Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2610.28264v1 Announce Type: cross Abstract: Oddly Satisfying Videos (OSVs) can evoke strong viewer responses through repetitive motion, visual tactility, and ASMR-like sounds despite containing...
arXiv:2610.27958v1 Announce Type: new Abstract: Hyperspectral images provide a rich representation of the underlying spectrum for each pixel, allowing for a pixel-wise classification/segmentation int...
arXiv:2610.28007v1 Announce Type: new Abstract: Pathology foundation models (PFMs) support a wide range of histopathology tasks on whole-slide images through multiple instance learning (MIL). Despite...
arXiv:2610.28346v1 Announce Type: new Abstract: Autonomous video editing from open-world resources remains a challenging problem in multimedia content creation. Retrieval-based methods are constraine...
arXiv:2601.07823v2 Announce Type: cross Abstract: Video world models have emerged as promising candidates for high-fidelity world models, offering the potential to synthesize high-quality videos capt...
MAS-Guard is a unified framework designed to safeguard both single- and multi-agent large language model (LLM) systems. It introduces a local Agent Guardrail (A-Guard) layer that uses task-specific experts to evaluate agentic events, and a system-level Control Plane that applies deterministic structural checks and semantic assessments to routes, capabilities, and message content. The framework enforces interventions—allow, reflection, confirmation, and block—across benchmarks, reducing attack success rates by 71.5‑86.3% while maintaining utility on benign tasks.
SciDiagBench is a new benchmark designed to evaluate scientific diagram generation in a way that mirrors human peer‑review practices. It provides a dataset of 165 diagrams from 10 domains, systematically classified into 11 distinct types such as graphical abstracts and technical roadmaps. The benchmark also introduces a preference‑guided calibration method that incorporates human preferences into type‑specific evaluation criteria, aiming to improve the reliability of automated assessments.
The paper investigates whether pretrained language models can act as lightweight auxiliary supervision for supervised vision training, without requiring paired image-text data or complex objectives. Two forms of language guidance—Explicit Language Guidance (ExLG) and Implicit Language Guidance (ImLG)—are evaluated across six generalization regimes, including in‑distribution, out‑of‑distribution, shortcut resistance, texture/shape bias, adversarial robustness, and continual learning. Results show that ExLG improves in‑distribution, low‑data performance, and continual learning retention, while ImLG is more effective in shortcut‑sensitive and adversarial settings, and both add minimal parameters and training overhead.
CovNav is a training‑free framework that uses Conformal Prediction to decide when an embodied agent should switch from fast, low‑latency LLM reasoning to slower, more accurate chain‑of‑thought reasoning during object‑goal navigation. By treating each spatial query as a binary decision and calibrating the fast LLM’s token probabilities, CovNav escalates only ambiguous queries to the slow mode, achieving higher success rates on Gibson and HM3D while invoking slow reasoning for just 12.4% of queries. Real‑robot tests on a TurtleBot 4 confirm the method’s practical benefits.
The paper investigates the use of Johnson‑Lindenstrauss (JL) transforms in Byzantine‑robust distributed optimization. By leveraging JL’s ability to preserve inner products and distances, worker messages are weighted based on their alignment with a trusted reference gradient, reducing communication costs while maintaining robustness. Experimental results support the theoretical claims.
arXiv:2610. 28314v1 Announce Type: cross Abstract: Prior-data fitted networks (PFNs) have recently gained popularity as an end-to-end approach to amortized Bayesian inference.
PACE is a trajectory inference framework that reconstructs continuous transport dynamics from destructive single‑cell time‑course snapshots. It uses a state‑ and time‑dependent anisotropic Riemannian metric to guide transport along tangent directions, alternates between refining cross‑time couplings and fitting neural bridges, and distills the learned dynamics into a global velocity field. Across multiple datasets, PACE outperforms existing methods, reducing key distance metrics by an average of 23.7% and improving RNA‑velocity alignment by 15.4%.
The paper introduces LLM‑Attack Suite, a unified framework for evaluating privacy risks in split and distributed large language models. It records attacker choices—such as message types, model weights, datasets, and recovery metrics—to enable consistent comparison and extension of privacy evaluations. Experiments on GPT‑2 demonstrate that providing gradients during training significantly boosts token and document recovery rates, illustrating how attacker access levels influence privacy outcomes.
The paper introduces a multimodal framework that jointly processes full‑video transcripts and speech waveforms to recognize ambivalence and hesitancy (A/H) at the video level. It uses pre‑trained multilingual BERT for text and wav2vec 2.0 for audio, projecting both into a shared space and applying two stacked bidirectional cross‑attention blocks to fuse linguistic and acoustic cues without explicit temporal alignment. The resulting representations are aggregated and combined via concatenation, element‑wise products, and absolute differences, achieving 72.19 % accuracy and 77.95 % A/H F1‑score on the Behavioural Ambivalence/Hesitancy dataset.
The paper "Fluent but Fallible: Benchmarking Hallucination and Factual Reliability in Large Language Models" evaluates ten advanced LLMs on a benchmark of 70 academic prompts across four complexity levels. Using five metrics—Accuracy, Precision, Recall, Response Deviation Index, and Hallucination Index—reviewers found that Perplexity and DeepSeek V3 achieved the highest overall accuracy, Gemini balanced recall and performance, LLaMA 3 resisted hallucinations, and Falcon had perfect precision but low recall and high deviation. The study highlights a trade‑off between response completeness and factual reliability, especially as prompt complexity increases.
The paper introduces a drill‑down evaluation framework that breaks down Text‑to‑SQL queries into simpler executable sub‑queries and natural‑language questions, enabling analysis of hallucination patterns along the generation path. It distinguishes recurrent hallucinations—those appearing in both early and final steps—from emergent hallucinations that only surface at the end. Experiments on five large language models and an expanded BIRD‑mini benchmark reveal varying recurrence and emergence probabilities across categories, with conversational history generally increasing recurrence and decreasing emergence.
The paper investigates how large audio‑language models (LALMs) integrate audio and text representations. By identifying cross‑modal neurons that encode both modalities, the authors demonstrate that audio and text representations align within a shared subspace across three LALMs and four acoustic attribute types. Causal interventions on these neurons confirm their necessity for prediction and show they can be manipulated to steer model outputs.
STEAM is a self‑training framework for unsupervised cross‑view geo‑localization that eliminates the need for image‑generation warm‑ups or clustering. It introduces a Stable Spatial‑Aware Module for bounded nonlinear spatial weighting, Elastic Matching with Bidirectional Top‑K Soft Matching and Dynamic Threshold Filtering, and Adaptive Purification that updates pseudo‑labels based on confidence, age, and expiration. On the University‑1652 dataset, STEAM outperforms other unsupervised methods in both retrieval directions and shows competitive results on SUES‑200 and DenseUAV.
CARI4D is a category‑agnostic method that reconstructs spatially and temporally consistent 4D human‑object interactions from monocular RGB videos at metric scale. It introduces a pose hypothesis selection algorithm that integrates predictions from foundation models, refines them via a learned render‑and‑compare paradigm, and reasons about contacts to satisfy physical constraints. Experiments demonstrate a 38% improvement on in‑distribution data and 36% on unseen data compared to prior work, and the model generalizes to unseen categories, enabling zero‑shot application to internet videos.
The paper investigates Fréchet means in orbit spaces arising from finite orthogonal group actions, a problem known to be NP-hard. It introduces spectral and semidefinite relaxations inspired by group synchronization, and provides a semidefinite certificate that can certify global optimality even when pairwise alignment fails. The authors demonstrate the effectiveness of their methods on Gaussian data with sign ambiguity and through numerical experiments on districting plan ensembles and time‑series clustering.