2/17/2026
Benchmarking AI-based data assimilation to advance data-driven global weather forecasting
Benchmark for evaluating AI-based data assimilation methods using real-world observations for weather forecasting.
2/17/2026
MixLinear: Extreme Low Resource Multivariate Time Series Forecasting with 0.1K Parameters
Lightweight multivariate time series forecasting model with 0.1K parameters for resource-constrained settings.
2/17/2026
Model-based Large Language Model Customization as Service
Privacy-preserving LLM customization service using differentially private synthesis for domain-specific fine-tuning.
2/17/2026
Denoising diffusion probabilistic models are optimally adaptive to unknown low dimensionality
Theoretical analysis showing denoising diffusion models adapt optimally to unknown low-dimensional data.
2/17/2026
Lynx: Enabling Efficient MoE Inference through Dynamic Batch-Aware Expert Selection
Dynamic batch-aware expert selection for efficient Mixture-of-Expert model inference during serving.
2/17/2026
Cautious Optimizers: Improving Training with One Line of Code
One-line modification to momentum-based optimizers improving stability and speed for transformer pretraining.
2/17/2026
Regularized Top-$k$: A Bayesian Framework for Gradient Sparsification
Bayesian framework for gradient sparsification in distributed training using error accumulation.
2/17/2026
Adaptive Width Neural Networks
Technique to learn unbounded neural network layer width during training without manual hyperparameter selection.
2/17/2026
SWIFT: Mapping Sub-series with Wavelet Decomposition Improves Time Series Forecasting
Lightweight time series forecasting model using wavelet decomposition for resource-constrained environments.
2/17/2026
Faster Adaptive Optimization via Expected Gradient Outer Product Reparameterization
Studies rotational equivariance in adaptive optimizers and proposes reparameterization for improved convergence.
2/17/2026
LO-BCQ: Block Clustered Quantization for 4-bit (W4A4) LLM Inference
Block clustered quantization technique for 4-bit weight and activation LLM inference without quantization-aware training.
2/17/2026
Fenchel-Young Variational Learning
Introduces Fenchel-Young losses as generalizations of KL divergence for variational learning methods.
2/17/2026
Less is More: Improving LLM Alignment via Preference Data Selection
Novel data selection approach for Direct Preference Optimization to improve LLM alignment by addressing parameter shrinkage from noisy preference data.
2/17/2026
Robust Multi-Objective Controlled Decoding of Large Language Models
RMOD inference algorithm aligns LLMs to multiple objectives via robust maximin game formulation achieving Nash equilibrium.
2/17/2026
Learning Rate Annealing Improves Tuning Robustness in Stochastic Optimization
Learning rate annealing schedules improve robustness to hyperparameter tuning in stochastic gradient optimization.
2/17/2026
Riemannian Denoising Diffusion Probabilistic Models
Riemannian Denoising Diffusion Probabilistic Models for learning distributions on submanifolds without requiring explicit geodesic information.
2/17/2026
Sparse Latent Factor Forecaster (SLFF) with Iterative Inference for Transparent Multi-Horizon Commodity Futures Prediction
Sparse Latent Factor Forecaster with iterative inference for commodity futures prediction addressing amortized inference deployment gap.
2/17/2026
Residual Feature Integration is Sufficient to Prevent Negative Transfer
Residual feature integration prevents negative transfer in transfer learning through simple architectural modifications.
2/17/2026
MoESD: Unveil Speculative Decoding's Potential for Accelerating Sparse MoE
MoESD applies speculative decoding to accelerate Mixture of Experts LLM inference without accuracy loss.
2/17/2026
On the Relation between Rectified Flows and Optimal Transport
Theoretical analysis of connections between rectified flows, flow matching, and optimal transport with invariance properties.
2/17/2026
Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay
Difficulty-targeted online data selection and rollout replay techniques improve data efficiency in LLM reinforcement learning fine-tuning.
2/17/2026
Enhancing Delta Compression in LLMs via SVD-based Quantization Error Minimization
SVD-based quantization for compressing fine-tuned LLM delta parameters with minimized quantization error for storage efficiency.
2/17/2026
HYPER: A Foundation Model for Inductive Link Prediction with Knowledge Hypergraphs
HYPER foundation model for inductive link prediction with knowledge hypergraphs handling novel entities and relation types.
2/17/2026
Calibrated Predictive Lower Bounds on Time-to-Unsafe-Sampling in LLMs
Novel safety measure for LLMs measuring generations needed to trigger unsafe responses, with calibrated predictive bounds for evaluation.
2/17/2026
NeuronSeek: On Stability and Expressivity of Task-driven Neurons
NeuronSeek uses symbolic regression to discover optimal neuron formulations and construct task-driven neural networks.
2/17/2026
Chain of Thought in Order: Discovering Learning-Friendly Orders for Arithmetic
Studies optimal ordering of intermediate steps in chain-of-thought reasoning for arithmetic tasks, improving learning efficiency.
2/17/2026
wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models
Weighted policy optimization method for reinforcement learning in diffusion-based LLMs, addressing intractable likelihood approximation.
2/17/2026
The Serial Scaling Hypothesis
Theoretical framework distinguishing inherently sequential problems that cannot be efficiently parallelized, relevant to LLM reasoning.
2/17/2026
FGBench: A Dataset and Benchmark for Molecular Property Reasoning at Functional Group-Level in Large Language Models
Benchmark dataset for evaluating LLM reasoning on molecular properties at functional group level for chemistry applications.
2/17/2026
Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign Prompts
Investigates spontaneous deception in LLMs on benign prompts, revealing trustworthiness risks in reasoning and planning tasks.
2/17/2026
MAVIS: Multi-Objective Alignment via Inference-Time Value-Guided Selection
Inference-time method for balancing multiple conflicting objectives in LLM outputs without expensive per-objective fine-tuning.
2/17/2026
Predicting the Order of Upcoming Tokens Improves Language Modeling
Token order prediction auxiliary objective improves language model performance, offering alternative to multi-token prediction for next-token training.
2/17/2026
ART: Adaptive Resampling-based Training for Imbalanced Classification
Adaptive resampling method for imbalanced classification that adjusts training data distribution based on class-wise learning difficulty.
2/17/2026
Online reinforcement learning via sparse Gaussian mixture model Q-functions
Online policy-iteration RL framework using sparse Gaussian mixture model Q-functions with interpretable exploration mechanisms.
2/17/2026
DiffusionNFT: Online Diffusion Reinforcement with Forward Process
Online reinforcement learning method for diffusion models addressing intractable likelihoods, enabling RLHF-style training without solver restrictions.
2/17/2026
Functional Scaling Laws in Kernel Regression: Loss Dynamics and Learning Rate Schedules
Analyzes scaling laws for loss dynamics and learning rate schedules in SGD on kernel regression, with implications for LLM training.
2/17/2026
The Rogue Scalpel: Activation Steering Compromises LLM Safety
Demonstrates that activation steering for LLM control can compromise safety mechanisms, causing models to comply with harmful requests.
2/17/2026
Effective Quantization of Muon Optimizer States
8-bit quantization technique for Muon optimizer states in LLM pre-training, reducing memory overhead while maintaining training efficiency.
2/17/2026
Better Hessians Matter: Studying the Impact of Curvature Approximations in Influence Functions
Studies how curvature approximations (GGN, K-FAC) in influence functions affect data attribution accuracy for deep learning models.
2/17/2026
LLM DNA: Tracing Model Evolution via Functional Representations
Method to trace evolutionary relationships between LLMs through functional representations, enabling better model management and understanding of fine-tuning/distillation lineages.
2/17/2026
OpenTSLM: Time-Series Language Models for Reasoning over Multivariate Medical Text- and Time-Series Data
OpenTSLM: time-series language models integrating multivariate medical time-series as native modality. Enables LLMs to handle temporal clinical data.
2/17/2026
Multi-scale Autoregressive Models are Laplacian, Discrete, and Latent Diffusion Models in Disguise
arXiv paper: Visual Autoregressive models reinterpreted as Laplacian latent pyramid with learned coarse-to-fine refinement. Formal analysis of design trade-offs.
2/17/2026
Where to Add PDE Diffusion in Transformers
Research on optimal placement of PDE diffusion layers in transformer architectures using heat equation-based smoothing for local geometric priors.
2/17/2026
Show HN: Claude Pilot – Claude Code is powerful. Pilot makes it reliable
Claude Pilot framework enforcing tests, context preservation, and quality standards for Claude Code to produce reliable production-grade code.
2/17/2026
Show HN: HJX – An AI-Native Web Language Unifying HTML, CSS and JavaScript
HJX: compiled UI language unifying HTML/CSS/JavaScript for AI code generation, reducing token fragmentation.
2/17/2026
The Journal for AI Generated Papers
Journal platform for AI-generated research papers featuring collaboration between human prompters and AI systems on OpenClaw framework.
2/17/2026
Show HN: CodeGraph CLI – Chat with your codebase using graph-augmented RAG
CodeGraph CLI tool combining tree-sitter AST parsing, SQLite dependency graphs, and LanceDB vector embeddings for code understanding via graph-augmented RAG.
2/17/2026
Show HN: How do you reliably benchmark sub-100ns code paths in Rust
Cuttlefish Rust library using CRDTs, io_uring, and SIMD for low-latency distributed systems coordination with sub-100ns code paths.
2/17/2026
Show HN: Neko – AI agent runtime that fits on a Raspberry Pi Zero 2W
Neko AI agent runtime designed for resource-constrained hardware like Raspberry Pi Zero 2W with markdown-based memory and session management.
2/17/2026
The quality cost of vibe coding: an audit of OpenClaw's 180K-star codebase
Code quality audit of OpenClaw, an AI agent framework that reached 180K GitHub stars, examining implementation practices and technical debt.