Alignment, interpretability, red-teaming, bias and privacy: the research on what these systems do when they misbehave.
Current image-to-video models achieve visual realism and physical plausibility, but reasoning about mental states remains unexplored. Actions are driven by belief, desire, and perception, requiring in...
Recent advances in distillation and flow-map models have enabled deterministic one- or few-step generation for high-quality data, facilitating a new branch of reward alignment approaches that directly...
Runtime error healing lets a crashed program continue by generating code that repairs its live runtime state. Recent work shows that LLMs can generate such healing code, but it is evaluated only on sm...
Legal language models require reward signals that capture not only answer correctness but also the multidimensional quality of legal responses. Existing reward methods, however, often rely on coarse-g...
Coding agents powered by large language models (LLMs) have shown remarkable abilities to autonomously reason about and achieve goals in the digital world. However, bringing this success to the physica...
Demographic fairness gaps in automatic speech recognition are almost always reported from a single training run. We fine-tune the Q-former projector and LoRA adapters of a speech LLM at five audio com...
arXiv:2609.37227v1 Announce Type: new Abstract: Inference-time steering adapts pretrained diffusion and flow-based models to new tasks, e.g., to generate samples from a conditional distribution or sa...
CST‑WM is a causally structured world model designed for embodied visual tracking, where a robot must keep a moving target visible and recover it after occlusion or drift. The model separates state into target‑evidence, robot, and observation branches, removing direct action‑to‑target‑evidence edges to prevent causal hallucination and instead letting actions influence evidence through robot motion and resulting views. Evaluated on EVT‑Bench, Habitat 3.0, and real‑world trials with a Unitree Go2 quadruped, CST‑WM outperforms reactive trackers and other world‑model baselines in following, distance control, safety, and re‑acquisition, achieving 20 of 30 successful real‑world recoveries versus 14 for TrackVLA.
PARSEE-VAD is a training‑free online video anomaly detection framework that separates semantic evidence acquisition from score‑state evolution. It uses Proposition‑Aware Reasoning to extract structured propositional evidence from the current causal window and selectively activates more specific queries, while Streaming Evidence Escalation maps this evidence into a compact score‑domain event state and propagates only the bounded state to maintain temporal continuity. Experiments on four benchmarks show strong performance with reduced specialist computation and sparse score‑state propagation, supporting a current‑first principle for streaming multimodal inference.
Quantum Fidelity Landscape-Guided Prior Calibration for Single-Circuit QGAN Image Generation explores a simpler pixel‑level, end‑to‑end approach to quantum generative adversarial networks (QGANs) that avoids patch‑based decomposition. The authors introduce the Quantum Fidelity Landscape (QFL) as an invariant pairwise‑fidelity structure preserved under shared unitary transformations, and use it to calibrate the quantum prior before adversarial training. Their BasicQGAN framework aligns the prior‑induced QFL with the data‑induced QFL, achieving stable, effective image generation on small‑scale grayscale datasets while requiring fewer qubits and trainable parameters than patch‑based quantum generators.
SCCM (Spherically Consistent Coarse Matching) improves dense feature correspondence on equirectangular projection (ERP) imagery by correcting topological, metric, and area distortions at the coarse matching stage. It augments a chart‑naive cross‑attention matcher with spherical positional attention and area‑aware covisibility priors, raising PCK@1° from 0.229 to 0.275 on Matterport3D while keeping the refiner unchanged. In the RoMa V1 framework, SCCM outperforms ERP‑native EDM and an ERP‑retrained RoMa V1, and it transfers zero‑shot to Stanford2D3D and outdoor Holo360D.
Back2Struct is a system that converts structured images—such as diagrams, charts, and flowcharts—into editable vector graphics code (SVG/XML). By predicting semantically rich, object-level SVG code rather than low-level pixel vectorization, it allows the generated graphics to be imported into tools like PowerPoint for easy editing, restyling, and reuse. The model is trained with supervised fine‑tuning and reward‑based learning that enforces syntactic validity, concise length, and visual fidelity to the input, leading to higher accuracy, editability, and user alignment compared to baselines.
ESMTrack is a fully end‑to‑end self‑supervised RGB‑T tracking framework that eliminates the need for costly modality‑aligned bounding boxes or offline pseudo‑label generation. It learns discriminative, temporally consistent representations using a grounding triplet loss on the initial annotated frame and a cross‑frame temporal triplet loss on unlabeled search frames, with reliable samples selected via forward‑backward consistency. A three‑branch architecture (fusion, RGB, thermal) and a modality decoupling mechanism mitigate modality dominance bias, enabling competitive state‑of‑the‑art performance, strong cross‑dataset generalization, and real‑time inference on five RGB‑T benchmarks.
The paper introduces Zero-Training Fisher Geometry Alignment (ZFGA), a closed‑form test‑time adaptation technique that aligns the Fisher geometry of test‑time feature embeddings with a reference geometry from clean data, improving robustness to covariate shift without altering model parameters. ZFGA is evaluated on CIFAR‑10‑C and ImageNet‑C with ResNet‑50, DINO ViT‑S/16, and CLIP ViT‑B/32, consistently outperforming zero‑shot inference and remaining stable across all three model families. Compared to other training‑free and gradient‑based methods, ZFGA uniquely avoids substantial performance degradation and requires only forward passes and matrix operations at inference time, making it a lightweight, deterministic alternative to optimization‑based test‑time adaptation.
The paper introduces a Lagrangian framework for managing a regenerative commons, framing the problem as a constrained Markov game with a specified depletion budget. It constructs policy sequences from unconstrained solutions, extending time‑average concepts to reset episodes with discounted rewards and terminal costs, and provides theoretical guarantees such as reward‑independent feasibility, cooperative feasibility, and approximate optimality. Experiments on a fishery model using constrained IPPO and MAPPO illustrate how depletion budgets influence stock retention, harvest rewards, and price adaptation.
FineSID introduces a new quantization framework for semantic identifier learning in generative recommendation systems. By replacing the traditional Top‑1 hard assignment with a soft, differentiable approach, it distributes gradient updates across all codewords, leading to balanced codebook optimization and reduced identifier collisions. Experiments on public benchmarks show that FineSID improves codebook utilization and recommendation accuracy without relying on complex initialization strategies.
arXiv:2506.01004v3 Announce Type: replace-cross Abstract: Unlike traditional video editing or inpainting, video semantic mixing fuses a reference concept with a moving target entity to produce a hybr...
The study investigates whether language models tailored to specific cognitive domains better align with corresponding brain systems. By prompting and fine‑tuning large language models into six domain experts—sensory, spatial, numerical, reasoning, social, and abstract—the authors find that each expert’s representations more closely match the brain region associated with its domain than other experts. This domain‑specific alignment holds across multiple base models and fMRI datasets, while overall prediction accuracy remains largely unchanged, indicating that regional alignment can be obscured when summarizing across the brain.
The paper introduces Dually Regularized AIL, a model‑free algorithm for adversarial imitation learning that jointly applies KL policy regularization and a quadratic reward penalty based on expert and learner occupancies. It proves fast convergence rates, achieving a ×O(1/K+1/N) bound on the regularized imitation gap in finite‑horizon MDPs with general function approximation, and establishes the first algorithm to attain ×O(1/ε) sample complexity in both expert demonstrations and online interactions for this regularized objective.
The paper introduces Alignment Forecasting, a method for predicting whether fine‑tuning a language model on a given dataset will increase specific alignment failures such as deception or sycophancy. It presents ALIGNMENTFORECASTBENCH, a benchmark of over 5,000 forecasting questions across many models, datasets, and failure modes, and shows that a simple forecasting scaffold using an LLM’s assessment of dataset bias can outperform baseline forecasters. The authors demonstrate that filtering out high‑risk training examples identified by the forecaster can improve alignment in multiple‑choice evaluations, though benefits in open‑ended conversations remain uncertain.