Novel kernel SGD algorithm using spherical radial basis functions and regularization strategy for large-scale supervised learning with general losses.
COMPASS benchmark evaluating LLM agents' ability to perform constrained optimization in multi-turn conversations for travel planning and scheduling tasks.
Study on entropy regularization with adaptive coefficients to improve LLM reasoning in reinforcement learning with verifiable rewards, addressing policy collapse.
Research on optimal hyperparameters (clipping bound and batch size) for differentially private transfer learning, addressing theory-practice gaps.
Systematic study demonstrating that enhancing LLM reasoning capabilities paradoxically increases tool hallucination in agent deployment scenarios.
Training-free in-context distillation with self-consistency cascades to reduce LLM agent inference costs while maintaining quality and iteration speed.
Approach using LLMs as lossless encoders/decoders for invertible problems like logic table to HDL conversion, mitigating hallucinations and omissions.
Framework teaching LLMs to predict chemical reaction mechanisms using arrow-pushing formalism notation for computer-assisted synthesis planning.
CadLLM training-free acceleration method for diffusion-based LLM inference using confidence-aware adaptive control of generation parameters.
Group Relative Policy Optimization method for improving consistency and reliability of LLM recommendations across semantically equivalent prompts.
Dynamic tool dependency retrieval method for LLM-based function calling agents, improving tool selection and reducing context length for on-device agents.
Federated learning approach for traffic prediction using adaptive prompts, enabling privacy-preserving collaborative training across distributed data.
CoMeT architecture enabling LLMs to process arbitrarily long sequences with constant memory and linear time complexity via efficient plug-in module.
Unsupervised domain adaptation for radioisotope identification in gamma spectroscopy using simulation-to-reality transfer learning.
Empirical analysis of adversarial attacks on safety-aligned LLMs showing polynomial-to-exponential growth in attack success rates with prompt injection.
Proves attention sinks are functionally necessary in softmax transformers for certain tasks, formalizing their role beyond optimization artifacts.
Continual learning framework for Fourier Neural Operators enabling model adaptation to new data distributions without retraining on prior data.
Dynamical analysis of MLP training through saddle structures, explaining vanishing gradients and overfitting phenomena.
Hybrid language models mixing RNNs and attention mechanisms, comparing performance advantages over pure transformers with theoretical and practical evidence.
Proposes Neural Computers that unify computation, memory, and I/O in learned runtime states, exploring general-purpose neural machine architectures.
Leave-one-out analysis for evaluating SVG generation beyond visual similarity, assessing structural editability and reusability.
Interpretable analysis of transformer in-context classification using permutation equivariance constraints to reveal layer-wise dynamics.
SOAR bridges SFT and RL stages in diffusion model training with self-correction mechanism for out-of-distribution denoising.
Explores goodness function design space for Forward-Forward algorithm, finding shape-sensitive functions outperform sum-of-squares.
Context-independent KV caching for LLMs eliminating recomputation overhead when reusing cached documents in new contexts.
Policy learning algorithm with regret and violation guarantees for agents operating under adversarial exogenous factors.
Completes financial KOL trading strategies from discourse by inferring unspecified execution decisions using structured analysis.
Evaluates LLM jury of frontier models scoring medical diagnoses against expert clinician panels on real-world hospital cases.
Systematic benchmark of 15 optimizers on 17 tabular datasets for training MLP-based models, examining alternatives to AdamW.
Chatbot using NLP and deep learning to answer FAQs in Amharic language for university students.
LLM-guided extractive summarization for processing code-mixed sexual harassment incident reports on citizen platforms.
Resource-efficient equivariant quantum convolutional neural networks for near-term quantum devices with limited computational resources.
ML-based preference elicitation algorithms for iterative combinatorial auctions addressing exponential bundle space problem.
Uses LLMs to generate synthetic data for market research and conjoint analysis, addressing scalability limitations of traditional surveys.
Constructive approach for building 2D CNNs to approximate Korobov functions with rigorous theoretical analysis.
Synthetic Discussion Generation framework uses LLMs to create simulated multi-participant discussions for cost-effective social science experiments.
Scalable multi-task learning for autonomous agents using Spiking Neural Networks with adaptive task-switching to reduce resource consumption.
LayerNorm tuning with concept drift guidance for multimodal metaphor identification in internet memes using efficient fine-tuning.
Investigates how LLMs effectively model sequential recommendations by analyzing preference intensity and temporal context factors.
Randomized SINDy presents a sequential machine learning algorithm for time-dependent dynamic data with PAC learning guarantees.
Compares statistical and neural mixed-effects models for predicting Parkinson's Disease progression using voice biomarkers.
TriagerX applies dual transformers to bug triaging, improving token relevance attention for software engineering tasks.
OjaKV compresses KV cache in long-context LLMs using context-aware online low-rank projection, reducing memory bottlenecks during inference.
MMAudioSep applies video-to-audio generative models for sound separation queried by video or text input.
PULSE transfers knowledge from rich sensor modalities to deployable sensors for embodied AI systems with sensor asymmetry constraints.
Analyzes CPU-centric execution characteristics of agentic AI systems, examining performance bottlenecks in tool-calling autonomous LLM agents.
Addresses one-sided conversation problem where only one speaker is recorded, reconstructing missing dialogue and generating summaries using conversational AI.
Optimizes sparse convolution operations in 3D point cloud networks by exploiting voxel coordinate properties.
Reviews machine learning approaches for plasma closure relations in fluid dynamics simulations.
Disco-RAG enhances RAG systems by incorporating discourse structure awareness, enabling LLMs to better synthesize knowledge from multiple retrieved passages.