Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
arXiv:2605. 21341v2 Announce Type: replace Abstract: We study functional bilevel optimization with conditional moments, where plug-in hypergradients can inherit first-order errors from nonparametrically learned conditional expectations.
GTASA: Ground Truth Annotations for Spatiotemporal Analysis, Evaluation and Training of Video Models presents GEST-Engine, an open‑source system that deterministically renders multi‑actor video scenarios while recording exhaustive ground truth—including 3D entity and camera states, pairwise spatial relations, event‑to‑frame mappings, instance segmentation, and long descriptions—at no extra annotation cost. The authors release a 938‑video dataset, GTASA, which offers the densest spatial‑relation coverage of any video dataset, with an entity‑pair relation graph at every frame that is roughly 84× denser than current state‑of‑the‑art collections. They demonstrate the dataset’s value by showing that pretraining on GTASA improves VideoLLaMA3 captioning, and by using it to probe frozen video encoders on 11 spatio‑temporal tasks, revealing that relational understanding remains near chance. whyItMatters":"The dataset provides a uniquely dense, exact 3D ground truth for spatio‑temporal video analysis, enabling measurable progress in training and evaluating video models on relational reasoning tasks."
The paper proposes a comprehensive data generation strategy to improve deep learning performance in classifying olive tree diseases. It combines traditional augmentation, CutMix, MixUp, and DCGAN-generated synthetic images to address data scarcity and class imbalance. Experiments with EfficientNetV2B0 show significant accuracy gains across three distribution scenarios, underscoring the importance of data diversity and balance.
SCOPE is a multi‑agent system that profiles news media outlets for bias, factuality, and credibility by coordinating agents for article acquisition, content analysis, evidence search, scoring, and report generation. It follows an MBFC‑style methodology and provides inspectable intermediate evidence and outputs via a structured interface. On a benchmark of 200 outlets, SCOPE outperforms LLM‑only, article‑only, search‑only, and media‑background‑check baselines, achieving the lowest bias and factuality MAE, perfect failed‑fact‑check detection, and the lowest factual error rate.
The paper introduces a unified framework for self‑driving laboratories (SDLs) that integrates experimental planning, automated execution, and measurement feedback. It comprises four components: LAM, a measurement‑conditioned policy that generates robot trajectories and typed instrument API calls; OpenSDL, an execution substrate that maps operations to heterogeneous devices and records causal events; Safe‑SDL, a runtime gate that performs pre‑dispatch checks and monitors execution with signal temporal logic; and L3‑Bench, a benchmark protocol evaluating physical outcomes, timing, reproducibility, safety, and fault recovery. Together, these elements aim to address coordination and traceability challenges in SDL workflows.
The paper introduces ReSeed, a lightweight post‑distillation technique that restores lost diversity in diffusion model distillation by adding a seed‑response bias adapter only to the first denoising step. By aligning the first step’s velocity with the teacher’s, ReSeed recovers most of the output diversity without altering later steps or adding target images. Experiments on SD3‑Medium, SD3.5‑Medium, FLUX.1‑dev, and Wan2.1‑T2V‑1.3B show a 48.3% increase in DINO diversity with negligible extra parameters and training data.
The paper introduces HausaSafeLM, a safety preference dataset for Hausa comprising 10,000 prompts (3,000 human‑corrected and 7,000 machine‑translated) across six harm categories. Using Qwen2.5‑7B‑Instruct, the authors compare supervised fine‑tuning (SFT) and SFT followed by Direct Preference Optimization (DPO) on 500 harmful and 200 benign Hausa prompts, finding that SFT improves refusal fluency and precision—especially with larger, noisier data—while DPO sharpens precision at the cost of linguistic quality. No single approach simultaneously maximizes safety, helpfulness, and fluency, and the study shows that refusal‑rate metrics alone are unreliable indicators of safety in low‑resource languages.
The paper introduces SCHEMA, a topology‑aware evaluation framework designed to assess hallucinations in large language model agents used for scientific research. SCHEMA builds scientific concept graphs from benchmark seeds and literature evidence, then generates graph‑grounded tasks such as claim verification, multi‑hop reasoning, open‑ended explanation, and experimental code generation. It evaluates agents using a trajectory hallucination pipeline that assigns a topology‑weighted severity score and a multi‑agent counterfactual attribution module that identifies the causal mechanisms behind failures. The study finds that hallucinations tend to cluster around highly connected knowledge hubs and that a model’s final accuracy can be misleading, as correct answers may arise from structurally flawed reasoning. "whyItMatters":"The findings highlight that terminal accuracy alone is insufficient for high‑stakes scientific applications, underscoring the need for mechanism‑level evaluation grounded in knowledge topology."
The paper introduces LAnR (Latent Abstraction for Retrieval-Augmented Generation), a unified framework where a single large language model jointly encodes, retrieves, and generates within its latent space. Instead of generating natural language queries, LAnR produces dense retrieval vectors from a designated [PRED] token and matches them against encoded document representations from the same model. An adaptive MLP control head decides when sufficient evidence has been retrieved, eliminating the need for a separate retriever and explicit token-level stopping reasoning. Experiments on six QA benchmarks show that LAnR outperforms existing RAG methods while improving inference efficiency by reducing retrieval calls and tightening model integration.
The paper introduces a hybrid framework that combines text analysis of aviation accident reports with physics-based models to extract and quantify causal chains of failure. Using a large language model, the authors build a causation knowledge graph from 56,202 reports and achieve 90.8% recovery of investigator-flagged causes on a 2,000-report test set. The framework calibrates occurrence models for physical factors and employs a leaky noisy-OR to estimate failure mode probabilities, demonstrating interpretability gains and modest predictive improvements for specific scenarios such as carburetor icing.
The study presents a multimodal machine‑learning framework that combines genomic sequences, epidemiological indicators, wastewater surveillance, and clinical rash images to detect measles and infer transmission pathways. In a three‑stage pipeline, the authors achieved high accuracy in classifying transmission linkage (AUC 0.924, F1 0.912), moderate success in predicting outbreak risk (AUC 0.707), and near‑perfect recall for aligning genomic and imaging embeddings (Recall@1 0.985). The results highlight the complementary strengths of genomic data for linkage prediction and wastewater data for outbreak‑risk forecasting.
The paper introduces Latent Visual Contrastive Reasoning, a branch-aware latent reasoning framework that generates multiple question-conditioned latent visual branches as soft hypotheses over image tokens. It trains these branches using bounding‑box evidence annotations and a soft contrastive branch‑evidence alignment objective, encouraging evidence‑relevant branches to align with annotated visual regions while contrasting against negative evidence. Experiments on perception‑intensive visual reasoning benchmarks demonstrate that the method improves performance over the corresponding backbone and single‑path latent reasoning baselines at both 3B and 7B scales.
arXiv:2610. 36097v1 Announce Type: cross Abstract: We study full-information online linear optimisation on the unit $\ell_p$ ball in $\mathbb{R}^d$, for fixed $2\le p<\infty$ and losses of dual norm at most one.
The paper introduces a framework to evaluate AlphaEarth embeddings for crop classification across spatial, temporal, and cross-regional dimensions. It compares a 64‑dimensional embedding space with traditional Sentinel‑2, NDVI, and harmonic features using field data from the US and India, finding that embeddings offer superior distribution‑aware separability and capture multi‑year crop rotation dynamics, while traditional features transfer better across continents. The study highlights a trade‑off between representational richness and cross‑region transferability, suggesting a hybrid approach for large‑scale crop classification.
The paper introduces Fog-based RetailLightNet (F RLN), an ultra‑lightweight deep learning model designed for real‑time video analytics in smart retail environments. By deploying the model on edge devices such as the NVIDIA Jetson Nano and using motion‑gated frame selection, the system reduces latency and computational load while maintaining high precision (0.988), recall (0.976), and F1‑score (0.976). The approach leverages publicly available retail datasets (SKU110K and RPC) and depth‑wise separable convolutions to enable efficient shelf monitoring and queue assessment.
The paper introduces MoME, a switchable unlearning framework for re-identification models that allows a single backbone to operate in either Full mode (standard inference) or Core mode (unlearned inference). MoME partitions the backbone into core parameters that retain retrieval capability and zero‑initialized parasitic parameters that store full‑data knowledge, training them with gradient isolation so the core never receives forget‑set gradients. The authors demonstrate that Core mode satisfies deletion requests with a simple flag flip, prove its independence from the forget set, and validate its effectiveness on Market‑1501, VeRi‑776, and RegDB using ResNet and CLIP backbones.
FastBEVFusion is a lightweight fusion system for real‑time 3D object detection that combines LiDAR and camera data without depth estimation. It projects modality‑specific features into semi‑aligned BEV spaces and fuses them with a dynamic cross‑attention module, achieving over 50 FPS on the nuScenes dataset. The method also demonstrates strong robustness to calibration errors.
The paper introduces API, a framework that adapts pretrained interpretable features for generative recommendation. By reusing sparse autoencoder‑derived feature indices as a reusable semantic ID vocabulary, API avoids additional vocabulary learning and preserves item distinctions even with limited token budgets. The Semantic Prediction Adapter learns user preferences from interaction histories, combining feature‑index and strength predictions to score candidates, while also providing explainable retrieval through feature contributions linked to pretrained semantic annotations.
ActiveAlign is a framework that uses active perception to reduce the Real2Sim2Real gap in robot manipulation. It introduces two modules: ActiveAlign-Waving, which performs hand‑eye calibration through active hand motions and differentiable rendering, and ActiveAlign-Touching, which refines object alignment with tactile feedback to achieve millimeter‑level precision. The approach is demonstrated by curating simulation environments for uncalibrated datasets and enabling a Unitree G1 robot to manipulate millimeter‑scale objects such as buttons and sliders.
ALADDIN is a one‑step face‑swapping method that distills a pre‑trained multi‑step diffusion model into a fast generator trained adversarially with identity and target‑preservation constraints. It offers a controllable trade‑off between identity transfer and attribute preservation via triplet pseudo‑label supervision, and it outperforms other one‑step methods in realism and target preservation while matching the best multi‑step model with far fewer evaluations and training pairs. The authors also uncover a bias in standard evaluation protocols and re‑evaluate baselines with an independent identity encoder.