Tag
This paper proposes a method to predict anticipatory outliers in embedding space that will become part of emerging topics, using geometric features and retrospective supervision, with high predictive performance on French news corpora.
SimLoss uses embedding-space contrastive supervision to enable single-pass fine-grained image captioning that matches multi-stage quality at much lower latency, with variants like SimLoss FFT achieving high precision.
EXPL-FR is a method that explains face recognition models by aligning vision-language embeddings to the recognition space, enabling label-free auditing of semantic attributes and model comparison without requiring architecture access.
This paper proposes a geometric filtering framework that selects high-quality LLM-generated samples by evaluating their Euclidean distance to real class examples in an embedding space, improving few-shot text classification performance by +2.61 percentage points over SMOTE.
ConceptFormer learns continuous latent concept representations for visual document retrieval, bridging visual evidence and semantic relevance without text intermediates, achieving significant improvements over baselines.
This paper introduces a geometry-aware adversarial attack framework that targets relational structure in contrastive embedding manifolds, showing that verification systems like Markmatch can be severely degraded by distorting pairwise similarities rather than decision boundaries.
An interactive map of GPT-2's token embedding space that lets users tap any token to explore its embeddings.
This paper treats books as points in sentence-embedding space and decomposes the transformation from original novel to sequel into interpretable axes, revealing a taxonomy of sequel types (formulaic, concentrated, compositional). It applies this method to thirteen author pairs and validates the geometry against authorial intent for Tom Sawyer to Huckleberry Finn.
Rotate2Think is a training-free method that improves language model reasoning by estimating a rotation from input embeddings to thinking embeddings and injecting a synthetic thinking vector at inference time, achieving accuracy gains across multiple benchmarks.
Geometric Latent Reasoning (GLR) introduces a geometric path-approximation method for latent reasoning in LLMs, enabling shorter generations while maintaining accuracy across mathematical reasoning benchmarks.
The paper introduces PASA, a robust watermarking algorithm for LLM-generated text that operates at the semantic level using latent embedding spaces to resist semantic-invariant attacks like paraphrasing.
Researchers introduce CSR-L and CS-MTEB benchmarks showing that code-switching queries degrade IR system performance by up to 27%, revealing embedding-space divergence that current multilingual techniques cannot fix.