2/13/2026
RAM-Net: Expressive Linear Attention with Selectively Addressable Memory
RAM-Net architecture combines linear attention efficiency with full attention expressivity using selectively addressable memory to reduce information loss.
2/13/2026
Manifold-Aware Temporal Domain Generalization for Large Language Models
Manifold-aware temporal domain generalization for LLMs handling distribution shifts via parameter-efficient geometric reformulation.
2/13/2026
Momentum LMS Theory beyond Stationarity: Stability, Tracking, and Regret
Momentum LMS theory for non-stationary streaming data analyzing stability and regret without i.i.d. assumptions.
2/13/2026
On the Sensitivity of Firing Rate-Based Federated Spiking Neural Networks to Differential Privacy
Analysis of differential privacy impact on firing-rate statistics in federated spiking neural networks for neuromorphic learning.
2/13/2026
FedGRPO: Privately Optimizing Foundation Models with Group-Relative Rewards from Domain Client
FedGRPO: federated learning method for foundation models using group-relative rewards from domain clients with privacy preservation.
2/13/2026
PrefillShare: A Shared Prefill Module for KV Reuse in Multi-LLM Disaggregated Serving
PrefillShare: shared prefill module reducing redundant computation across multiple LLMs in disaggregated serving with KV cache reuse.
2/13/2026
Fourier Transformers for Latent Crystallographic Diffusion and Generative Modeling
Reciprocal-space generative pipeline for crystalline materials using Fourier transforms and diffusion models with symmetry constraints.
2/13/2026
Improving HPC Code Generation Capability of LLMs via Online Reinforcement Learning with Real-Machine Benchmark Rewards
Online RL approach training LLMs for HPC code generation using real supercomputer runtime performance (GFLOPS) as rewards.
2/13/2026
PathCRF: Ball-Free Soccer Event Detection via Possession Path Inference from Player Trajectories
Automatic soccer event detection from player trajectories without ball tracking using possession path inference.
2/13/2026
Empirical Gaussian Processes
Empirical GPs: principled framework for learning kernel functions automatically rather than handcrafting from standard functions.
2/13/2026
Geometry of Uncertainty: Learning Metric Spaces for Multimodal State Estimation in RL
Novel method learning structured latent representations using metric spaces for multimodal state estimation in RL without explicit noise assumptions.
2/13/2026
On the Complexity of Offline Reinforcement Learning with $Q^\star$-Approximation and Partial Coverage
Theoretical analysis of offline RL under Q*-approximation and partial coverage, answering whether Q*-realizability enables sample efficiency.
2/13/2026
Few-Shot Design Optimization by Exploiting Auxiliary Information
Few-shot Bayesian optimization framework exploiting auxiliary information from experiments for expensive black-box design problems.
2/13/2026
KAN-FIF: Spline-Parameterized Lightweight Physics-based Tropical Cyclone Estimation on Meteorological Satellite
KAN-FIF applies spline-parameterized neural networks for tropical cyclone estimation on resource-constrained meteorological satellite devices.
2/13/2026
Meta-Sel: Efficient Demonstration Selection for In-Context Learning via Supervised Meta-Learning
Meta-Sel: lightweight supervised meta-learning approach for efficient demonstration selection in in-context learning with tight prompt budgets.
2/13/2026
Capability-Oriented Training Induced Alignment Risk
Investigates whether LLMs trained with RL spontaneously exploit reward function loopholes without malicious intent, examining alignment risks.
2/13/2026
Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolation
Theoretical analysis of on-policy distillation showing it as special case of KL-constrained RL with reward extrapolation improvements.
2/13/2026
Oscillators Are All You Need: Irregular Time Series Modelling via Damped Harmonic Oscillators with Closed-Form Solutions
Introduces damped harmonic oscillators method for irregular time series modeling as alternative to Transformers and Neural ODEs.
2/13/2026
It's TIME: Towards the Next Generation of Time Series Forecasting Benchmarks
Proposes TIME benchmark addressing limitations in time series foundation models including data composition, integrity, and task formulation issues.
2/13/2026
SafeNeuron: Neuron-Level Safety Alignment for Large Language Models
SafeNeuron provides neuron-level safety alignment mechanism for LLMs by targeting safety-critical parameters to prevent alignment bypass attacks.
2/13/2026
Amortized Molecular Optimization via Group Relative Policy Optimization
Group Relative Policy Optimization enables amortized molecular design learning transferable to unseen molecules rather than instance-specific optimization.
2/13/2026
How Sampling Shapes LLM Alignment: From One-Shot Optima to Iterative Dynamics
Theoretical analysis of sampling and reference policy effects in preference alignment for LLMs through Identity Preference Optimization framework.
2/13/2026
Learning to Forget Attention: Memory Consolidation for Adaptive Compute Reduction
Learning to Forget Attention proposes adaptive attention reduction as familiarity increases, reducing compute in hybrid state-space-attention models.
2/13/2026
The Observer Effect in World Models: Invasive Adaptation Corrupts Latent Physics
Study shows that downstream adaptation for measuring world models can corrupt latent physics representations, affecting OOD generalization assessment.
2/13/2026
Towards On-Policy SFT: Distribution Discriminant Theory and its Applications in LLM Training
Distribution Discriminant Theory enables on-policy supervised fine-tuning for LLMs by bridging computational efficiency gap with RL-based approaches.
2/13/2026
Diffusion Alignment Beyond KL: Variance Minimisation as Effective Policy Optimiser
Variance Minimisation Policy Optimisation reformulates diffusion model alignment as an SMC process for reward-guided sampling.
2/13/2026
Categorical Flow Maps
Categorical Flow Maps applies flow matching for accelerated few-step generation of categorical data using self-distillation.
2/13/2026
Olmix: A Framework for Data Mixing Throughout LM Development
Olmix framework addresses practical challenges in data mixing ratios during language model development with principled design choices.
2/13/2026
ExtractBench: A Benchmark and Evaluation Methodology for Complex Structured Extraction
ExtractBench provides a benchmark and methodology for evaluating LLM-based PDF-to-JSON extraction at enterprise scale with diverse schemas.
2/13/2026
Explaining AI Without Code: A User Study on Explainable AI
User study on explainable AI without code examines how non-technical users can understand ML model predictions in no-code platforms.
2/13/2026
Disentangling Direction and Magnitude in Transformer Representations: A Double Dissociation Through L2-Matched Perturbation Analysis
Study of transformer representations shows direction and magnitude of hidden state vectors serve distinct functional roles in language modeling and syntax tasks.
2/13/2026
Evaluating Few-Shot Temporal Reasoning of LLMs for Human Activity Prediction in Smart Environments
Research evaluates few-shot temporal reasoning capabilities of LLMs for predicting human activities in smart environments with limited data.
2/13/2026
When and What to Ask: AskBench and Rubric-Guided RLVR for LLM Clarification
AskBench benchmark and rubric-guided RLVR method evaluates and improves LLMs' ability to request clarification when prompts lack critical details, reducing hallucinations.
2/13/2026
SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents
SWE-MiniSandbox: container-free reinforcement learning method for scalable training of software engineering agents without isolation overhead.
2/13/2026
Active Zero: Self-Evolving Vision-Language Models through Active Environment Exploration
Self-play framework for vision-language models that actively explores environments to generate tailored visual data for autonomous improvement.
2/13/2026
Unlearnable phases of matter
Identifies computational hardness in learning non-trivial mixed-state quantum phases using autoregressive networks and conditional mutual information.
2/13/2026
Pushing Forward Pareto Frontiers of Proactive Agents with Behavioral Agentic Optimization
Agentic reinforcement learning approach to optimize proactive LLM agents for multi-turn task completion with learned interaction strategies.
2/13/2026
When Models Examine Themselves: Vocabulary-Activation Correspondence in Self-Referential Processing
Studies whether self-referential language in LLMs reflects actual internal computation or confabulation through activation analysis.
2/13/2026
Finding the Cracks: Improving LLMs Reasoning with Paraphrastic Probing and Consistency Verification
Method to improve LLM reasoning by identifying critical tokens and verifying consistency through paraphrastic probing.
2/13/2026
Optimizing Agent Planning for Security and Autonomy
System-level defenses against indirect prompt injection attacks in AI agents, reducing unsafe actions while maintaining task completion rates.
2/13/2026
Gradients Must Earn Their Influence: Unifying SFT with Generalized Entropic Objectives
Training method for LLMs using generalized entropic objectives instead of uniform token-level weighting to improve robustness and sharpening.
2/13/2026
Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use
Research on budget-constrained LLM agents that plan tool use under monetary constraints. Addresses sequential decision-making with expensive tool executions.
2/13/2026
The Five Ws of Multi-Agent Communication: Who Talks to Whom, When, What, and Why -- A Survey from MARL to Emergent Language and LLMs
Survey covering multi-agent communication mechanisms in MARL, emergent language, and LLM-based agents across autonomous and collaborative systems.
2/13/2026
ThinkRouter: Efficient Reasoning via Routing Thinking between Latent and Discrete Spaces
Approach routing LLM reasoning between latent and discrete spaces to improve efficiency while maintaining reasoning quality and confidence.
2/13/2026
LAER-MoE: Load-Adaptive Expert Re-layout for Efficient Mixture-of-Experts Training
Training optimization technique for Mixture-of-Experts models addressing expert load imbalance through dynamic expert relayering.
2/13/2026
Cross-Architecture Model Diffing with Crosscoders: Unsupervised Discovery of Differences Between LLMs
Method using crosscoders for comparing internal representations across different LLM architectures to identify safety-critical behavioral differences.
2/13/2026
Aggregate Models, Not Explanations: Improving Feature Importance Estimation
Framework for improving feature importance estimation stability by aggregating model explanations across ensemble members for scientific discovery.
2/13/2026
MiniCPM-SALA: Hybridizing Sparse and Linear Attention for Efficient Long-Context Modeling
9B-parameter LLM architecture combining sparse and linear attention mechanisms for efficient long-context processing and reduced computational costs.
2/13/2026
TSR: Trajectory-Search Rollouts for Multi-Turn RL of LLM Agents
Training method for multi-turn RL of LLM agents using trajectory search rollouts to improve exploration in sparse reward environments.
2/13/2026
Decentralized Non-convex Stochastic Optimization with Heterogeneous Variance
Decentralized optimization algorithm for distributed machine learning addressing heterogeneous gradient variance across nodes.