Research paper proposing temporal, confidence-weighted, relational metadata for vector embeddings in RAG systems to improve accuracy on versioned queries.
pAI/MSc: Open-source multi-agent system for automating academic ML research workflows from hypothesis to manuscript draft with human guidance.
Auto-ART framework synthesizes adversarial robustness research across 2020-2026 and provides automated testing protocols to detect gradient masking.
Introduces AAC, a differentiable landmark compression module for shortest-path heuristics with admissible-by-construction guarantees.
V-tableR1 framework uses process-supervised RL to train multimodal LLMs for rigorous multi-step table reasoning with verifiable rewards.
Proposes semantic stratification method to improve retrieval evaluation for RAG systems by addressing bias in query set construction and metric reliability.
Integrates working memory constraints into transformers via attention variants, evaluates on language understanding tasks.
Identifies convergent evolution of periodic number representations across different LLM architectures and training methods.
Analyzes value function interference and overestimation issues in multi-objective reinforcement learning algorithms.
Optimizes RMSNorm computation in LLMs by eliminating normalization weights through matrix folding for parallel execution.
Extends certified unlearning methods from convex models to deep neural networks with theoretical guarantees.
Demonstrates vulnerabilities in machine unlearning verification strategies used to validate data removal from models.
Proposes mixed-precision quantization for LLMs that varies precision across output features for efficient compression.
Studies best policy identification in preference-based RL as alternative to learned reward models for alignment.
Introduces recency bias mechanism for transformer attention in time-series forecasting via reweighted attention scores.
Proposes PODS to decouple rollout generation from policy updates in LLM reinforcement learning with verifiable rewards.
Improves advantage estimation in language model RL by applying Kalman filtering to Group Relative Policy Optimization.
Examines interpretability definitions in scientific ML and barriers to integrating neural network findings into scientific knowledge.
Extends stochastic interpolants framework to latent space with joint optimization of encoder-decoder for generative modeling.
Theoretical analysis of transformer expressivity and learnability through the lens of universal simulators of attention mechanisms.
Theoretical analysis of offline reinforcement learning in average-reward MDPs with sample complexity guarantees under single-policy data coverage.
Dataset and evaluation methodology for assessing fairness in federated learning at client level, addressing discrimination across distributed clients.
POLIS framework enabling heterogeneous LLM agents to accumulate knowledge through interaction and emergent communication, modeling cumulative cultural evolution.
Federated learning framework addressing privacy and stragglers by supporting user opt-out and handling missing data from heterogeneous devices.
KANMixer architecture using Kolmogorov-Arnold Networks for long-term time series forecasting with improved expressivity over MLP and Transformer baselines.
Research on diffusion models that incorporate local spatial structure into score functions for better generative modeling of spatially structured data.
Explanation method for Graph Neural Networks using LLMs to generate interpretable narratives from node-level decisions and graph context.
WISCA improves LLM training by systematically optimizing weight patterns through weight scaling transitions during transformer training.
Evaluates calibration metrics and recalibration methods for uncertainty estimation in data-driven regression models for safety-critical applications.
EvolveSignal uses LLM-powered coding agent to automatically discover traffic signal control strategies, replacing manual hand-crafted formulas.
Multi-Level Optimal Transport method for symmetric representational alignment across model layers and different network depths.
Distributional framework for offline inverse RL capturing richer expert behavior by modeling uncertainty over rewards and return distributions.
Applies reinforcement learning with GRPO to subject-driven image generation, addressing fidelity-editability trade-off via synergistic reward alignment.
Investigates how transformers learn latent structure through staged dynamics using controlled Alchemy benchmark tasks.
Hybrid-AIRL combines adversarial inverse RL with supervised expert guidance, evaluated on poker domain with imperfect information.
Unified theoretical framework for sparse dictionary learning in mechanistic interpretability, analyzing piecewise biconvexity and spurious minima in neural networks.
Novel framework combining mechanistic interpretability with gradient ascent for interpretable persona control in LLMs, improving on unscalable prompt engineering.
CEDAR demonstrates automated data science task solving using LLM agents via context engineering, addressing task complexity, data size, and computational constraints.
Theoretical analysis of language identification and generation without realizability assumptions, establishing statistical rates under relaxed conditions.
Few-shot learning framework for T cell receptor repertoire analysis using dynamic kernel codes and prototype-based parameterizations for disease detection.
Analysis of Rashomon sets in federated learning to understand model multiplicity and decision boundary instabilities affecting transparency and robustness.
Faithfulness-based explainability framework for analyzing generative diffusion models in medical MRI synthesis applications.
Study analyzing explainability in federated learning systems with differential privacy layers, examining privacy-interpretability trade-offs.
Training-free retrieval-augmented generation framework for flood damage nowcasting using reinforced reasoning with geographically local context retrieval.
Method for applying LLMs to graph problems using human-interpretable graph encodings that address the mismatch between text-based models and graph reasoning.
Adaptive model selection framework for demand forecasting that accounts for horizon-induced performance degradation in inventory planning.
SphUnc framework combining hyperspherical representation learning with structural causal modeling for uncertainty decomposition in multi-agent systems.
Study on why multimodal models underperform unimodal counterparts in context-aided forecasting, investigating context quality issues in datasets.
Framework using optimal transport and distributionally robust optimization to analyze ML model vulnerabilities through constrained data perturbations for explainability.
AutoML approach combining deep unfolding with learnable proximal gradient descent for wireless optimization.