Retrieval pipelines, vector search, chunking and reranking: how models are grounded in a corpus instead of their weights.
arXiv:2609.18898v1 Announce Type: new Abstract: Training-free weighted aggregation is widely used to lift 2D semantic features onto 3D Gaussians for open-vocabulary scene understanding, yet its theor...
The paper audits vision‑language models (CLIP) for gender bias using 1,500 artworks from the Metropolitan Museum of Art, focusing on zero‑shot logit differences for prompts like "masterpiece," "quality," and "influence." Unadjusted results show no significant gender effect and statistical equivalence across models, but high residual variance suggests that global zero‑shot metrics are largely noise‑dominated and may miss fine‑grained biases. The study underscores the need for multivariate confound control, equivalence testing, and provenance auditing when evaluating AI fairness in cultural heritage data.
The paper investigates how to efficiently repair stale key-value (KV) caches in retrieval‑augmented generation systems after document edits. It proposes a budgeted in‑place recomputation approach and evaluates training‑free position‑selection policies on a factual RAG benchmark. Across three model families, a contiguous edit‑local window consistently recovers most of the post‑edit answer quality while being 13–21 times faster than a full re‑prefill, though its effectiveness diminishes when answer‑bearing text moves downstream.
arXiv:2601.11262v2 Announce Type: replace-cross Abstract: Music Cover Retrieval, also known as Version Identification, aims to recognize distinct renditions of the same underlying musical work, a tas...
ReDIL-GNN is a framework for resynthesis domain‑incremental learning in circuit graph neural networks. It adapts a fixed prediction or representation head as new synthesis styles appear and evaluates retention across all previously seen domains. The method introduces the Resynthesis Adaptability Index (RAI), a pre‑adaptation score that combines adaptation need, source‑equivalence recoverability, structural coverage, and update compatibility to decide whether to adapt, reuse, or defer updates.
CALIPER is a model‑free RGB‑D framework that performs fine‑grained recognition of visually similar industrial parts by combining support‑based appearance matching with metric size evidence. Each class is onboarded from a single turntable RGB‑D video and a few labeled real images, enabling 3D reconstruction for appearance support and depth‑aligned size profiling. At inference, a YOLOv8n‑seg model localizes parts, a frozen DINOv2 backbone with an episodically trained embedding head matches support, and margin‑conditioned metric fusion selectively uses size evidence for ambiguous cases, achieving high accuracy on 18 parts and robust enrollment of unseen screws without retraining.
NeMo Data Designer (NDD) is an open‑source framework for generating multimodal synthetic data. It uses a declarative configuration format that lets users define dataset columns—text, code, structured outputs, images, embeddings, and statistical samplers—to steer diversity. The system supports a preview‑and‑revision workflow, dependency resolution, and retry logic, and can be extended via plugins. Case studies demonstrate its use for structured, agentic, multimodal, and domain‑specialized tasks, including datasets for Nemotron model development and enterprise deployments.
The paper introduces REASONS, a benchmark comprising 12,723 sentence-level citation instances across 12 arXiv subject categories, to evaluate scientific citation attribution by large language models. It proposes a dual-metric framework—Abstention Rate (AR) and Hallucination Rate (HR)—to assess the trade-off between reliability and responsiveness. Experiments on proprietary and open-source LLMs under various prompting and retrieval settings show that advanced Retrieval-Augmented Generation (RAG) reduces hallucinations but may increase abstention, while retrieval-augmented variants often maintain near-zero abstention. Human evaluation reveals a high ratio of factual hallucinations to acceptable paraphrases, underscoring the need for systems that can appropriately abstain under uncertainty.
arXiv:2605.19309v4 Announce Type: replace Abstract: Document Layout Analysis (DLA) pipelines provide structured page representations for retrieval-augmented generation, long-document question answeri...
The paper evaluates three approaches for emotion recognition in conversation— a low‑cost stacked ensemble, an off‑the‑shelf LLM prompt, and a confidence‑gated hybrid that escalates only uncertain ensemble predictions to the LLM. Across three datasets (IEMOCAP, MELD, CMU‑MOSI), the hybrid consistently outperforms each pure system, achieving higher weighted F1 scores while routing most traffic through the inexpensive ensemble. This results in significant cost savings (≈$10‑85 per million utterances) and provides an interpretable escalation signal tied to emotion or sentiment shifts.
MINT (Multimodal Imaging-to-Speech Knowledge Transfer) is a three-stage framework that transfers MRI-derived biomarkers to speech representations for early Alzheimer’s screening. An MRI teacher creates a compact embedding space for CN‑versus‑MCI classification, and a residual projection head aligns speech features to this space using a geometric loss, allowing imaging‑free inference. Experiments on ADNI‑4 show that aligned speech matches speech baselines, while multimodal fusion outperforms MRI alone, and ablations highlight dropout regularization and self‑supervised pretraining as key design choices.
TACTICS is a method for selecting evaluation samples in machine translation that explicitly optimizes for coverage of rare linguistic categories, document-level coherence, and distributional fidelity to the full corpus. It builds a hierarchical taxonomy from a locale style guide, classifies segments, and chooses a fixed-budget subset that better represents the full range of phenomena a system must handle. Compared to random, lexical, or embedding-based selection, TACTICS improves coverage of rare categories and yields more accurate system rankings with fewer segments.
The paper argues that computational text‑based ideal point estimation (CT‑IPE) methods should be viewed as configurable measurement pipelines rather than fixed estimators. It presents a large‑scale comparative experiment involving 17 CT‑IPE algorithms, 5,537 runs, and about 4.25 million left‑right position estimates, and describes shared infrastructure that enables joint execution of these heterogeneous methods. Sensitivity analyses reveal that most algorithms exhibit low hyperparameter sensitivity (ICC < .10), with any remaining sensitivity concentrated in a few key researcher choices such as the language or embedding model, seed keyword lists, and number of topics.
The paper investigates hyperbolic graph representation learning applied to biomedical knowledge graphs for Mendelian-disease differential diagnosis. It shows that hyperbolic embeddings outperform Euclidean baselines on isolated ontology subgraphs while requiring fewer dimensions. In a link-prediction task, hyperbolic models rank candidate diseases for patients, indicating they can leverage hierarchical structure in heterogeneous patient-level graphs.
This paper proposes a Retrieval-Augmented Generation (RAG) framework for scientific image quality assessment, designed to simultaneously address both the understanding track (SIQA-U) and the scoring t...
Conditioning a video generator on multiple images requires preserving appearance while associating each reference with its intended role. We present MSR (Multiple Subject Reference), a slot-aware cond...
Cross-view object geo-localization (CVOGL) locates a target in satellite imagery using drone or street-view queries. Existing methods train separate detectors for each viewpoint, leading to parameter...
Large language model (LLM) agents augmented by tools can automate complex, multi-step tasks, such as web navigation, code generation, and workflow orchestration, by acting on external systems through...
The paper investigates Mandarin Chinese reduplicative constructions that repeat two-character base words or their constituents, such as expressions meaning ‘in good health’ or ‘discuss a bit’. Using Tencent word embeddings, the study demonstrates that distributional semantics can recover known semantic and grammatical properties of these reduplications, revealing clear semantic and pragmatic differentiation between the two patterns. Procrustes analysis shows that the overall organization of the base-word space is largely preserved in the reduplication space, with local mismatches indicating discourse-pragmatic reorganization.
The study evaluates the use of frozen geospatial foundation embeddings (AlphaEarth) for mapping cultivated versus non‑cultivated land in Maine. Using 192 spatially separated patches and USDA Cropland Data Layer labels, a lightweight classifier achieved 93.7% overall accuracy without fine‑tuning, and a nearest‑class‑centroid rule reached 90.2%. A balanced sample of 60,000 labeled pixels was nearly as effective as the full 8.6 million‑pixel pool, and classifiers trained in one year remained accurate across 2018‑2023. In a blind human validation of 385 points, the AlphaEarth‑plus‑random‑forest map matched 95.3% of the consensus, outperforming the CDL reference (91.7%).