Retrieval pipelines, vector search, chunking and reranking: how models are grounded in a corpus instead of their weights.
arXiv:2606.11350v2 Announce Type: replace Abstract: Retrieval-augmented generation (RAG) degrades when applied to large, heterogeneous document collections. In RAG systems, dense similarity becomes l...
arXiv:2610.15158v2 Announce Type: replace Abstract: We study neuro-symbolic natural language inference (NLI) with independently translated propositional representations of the premise and hypothesis....
arXiv:2610.20158v1 Announce Type: new Abstract: Self-supervised joint embedding learning aligns representations of augmented views while regularizing their distribution to prevent collapse. Regulariz...
arXiv:2610.20240v1 Announce Type: new Abstract: A multiplayer world model must predict how the independent actions of several players change a shared environment, including changes outside a player's...
arXiv:2604.03414v2 Announce Type: replace Abstract: Video Large Language Models (Video LLMs) achieve strong performance on video understanding tasks but suffer from high inference costs due to the la...
TopicMerge is a multi-stage pipeline that combines efficient embedding-based clustering with Large Language Model (LLM) guidance to produce high-quality topic models. The method first over-segments a corpus into fine-grained micro-clusters, then iteratively merges them using an LLM that evaluates candidate pairs based on local context. Experiments on Bills and Wiki datasets demonstrate state‑of‑the‑art clustering performance while cutting LLM calls by more than 92% compared to document‑level approaches.
The paper investigates how large language model (LLM) workflows and agents make binary decisions—such as allowing or blocking an input—when human deferral is not possible. It proposes using System One decision scores that default to a conservative fallback in uncertain cases and control false omission risk within the subset of inputs that receive the utility‑maximizing action. Experiments on six datasets spanning moderation, safety, and reranking show that naive threshold baselines often exceed target risk, whereas a Learn‑Then‑Test approach achieves only 7.8% risk, demonstrating the limitations of probability‑like interfaces and the benefits of risk‑budgeted selection.
The paper introduces Contrastive Uncertainty Learning (CUL), a supervised framework that leverages voxel‑wise predictive uncertainty to guide representation learning for multi‑structure segmentation in lower‑pelvic MRI. Using Monte Carlo Dropout, CUL identifies uncertain voxels and shapes the embedding space so that low‑uncertainty, correctly predicted features become stable while high‑uncertainty or misclassified features are repelled, preventing feature collapse at challenging boundaries. Evaluated on 589 prostate MRI scans with a ViT‑based UperNet decoder, CUL achieves comparable Dice scores to a baseline while significantly reducing surface distance for most structures, and ablation studies confirm the necessity of both uncertainty‑guided sampling and the contrastive objective.
The paper introduces COUNT, a method that builds n‑gram embeddings without adding learned parameters by storing precomputed n‑gram counts on host memory. By weighting the language model’s own unembeddings with corpus frequencies, COUNT can be added as an extra input embedding during pretraining, improving validation loss equivalent to a 57% increase in compute for a 1B‑parameter model. Additionally, COUNT enables efficient GPU memory usage and allows domain‑specific n‑gram tables to be swapped in at inference time for training‑free adaptation.
The paper introduces the Constitutive Parameterized Deep Energy Method (CPDEM), a physics‑driven neural network that embeds stochastic material parameters directly into the strain energy density functional. By training via expected energy minimization over the parameter domain, CPDEM learns a continuous solution manifold, enabling zero‑shot, real‑time inference of displacement fields for any material parameters without retraining or dataset generation. The method is validated on linear elasticity, finite‑strain hyperelasticity, and highly nonlinear contact mechanics, marking the first purely physics‑driven deep learning approach that efficiently handles continuous multi‑parameter variations in solid mechanics.
The paper presents a hybrid quantum‑classical pipeline that uses neutral‑atom reservoir computing for medical image classification, specifically binary polyp detection. It incorporates a guided auto‑encoder to reduce dimensionality and a differentiable surrogate model to bypass the gradient barrier of quantum measurements, enabling end‑to‑end backpropagation. The resulting compact latent representations are encoded as pulse detuning parameters in a Rydberg Hamiltonian, producing quantum embeddings that feed a linear classifier, and simulations show improved performance over traditional PCA or unguided auto‑encoder methods.
The paper presents a basket‑choice model that learns embeddings to capture interactions among products, prices, and stores, using random negative sampling to handle large item sets. Experiments on synthetic data demonstrate that sampling based on the actual assortment improves performance over existing methods. The model is validated on millions of transactions from Fortenova Group, showing potential to aid category managers in retail decision‑making.
The paper presents a tutorial on enhancing quantum amplitude embedding through data re-uploading, a technique that repeatedly encodes input data between trainable blocks to increase model expressiveness. It demonstrates that each additional upload raises the reachable polynomial order by two and provides simulation results showing significant improvements in test coefficient of determination for higher-order targets. The study offers practical guidance on selecting the number of uploads based on the polynomial order required for a given task.
The paper introduces FALLNEST, a data‑centric framework that condenses large, imbalanced inertial measurement unit (IMU) datasets into tiny, class‑balanced synthetic training sets for wearable fall detection. FALLNEST builds on dual‑domain time‑series condensation, adding a fall‑specific 7‑channel input, class‑balanced K‑Means initialization, normalized gradient matching with clipping, MMD‑based embedding alignment, and multi‑view Fourier‑domain augmentation. Experiments on SisFall, KFall, and MobiFall show that a 102K‑parameter TinyCNN trained on only 32 to 146 synthetic windows (≈1% of the original data) achieves 93.5–95.8% mean accuracy and 0.89–0.98 PR‑AUC with low variance.
arXiv:2607.01455v4 Announce Type: replace-cross Abstract: Language models learn continuous programs over discrete symbols, with the embedding table and LM-head acting as the read/write interface betw...
arXiv:2610.14730v1 Announce Type: cross Abstract: Adapting embedding models to new tasks requires training data that reflect task-specific information needs and relevance criteria. However, manual da...
arXiv:2610.19133v1 Announce Type: new Abstract: Indirect prompt injection uses untrusted retrieved content, tool outputs, or agent messages to override an LLM's intended behavior. We show that defens...
arXiv:2610.19892v1 Announce Type: new Abstract: Vision-language models suffer significant accuracy drops under common corruptions. While test-time adaptation (TTA) successfully mitigates these shifts...
arXiv:2605.16026v2 Announce Type: replace-cross Abstract: Multilingual speech translation frontends built on speech large language models typically condition on language identity, as a learned embedd...
arXiv:2610.17582v1 Announce Type: cross Abstract: Entities in real-world systems often evolve over time: users shift between information consumption patterns, researchers migrate between fields, and...