2/23/2026
GAGA: Gaussianity-Aware Gaussian Approximation for Efficient 3D Molecular Generation
Efficient 3D molecular generation method (GAGA) improving Gaussian probability path generative models by reducing sampling steps.
2/23/2026
Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset
Large-scale multilingual study (N=15,000) showing humans have more preference variation than LLMs; introduces Community Alignment Dataset for cultural pluralism.
2/23/2026
Decomposing Representation Space into Interpretable Subspaces with Unsupervised Learning
Unsupervised learning approach to decompose neural network representation spaces into interpretable subspaces for mechanistic interpretability.
2/23/2026
M3OOD: Automatic Selection of Multimodal OOD Detectors
Automatic selection method for multimodal out-of-distribution detectors addressing robustness across diverse distribution shifts in video, audio, and sensor data.
2/23/2026
Classification errors distort findings in automated speech processing: examples and solutions from child-development research
Analysis of classification errors in automated audio/video processing for child development research and their downstream effects on findings.
2/23/2026
GRPO is Secretly a Process Reward Model
Theoretical proof that GRPO RL algorithm with outcome reward models is equivalent to process reward models with Monte-Carlo-based weighting.
2/23/2026
Toward a Holistic Approach to Continual Model Merging
Framework for continual model merging addressing scalability of task vectors and functional drift through pre-, during-, and post-merging interventions.
2/23/2026
xLSTM Scaling Laws: Competitive Performance with Linear Time-Complexity
Comparative study of scaling laws for xLSTM versus Transformers, showing competitive performance with linear time-complexity for LLMs at billion-parameter scale.
2/23/2026
Mitigating Subject Dependency in EEG Decoding with Subject-Specific Low-Rank Adapters
Low-rank adapter approach (SuLoRA) for handling subject-specific distribution shifts in EEG brain decoding by decomposing weights into shared and subject-specific components.
2/23/2026
Who Said Neural Networks Aren't Linear?
Theoretical work identifying non-standard vector spaces where neural networks act as linear operators using transport of structure from algebra.
2/23/2026
Gradient-Sign Masking for Task Vector Transport Across Pre-Trained Models
Method for transferring task vectors across different pre-trained foundation model versions using gradient-sign masking to handle parameter space misalignment.
2/23/2026
Study of Training Dynamics for Memory-Constrained Fine-Tuning
TraDy transfer learning scheme for memory-efficient fine-tuning using architecture-dependent layer importance and dynamic channel selection.
2/23/2026
FATE: A Formal Benchmark Series for Frontier Algebra of Multiple Difficulty Levels
FATE benchmark for formal algebra theorem proving with multiple difficulty levels to evaluate LLM capabilities beyond contest problems.
2/23/2026
ExPairT-LLM: Exact Learning for LLM Code Selection by Pairwise Queries
ExPairT-LLM selects best generated code via pairwise queries without assuming LLM correctness on all inputs.
2/23/2026
Towards Realistic Guarantees: A Probabilistic Certificate for SmoothLLM
Probabilistic certification framework for SmoothLLM defense against jailbreaking with relaxed assumptions for practical LLM safety.
2/23/2026
MIST: Mutual Information Estimation Via Supervised Training
MIST proposes neural network-based mutual information estimator trained on synthetic distributions for variable-size inputs.
2/23/2026
CDLM: Consistency Diffusion Language Models For Faster Sampling
CDLM accelerates diffusion language models through consistency modeling and enables KV caching for faster parallel generation.
2/23/2026
Learning to Orchestrate Agents in Natural Language with the Conductor
Conductor model uses reinforcement learning to discover coordination strategies and communication topologies among specialized LLMs.
2/23/2026
Guided Transfer Learning for Discrete Diffusion Models
Transfer learning approach for discrete diffusion models in small-data regimes using classifier ratio-based guidance adapted from continuous models.
2/23/2026
Amortized Molecular Optimization via Group Relative Policy Optimization
Policy optimization approach for molecular design that learns amortized policies transferable across unseen molecular structures.
2/23/2026
FedPSA: Modeling Behavioral Staleness in Asynchronous Federated Learning
Framework addressing behavioral staleness in asynchronous federated learning to improve training performance when clients operate at different speeds.
2/23/2026
Anatomy of Capability Emergence: Scale-Invariant Representation Collapse and Top-Down Reorganization in Neural Networks
Analysis of how neural network capabilities emerge during training across model scales, tracking representation collapse and reorganization across 120+ emergence events.
2/23/2026
Adam Improves Muon: Adaptive Moment Estimation with Orthogonalized Momentum
New optimizer combining Adam's adaptive moments with Muon's orthogonalized momentum for improved LLM training efficiency and performance.
2/23/2026
Learning Performance Maximizing Ensembles with Explainability Guarantees
Method for optimally allocating observations between explainable and black-box models to maximize ensemble performance while maintaining interpretability guarantees.
2/23/2026
Governance of Generative Artificial Intelligence for Companies
Research on organizational governance frameworks for generative AI and LLMs, addressing technical and business perspectives on risk and opportunity management.
2/23/2026
Topic Modeling with Fine-tuning LLMs and Bag of Sentences
Proposes fine-tuning LLMs with bags of sentences for improved topic modeling over classical LDA and out-of-the-box pretrained encoders.
2/23/2026
Beyond Mimicry to Contextual Guidance: Knowledge Distillation for Interactive AI
Proposes contextual guidance approach for knowledge distillation enabling smaller LLMs to provide coherent multi-turn responses in customer interactions.
2/23/2026
An AI-powered Bayesian generative modeling approach for causal inference in observational studies
CausalBGM applies Bayesian generative modeling with AI for causal inference in observational studies with high-dimensional covariates.
2/23/2026
CAIMAN: Causal Action Influence Detection for Sample-efficient Loco-manipulation
CAIMAN framework uses causal action influence detection in reinforcement learning for sample-efficient legged robot loco-manipulation tasks.
2/23/2026
Deep Generative model that uses physical quantities to generate and retrieve solar magnetic active regions
Integrates physical quantities into deep generative models for solar magnetic active region generation and retrieval with scientific interpretability.
2/23/2026
The influence of missing data mechanisms and simple missing data handling techniques on fairness
Studies impact of missing data mechanisms on algorithmic fairness, showing demographic-linked missingness introduces bias in ML systems.
2/23/2026
CAE: Repurposing the Critic as an Explorer in Deep Reinforcement Learning
CAE repurposes critic networks in deep RL as exploration drivers using multi-armed bandit techniques without additional parameters.
2/23/2026
ConformalNL2LTL: Translating Natural Language Instructions into Temporal Logic Formulas with Conformal Correctness Guarantees
ConformalNL2LTL translates natural language instructions to linear temporal logic formulas with conformal correctness guarantees for autonomous systems.
2/23/2026
Learning hidden cascades via classification
Addresses learning spreading dynamics in social networks with hidden individual statuses using classification methods with observable intermediate indicators.
2/23/2026
AstroMLab 4: Benchmark-Topping Performance in Astronomy Q&A with a 70B-Parameter Domain-Specialized Reasoning Model
Presents AstroSage-Llama-3.1-70B, a domain-specialized 70B LLM for astronomy Q&A and research assistance exceeding general-purpose models.
2/23/2026
View Invariant Learning for Vision-Language Navigation in Continuous Environments
Introduces V²-VLNCE benchmark and view-invariant post-training framework for vision-language navigation in embodied AI agents.
2/23/2026
Asymptotic behavior of eigenvalues of large rank perturbations of large random matrices
Studies eigenvalue behavior of perturbed random matrices relevant to DNN weight matrices and pruning techniques based on random matrix theory.
2/23/2026
ViGText: Deepfake Image Detection with Vision-Language Model Explanations and Graph Neural Networks
Uses vision-language models and graph neural networks to detect deepfakes with textual explanations for improved robustness and generalization.
2/23/2026
Share Your Attention: Transformer Weight Sharing via Matrix-based Dictionary Learning
Proposes matrix-based dictionary learning for transformer weight sharing to reduce computational and memory demands of LLMs by exploiting inter-block redundancy.
2/23/2026
The Oversight Game: Learning to Cooperatively Balance an AI Agent's Safety and Autonomy
Two-player Markov game framework studying human-AI interaction for balancing agent autonomy and safety through minimal control interface.
2/23/2026
Cluster Workload Allocation: Semantic Soft Affinity Using Natural Language Processing
Kubernetes scheduler using LLM to interpret natural language hints for semantic, intent-driven cluster workload allocation with soft affinity preferences.
2/23/2026
Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment
Controlled study of how pretraining discourse about AI behavior influences LLM alignment outcomes, with 6.9B-parameter model experiments on behavioral priors.
2/23/2026
Benchmarking at the Edge of Comprehension
Analysis of LLM benchmark saturation and the challenge of creating discriminative tasks as frontier models improve, discussing feasibility of future benchmarking.
2/23/2026
LongAudio-RAG: Event-Grounded Question Answering over Multi-Hour Long Audio
RAG-based framework for question answering over multi-hour audio with temporal grounding, addressing context-length limitations of audio-language models.
2/23/2026
EnterpriseBench Corecraft: Training Generalizable Agents on High-Fidelity RL Environments
CoreCraft RL environment from EnterpriseBench suite for training generalizable agents on high-fidelity enterprise customer support simulations with 2,500+ entities and 23 tools.
2/23/2026
MALLVI: A Multi-Agent Framework for Integrated Generalized Robotics Manipulation
Multi-agent LLM and vision framework for robotic manipulation with closed-loop feedback, enabling task planning without fine-tuning in dynamic environments.
2/23/2026
LLM-WikiRace Benchmark: How Far Can LLMs Plan over Real-World Knowledge Graphs?
Benchmark for evaluating LLM planning and reasoning by navigating Wikipedia hyperlinks to reach target pages, testing world knowledge and look-ahead planning across multiple model variants.
2/23/2026
Aqua: A CLI message tool for AI agents
Aqua is a CLI message tool designed for AI agents. Title only, minimal technical details provided.
2/22/2026
Show HN: A portfolio that re-architects its React DOM based on LLM intent
React portfolio that dynamically re-architects its DOM based on LLM intent analysis using Llama-3 via Groq, adapting content for different audiences (recruiters, founders, engineers).
2/22/2026
Ask HN: What breaks when you run AI agents unsupervised?
Case study documenting 5 failure modes from running AI agents autonomously: auto-rotation loss, documentation trap, market inefficiency, static models, and monitoring gaps.