Automated dataset construction and query-side adaptation for multi-tenant search systems using dark data from query logs without full corpus re-indexing.
Self-evolution approach with iterative generate-verify-refine cycles for code optimization, improving exploration efficiency and experience utilization to discover superior algorithmic complexity.
Budget-conscious data acquisition strategy for allocating annotation budgets between ground-truth labels and pairwise preferences using semi-parametric inference.
Framework enabling LLMs to explore code sandboxes for general agentic intelligence without fine-tuning, leveraging file systems and external resources for non-code tasks.
Examines self-preference bias in LLM evaluators, identifying methodological confounds that distort measurement of narcissism in model-based automated evaluation workflows.
Analyzes human susceptibility to LLM hallucinations and disinformation using structural causal models across GPT-4, Llama-2, and other foundation models with 918 evaluations.
Combines 3D Gaussian splatting with physics engines to predict physically plausible video dynamics from visual data, reducing computational costs.
Neural networks evaluated against exact posteriors using normalizing flows as oracles to study scaling laws, aleatoric uncertainty, and epistemic error on realistic images.
CardinalGraphFormer applies graph transformers with cardinality-preserving attention for molecular property prediction with limited labeled data.
Data-driven silicon sociology framework analyzing social structure formation in large-scale autonomous LLM agent ecosystems and multi-agent communities.
Studies reinforcement learning for autonomous cyber defense agents, exploring reward function design beyond dense engineering in complex network environments.
CoSA adapts LLMs via compressed sensing-based parameter-efficient fine-tuning, extending beyond low-rank decomposition for task-specific adaptation.
Systematic empirical study of context engineering for LLM agents operating SQL generation across 9,649 experiments, evaluating schema formats at scale.
Analysis of Group Relative Probability Optimization (GRPO) limitations in reinforcement learning with verifiable rewards, identifying implicit advantage symmetry constraints.
HuMI framework enables humanoid whole-body manipulation learning from robot-free demonstrations without teleoperation or complex reward engineering.
MAU-GPT uses anomaly-aware expert adaptation for multi-type industrial anomaly detection, with new comprehensive dataset spanning multiple manufacturing domains.
Fin-RATE benchmarks LLM performance on financial document analysis requiring multi-document synthesis across reporting periods and corporate entities.
First quantitative theory predicting neural scaling law exponents for LLMs based on natural language statistics, derived from data-limited scaling analysis.
Generative Reasoning Re-ranker applies LLMs to recommendation reranking via reinforcement learning, exploring zero-shot and in-context learning paradigms.
STaR enables scalable long-horizon memory for mobile robots with task-conditioned retrieval supporting agentic planning and reasoning over multimodal data.
Anagent enhances scientific table and figure analysis by integrating multimodal knowledge and domain-specific reasoning for complex document interpretation.
Automated black-box pipeline detects unverbalized biases in LLM chain-of-thought reasoning without predefined categories or hand-crafted datasets.
Aletheia is a math research agent that iteratively generates, verifies, and revises solutions, advancing from competition-level problem-solving to autonomous professional mathematics research.
Control Reinforcement Learning trains policies to steer LLMs via interpretable sparse autoencoder features, producing intervention logs showing which features change model outputs.
Unified vision-language framework for detecting and grounding forgery across interleaved text, images, and videos using balanced RL.
Security analysis of mobile AI agents operating as intent-centric systems, identifying vulnerabilities in screen-as-interface paradigm.
Vision-language model framework enabling iterative visual reasoning through image manipulation for improved fine-grained understanding.
LLM-based approach for automated optimization modeling with error-driven data generation to improve training data quality.
Vector quantization compression method for Mixture of Experts LLMs using KLT-guided SVD for resource-constrained deployment.
Analysis of spectral anisotropy in LLM gradient signals and proposed optimizer improvements to better utilize tail directions.
Lightweight KAN-based classifier for GNSS interference detection on edge devices running generative AI applications.
Multi-agent RL approach for chiplet placement optimization in 2.5D circuits, addressing thermal and wirelength constraints.
Time-TK combines transformers and Kolmogorov-Arnold Networks for time series forecasting, addressing information bottleneck in long sequences.
MELINOE method enabling memory-efficient inference for Mixture-of-Experts models through fine-tuning with expert offloading.
Foundation model for ambient intelligence using WiFi signals for human presence, activity, and physiology detection in smart environments.
Defense method against downstream-agnostic adversarial examples for pre-trained SSL encoders without task-specific fine-tuning.
UltraLIF framework for fully differentiable spiking neural networks using ultradiscretization and max-plus algebra as alternative to surrogate gradients.
First systematic study of multi-disciplinary LLM fine-tuning across five scientific domains, analyzing learning dynamics and cross-domain knowledge synergy.
Uses protein language model embeddings to improve generative molecular dynamics for predicting molecular properties and distributions.
Investigates knowledge transfer from pretraining to supervised fine-tuning in language models, analyzing capability persistence and benchmark reliability.
Credal concept bottleneck models that structurally separate epistemic and aleatoric uncertainty components in predictive models.
RL-based data rewriting agent for stable off-policy supervised fine-tuning of LLMs when handling distribution-shifted downstream data.
Evaluates memory structures in LLM-based agents with focus on long-term memory frameworks for knowledge storage and reasoning capability assessment.
Mathematical framework analyzing linear representation hypothesis in language models, determining how many neurons needed to linearly represent and access features.
Introduces HiFloat4, a block floating-point format for efficient language model inference with 4.5 bits per value and three-level scaling hierarchy.
Proposes Predictive Associative Memory (PAM), a neural architecture using temporal co-occurrence for memory retrieval instead of similarity-based approaches.
Multi-objective combinatorial optimization reformulated as online learning with decomposed decision space and regret-bounded guarantees.
Hybrid Transformer-SSM model using retrieval-aware distillation to preserve in-context retrieval capabilities while improving efficiency.
Provably efficient algorithms for robust MDPs with general policy parameterization, reducing to entropy-regularized formulations.
Mathematical framework explaining LLM generalization through sparse low-dimensional manifold geometry of activation states rather than parameter count.