Every entry carries a structured overview, a code link where one exists, and the reproduction evidence collected so far. Search by keyword or by meaning.
Looking for an answer rather than a list? Ask AutoXiv across the corpus →
Showing 21–40 of 103
The paper proposes Supplement Generation Training (SGT), which trains smaller LLMs to generate supplemental text that improves larger foundation models' performance on agentic tasks without retraining the large models. This approach reduces computational costs and enables efficient adaptation to new tasks and domains.
This paper proposes Tokenised Flow Matching for Posterior Estimation (TFMPE), a method that reduces simulator evaluation costs in hierarchical simulation-based inference by learning per-site neural surrogates and assembling synthetic observations. The approach is validated on infectious disease and computational fluid dynamics models with improved computational efficiency.
COMPASS is a data-centric framework that adapts large language models to target languages using parameter-efficient fine-tuning with semantic-aware sampling of multilingual data. It minimizes negative cross-lingual interference while maximizing positive transfer, and includes a continual learning variant to handle production data distribution shifts.
This paper applies Generative Flow Networks (GFlowNets) to adapt digital twin models of natural systems by learning to sample simulator parameters proportional to their agreement with observations. The approach handles the fundamental challenge that sparse observations often leave multiple plausible parameter configurations, preserving uncertainty rather than finding a single optimal solution.
This paper provides a comprehensive structured analysis of adversarial robustness literature and introduces Auto-ART, an open-source framework that standardizes robustness evaluation across 50+ attacks, detects gradient masking, and maps compliance to regulatory frameworks. It addresses fragmentation in the field by combining meta-scientific synthesis with practical engineering tools.
MGDA-Decoupled is a geometry-based multi-objective optimization algorithm for aligning LLMs that balances conflicting objectives like helpfulness and truthfulness without requiring reinforcement learning or explicit reward models. It operates within the Direct Preference Optimization framework and achieves higher win rates than existing methods by accounting for each objective's convergence dynamics.
This paper systematically studies transformer compression across GPT-2 and Mistral 7B, identifying five structural properties that explain why certain compression techniques fail, particularly showing that high-variance activations are not predictively important and that block-level approximations suffer from distribution shift. The findings suggest fundamental limits to static compression and motivate adaptive, per-token approaches instead.
This paper proposes a feature whitening approach that decorrelates anatomically grouped brain regions to improve the interpretability of linear neuroimaging models while maintaining predictive performance. The method helps disentangle overlapping information across correlated brain measures, making model weights more clinically meaningful for identifying biomarkers in psychiatric disorders.
This paper introduces GRPO-VPS, which enhances Group Relative Policy Optimization by adding verifiable process supervision that tracks the model's confidence in correct answers throughout reasoning steps. This enables more targeted credit assignment and sample-efficient policy updates without requiring critic models or auxiliary supervision.
The paper introduces a multi-scale metric for strings that combines angle distances between n-gram count vectors at different scales using exponential weights. It provides algorithms, theoretical guarantees, and empirical validation for clustering applications.
This paper proposes Occupancy Reward Shaping (ORS), a method that extracts temporal and geometric information from learned world models to create reward functions that improve credit assignment in offline goal-conditioned reinforcement learning. The approach uses optimal transport on occupancy measures and demonstrates 2.2x performance improvements across diverse tasks including real-world Tokamak control.
pAI/MSc is an open-source multi-agent system designed to assist researchers in ML theory by automating the conversion of hypotheses into manuscript drafts, significantly reducing manual steering effort. The system emphasizes human-in-the-loop collaboration rather than fully autonomous research.
SmartVector augments vector embeddings with temporal awareness, confidence decay, and relational properties to improve retrieval-augmented generation accuracy on versioned knowledge. The framework doubles accuracy on versioned queries and reduces stale answers by 62% compared to standard cosine similarity retrieval.
PINA is a differentially private clustered federated learning framework that uses low-rank adapters and compressed sketches for privacy-preserving initialization, followed by normality-driven aggregation to improve convergence. It achieves 2.9% higher accuracy than state-of-the-art DP-FL methods while maintaining formal privacy guarantees.
This paper establishes a mathematical correspondence between state space models (like S4) and nonlinear oscillator networks, deriving an exact operator expression for S4D's forward pass. This provides analytical interpretability for modern sequence models by revealing how information propagates as waves in a ring network topology.
This paper proposes a hierarchical multi-agent deep reinforcement learning framework that enables distributed energy resources to participate in both peer-to-peer retail trading and wholesale electricity markets through coordinated prosumer aggregation. The approach uses a Stackelberg game to optimize market performance and grid flexibility.
This paper proposes a graph neural network framework to analyze assurance cases (structured safety argument documents) for link prediction between argument elements and detection of LLM-generated versus human-authored cases. The work reveals that LLM-generated assurance cases have distinct structural patterns and highlights limitations in GNN explanation methods.
This paper proposes Vine Denoising Copula (VDC), an amortized approach to vine copulas that trains a single bivariate denoising model and reuses it across all vine edges, enabling fast and tractable high-dimensional density and information estimation. The method combines neural flexibility with classical copula interpretability while achieving significant computational speedups.
This paper provides the first systematic theoretical analysis of Bayesian mixture-of-experts models with softmax gating, establishing posterior contraction rates for density estimation, convergence guarantees for parameter estimation, and strategies for selecting the number of experts.
This paper presents an efficient symbolic computation algorithm for determining whether causal effects can be identified from observational data with latent confounders in linear structural causal models. The algorithm finds the lowest-degree identifying formulas in quasi-polynomial time, making causal identifiability analysis practical for larger problems.