Applies policy gradient with decision transformers for reliable path planning in stochastic transportation networks with uncertain travel times.
Uses LLM reasoning to improve decision tree induction for tabular data, balancing interpretability with performance over black-box foundation models.
Examines data leakage issues in ML models for bearing fault diagnosis, highlighting methodological flaws preventing real-world generalization in condition monitoring.
Method to train small open-weight advisor models that generate dynamic prompts to improve black-box LLM performance. Demonstrates 27.4% improvement on GPT-5.2 tax tasks.
Graph-based gradient boosting approach for insurance fraud detection using inductive inference on networked claim data.
Strategy for efficiently pre-training large LLMs by orthogonally expanding Mixture-of-Experts parameters from existing checkpoints.
Post-training method enabling LLMs to improve reasoning without external rewards through self-generation and self-training.
LLM-based semantic optimization approach for expensive black-box problems incorporating domain knowledge and heuristics.
Regularization technique for neural network pruning that improves robustness under aggressive sparsity levels.
Analysis of optimization dynamics comparing preference optimization methods like DPO and PPO for LLM alignment.
Graph-regularized sparse autoencoders for extracting and steering safety behaviors in LLMs during inference.
LLM agent that learns graph construction rules from topology data without feature access or task knowledge for CAD applications.
Language model based approach for predicting resource consumption and runtime in cloud workflows using semi-structured job configurations.
Method for extreme 1-bit quantization of LLMs using output alignment techniques to enable deployment on resource-constrained devices.
Studies representation misdirection in LLM unlearning revealing controllable side behaviors via latent vectors
Foundation model for zero-shot feature importance estimation in tabular data without model access
Fingerprinting method to detect when student LLMs trained on teacher model outputs via distillation
Vector quantized attention layer balancing efficiency and performance for long context language models
Bit-plane decomposition quantization enabling 2-3 bit inference for large language models
Geometric anchoring approach for DPO addressing reference policy miscalibration in LLM alignment
PEPO algorithm for Direct Preference Optimization mitigating over-optimization without reward models
Unified system combining reinforcement learning with adaptive speculative decoding for efficient LLM serving
Chain-of-thought compression method for LLMs achieving high-fidelity reasoning with extreme compression ratios
Explains neural network compression using algorithmic complexity lens via mosaic-of-motifs pattern analysis
Neural PDE surrogates using spectral methods to reduce error accumulation in long-horizon predictions
Layer gradient analysis method for identifying optimal layers to edit in LLMs, improving knowledge editing performance across samples.
Stock prediction combining node transformers with BERT sentiment analysis. Financial ML application with limited novelty.
Study of how LLMs detect and identify injected steering vectors in residual streams, revealing introspective awareness mechanisms.
Research on internalizing skills within LLM agents through in-context reinforcement learning, reducing token overhead and retrieval noise.
DMax paradigm for efficient parallel decoding in diffusion language models via progressive self-refinement, reducing error accumulation.
Research on KV cache offloading to reduce memory and latency bottlenecks for long-context LLM inference tasks.
Benchmark for evaluating LLMs' ability to reason about differential privacy algorithms, lowering barrier for DP application design.
Research on unlearning sensitive information from LLMs via controllable alignment prompting without modifying weights, addressing regulatory compliance.
Benchmark comparing model scaling effects across 26 molecular property prediction tasks, finding compact and task-specific models often outperform large foundation models.
Gradient-based acquisition criterion for pool-based active learning derived from generalization bounds.
Multi-scale MLP architecture for long-term time series forecasting capturing patterns at multiple temporal resolutions.
System for serving mixture-of-experts LLMs on prefill-only workloads without redundant distributed execution overhead.
Investigation of why transformers fail at counting tasks and demonstration that failures stem from output representation rather than internal counting inability.
Physics-embedded rectified flow generative model for reconstructing PDE-governed spatiotemporal fields with uncertainty quantification from sparse measurements.
Active learning acquisition rule using pretrained MLIP latent representations to reduce quantum chemical labeling costs in molecular potential training.
State vector space partitioning method to distill black-box RL policies into interpretable linear subpolicies using support vector machines.
Pass-rate control method for binary-reward RL in software engineering tasks, steering trajectories toward maximally informative regimes for policy optimization.
Technique for estimating expected output of wide random MLPs at initialization without sampling, using activation distribution approximations.
Domain transfer method for simulation-based inference under model misspecification using unlabeled real-world data.
Universal semi-supervised learning framework handling arbitrary unlabeled data distributions using simplex anchored structural inference.
Theoretical analysis of how models learn sparse data regimes where the manifold hypothesis does not apply.
Meta-learning approach framing root cause analysis in complex systems as Bayesian inference over synthetic causal model priors.
Technique using spectral orthogonal gradient projection to address catastrophic forgetting in LLM continual learning through orthogonal parameter updates.
Method for accelerating zeroth-order optimization in LLM fine-tuning via partial orthogonalization, enabling efficient adaptation on edge devices without backpropagation.
Research on how FFN architecture choices in Transformers affect attention mechanisms, comparing dense, gated, and sparse mixture-of-experts designs on algorithmic tasks.