Research on Physics-Informed Neural Networks reducing spatial derivative complexity from O(d^k) to O(1) using stochastic dimension-free estimators.
MazeBench evaluates whether multimodal models solve visual mazes through genuine planning or token-space search, finding high accuracy scores misleading.
Research on survival analysis for medical applications using prior-fitted Bayesian networks on tabular data with censoring.
Online gradient-based data selection and reweighting framework for LLM fine-tuning with optimizer-aware design.
SeqComm-DFL: Multi-agent learning framework with value-aware sequential communication optimized for task performance.
PRiMeFlow: Flow matching approach for modeling genetic and drug perturbation effects on single-cell gene expression.
Observable neural ODEs framework for continuous-time causal forecasting with identifiability via control-theoretic observability.
DPLM-Evo: Discrete diffusion protein language model capturing evolutionary constraints for protein design and generation.
Behavioral evaluation framework for agentic LLM-based stock prediction systems using closed-loop RL feedback and multi-dimensional metrics.
Asymmetric on-policy distillation: Token-level approach addressing variance, vanishing gradients, and exploration issues in LLM training.
Study of when neural networks fail at OOD generalization, decoupling feature learning from data-generating-process identifiability.
Analysis of LLM scaling laws showing tokens-per-parameter ratio affects extrapolation robustness; demonstrates collinearity causes ill-conditioning.
FactoryNet: First large-scale industrial time-series pretraining corpus with 51M datapoints for foundation models with cross-embodiment transfer.
Study of MXFP4 quantization for full-pipeline FP4 training of LLMs, analyzing divergence in forward/backward passes on native FP4 hardware.
Key-Value Means: A block-recurrent attention mechanism enabling O(N) transformers with fixed or expandable memory for long-context tasks.
Consolidation-Expansion Operator Mechanics: A framework for understanding adaptive learning systems through order-gap formalism.
Metis: A framework using policy optimization to automate jailbreaking of LLMs via red teaming, treating attacks as POMDP inference.
V4FinBench: A benchmark with 1M+ company-year records for evaluating tabular foundation models and LLMs on corporate bankruptcy prediction.
Autonomous AI system for data-centric machine learning that searches, adapts, and validates datasets to improve model performance.
Trajectory matching policy optimization for diffusion model alignment that prevents reward hacking through probability distribution constraints.
Self-supervised architecture for rare event prediction in multivariate time series using causal transformers and joint-embedding predictive learning.
Lightweight method for selecting targeted synthetic training data by scoring samples on boundary proximity, uncertainty, and real-data density.
Leverages test-time scaling comparative information as self-training signal in dual judgment space for improved code generation.
Proposes ZeNO, a gradient-free optimization method for reward alignment in diffusion and flow models using noise-space optimization.
Uses agentic program search to apply test-time compute to frozen embedding models, optimizing inference programs for dense retrieval.
Demonstrates data curation alone improves vision-language model performance by 11.7pp without architecture or compute changes.
Investigates block size effects in reinforcement learning for diffusion large language models with semi-autoregressive generation.
Proposes FAMeX algorithm for AI explainability using graph-theoretic feature association maps.
Low-rank approximation framework for compressing transformer attention layers by analyzing architectural characteristics rather than individual layer outputs.
Explores learned continue-thinking tokens to extend reasoning steps and improve LLM performance through test-time compute scaling.
Proposes Dilated Unmasking Scheduler for faster non-autoregressive text generation in masked diffusion language models by scheduling token unmasking to avoid sequential behavior.
Investigates whether activation verbalization methods (using LLMs to describe internal representations) reveal privileged information or just input information.
O3 method for sample-efficient optimization within generative model outputs (diffusion/flow models) for task-specific criteria.
Aligns video foundation models with human social perception using geometric supervision of behavioral features.
Greedy algorithm for selecting which source tasks and sample sizes to transfer in multi-source transfer learning for linear models.
Evaluates Minimum Bayes Risk decoding versus beam search for automatic speech recognition tasks.
Uses LLMs to explain what linguistic features (e.g., city names) predict deceptive reviews, leveraging language models for feature interpretation.
FireScope-Bench dataset and method for predicting wildfire risk using chain-of-thought reasoning with satellite imagery and climate data.
PersonalAlign framework aligns GUI agents with implicit user intents using long-term user records as context for improved personalization.
Sockpuppetting combines prefill attacks with optimization to jailbreak LLMs more effectively by ensembling multiple prefill variants.
PolySHAP improves KernelSHAP algorithm for computing Shapley values in explainable AI by using polynomial regression instead of linear approximation to reduce computational cost.
Mechanistic interpretability of transformer token associations through gradient analysis revealing semantic learning mechanisms.
SynCABEL: LLM-generated synthetic training data framework for biomedical entity linking without manual annotation.
Multi-LLM context learning method (M2CL) addressing discussion inconsistency in multi-agent collaborative problem-solving.
Hierarchical vision-language learning for fine-grained understanding of long captions using part-to-whole composition analysis.
DiscoverLLM: LLM framework for discovering user intents through interactive exploration rather than clarification questions.
ZKBoost: zero-knowledge proof protocol for verifying XGBoost model training integrity without revealing data or parameters.
Uncertainty-aware policy steering for robot deployment using learned verifiers and vision-language models to select safe actions.
Interactive Benchmarks: unified evaluation paradigm assessing model reasoning through adaptive information acquisition rather than fixed benchmarks.
Benchmark and evaluation framework for multi-party sequential negotiation games with configurable generator and baseline solvers.