PIS: Linking Importance Sampling and Attention Mechanisms for Efficient Prompt Compression

Large language models (LLMs) have achieved remarkable progress, demonstrating unprecedented capabilities across various natural language processing tasks. However, the high costs associated with such exceptional performance limit the widespread adoption of LLMs, highlighting the need for prompt compression. Existing prompt compression methods primarily rely on heuristic truncation or abstractive summarization techniques, which fundamentally overlook the intrinsic mechanisms of LLMs and lack a systematic evaluation of token importance for generation. In this work, we introduce Prompt Importance Sampling (PIS), a novel compression framework that dynamically compresses prompts by sampling important tokens based on the analysis of attention scores of hidden states. PIS employs a dual-level compression mechanism: 1) at the token level, we quantify saliency using LLM-native attention scores and implement adaptive compression through a lightweight 9-layer reinforcement learning (RL) network; 2) at the semantic level, we propose a Russian roulette sampling strategy for sentence-level importance sampling. Comprehensive evaluations across multiple domain benchmarks demonstrate that our method achieves state-of-the-art compression performance. Notably, our framework serendipitously enhances reasoning efficiency through optimized context structuring. This work advances prompt engineering by offering both theoretical grounding and practical efficiency in context management for LLMs.

Learning over von Mises-Fisher Distributions via a Wasserstein-like Geometry

We introduce a novel, geometry-aware distance metric for the family of von Mises-Fisher (vMF) distributions, which are fundamental models for directional data on the unit hypersphere. Although the vMF distribution is widely employed in a variety of probabilistic learning tasks involving spherical data, principled tools for comparing vMF distributions remain limited, primarily due to the intractability of normalization constants and the absence of suitable geometric metrics. Motivated by the theory of optimal transport, we propose a Wasserstein-like distance that decomposes the discrepancy between two vMF distributions into two interpretable components: a geodesic term capturing the angular separation between mean directions, and a variance-like term quantifying differences in concentration parameters. The derivation leverages a Gaussian approximation in the high-concentration regime to yield a tractable, closed-form expression that respects the intrinsic spherical geometry. We show that the proposed distance exhibits desirable theoretical properties and induces a latent geometric structure on the space of non-degenerate vMF distributions. As a primary application, we develop the efficient algorithms for vMF mixture reduction, enabling structure-preserving compression of mixture models in high-dimensional settings. Empirical results on synthetic datasets and real-world high-dimensional embeddings, including biomedical sentence representations and deep visual features, demonstrate the effectiveness of the proposed geometry in distinguishing distributions and supporting interpretable inference. This work expands the statistical toolbox for directional data analysis by introducing a tractable, transport-inspired distance tailored to the geometry of the hypersphere.

SentenceKV: Efficient LLM Inference via Sentence-Level Semantic KV Caching

Large language models face significant computational and memory challenges when processing long contexts. During inference, efficient management of the key-value (KV) cache, which stores intermediate activations for autoregressive generation, is critical to reducing memory overhead and improving computational efficiency. Traditional token-level efficient KV caching methods overlook semantic information, treating tokens independently without considering their semantic relationships. Meanwhile, existing semantic-preserving KV cache management approaches often suffer from substantial memory usage and high time-to-first-token. To address these limitations, we propose SentenceKV, a novel sentence-level semantic KV caching approach designed to enhance inference efficiency while preserving semantic coherence. During prefilling, SentenceKV groups tokens based on sentence-level semantic similarity, compressing sentence representations into concise semantic vectors stored directly on the GPU, while individual KV pairs are offloaded to CPU. During decoding, SentenceKV generates tokens by selectively retrieving semantically relevant sentence-level KV entries, leveraging the semantic similarity between the prefilling-stage semantic vectors and decoding-stage queries. This ensures efficient and contextually accurate predictions, minimizing the loading of redundant or irrelevant data into GPU memory and significantly reducing memory overhead while maintaining stable inference latency, even for extremely long contexts. Extensive evaluations on benchmarks including PG-19, LongBench, and Needle-In-A-Haystack demonstrate that SentenceKV significantly outperforms state-of-the-art methods in both efficiency and memory usage, without compromising model accuracy.

Text Compression for Efficient Language Generation

We challenge the prevailing assumption that LLMs must rely fully on sub-word tokens for high-quality text generation. To this end, we propose the "Generative Pretrained Thoughtformer" (GPTHF), a hierarchical transformer language model capable of text generation by compressing text into sentence embeddings and employing a sentence attention mechanism. GPTHF retains GPT's architecture, modifying only token interactions via dynamic sparse attention masks. Our experiments show that GPTHF achieves an up to an order of magnitude improvement in FLOPs efficiency and a threefold increase in runtime speed compared to equally-sized GPT models in the low-size regime. This is achieved through a unique generation method that caches and reuses sentence embeddings, allowing significant portions of the input to bypass large parts of the network.

TROVE: A Challenge for Fine-Grained Text Provenance via Source Sentence Tracing and Relationship Classification

LLMs have achieved remarkable fluency and coherence in text generation, yet their widespread adoption has raised concerns about content reliability and accountability. In high-stakes domains such as healthcare, law, and news, it is crucial to understand where and how the content is created. To address this, we introduce the Text pROVEnance (TROVE) challenge, designed to trace each sentence of a target text back to specific source sentences within potentially lengthy or multi-document inputs. Beyond identifying sources, TROVE annotates the fine-grained relationships (quotation, compression, inference, and others), providing a deep understanding of how each target sentence is formed. To benchmark TROVE, we construct our dataset by leveraging three public datasets covering 11 diverse scenarios (e.g., QA and summarization) in English and Chinese, spanning source texts of varying lengths (0-5k, 5-10k, 10k+), emphasizing the multi-document and long-document settings essential for provenance. To ensure high-quality data, we employ a three-stage annotation process: sentence retrieval, GPT provenance, and human provenance. We evaluate 11 LLMs under direct prompting and retrieval-augmented paradigms, revealing that retrieval is essential for robust performance, larger models perform better in complex relationship classification, and closed-source models often lead, yet open-source models show significant promise, particularly with retrieval augmentation.

IPCGRL: Language-Instructed Reinforcement Learning for Procedural Level Generation

Recent research has highlighted the significance of natural language in enhancing the controllability of generative models. While various efforts have been made to leverage natural language for content generation, research on deep reinforcement learning (DRL) agents utilizing text-based instructions for procedural content generation remains limited. In this paper, we propose IPCGRL, an instruction-based procedural content generation method via reinforcement learning, which incorporates a sentence embedding model. IPCGRL fine-tunes task-specific embedding representations to effectively compress game-level conditions. We evaluate IPCGRL in a two-dimensional level generation task and compare its performance with a general-purpose embedding method. The results indicate that IPCGRL achieves up to a 21.4% improvement in controllability and a 17.2% improvement in generalizability for unseen instructions. Furthermore, the proposed method extends the modality of conditional input, enabling a more flexible and expressive interaction framework for procedural content generation.

Statistical Analysis of Sentence Structures through ASCII, Lexical Alignment and PCA

While utilizing syntactic tools such as parts-of-speech (POS) tagging has helped us understand sentence structures and their distribution across diverse corpora, it is quite complex and poses a challenge in natural language processing (NLP). This study focuses on understanding sentence structure balance - usages of nouns, verbs, determiners, etc - harmoniously without relying on such tools. It proposes a novel statistical method that uses American Standard Code for Information Interchange (ASCII) codes to represent text of 11 text corpora from various sources and their lexical category alignment after using their compressed versions through PCA, and analyzes the results through histograms and normality tests such as Shapiro-Wilk and Anderson-Darling Tests. By focusing on ASCII codes, this approach simplifies text processing, although not replacing any syntactic tools but complementing them by offering it as a resource-efficient tool for assessing text balance. The story generated by Grok shows near normality indicating balanced sentence structures in LLM outputs, whereas 4 out of the remaining 10 pass the normality tests. Further research could explore potential applications in text quality evaluation and style analysis with syntactic integration for more broader tasks.

Task-agnostic Prompt Compression with Context-aware Sentence Embedding and Reward-guided Task Descriptor

The rise of Large Language Models (LLMs) has led to significant interest in prompt compression, a technique aimed at reducing the length of input prompts while preserving critical information. However, the prominent approaches in prompt compression often require explicit questions or handcrafted templates for compression, limiting their generalizability. We propose Task-agnostic Prompt Compression (TPC), a novel framework that generalizes compression across tasks and domains without requiring input questions or templates. TPC generates a context-relevant task description using a task descriptor trained on a curated dataset of context and query pairs, and fine-tuned via reinforcement learning with a reward function designed to capture the most relevant information. The task descriptor is then utilized to compute the relevance of each sentence in the prompt to generate the compressed prompt. We introduce 3 model sizes (Base, Large, and Huge), where the largest model outperforms the existing state-of-the-art methods on LongBench and ZeroSCROLLS benchmarks, and our smallest model performs comparable to the existing solutions while being considerably smaller.

Variable-frame CNNLSTM for Breast Nodule Classification using Ultrasound Videos

The intersection of medical imaging and artificial intelligence has become an important research direction in intelligent medical treatment, particularly in the analysis of medical images using deep learning for clinical diagnosis. Despite the advances, existing keyframe classification methods lack extraction of time series features, while ultrasonic video classification based on three-dimensional convolution requires uniform frame numbers across patients, resulting in poor feature extraction efficiency and model classification performance. This study proposes a novel video classification method based on CNN and LSTM, introducing NLP's long and short sentence processing scheme into video classification for the first time. The method reduces CNN-extracted image features to 1x512 dimension, followed by sorting and compressing feature vectors for LSTM training. Specifically, feature vectors are sorted by patient video frame numbers and populated with padding value 0 to form variable batches, with invalid padding values compressed before LSTM training to conserve computing resources. Experimental results demonstrate that our variable-frame CNNLSTM method outperforms other approaches across all metrics, showing improvements of 3-6% in F1 score and 1.5% in specificity compared to keyframe methods. The variable-frame CNNLSTM also achieves better accuracy and precision than equal-frame CNNLSTM. These findings validate the effectiveness of our approach in classifying variable-frame ultrasound videos and suggest potential applications in other medical imaging modalities.

State Space Models for Extractive Summarization in Low Resource Scenarios

Extractive summarization involves selecting the most relevant sentences from a text. Recently, researchers have focused on advancing methods to improve state-of-the-art results in low-resource settings. Motivated by these advancements, we propose the MPoincareSum method. This method applies the Mamba state space model to generate the semantics of reviews and sentences, which are then concatenated. A Poincare compression is used to select the most meaningful features, followed by the application of a linear layer to predict sentence relevance based on the corresponding review. Finally, we paraphrase the relevant sentences to create the final summary. To evaluate the effectiveness of MPoincareSum, we conducted extensive experiments using the Amazon review dataset. The performance of the method was assessed using ROUGE scores. The experimental results demonstrate that MPoincareSum outperforms several existing approaches in the literature