3/19/2026
A Comedy of Estimators: On KL Regularization in RL Training of LLMs
Analysis of KL divergence estimators used in RL training of LLMs, evaluating different approximation methods for reverse KL regularization.
3/19/2026
VL-RouterBench: A Benchmark for Vision-Language Model Routing
Benchmark for evaluating vision-language model routing systems across quality and cost dimensions using real inference logs.
3/19/2026
Evaluating Feature Dependent Noise in Preference-based Reinforcement Learning
Study on feature-dependent noise in preference-based reinforcement learning, examining how observation-dependent uncertainty affects learning.
3/19/2026
EpiQAL: Benchmarking Large Language Models in Epidemiological Question Answering for Enhanced Alignment and Reasoning
Diagnostic benchmark for evaluating epidemiological reasoning in large language models, focusing on evidence-grounded inference over clinical knowledge.
3/19/2026
APEX-SWE
Benchmark for evaluating frontier AI models on real-world software engineering tasks including integration and end-to-end system construction.
3/19/2026
GIFT: Reconciling Post-Training Objectives via Finite-Temperature Gibbs Initialization
Reformulation of supervised fine-tuning to reconcile post-training objectives in large reasoning models using Gibbs initialization before reinforcement learning.
3/19/2026
Speculative Decoding: Performance or Illusion?
Systematic evaluation of speculative decoding acceleration techniques on production-grade vLLM inference engine, assessing real-world effectiveness.
3/19/2026
Improving Low-Resource Machine Translation via Round-Trip Reinforcement Learning
Self-supervised reinforcement learning approach for improving low-resource machine translation using round-trip bootstrapping with NLLB models.
3/19/2026
YOLO26: An Analysis of NMS-Free End to End Framework for Real-Time Object Detection
Analysis of YOLO26 object detection framework that eliminates NMS post-processing for native end-to-end learning strategy.
3/19/2026
CARPE: Context-Aware Image Representation Prioritization via Ensemble for Large Vision-Language Models
Method to improve vision-centric capabilities of multimodal LLMs through context-aware image representation prioritization via ensemble approach.
3/19/2026
Failing on Bias Mitigation: A Case Study on the Challenges of Fairness in Government Data
Case study examining why bias mitigation techniques fail on government data, using crime rate prediction with Bristol City Council data as test case.
3/19/2026
Unsupervised Decomposition and Recombination with Discriminator-Driven Diffusion Models
Unsupervised method for decomposing complex data into factorized components using diffusion models, enabling reusable component recombination for images and robotic videos.
3/19/2026
PAND: Prompt-Aware Neighborhood Distillation for Lightweight Fine-Grained Visual Classification
Knowledge distillation framework from vision-language models into lightweight networks for fine-grained visual classification using prompt-aware semantic calibration.
3/19/2026
Learning Adaptive Distribution Alignment with Neural Characteristic Function for Graph Domain Adaptation
Graph domain adaptation method using neural characteristic functions to align distributional shifts across labeled source and unlabeled target graphs.
3/19/2026
interwhen: A Generalizable Framework for Verifiable Reasoning with Test-time Monitors
Framework for test-time verification of LLM reasoning traces using single-trajectory monitoring to catch errors without exploring multiple branches.
3/19/2026
No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection
Zero-shot video anomaly detection framework using multimodal large language models without requiring real anomaly examples, addressing dataset diversity limitations.
3/19/2026
Hyperparameter Trajectory Inference with Conditional Lagrangian Optimal Transport
Method for inferring hyperparameter trajectories using optimal transport to adapt neural network behavior post-deployment without expensive retraining.
3/19/2026
Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning
Research on continual learning in pretrained Vision-Language-Action models for robot policy learning, showing resistance to catastrophic forgetting when acquiring new skills.
3/19/2026
SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration
Benchmark evaluating LLM agent capabilities in long-term codebase maintenance via continuous integration, beyond static bug fixing.
3/19/2026
Thin Keys, Full Values: Reducing KV Cache via Low-Dimensional Attention Selection
Technique to reduce KV cache in transformers by using low-dimensional attention selection for keys while maintaining full-dimensional values.
3/19/2026
LUMINA: LLM-Guided GPU Architecture Exploration via Bottleneck Analysis
LLM-guided GPU architecture exploration for LLM inference workloads via bottleneck analysis and design space optimization.
3/19/2026
Real-World AI Evaluation: How FRAME Generates Systematic Evidence to Resolve the Decision-Maker's Dilemma
FRAME methodology for systematic evaluation of AI systems in real-world organizational contexts beyond abstract capability measures.
3/19/2026
Ethical Fairness without Demographics in Human-Centered AI
Method for achieving ethical fairness in AI systems without relying on demographic attributes in human-centered applications.
3/19/2026
"I'm Not Reading All of That": Understanding Software Engineers' Level of Cognitive Engagement with Agentic Coding Assistants
Study on software engineers' cognitive engagement with agentic coding assistants, examining over-reliance risks and critical thinking impacts.
3/19/2026
Open Biomedical Knowledge Graphs at Scale: Construction, Federation, and AI Agent Access with Samyama Graph Database
Open-source biomedical knowledge graphs (pathways, clinical trials, drug interactions) with federated access and AI agent integration.
3/19/2026
OMNIFLOW: A Physics-Grounded Multimodal Agent for Generalized Scientific Reasoning
Multimodal LLM agent for physics-informed scientific reasoning, combining language models with PDE solving without domain-specific fine-tuning.
3/19/2026
100x Cost & Latency Reduction: Performance Analysis of AI Query Approximation using Lightweight Proxy Models
Study on using lightweight proxy models to reduce cost and latency of AI queries in SQL, achieving 100x improvements.
3/19/2026
Resource Consumption Threats in Large Language Models
Survey of resource consumption threats in LLMs, covering efficiency issues affecting service capacity, latency, and API costs.
3/19/2026
VisBrowse-Bench: Benchmarking Visual-Native Search for Multimodal Browsing Agents
Benchmark for visual-native search in multimodal browsing agents, evaluating MLLM visual reasoning over web pages.
3/19/2026
Attention-guided Evidence Grounding for Spoken Question Answering
End-to-end spoken question answering framework using attention guidance from speech LLMs for cross-modal alignment without ASR.
3/19/2026
EngGPT2: Sovereign, Efficient and Open Intelligence
Italian open-source LLM with 16B parameters achieving competitive performance on benchmarks while requiring fraction of inference power.
3/19/2026
When Openclaw Agents Learn from Each Other: Insights from Emergent AI Agent Communities for Human-AI Partnership in Education
Studies emergent learning behaviors in ecosystem of 167,000 AI agents interacting as peers across platforms without intervention.
3/19/2026
DexGrasp-Zero: A Morphology-Aligned Policy for Zero-Shot Cross-Embodiment Dexterous Grasping
Zero-shot cross-embodiment dexterous grasping policy using morphology-aligned approach for diverse robotic hands.
3/19/2026
Federated Multi Agent Deep Learning and Neural Networks for Advanced Distributed Sensing in Wireless Networks
Overview of multi-agent deep learning and federated training for distributed sensing in 5G/6G wireless networks.
3/19/2026
Multi-Agent Reinforcement Learning for Dynamic Pricing: Balancing Profitability,Stability and Fairness
Empirical evaluation of multi-agent RL algorithms (MAPPO, MADDPG) for dynamic pricing in competitive markets.
3/19/2026
From Language to Action in Arabic: Reliable Structured Tool Calling via Data-Centric Fine-Tuning
Production framework for reliable Arabic function-calling models enabling agentic AI systems through data-centric fine-tuning.
3/19/2026
HoloByte: Continuous Hyperspherical Distillation for Tokenizer-Free Modeling
Tokenizer-free sequence modeling using continuous hyperspherical distillation to preserve byte-level continuity without discrete quantization.
3/19/2026
MHPO: Modulated Hazard-aware Policy Optimization for Stable Reinforcement Learning
Proposes modulated hazard-aware policy optimization to improve training stability in GRPO-based reinforcement learning frameworks.
3/19/2026
Integrating Inductive Biases in Transformers via Distillation for Financial Time Series Forecasting
Addresses transformer limitations for financial time-series forecasting by integrating inductive biases via distillation.
3/19/2026
Transformers Can Learn Rules They've Never Seen: Proof of Computation Beyond Interpolation
Tests whether transformers can learn unseen rules beyond interpolation using controlled experiments with symbolic derivations.
3/19/2026
Do Understanding and Generation Fight? A Diagnostic Study of DPO for Unified Multimodal Models
Systematic study of DPO alignment on unified multimodal models, finding generation quality resists alignment while understanding improves.
3/19/2026
SCE-LITE-HQ: Smooth visual counterfactual explanations with generative foundation models
Proposes counterfactual explanation method using generative foundation models for interpreting neural network predictions in visual domains.
3/19/2026
Early Quantization Shrinks Codebook: A Simple Fix for Diversity-Preserving Tokenization
Investigates vector quantization in generative models, proposing early quantization to address codebook diversity issues in tokenization for LLMs and diffusion models.
3/19/2026
PRISM: Demystifying Retention and Interaction in Mid-Training
PRISM empirical study of mid-training for LLMs across 7 models showing consistent +15 to +40 point gains on benchmarks with 27B tokens.
3/19/2026
CircuitBuilder: From Polynomials to Circuits via Reinforcement Learning
Applies reinforcement learning with AlphaZero-style training to discover efficient arithmetic circuits for polynomial computation.
3/19/2026
SENSE: Efficient EEG-to-Text via Privacy-Preserving Semantic Retrieval
Proposes privacy-preserving EEG-to-text system using semantic retrieval instead of fine-tuning LLMs for brain-computer interfaces.
3/19/2026
Contextual Preference Distribution Learning
Proposes pipeline to learn context-dependent preference distributions for risk-averse decision-making via inverse optimization.
3/19/2026
REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge
Introduces regression-aware RL method for LLM-as-a-Judge that leverages ordinal structure in scoring tasks.
3/19/2026
Noise-Response Calibration: A Causal Intervention Protocol for LLM-Judges
Proposes calibration protocol for LLM-judges using controlled noise interventions to improve reliability in low-label settings.
3/19/2026
Domain-informed explainable boosting machines for trustworthy lateral spread predictions
Develops domain-informed framework for explainable boosting machines ensuring physical consistency in natural hazard prediction.