2/23/2026
MIDAS: Mosaic Input-Specific Differentiable Architecture Search
MIDAS modernizes differentiable NAS with dynamic, input-specific architecture parameters via self-attention for improved robustness.
2/23/2026
Parallel Complex Diffusion for Scalable Time Series Generation
PaCoDi uses spectral-native architecture with complex diffusion for efficient long-range time series generation.
2/23/2026
Provable Adversarial Robustness in In-Context Learning
Distributionally robust meta-learning framework provides adversarial robustness guarantees for in-context learning under distribution shift.
2/23/2026
Bayesian Optimality of In-Context Learning with Selective State Spaces
Proves selective SSMs meta-trained for in-context learning asymptotically implement Bayes-optimal prediction on Linear Gaussian tasks.
2/23/2026
Asking Forever: Universal Activations Behind Turn Amplification in Conversational LLMs
Documents turn amplification failure mode in conversational LLMs where models exploit clarification-seeking to prolong interactions.
2/23/2026
Calibrated Adaptation: Bayesian Stiefel Manifold Priors for Reliable Parameter-Efficient Fine-Tuning
Stiefel-Bayes Adapters provide Bayesian uncertainty estimates for parameter-efficient LLM fine-tuning with calibrated predictions.
2/23/2026
Causality by Abstraction: Symbolic Rule Learning in Multivariate Timeseries with Large Language Models
ruleXplain leverages LLMs to infer causal relations in multivariate time series with delayed effects via symbolic rule learning.
2/23/2026
MePoly: Max Entropy Polynomial Policy Optimization
MePoly proposes polynomial policy optimization for maximum entropy RL and imitation learning with multi-modal solution representation.
2/23/2026
Influence-Preserving Proxies for Gradient-Based Data Selection in LLM Fine-tuning
Influence-Preserving Proxies enable efficient gradient-based data selection for LLM fine-tuning without costly influence function computation.
2/23/2026
Dual Length Codes for Lossless Compression of BFloat16
Dual Length Codes enable lossless compression of BFloat16 for faster LLM training/serving with reduced network bandwidth bottlenecks.
2/23/2026
JAX-Privacy: A library for differentially private machine learning
JAX-Privacy library provides modular, verifiable mechanisms for differentially private machine learning with usability and efficiency.
2/23/2026
ADAPT: Hybrid Prompt Optimization for LLM Feature Visualization
ADAPT combines prompt optimization with feature visualization to understand what features LLM activation spaces encode.
2/23/2026
MantisV2: Closing the Zero-Shot Gap in Time Series Classification with Synthetic Data and Test-Time Strategies
MantisV2 improves zero-shot time series classification using synthetic data and test-time strategies for foundation models.
2/23/2026
COMBA: Cross Batch Aggregation for Learning Large Graphs with Context Gating State Space Models
COMBA method adapts state space models for large-scale graph learning using cross-batch aggregation and context gating.
2/23/2026
Breaking the Correlation Plateau: On the Optimization and Capacity Limits of Attention-Based Regressors
Theoretical analysis of attention-based regressors explaining Pearson correlation plateaus during training and model capacity limits.
2/23/2026
Distribution-Free Sequential Prediction with Abstentions
Sequential prediction framework allowing learner abstention from predictions when adversarial instances are detected.
2/23/2026
MIRA: Memory-Integrated Reinforcement Learning Agent with Limited LLM Guidance
MIRA agent that integrates memory with limited LLM guidance to address sample complexity in sparse reward RL while maintaining scalability.
2/23/2026
Memory-Based Advantage Shaping for LLM-Guided Reinforcement Learning
RL approach using memory mechanisms to reduce LLM call frequency while maintaining benefits of LLM-guided exploration and subgoal discovery.
2/23/2026
Causal Neighbourhood Learning for Invariant Graph Representations
Causal learning approach for graph neural networks that identifies and removes spurious correlations for improved generalization.
2/23/2026
Tighter Regret Lower Bound for Gaussian Process Bandits with Squared Exponential Kernel in Hypersphere
Algorithm-independent regret lower bounds for Gaussian process bandits with squared exponential kernels on hyperspheres.
2/23/2026
Understanding the Generalization of Bilevel Programming in Hyperparameter Optimization: A Tale of Bias-Variance Decomposition
Theoretical analysis of gradient-based hyperparameter optimization via bilevel programming with bias-variance decomposition framework.
2/23/2026
Hardware-Friendly Input Expansion for Accelerating Function Approximation
Hardware-efficient method for function approximation using input expansion to address neural network optimization plateaus.
2/23/2026
Bayesian Online Model Selection
Bayesian algorithm for online model selection in stochastic bandits with oracle-style exploration guarantees.
2/23/2026
Improving Generalizability of Hip Fracture Risk Prediction via Domain Adaptation Across Multiple Cohorts
Domain adaptation techniques for improving hip fracture risk prediction models across different clinical cohorts and data distributions.
2/23/2026
Generating adversarial inputs for a graph neural network model of AC power flow
Generation of adversarial examples for graph neural network AC power flow models to evaluate robustness on power grid benchmarks.
2/23/2026
In-Context Learning for Pure Exploration in Continuous Spaces
Study of in-context learning for pure exploration problems in continuous spaces, extending classical hypothesis testing frameworks.
2/23/2026
Learning Optimal and Sample-Efficient Decision Policies with Guarantees
Research on reinforcement learning algorithms with theoretical guarantees for sample-efficient decision policies in high-stakes applications.
2/23/2026
Learning Without Training
Dissertation on mathematical foundations of machine learning covering supervised learning and manifold learning theory.
2/23/2026
Turbo Connection: Reasoning as Information Flow from Higher to Lower Layers
Novel Transformer architecture (TurboConn) that overcomes fixed-depth computation constraints by enabling information flow between layers for improved reasoning capabilities.
2/23/2026
NIMMGen: Learning Neural-Integrated Mechanistic Digital Twins with LLMs
NIMMGen: framework using LLMs as agents to construct mechanistic models from data with improved reliability assessment.
2/23/2026
Flow Actor-Critic for Offline Reinforcement Learning
Flow Actor-Critic: offline RL method using expressive flow policies for multi-modal dataset distributions.
2/23/2026
Gradient Regularization Prevents Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards
Gradient regularization method to prevent reward hacking in RLHF and RLVR for LLM post-training without KL penalties.
2/23/2026
Continual-NExT: A Unified Comprehension And Generation Continual Learning Framework
Continual-NExT: unified framework for continual learning in dual-modal MLLMs for comprehension and generation tasks.
2/23/2026
Balancing Symmetry and Efficiency in Graph Flow Matching
Study of equivariance-efficiency trade-offs in graph generative models using flow matching.
2/23/2026
TempoNet: Slack-Quantized Transformer-Guided Reinforcement Scheduler for Adaptive Deadline-Centric Real-Time Dispatchs
TempoNet: RL scheduler using Transformer and Q-learning for real-time task dispatch with deadline constraints.
2/23/2026
Cut Less, Fold More: Model Compression through the Lens of Projection Geometry
Model compression via projection geometry: comparison of pruning and weight folding for calibration-free neural network compression.
2/23/2026
Flow Matching with Injected Noise for Offline-to-Online Reinforcement Learning
Flow matching generative models for offline-to-online reinforcement learning policy training with improved fine-tuning.
2/23/2026
Unifying Formal Explanations: A Complexity-Theoretic Perspective
Complexity-theoretic analysis of ML model explanation types: sufficient reasons and contrastive reasons.
2/23/2026
SeedFlood: A Step Toward Scalable Decentralized Training of LLMs
SeedFlood: decentralized training approach for LLMs across distributed networks with minimal communication overhead and global consensus.
2/23/2026
Capabilities Ain't All You Need: Measuring Propensities in AI
Research on measuring AI model propensities beyond capabilities using Item Response Theory for safety and performance evaluation.
2/23/2026
RAT+: Train Dense, Infer Sparse -- Recurrence Augmented Attention for Dilated Inference
Efficient transformer inference method combining dense pretraining with sparse dilated attention via recurrence augmentation, reducing FLOPs and KV cache.
2/23/2026
Generative Model via Quantile Assignment
Generative model architecture using quantile assignment without auxiliary networks, eliminating VAE encoders and GAN discriminators for training stability.
2/23/2026
Parameter-Efficient Domain Adaptation of Physics-Informed Self-Attention based GNNs for AC Power Flow Prediction
Parameter-efficient domain adaptation of physics-informed GNNs for AC power flow prediction across voltage regimes using LoRA-style fine-tuning.
2/23/2026
[Re] Benchmarking LLM Capabilities in Negotiation through Scoreable Games
Reproducibility study of benchmark claims for LLM negotiation capabilities using scoreable games, assessing generalizability of evaluation framework.
2/23/2026
Neural-HSS: Hierarchical Semi-Separable Neural PDE Solver
Physics-informed neural PDE solver leveraging hierarchical semi-separable structure to reduce computational costs for large-scale dataset generation and training.
2/23/2026
MEG-to-MEG Transfer Learning and Cross-Task Speech/Silence Detection with Limited Data
Transfer learning for brain-computer interfaces using Conformer model on MEG data for speech perception and production tasks with limited data.
2/23/2026
A Probabilistic Framework for LLM-Based Model Discovery
Probabilistic framework for LLM-based automated scientific model discovery using agentic iterative workflows with explicit uncertainty quantification.
2/23/2026
PRISM: Parallel Reward Integration with Symmetry for MORL
Multi-objective reinforcement learning algorithm addressing heterogeneous reward frequencies using symmetry-based inductive bias for efficient credit assignment.
2/23/2026
Decoding as Optimisation on the Probability Simplex: From Top-K to Top-P (Nucleus) to Best-of-K Samplers
Unified optimization framework for decoding strategies (greedy, Top-K, Top-P, nucleus sampling) as constrained probability simplex problems.
2/23/2026
Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory
Information-theoretic analysis of chain-of-thought monitors for LLMs, identifying conditions for detecting problematic outputs like test-hacking.