4/21/2026
NI Sampling: Accelerating Discrete Diffusion Sampling by Token Order Optimization
Token order optimization for discrete diffusion language models to accelerate sampling speed beyond traditional autoregressive decoding.
4/21/2026
Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts
BAR framework trains independent domain experts separately via supervised finetuning and RL, composes via mixture-of-experts for modular LLM post-training.
4/21/2026
Barrier-enforced multi-objective optimization for direct point and sharp interval forecasting
Multi-objective neural network framework for point and interval forecasting with non-crossing prediction intervals.
4/21/2026
Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data
Addresses mode collapse in RL training for LLM reasoning on saturated benchmarks. Proposes Constrained Uniform Top-K Sampling for improved policy learning.
4/21/2026
IDOBE: Infectious Disease Outbreak forecasting Benchmark Ecosystem
Benchmark ecosystem for epidemic forecasting using statistical and ML models. Specialized application domain with limited relevance.
4/21/2026
A Note on TurboQuant and the Earlier DRIVE/EDEN Line of Work
Technical note clarifying relationship between TurboQuant and earlier DRIVE/EDEN quantization schemes for neural networks.
4/21/2026
Latent Phase-Shift Rollback: Inference-Time Error Correction via Residual Stream Monitoring and KV-Cache Steering
Inference-time error correction for LLMs via residual stream monitoring and KV-cache steering to prevent mid-generation reasoning errors.
4/21/2026
When Can LLMs Learn to Reason with Weak Supervision?
Empirical study on when LLMs can learn reasoning with weak supervision signals, examining reinforcement learning with verifiable rewards across model families and domains.
4/21/2026
Bounded Ratio Reinforcement Learning
Introduces Bounded Ratio Reinforcement Learning framework bridging trust region methods and PPO's clipped objective for on-policy RL.
4/21/2026
Sessa: Selective State Space Attention
Proposes Sessa, selective state-space attention mechanism addressing information dilution in diffuse attention by improving token influence in long contexts.
4/21/2026
TokenChain: A Discrete Speech Chain via Semantic Token Modeling
Presents TokenChain, discrete speech chain coupling semantic-token ASR with two-stage TTS enabling feedback across text interface.
4/21/2026
How Robustly do LLMs Understand Execution Semantics?
Studies robustness of LLM code understanding via program-output prediction, finding open-source reasoning models outperform closed models on execution semantics.
4/21/2026
Decoding AI Tutor Effects for Educational Measurement: Temporal, Multi-Outcome, and Behavior-Cognitive Analysis
Proposes AI agent prototype for exploring AI-assisted learning with temporal interaction analysis and behavioral-cognitive profiling of learners.
4/21/2026
Reciprocal Co-Training (RCT): Coupling Gradient-Based and Non-Differentiable Models via Reinforcement Learning
Introduces reciprocal co-training framework coupling gradient-based LLMs with non-differentiable models like Random Forests via reinforcement learning.
4/21/2026
Data Mixing for Large Language Models Pretraining: A Survey and Outlook
Survey on data mixing strategies for LLM pretraining, optimizing domain-level sampling weights to improve training efficiency and generalization under budget constraints.
4/21/2026
CoLLM: A Unified Framework for Co-execution of LLMs Federated Fine-tuning and Inference
Unified framework for federated fine-tuning and inference of LLMs on edge devices, optimizing parameter-efficient training and low-latency deployment.
4/21/2026
ICAT: Incident-Case-Grounded Adaptive Testing for Physical-Risk Prediction in Embodied World Models
Proposes ICAT framework to evaluate physical risk prediction in video-generative world models, grounding testing in real incidents for safe embodied planning.
4/21/2026
Heterogeneous Self-Play for Realistic Highway Traffic Simulation
Presents PHASE, heterogeneous self-play method for realistic highway traffic simulation enabling rare safety-critical scenario generation for autonomous vehicle evaluation.
4/21/2026
Cooperative Coevolution versus Monolithic Evolutionary Search for Semi-Supervised Tabular Classification
Proposes cooperative coevolutionary method for semi-supervised tabular classification in extreme low-label regime, comparing evolutionary vs monolithic approaches.
4/21/2026
Modeling User Exploration Saturation: When Recommender Systems Should Stop Pushing Novelty
Studies when recommender systems should reduce novelty promotion by modeling user exploration saturation to improve recommendation fairness.
4/21/2026
Injecting Structured Biomedical Knowledge into Language Models: Continual Pretraining vs. GraphRAG
Compares continual pretraining versus GraphRAG for injecting structured biomedical knowledge from UMLS into language models for domain specialization.
4/21/2026
Safety, Security, and Cognitive Risks in State-Space Models: A Systematic Threat Analysis with Spectral, Stateful, and Capacity Attacks
First systematic threat analysis of State-Space Models (S4, Mamba, etc.) identifying spectral, stateful, and capacity attacks for safety-critical applications.
4/21/2026
Method for Aggregating Unstructured Data Using Large Language Models
Presents method for automated collection and aggregation of unstructured web data using LLMs, addressing webpage structure changes and dynamic content loading.
4/21/2026
Quantifying how AI Panels improve precision
Derives formula to estimate precision bounds for AI-based job application screening, analyzing bias and reliability of single AI systems.
4/21/2026
Fuzzy Encoding-Decoding to Improve Spiking Q-Learning Performance in Autonomous Driving
Proposes fuzzy encoder-decoder architecture for spiking Q-networks in autonomous driving to reduce information loss and improve value function representation.
4/21/2026
Thermal-GEMs: Generalized Models for Building Thermal Dynamics
Transfer learning models for building thermal dynamics using multi-source data to enable energy-efficient operation and fault detection.
4/21/2026
SAND: The Challenge on Speech Analysis for Neurodegenerative Disease Assessment
SAND challenge dataset and algorithms for speech-based detection of neurodegenerative diseases like ALS.
4/21/2026
FM-CAC: Carbon-Aware Control for Battery-Buffered Edge AI via Time-Series Foundation Models
FM-CAC framework using time-series foundation models for carbon-aware control of battery-buffered edge AI deployments.
4/21/2026
FairLogue: Evaluating Intersectional Fairness across Clinical Machine Learning Use Cases using the All of Us Research Program
FairLogue toolkit for intersectional fairness auditing of clinical ML models using All of Us dataset across combined demographic groups.
4/21/2026
SynopticBench: Evaluating Vision-Language Models on Generating Weather Forecast Discussions of the Future
SynopticBench evaluates vision-language models on generating future weather forecast discussions from meteorological data.
4/21/2026
EchoChain: A Full-Duplex Benchmark for State-Update Reasoning Under Interruptions
EchoChain benchmark for evaluating full-duplex state-update reasoning in voice assistants under mid-speech interruptions.
4/21/2026
Deep Hierarchical Knowledge Loss for Fault Intensity Diagnosis
Deep hierarchical knowledge loss framework for fault intensity diagnosis that captures dependencies among target classes in manufacturing.
4/21/2026
Modelling Gas-Phase Reaction Kinetics with Guided Particle Diffusion Sampling
Physics-guided diffusion sampling method for modeling gas-phase reaction kinetics from sparse PDE observations.
4/21/2026
Saccade Attention Networks: Using Transfer Learning of Attention to Reduce Network Sizes
Saccade Attention Networks using sparse attention patterns to reduce transformer sizes and computation by learning selective focus.
4/21/2026
Aletheia: Physics-Conditioned Localized Artifact Attention (PhyLAA-X) for End-to-End Generalizable and Robust Deepfake Video Detection
PhyLAA-X deepfake detector combining semantic artifact learning with physical invariants for improved cross-generator robustness.
4/21/2026
An Uncertainty-Aware Loss Function Incorporating Fuzzy Logic: Application to MRI Brain Image Segmentation
Novel loss function integrating fuzzy logic for improved deep learning performance on MRI brain image segmentation tasks.
4/21/2026
NL2SQLBench: A Modular Benchmarking Framework for LLM-Enabled NL2SQL Solutions
NL2SQLBench: First modular benchmarking framework for evaluating LLM-based natural language to SQL solutions systematically.
4/21/2026
Gradient-Free Continual Learning in Spiking Neural Networks via Inter-Spike Interval Regularization
Gradient-free continual learning method for spiking neural networks using inter-spike interval regularization, compatible with neuromorphic hardware.
4/21/2026
From Handwriting to Structured Data: Benchmarking AI Digitisation of Handwritten Forms
Benchmark comparing 17 frontier multimodal LLMs and open-source models on handwritten medical form digitization, with latest models reaching 85% accuracy.
4/21/2026
Learning-Based Sparsification of Dynamic Graphs in Robotic Exploration Algorithms
Transformer-based framework using PPO to prune graph structures in robotic exploration, reducing redundant information during frontier-based planning.
4/21/2026
BARD: Bridging AutoRegressive and Diffusion Vision-Language Models Via Highly Efficient Progressive Block Merging and Stage-Wise Distillation
Framework converting autoregressive vision-language models to diffusion-based models via progressive block merging and distillation.
4/21/2026
Penny Wise, Pixel Foolish: Bypassing Price Constraints in Multimodal Agents via Visual Adversarial Perturbations
Security analysis of multimodal agents showing visual adversarial perturbations can override price constraints in transactions.
4/21/2026
Operationalizing Fairness in Text-to-Image Models: A Survey of Bias, Fairness Audits and Mitigation Strategies
Survey of bias evaluation and mitigation strategies for text-to-image generation models with operational fairness definitions.
4/21/2026
Understanding Tool-Augmented Agents for Lean Formalization: A Factorial Analysis
Systematic analysis of tool-augmented agents for translating natural language mathematics to Lean 4 formal proofs.
4/21/2026
SafeLM: Unified Privacy-Aware Optimization for Trustworthy Federated Large Language Models
Federated learning framework for LLM safety addressing privacy, security, misinformation, and adversarial robustness.
4/21/2026
AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation
LLM agent framework for kernel generation in Triton using failure-driven adaptation and diversity-preserving search.
4/21/2026
Agentic Risk-Aware Set-Based Engineering Design
Multi-agent LLM framework for early-stage engineering design with human-in-the-loop control for aerodynamic optimization.
4/21/2026
Surgical Repair of Insecure Code Generation in LLMs
Analysis of security vulnerability gap in LLM code generation and mechanistic approach to improve code safety.
4/21/2026
Scalable and Adaptive Parallel Training of Graph Transformer on Large Graphs
Scalable parallel training framework for graph transformers on large-scale graphs across multiple GPUs.
4/21/2026
Debate as Reward: A Multi-Agent Reward System for Scientific Ideation via RL Post-Training
Multi-agent reward system using debate mechanism for scientific ideation via LLM reinforcement learning post-training.