finding that inductive, data‑driven schemas yield higher inter‑model agreement while remaining valid. Experiments on 100 Israeli parliament transcripts support the reliability of these data‑driven schemas for large‑scale political discourse analysis.
The study compares two fine‑tuning paradigms for detecting hallucinations in large language models: a classifier‑head (“cls”) and a sequence‑to‑sequence (“s2s”) approach. “Cls” consistently outperforms “s2s” because it induces a clear two‑cluster geometry in hidden states, enabling stronger class separation and linear decodability. Analysis of training dynamics reveals a narrow upper‑middle “InfoWindow” where task information emerges; truncating layers above this window removes 31% of the network without loss of accuracy and reduces latency by 1.4×, a strategy that transfers across multiple models and tasks.
The paper proposes an autonomous construction system that combines a large language model (ChatGPT‑5) with a reinforcement learning algorithm to control a mobile crane during retaining wall construction. An enhanced soft actor‑critic model plans crane actions, while robot kinematics ensure these actions are executable. The system also uses a custom embedding model to extract relevant construction data and generates progress updates, safety information, and reports.
MedPAO-Assist is an edge‑deployable, agentic AI framework that structures radiology reports from radiologist dictation and voice commands using only about 3.3 GB of RAM and an average latency of 31 seconds on a standard laptop CPU. It supports four imaging modalities—Chest X‑ray, Chest CT, Abdomen CT, and Brain MRI—and operates entirely with a 0.6 B‑parameter small language model, avoiding cloud APIs or high‑end GPUs. The system follows an Extract‑Ground‑Generate pipeline, performing concept extraction, SNOMED‑CT grounding, structured report generation, and a speech‑driven human‑in‑the‑loop correction phase, enabling a fully “no‑touch” reporting workflow that requires no keyboard input or internet connectivity.
MAMath-Bench is a new benchmark that evaluates multimodal mathematical reasoning in multi‑agent systems (MASs). It contains 1,003 problems across 11 categories, multiple difficulty levels, and modalities such as text, images, and audio, with fine‑grained milestone annotations for correctness, reasoning‑process reliability, and robustness to adversarial perturbations. The study finds that while collaboration can improve intermediate reasoning coverage, it does not consistently boost final‑answer accuracy and can increase redundancy, and current systems remain vulnerable to subtle adversarial errors.
arXiv:2610.27965v1 Announce Type: new Abstract: Can humans and artificial intelligence cooperate such that their combination yields greater accuracy than either agent alone? This result, termed \emph...
arXiv:2610.28522v1 Announce Type: new Abstract: Decision models support routing, guardrails, and output verification in LLM systems. However, item accuracy alone cannot establish whether models make...