Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
The paper introduces an Adaptive Duration Model designed to predict phoneme-level duration distributions directly from text, improving alignment in text-to-speech systems. It outperforms existing baseline models in precision and adaptability to varying conditions, enhancing phoneme-level alignment accuracy. The model also strengthens zero‑shot TTS performance by reducing sensitivity to mismatches between prompt and input audio.
The study reports a mixed‑methods pilot evaluation of an AI‑orchestrated research workflow in which a human researcher coordinated multiple LLM‑based agents to perform data extraction, corpus construction, artifact generation, and artifact evaluation. Using the generation and assessment of multiple‑choice questions (MCQs) as a testbed, the team collected 1,071 SAT Math MCQs and employed LLM agents to extract questions from PDFs, retrieve and convert open textbooks into structured representations, align each MCQ with relevant textbook content, generate new MCQs under specified difficulty and cognitive levels, and evaluate both original and generated MCQs using a 24‑criterion quality framework. Across all evaluations, average MCQ quality was high, but criterion‑level analysis and equivalence testing show that generated MCQs are not fully comparable to expert‑vetted baseline questions, with persistent gaps in skill depth, cognitive engagement, difficulty calibration, and metadata alignment. "whyItMatters":"The findings highlight how AI‑orchestrated workflows shift researcher labor toward specification, orchestration, verification, and governance, underscoring emerging AI research operations skills needed for AI‑empowered research pipelines."
The paper introduces CAIA, a framework that enhances EEG-based visual decoding by adaptively blurring visual features to match neural granularity and applying an information-constrained alignment on EEG signals. It incorporates cognitive-guided adaptive blurring, spectral sparsity regularization, and a distribution-aware boundary calibration loss to improve signal-to-noise ratio and alignment robustness. Experiments show CAIA outperforms previous methods in both subject-dependent and subject-independent zero-shot brain-to-image retrieval tasks.
The study presents an unsupervised domain adaptation model that classifies normal versus abnormal retinal development using optical coherence tomography (OCT) scans from multiple devices and centres. Trained on labelled data from one OCT system and unlabelled data from another, the ResNet50‑based Domain Adversarial Neural Network achieved accuracies between 87.75% and 96.06% across six source‑to‑target pairings and maintained performance on an unseen Heidelberg Spectralis dataset. The approach demonstrates that OCT classification can be transferred across devices without requiring labelled target data, potentially enabling more consistent assessment of retinal development and large‑scale analysis of existing imaging datasets.
The study examines how temporal resampling (binning) of irregular clinical data affects the safety and performance of offline reinforcement learning (ORL) models in healthcare. By analyzing MIMIC‑IV insulin infusion records, the authors show that binning can reverse causal order and obscure critical events. In simulated trials with virtual type‑1 diabetes patients, ORL models trained on binned data performed up to 60% worse than those trained on unprocessed data, yet retrospective evaluation on resampled data overestimated their returns.
The paper introduces 3DTCR, a vortex‑following generative framework that reconstructs high‑resolution, three‑dimensional tropical cyclone (TC) structures from coarse global AI forecasts. By training on a 3‑km, TC‑centered dataset generated with the WRF model, 3DTCR learns a coarse‑to‑fine mapping and adapts it to forecast inputs, producing detailed TC fields in minutes on a single GPU. Evaluations on 2024 western North Pacific forecasts show significant reductions in RMSE and mean bias error for maximum 10‑m wind speed compared to both FuXi and ECMWF‑HRES, while capturing key wind gradients and vortex asymmetries.
SUPRA-Tex is a new method for single‑image 3D generation that dresses a 3D mesh with relightable materials while preserving the visual details of a reference image. It decouples visible‑material estimation, surface alignment, and observation‑guided 3D completion, allowing a 3D generator trained on limited data to focus on preserving visible evidence and completing unobserved surfaces. The approach uses an MLLM‑based image editor to predict lighting‑separated PBR maps, projects them onto the visible surface, encodes them with a pretrained material VAE, and employs an interleaved flow transformer to generate complete material latents, achieving state‑of‑the‑art performance on geometry‑conditioned texturing and end‑to‑end asset generation.
BD-HazardVLM is a diagnostic benchmark of 500 Bangladesh road-scene images designed to probe vision‑language models’ defensive‑driving reasoning in weakly structured Global South traffic. The benchmark includes annotations for hazard presence, category, risk timing, severity, defensive actions, visible evidence, and bias traps, and evaluates models such as Qwen2.5‑VL‑7B, InternVL3‑8B, and MiniCPM‑V‑4.6. Results show high binary hazard‑presence accuracy but poor fine‑grained reasoning, with category accuracy below 40 %, risk‑timing accuracy under 13 %, and defensive‑action matching under 2 %. whyItMatters":"The study demonstrates that coarse hazard recognition does not guarantee reliable fine‑grained defensive‑driving reasoning, highlighting the need for geographically diverse safety evaluation of vision‑language models."
The paper introduces Recursive Reward Improvement (RRI), a method that enhances recursive self‑improvement loops by using an adversary to expose and fix weaknesses in model‑generated rubrics for open‑ended tasks. In RRI, an adversary crafts a rubric‑gaming answer, an honest answer is generated without rubric exposure, and up to two repair steps adjust the rubric so the honest answer scores higher. Experiments on HealthBench Professional show that RRI boosts adjusted scores by up to 12.4 points on a 27B model and yields significant gains on difficult and writing tasks, while also improving performance on unseen benchmarks PRBench Hard and ProfBench.
The paper proposes Local Appearance Purification (LAP), a method that refines the alignment target for diffusion models by removing locally predictable content and focusing on the residual. LAP improves generation quality across multiple pretrained encoders, notably reducing FID from 53.1 to 42.3 on ImageNet-256 with MAE. The authors provide a theoretical framework linking alignment effectiveness to the information emphasized by the target and introduce a constant‑readout test to predict purification benefits.
The paper derives closed‑form bounds on the performance of a drone‑mesh interception system before any learning policy is applied. It shows that, due to a separation constraint, a mesh of agents can cover at most 90.69 % of a plane (74.05 % of a volume) and that optimal hexagonal packing achieves this limit. The authors validate these theoretical ceilings with Monte‑Carlo simulations and argue that such bounds should be reported alongside any learning‑based results.
The study investigates how a large language model (LLaMA2‑7B) aligns with human brain activity during naturalistic language comprehension by using SCRIBE, an interpretable semantic representation of 65 neurocognitively grounded properties. Variance partitioning on fMRI data from seven participants reveals that shared variance between LLaMA2‑7B and SCRIBE dominates across the brain, yet both representations also contribute unique variance. Further analysis shows that concrete semantic properties correlate more strongly with shared variance, while abstract properties do not show increased unique variance, and ablation experiments demonstrate that multiple semantic dimensions cumulatively influence both shared and unique brain-predictive variance.
Data World proposes a multi‑agent engine that generates synthetic data environments, tasks, and evaluation rubrics by first specifying the underlying world rather than curating real files. The system produces thousands of training tasks across diverse domains and enables independent audits of target recoverability, while human curation ensures realistic question selection. Benchmark results show that models trained on these synthesized tasks still face significant challenges, yet RL‑zero training on the synthetic data yields measurable performance gains over baseline models.
PROBE is an algorithm that trains language‑model auditors to probe deployed tool‑using agents for policy violations through realistic, multi‑turn, tool‑mediated dialogue. By framing auditing as a two‑player Markov game and using a composite reward that balances violation discovery, behavioral plausibility, conversation progress, and policy diversity, PROBE outperforms zero‑shot adversaries in uncovering broader and higher‑yield policy violations. The study compares GRPO, DPO, and GFlowNet optimizations, demonstrating that multi‑turn behavioral auditing is a distinct and necessary axis of agent evaluation, complementing existing capability and safety benchmarks.
X‑RIC is a new framework that embeds agentic AI into O‑RAN architectures to enhance the RAN Intelligent Controller (RIC). It combines reactive reinforcement learning with strategic reasoning via automated behavioral cloning, a telecom language model, and closed‑loop co‑adaptation, and introduces BARX, a benchmark for evaluating agentic control. In simulations and testbed deployments, X‑RIC cuts operational costs by 89 %, meets SLA compliance, and converges 75 % faster with fewer training steps and lower variance.
The paper introduces Fractional Power-of-Two (FPoT) quantization, a base‑√2 weight grid that enables 4‑bit weight quantization for large language models while keeping a fully integer shift‑add datapath. FPoT combines a fractional PoT grid with uniform integer activations, a calibration‑time grid‑alignment refinement, and a lightweight 1‑adder √2 approximation to reduce gate count compared to a 4‑bit multiplier. On Llama‑3‑8B, FPoT boosts zero‑shot commonsense reasoning by 8.72% over the standard PoT baseline and stays within 0.4% of uniform integer performance, with benefits extending across seven models and three bit‑widths.
The paper introduces the Redistribution Sensitivity Score (RSS) to detect consistent changes in sparse feature activations of Sparse Autoencoders (SAEs) when prompts are transformed. RSS identifies large, directionally consistent feature responses and evaluates their impact on downstream readouts across 27 pretrained SAEs from eight language models. Experiments show that combining RSS support erosion with the clean margin predicts recurrent safety readout failures with high AUROC, demonstrating RSS’s utility in quantifying representation changes without conflating them with failure.
The paper investigates when quantum machine learning can enhance predictive performance by fusing classical and quantum representations in multi‑view learning. It introduces a framework that preserves full classical inputs while adding simulated quantum features, and uses concatenation and attention to combine the views. Experiments on synthetic angular‑polynomial data and real datasets (Wine, Wisconsin Diagnostic Breast Cancer, Fashion‑MNIST, Steel Plates Faults) show that fusion improves predictions for targets that depend on complementary classical and quantum relationships, while classical‑only baselines perform best on other tasks.
The paper introduces query‑guided grounding methods for video understanding, covering object‑level, relationship‑level, and moment‑level tasks. For object grounding, QDETRv and TOViT enable one‑shot and few‑shot localization of unseen objects using query‑context fusion, memory modules, and language‑aligned vision encoders. Relationship grounding is addressed with a few‑shot referring framework that optimizes a T‑partite random field, while moment grounding is tackled by the Moment Alignment Transformer (MATR), which employs dynamic time warping and self‑supervised pretraining for video‑to‑video and text‑to‑video retrieval.
arXiv:2610. 21964v1 Announce Type: cross Abstract: Neural networks often learn low-frequency components before high-frequency ones --- a phenomenon called spectral bias or termed F-principle.