Research
On-device research index

arXiv research

A locally-built, LLM-digested index of recent arXiv papers in quant finance, geometry/topology, and statistical ML — keyword search served straight from SQLite on this machine.

168,657 papers · 148 categories

Trend · papers per month

3877115153 · Jun 202019922001200920172026
48 results for sparse tokens

Attention mechanism learns to focus on sparse tokens efficiently.

problem Detecting weak, rare, and sparsely located features in long sequences.
method Theoretical analysis and training of a single-layer attention classifier in a sparse-token classification model.
result A single-layer attention classifier can achieve vanishing test error with logarithmic signal strength growth, unlike linear classifiers requiring linear growth.

Gradient descent biases linear models in next-token prediction towards data entropy.

problem Optimization bias in next-token prediction models.
method Analysis of gradient descent on linear models with sparse conditional distributions.
result Gradient descent selects parameters that equate token logits differences to log-odds in the data subspace.

This study examines whether tokenized assets improve liquidity and finds significant differences across categories.

problem Improving liquidity for real-world assets through tokenization.
method Examined tokenized real-world assets using Ethereum-based data, measuring liquidity through turnover, active addresses, and active-month indicator.
result Gold-backed tokens show more persistent on-chain activity than Treasury and private-credit-related products, but asset value alone does not reliably predict liquidity.

DRFormer uses dynamic tokenization and multi-scale transformer to forecast long time series.

problem Forecasting long-term time series data across diverse scales.
method Dynamic tokenizer, multi-scale transformer, dynamic sparse learning, rotary position encoding.
result DRFormer outperforms existing methods in forecasting accuracy.

MAESTRO improves multimodal learning for dynamic time series with adaptive attention and robustness.

problem Challenges in multimodal learning, especially in healthcare and daily living.
method Dynamic intra- and cross-modal interactions, symbolic tokenization, adaptive attention budgeting, sparse cross-modal attention, MoE mechanism.
result Average relative improvements of 4% and 8% over existing multimodal and multivariate approaches, respectively, under complete observations.

CoT improves transformer sample efficiency by reducing input token dependencies and attention sparsity.

problem Transformer sample inefficiency in simple tasks.
method Demonstrated through parity-learning setup, showing CoT reduces required samples from exponential to polynomial.
result Transformer learns function within polynomial samples with CoT, requiring exponential samples without CoT.

New analysis shows interpretability doesn't guarantee steering utility in LLMs.

problem Does higher interpretability lead to better steering utility in large language models?
method Trained 90 SAEs across three LLMs, evaluated interpretability and steering utility, used Kendall's rank coefficients for analysis.
result Interpretability is only weakly associated with steering utility, and features selected by Delta Token Confidence improve steering performance.

The paper analyzes dynamics of momentum in high dimensions with sparse updates.

problem Theoretical analysis of momentum dynamics in high-dimensional sparse settings.
method Theoretical analysis of two models: least squares with sparse inputs and logistic regression with a rare class.
result Characterization of high-dimensional limits of momentum dynamics and phase structure.

This paper analyzes shallow ViTs, providing sample complexity and SGD behavior insights.

problem Theoretical understanding of shallow ViTs, especially their sample complexity and SGD behavior.
method Data model with label-relevant and label-irrelevant tokens, theoretical analysis of shallow ViT training.
result Characterization of sample complexity for zero generalization error in shallow ViTs.

We propose a new method for learning word representations using hierarchical regularization in sparse coding inspired by the linguistic study of word meanings. We show an efficient learning algorithm based on stochastic proximal methods that is significantly faster than previous approaches, making it possible to perfor…

2014-06-08abs ↗pdf ↗

Transformer attention layers solve single-location regression tasks.

problem Understanding token-wise sparsity and internal linear representations in attention-based models.
method Introduce single-location regression task and a simplified predictor based on self-attention layers.
result Transformer attention layers are asymptotically Bayes optimal and can learn underlying structures effectively.

Sparse attention model reduces long-context inference time with exponential accuracy guarantees.

problem Efficiently processing long-context queries in large language models.
method Formalizes attention as a projection onto key vectors, analyzes entropic relaxation, and introduces Vashista Sparse Attention.
result Sparse attention concentrates on a constant-size active face, leading to exponential decay of inactive tokens' mass and linear scaling of active face error.

Transformer models waste resources on long-context tasks.

problem Redundant attention computations in Transformer models for long-context tasks.
method Reformulate sequence modeling as supervised learning, analyze attention sparsity, formulate attention optimization as linear coding problem, propose Dynamic Group Attention.
result DGA reduces computational costs while maintaining performance.

Prototype model improves model auditing and understanding.

problem Auditing and understanding modern language models is expensive and approximate.
method Introduced a sparse, non-negative mixture of learned prototypes trained with clustering objectives.
result Prototype models either surpass or remain within 2.5 percentage points of dense baselines on downstream tasks.

This study explains RL training dynamics in LLMs, focusing on token-level optimization and reasoning pattern reshaping.

problem Understanding the training dynamics of RL in LLMs to improve their reasoning capabilities.
method Empirical analysis and theoretical modeling of RL training process, focusing on reasoning patterns and token optimization.
result RL primarily optimizes a sparse subset of critical tokens, reshaping reasoning pattern distributions and affecting model performance.

TROLL improves RL for LLMs by replacing clipping with a trust region projection.

problem Clipping in RL for LLMs causes instability and suboptimal performance.
method TROLL uses a discrete differentiable trust region projection to replace clipping, balancing computational cost and effectiveness.
result TROLL consistently outperforms PPO-like clipping in training speed, stability, and final success rates.

To scale non-parametric extensions of probabilistic topic models such as Latent Dirichlet allocation to larger data sets, practitioners rely increasingly on parallel and distributed systems. In this work, we study data-parallel training for the hierarchical Dirichlet process (HDP) topic model. Based upon a representati…

2019-06-06abs ↗pdf ↗

Two new estimators reduce costs and improve accuracy for EHR outcome prediction.

problem Sparse estimate distributions, high computational cost, and high sampling variance in EHR outcome prediction.
method Proposed SCOPE and REACH estimators that leverage next-token probability distributions.
result SCOPE and REACH match Monte Carlo accuracy with token reductions of 2.5-3.4 times and variance guarantees.

Sparse text alignments learned via optimal transport improve model explainability.

problem Building self-explaining models by selecting relevant text pieces.
method Employing optimal transport to find minimal cost alignments, introducing constrained variants for sparsity.
result Sparse and interpretable alignments achieved, preserving prediction accuracy.

Neural sequence generation is typically performed token-by-token and left-to-right. Whenever a token is generated only previously produced tokens are taken into consideration. In contrast, for problems such as sequence classification, bidirectional attention, which takes both past and future tokens into consideration, …

2019-08-16abs ↗pdf ↗

This work explains the structural origins of attention sinks in LLMs.

problem Initial tokens disproportionately monopolize attention scores in LLMs.
method Traced to self-attention's value aggregation process and FFN layer activations.
result Attention sinks form due to variance discrepancy and dimension disparity.

This research improves capital efficiency and impermanent loss in cryptocurrency markets using multi-token trading pools.

problem Poor impermanent loss and capital efficiency in automated market makers.
method Analysis and construction of a multi-token token proactive market maker (MPMM).
result MPMM shows better impermanent loss and capital efficiency than comparable market makers.

Paper introduces a method to assess liquidity risk in meme tokens using entity-linked address analysis.

problem High market volatility and vulnerability to manipulation in meme tokens.
method Multi-dimensional approach integrating fund flow analysis, behavioral similarity, and anomalous transaction detection.
result Significant disparities between apparent and actual liquidity in meme token markets.

DOS improves language model generation by considering inter-token dependencies.

problem Lack of sequence-level information and inter-token dependencies in existing decoding strategies.
method Dependency-Oriented Sampler (DOS) that uses attention matrices to approximate inter-token dependencies.
result DOS consistently achieves superior performance on code generation and mathematical reasoning tasks.

Mobile V-MoEs scale down ViTs for resource-constrained vision tasks.

problem Scaling down Vision Transformers for resource-constrained applications.
method Sparse Mixture-of-Experts (MoEs) applied to entire images, with a stable training procedure.
result Mobile V-MoEs achieve better performance-efficiency trade-offs than dense ViTs.

Proving that next-token prediction makes language models generate coherent long documents.

problem Understanding why language models generate coherent documents despite focusing on next-token prediction.
method Proving the power of next-token prediction in learning longer-range structure using Recurrent Neural Networks (RNN).
result Optimizing next-token prediction in RNNs yields a model that closely approximates the training distribution, even for long-range coherence.

Moirai-MoE improves time series forecasting by automatically specializing tokens without human-defined frequency.

problem Unified training on time series data remains challenging due to heterogeneity and non-stationarity.
method Uses sparse mixture of experts (MoE) within Transformers to automatically specialize tokens for diverse time series patterns.
result Moirai-MoE outperforms existing foundation models in both in-distribution and zero-shot scenarios.

Hydra boosts efficiency for long-context reasoning in resource-constrained settings.

problem Quadratic complexity of transformers limits long-context reasoning in resource-constrained systems.
method Hydra uses a modular architecture with adaptive routing between sparse global attention, mixture-of-experts, and dual memories.
result Hydra achieves significant throughput and accuracy improvements for long-context reasoning.

QA-Token improves tokenization for noisy data, boosting model performance.

problem Tokenization ignores data quality, limiting model effectiveness on noisy corpora.
method QA-Token combines signal quality with vocabulary construction through bilevel optimization and reinforcement learning.
result QA-Token achieves state-of-the-art performance on genomic and financial datasets.

Minimal token perturbations reveal how Transformer models process information.

problem Understanding information propagation in Transformer models for interpretability.
method Study of minimal token perturbations on embedding space.
result Rare tokens cause larger shifts, and input information mixes deeper.

LLM-as-a-service prices vary arbitrarily due to tokenization multiplicity.

problem Arbitrary price variation in LLM-as-a-service due to multiple tokenizations of the same output.
method Introduce canonical generation to restrict LLMs to unique tokenizations and develop an efficient sampling algorithm.
result Our sampling algorithm for canonical generation solves tokenization multiplicity and maintains comparable performance and runtime to standard sampling.

SPARTAN learns sparse interaction graphs between objects in scenes.

problem Capturing structured interactions between entities in complex settings remains challenging.
method SPARTAN is a Transformer-based world model that learns sparse, context-dependent interaction graphs through sparsity regularization on attention patterns.
result SPARTAN achieves significantly improved few-shot adaptation to dynamics changes and robustness against distractors.