2/24/2026
Layer Collapse Can be Induced by Unstructured Pruning
Study demonstrating unstructured neural network pruning can induce structural effects and layer collapse, shortening computational paths.
2/24/2026
Improving Discrete Optimisation Via Decoupled Straight-Through Estimator
Improved straight-through estimator for discrete optimization decoupling forward-pass stochasticity from backward-pass gradient dispersion in neural networks.
2/24/2026
Robust Time Series Causal Discovery for Agent-Based Model Validation
Robust causal discovery method for agent-based model validation using cross-validation approach to improve accuracy on complex time series data.
2/24/2026
vCache: Verified Semantic Prompt Caching
Semantic prompt caching system using variable similarity thresholds in vector databases to reduce LLM inference latency and cost beyond static threshold approaches.
2/24/2026
The Curse of Depth in Large Language Models
Research on 'Curse of Depth' phenomenon where ~50% of LLM layers underperform, analyzing root causes across Llama, Mistral, DeepSeek, Qwen families with theoretical and empirical analysis.
2/24/2026
AdaGC: Improving Training Stability for Large Language Model Pretraining
AdaGC: gradient clipping method stabilizing LLM pretraining by addressing multi-factor loss spikes.
2/24/2026
Test-Time Training Provably Improves Transformers as In-context Learners
Test-time training provably enhances transformer in-context learning through gradient-based weight updates.
2/24/2026
Noise-Aware Generalization: Robustness to In-Domain Noise and Out-of-Domain Generalization
Noise-Aware Generalization: training methods handling label noise and domain shifts simultaneously.
2/24/2026
Towards A Universal Graph Structural Encoder
Universal graph encoder for learning transferable structural representations across diverse graph domains.
2/24/2026
GraphOmni: A Comprehensive and Extensible Benchmark Framework for Large Language Models on Graph-theoretic Tasks
GraphOmni: comprehensive benchmark for evaluating LLM reasoning on graph-theoretic tasks in natural language.
2/24/2026
Think2SQL: Reinforce LLM Reasoning Capabilities for Text2SQL
Think2SQL: reinforcement learning approach improving LLM reasoning for complex text-to-SQL generation.
2/24/2026
Learning to Rank Critical Road Segments via Heterogeneous Graphs with Origin-Destination Flow Integration
HetGL2R: heterogeneous graph framework for ranking critical road segments using origin-destination flows.
2/24/2026
FedSDAF: Leveraging Source Domain Awareness for Enhanced Federated Domain Generalization
Federated domain generalization leveraging source domain features for improved generalization.
2/24/2026
GRILL: Restoring Gradient Signal in Ill-Conditioned Layers for More Effective Adversarial Attacks on Autoencoders
GRILL: adversarial attack method for autoencoders addressing ill-conditioned gradient propagation.
2/24/2026
Performance Estimation in Binary Classification Using Calibrated Confidence
Performance estimation for binary classifiers without ground truth labels using calibrated confidence.
2/24/2026
CodePDE: An Inference Framework for LLM-driven PDE Solver Generation
CodePDE: LLM-based inference framework generating code for solving partial differential equations.
2/24/2026
FairSHAP: Preprocessing for Fairness Through Attribution-Based Data Augmentation
FairSHAP: preprocessing framework using Shapley values for identifying and mitigating bias in ML models.
2/24/2026
Heterogeneity-Aware Client Sampling for Optimal and Efficient Federated Learning
Federated learning client sampling strategy addressing heterogeneous communication and computational capabilities.
2/24/2026
The Spacetime of Diffusion Models: An Information Geometry Perspective
Information geometry analysis of diffusion model latent spaces, examining geodesic decoding and stochastic decoders.
2/24/2026
QiMeng-CodeV-R1: Reasoning-Enhanced Verilog Generation
QiMeng-CodeV-R1 applies RLVR to Verilog generation from natural language, introducing verifiable reward signal for hardware design automation.
2/24/2026
It Takes a Good Model to Train a Good Model: Generalized Gaussian Priors for Optimized LLMs
Shows LLM weights follow generalized Gaussian distributions and proposes end-to-end optimization framework for improved training efficiency.
2/24/2026
E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models
E-BATS enables efficient test-time adaptation for speech models under domain shift without backpropagation, reducing memory overhead.
2/24/2026
Predicting New Research Directions in Materials Science using Large Language Models and Concept Graphs
Uses LLMs and concept graphs to extract scientific concepts from materials science abstracts and discover novel research connections.
2/24/2026
MoNE: Replacing Redundant Experts with Lightweight Novices for Structured Pruning of MoE
MoNE prunes MoE models by replacing redundant experts with lightweight novices, reducing memory overhead while maintaining performance.
2/24/2026
Symbolic Branch Networks: Tree-Inherited Neural Models for Interpretable Multiclass Classification
Symbolic Branch Networks inherit architecture from decision tree ensembles to create interpretable neural models for multiclass classification.
2/24/2026
Sampling-aware Adversarial Attacks Against Large Language Models
Demonstrates sampling-aware adversarial attacks against LLMs that leverage stochastic sampling to more accurately assess robustness.
2/24/2026
Graph Neural Networks Powered by Encoder Embedding for Improved Node Learning
Improves GNN node learning performance using statistically grounded graph encoder embedding initialization for faster convergence.
2/24/2026
Modular Delta Merging with Orthogonal Constraints: A Scalable Framework for Continual and Reversible Model Composition
MDM-OC enables scalable, reversible model composition with orthogonal constraints to prevent task interference and catastrophic forgetting.
2/24/2026
MolReasoner: Toward Effective and Interpretable Reasoning for Molecular LLMs
MolReasoner enhances LLM reasoning for molecular tasks with domain-specific semantics, reducing hallucinations and improving interpretability.
2/24/2026
GEDAN: Learning the Edit Costs for Graph Edit Distance
GEDAN learns optimal edit costs for graph edit distance computation, improving upon NP-hard approximation methods with learnable cost functions.
2/24/2026
Shuffle-R1: Efficient RL framework for Multimodal Large Language Models via Data-centric Dynamic Shuffle
Shuffle-R1 framework improves RL efficiency for multimodal LLMs through data-centric dynamic shuffling to address advantage collapsing and rollout silencing.
2/24/2026
Biased Local SGD for Efficient Deep Learning on Heterogeneous Systems
Proposes biased local SGD for efficient parallel neural network training on heterogeneous computing systems with varying resource availability.
2/24/2026
AFABench: A Generic Framework for Benchmarking Active Feature Acquisition
AFABench framework for benchmarking active feature acquisition methods that dynamically select informative features under acquisition cost constraints.
2/24/2026
Bootstrapping Task Spaces for Self-Improvement
Presents Exploratory Iteration (ExIt), RL methods enabling agents to self-improve through iterative refinement without fixed iteration limits.
2/24/2026
Sequential Data Augmentation for Generative Recommendation
Data augmentation strategies for generative recommendation systems improving generalization in sequential user behavior prediction.
2/24/2026
Wonder Wins Ways: Curiosity-Driven Exploration through Multi-Agent Contextual Calibration
Multi-agent RL with curiosity-driven exploration using contextual calibration to distinguish novelty from environmental stochasticity.
2/24/2026
DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models
DriftLite: Training-free particle-based approach for inference-time diffusion model adaptation to new distributions.
2/24/2026
Lightweight error mitigation strategies for post-training N:M activation sparsity in LLMs
Error mitigation methods for post-training N:M activation sparsity in LLMs enabling dynamic input-adaptive compression.
2/24/2026
SpinGPT: A Large-Language-Model Approach to Playing Poker Correctly
SpinGPT applies LLM approach to poker strategy, addressing CFR computational limits in multi-player game settings.
2/24/2026
Effective Quantization of Muon Optimizer States
8-bit blockwise quantization of Muon optimizer states reducing memory overhead for large-scale LLM pretraining.
2/24/2026
Polychromic Objectives for Reinforcement Learning
Polychromic objectives for RL fine-tuning preventing policy collapse and preserving diversity in pretrained model behaviors.
2/24/2026
Diffusion Alignment as Variational Expectation-Maximization
Diffusion Alignment as Variational EM framework addressing reward over-optimization and mode collapse in diffusion model alignment.
2/24/2026
On Predictability of Reinforcement Learning Dynamics for Large Language Models
Analysis of RL-induced parameter dynamics in LLMs revealing rank-1 dominance in reasoning improvements and predictability of training trajectories.
2/24/2026
The Unseen Frontier: Pushing the Limits of LLM Sparsity with Surrogate-Free ADMM
Surrogate-free ADMM method for LLM pruning achieving >50% sparsity without accuracy degradation, breaking through conventional compression limits.
2/24/2026
Scaling Laws Revisited: Modeling the Role of Data Quality in Language Model Pretraining
Scaling law formalization incorporating data quality parameter for language model pretraining, extending traditional model/dataset size relationships.
2/24/2026
KVComm: Enabling Efficient LLM Communication through Selective KV Sharing
KVComm: Communication framework for multi-agent LLM systems using selective key-value sharing instead of natural language or hidden states.
2/24/2026
TROLL: Trust Regions improve Reinforcement Learning for Large Language Models
TROLL: Trust region-based RL method improving upon PPO clipping for LLM fine-tuning, achieving more stable and optimal reward-based training.
2/24/2026
Enhancing Reasoning for Diffusion LLMs via Distribution Matching Policy Optimization
Novel RL algorithm for diffusion LLMs using distribution matching policy optimization to improve reasoning capabilities and match autoregressive LLM performance.
2/24/2026
Medical Interpretability and Knowledge Maps of Large Language Models
Systematic interpretability study of five LLMs' medical knowledge using activation analysis and layer lesioning techniques.
2/24/2026
Lean Finder: Semantic Search for Mathlib That Understands User Intents
Semantic search engine for Lean theorem prover mathlib using intent-aware ranking for theorem discovery.