Hexagonal coordinate system for efficient world models in adaptive AI, inspired by grid cells in human brain. Mathematical framework for rotational symmetry and low-cost spatial computation.
RubricBench benchmark for evaluating rubric-guided LLM reward models against human standards, addressing discriminative complexity in alignment evaluation.
Nano-EmoX framework unifying multimodal emotional intelligence across perception, understanding, and interaction levels with cognitively-inspired hierarchy.
Geometric theory formalizing alignment tax as projection in representation space, deriving Pareto frontier for safety-capability tradeoffs in LLMs.
SDK providing LLM-agents structured data access to scientific literature via agentic interface, reducing token consumption and improving retrieval efficiency.
Theory of Code Space benchmark evaluating whether AI code agents understand software architecture through multi-file codebase exploration in procedurally generated environments.
IDER method addressing catastrophic forgetting in continual learning through idempotent experience replay with uncertainty calibration.
FastCode system for efficient repository-scale code reasoning using selective context retrieval and compression for cost-effective LLM-based software engineering.
Proposes QIME framework for interpretable biomedical text embeddings using ontology-grounded natural language questions for clinical decision-making.
Compares numerical methods with physics-informed neural networks for solving direct and inverse PDE problems in physical/engineering systems.
Studies policy diversity in ensemble policy gradient methods for large-scale RL, analyzing exploration-exploitation tradeoffs across parallel environments.
Hyperparameter trajectory inference framework using conditional Lagrangian optimal transport to enable post-deployment hyperparameter adjustments without retraining.
Speech bandwidth extension method using conditional flow matching in neural codec latent space for improved clarity and intelligibility.
Evaluates multimodal GUI agents' ability to identify and execute toggle controls, revealing reliability bottlenecks in ubiquitous GUI interaction.
Reinterprets LLM softmax as energy-based model to track 'energy spills' during decoding, correlating them with factual errors and biases.
Proposes ANSE method using Bayesian active noise selection with attention mechanisms to improve video diffusion quality by selecting optimal initial noise seeds.
Research on classifier-free guidance scale annealing in diffusion models to improve image quality and prompt alignment convergence during sampling.
WebDevJudge benchmark evaluates LLMs-as-judges for web development quality assessment, testing reliability on open-ended tasks with dynamic environments.
RxnNano trains compact LLMs for chemical reaction prediction using hierarchical curriculum learning, emphasizing chemical intuition over parameter scaling.
ATPO uses hierarchical reinforcement learning to optimize LLM behavior for multi-turn medical dialogues with incomplete information.
Analysis of MoE compression methods identifies router-expert mismatch as key degradation factor; proposes calibration approach for efficient model deployment.
Research on self-play loops in LLMs showing sustainable self-evolution requires learnable information gain, not just more synthetic data generation.
NExT-Guard provides training-free safeguarding for streaming LLM deployments without requiring token-level annotations or supervision.
MedFeat integrates LLM domain knowledge into feature engineering for clinical tabular prediction, balancing model characteristics with feature importance signals.
Audit of MedCalc-Bench clinical calculator benchmark reveals implementation issues and proposes open-book evaluation methodology for more accurate LLM assessment.
ML research paper using correspondence analysis, clustering, and classification to model wildfire evacuation behavior from survey data.
Geometric theory of catastrophic forgetting in LoRA through gradient subspace interactions, deriving quantitative forgetting formula.
Unsupervised reward modeling scaling via preference learning on web document prefixes/suffixes, reducing human annotation costs.
Efficient RNN architecture with selective state updates for long-range sequence modeling, reducing unnecessary computation on static inputs.
Neural Paging architecture enabling Turing-complete agents by learning hierarchical context window management policies.
Study of safety training persistence in multi-step agentic LLM settings when optimizing for helpfulness, comparing DPO effects.
Generalized discrete diffusion model with self-correction during pretraining using uniform-absorbing objective.
Principled mathematical framework for reward modeling leveraging ordinal preference feedback from human annotators for LLM alignment.
Personalized federated learning approach using kernel mean embeddings to learn inter-agent weight combinations without raw data sharing.
Framework for automatic specification generation to improve neural network verification tool adoption by supporting higher-level semantic constraints.
CUDABench benchmark for evaluating LLM text-to-CUDA code generation with performance assessment metrics for GPU kernels.
Method for steering LLM behavior via representation manipulation that accounts for heterogeneous concept encoding across embedding spaces.
Theoretical analysis of length generalization bounds for transformers on CRASP language class, addressing model generalization guarantees.
Label-guided distance scaling method for few-shot text classification, improving meta-learner effectiveness with selective label guidance.
Quantum-enhanced LoRA fine-tuning method for few-shot AI-generated content detection, combining quantum neural networks with low-rank adaptation.
Preconditioning techniques for flow matching and score-based diffusion to improve optimization by handling ill-conditioned covariance matrices.
Diffusion-based model predictive control with discrete denoising for game playing, tested on Tetris with feasibility constraints and critic alignment.
Proposes rigidity-aware geometric pretraining for protein design and conformational ensembles using global geometric representations.
Studies personalized multi-agent average reward TD learning with joint linear approximation, inspired by federated learning approaches.
Analyzes temperature parameter selection in knowledge distillation and its interaction with optimizer, pretraining, and finetuning choices.
Introduces loss-level spectral regularization using Fourier and wavelet-domain losses to improve diffusion model training without architecture changes.
Studies computational reducibility in neural solvers for graph combinatorial optimization, enabling model generalization across task distributions.
MUSE is an open-source platform for multimodal safety evaluation of LLMs with cross-modal payload generation and multi-turn attack algorithms.
Proves selection theorems showing that low average-case regret forces AI agents to develop internal world models or belief states for robust decision-making.
ParEVO uses LLM-based agentic evolution to synthesize parallel code for irregular data structures, addressing limitations of standard models on concurrent programming.