3/9/2026
Theory of Code Space: Do Code Agents Understand Software Architecture?
Theory of Code Space: Benchmark evaluating whether code agents maintain coherent architectural understanding when exploring multi-file codebases.
3/9/2026
"When to Hand Off, When to Work Together": Expanding Human-Agent Co-Creative Collaboration through Concurrent Interaction
Framework enabling AI agents to achieve collaborative context awareness by interpreting concurrent user actions on shared creative artifacts in real-time.
3/9/2026
Fine-Tuning and Evaluating Conversational AI for Agricultural Advisory
Hybrid LLM architecture fine-tuned for agricultural advisory with domain-specific knowledge to provide accurate, actionable recommendations.
3/9/2026
Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations
Empirical evaluation of LLM robustness to five types of chain-of-thought perturbations: math errors, unit conversion, sycophancy, skipped steps, and reasoning corruption.
3/9/2026
Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion
Phys4D: Three-stage pipeline ensuring physics consistency in 4D world models generated from video diffusion models through iterative refinement.
3/9/2026
Large-Language-Model-Guided State Estimation for Partially Observable Task and Motion Planning
Framework combining LLM common-sense reasoning with robot planning in partially observable environments for task and motion planning.
3/9/2026
Measuring AI R&D Automation
Analysis of AI R&D automation extent and effects, proposing empirical frameworks beyond capability benchmarks to measure real-world AIRDA impact.
3/9/2026
Simulating Meaning, Nevermore! Introducing ICR: A Semiotic-Hermeneutic Metric for Evaluating Meaning in LLM Text Summaries
ICR: Framework integrating semiotics and hermeneutics to evaluate meaning in LLM-generated text summaries beyond traditional metrics.
3/9/2026
vLLM Semantic Router: Signal Driven Decision Routing for Mixture-of-Modality Models
vLLM Semantic Router: Inference framework for intelligent request routing across diverse multimodal LLM deployments using composable signal orchestration.
3/9/2026
Traversal-as-Policy: Log-Distilled Gated Behavior Trees as Externalized, Verifiable Policies for Safe, Robust, and Efficient Agents
Traversal-as-Policy: Framework distilling LLM agent execution logs into verifiable Gated Behavior Trees for safe, robust autonomous agents with explicit policy control.
3/9/2026
VDCook:DIY video data cook your MLLMs
VDCook: A configurable platform for constructing video datasets for multimodal LLMs via natural language queries, combining real retrieval with controlled synthesis.
3/9/2026
IntSeqBERT: Learning Arithmetic Structure in OEIS via Modulo-Spectrum Embeddings
IntSeqBERT: Transformer encoder with modulo-spectrum embeddings for predicting integer sequences in OEIS, handling out-of-vocabulary large values.
3/9/2026
A Novel Hybrid Heuristic-Reinforcement Learning Optimization Approach for a Class of Railcar Shunting Problems
Hybrid heuristic-reinforcement learning approach for optimizing railcar shunting operations in freight yards.
3/9/2026
Bias In, Bias Out? Finding Unbiased Subnetworks in Vanilla Models
Method for extracting fair and unbiased subnetworks from standard trained models without additional data or complex procedures.
3/9/2026
On the Value of Tokeniser Pretraining in Physics Foundation Models
Study of tokenizer pretraining impact on physics foundation models for emulation and simulation of multiphysics phenomena.
3/9/2026
Improved Scaling Laws via Weak-to-Strong Generalization in Random Feature Ridge Regression
Analysis of weak-to-strong generalization in random feature ridge regression, showing improved scaling laws when strong models train on weak labels.
3/9/2026
Warm Starting State-Space Models with Automata Learning
Formal correspondence between Moore machines and state-space models, using automata learning for warm-starting SSM training.
3/9/2026
Revisiting the (Sub)Optimality of Best-of-N for Inference-Time Alignment
Theoretical analysis of best-of-N sampling for LLM alignment, examining statistical optimality and vulnerability to reward model exploits.
3/9/2026
MIRACL: A Diverse Meta-Reinforcement Learning for Multi-Objective Multi-Echelon Combinatorial Supply Chain Optimisation
Meta-reinforcement learning approach for multi-objective supply chain optimization in dynamic environments with hierarchical learning structure.
3/9/2026
TML-Bench: Benchmark for Data Science Agents on Tabular ML Tasks
Benchmark for evaluating autonomous data science agents on Kaggle-style tabular ML tasks with time constraints, testing 10 open-source LLMs.
3/9/2026
Bridging Domains through Subspace-Aware Model Merging
Research on merging task-specific models for domain generalization, analyzing parameter competition and singular value decomposition of task matrices.
3/9/2026
MoE Lens -- An Expert Is All You Need
Analysis of expert specialization in Mixture of Experts models through routing patterns and early decoding framework to understand inference behavior.
3/9/2026
Self-Auditing Parameter-Efficient Fine-Tuning for Few-Shot 3D Medical Image Segmentation
Parameter-efficient fine-tuning method for adapting foundation models to medical imaging tasks with limited data, including automated adapter configuration.
3/9/2026
Test-Time Adaptation via Many-Shot Prompting: Benefits, Limits, and Pitfalls
Study on test-time adaptation for LLMs using many-shot prompting, analyzing benefits, limits, and failure modes of in-context learning with large demonstration sets.
3/9/2026
Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning
Research on using LLM reasoning with reinforcement learning for molecular optimization tasks, addressing limitations of fine-tuning on reference molecules without optimization trajectories.
3/9/2026
Weak-SIGReg: Covariance Regularization for Stable Deep Learning
Weak-SIGReg covariance regularization technique for stabilizing deep learning training in low-data regimes and architectures like Vision Transformers.
3/9/2026
Omni-Masked Gradient Descent: Memory-Efficient Optimization via Mask Traversal with Improved Convergence
Omni-Masked Gradient Descent (OMGD) memory-efficient optimization via mask traversal with improved convergence for training large language models.
3/9/2026
EvoESAP: Non-Uniform Expert Pruning for Sparse MoE
EvoESAP non-uniform expert pruning method for Sparse Mixture-of-Experts language models with evolved layer-wise sparsity allocation.
3/9/2026
Preventing Learning Stagnation in PPO by Scaling to 1 Million Parallel Environments
Identifies and addresses performance plateaus in PPO by scaling to 1M parallel environments, showing sample-based loss estimation issues during training.
3/9/2026
Improved high-dimensional estimation with Langevin dynamics and stochastic weight averaging
Analysis of Langevin dynamics and stochastic weight averaging for high-dimensional estimation in tensor PCA and single-index models.
3/9/2026
Dynamic Momentum Recalibration in Online Gradient Learning
Dynamic Momentum Recalibration method for SGD variants that adaptively adjusts momentum coefficients to balance bias-variance tradeoff in gradient updates.
3/9/2026
DQE: A Semantic-Aware Evaluation Metric for Time Series Anomaly Detection
DQE semantic-aware evaluation metric for time series anomaly detection addressing bias, false alarms, and threshold sensitivity in existing metrics.
3/9/2026
Partial Policy Gradients for RL in LLMs
Partial Policy Gradients method for reinforcement learning in LLMs that optimizes subsets of future rewards to improve policy gradient accuracy.
3/9/2026
Predictive Coding Graphs are a Superset of Feedforward Neural Networks
Theoretical proof that Predictive Coding Graphs form a mathematical superset of feedforward neural networks, bridging neuroscience and ML.
3/9/2026
FedSCS-XGB -- Federated Server-centric surrogate XGBoost for continual health monitoring
Federated learning protocol using XGBoost surrogates for distributed health monitoring from wearable sensor data in spinal cord injury patients.
3/9/2026
DC-Merge: Improving Model Merging with Directional Consistency
DC-Merge method for merging multiple task-adapted models while maintaining directional consistency in singular spaces and preserving task knowledge.
3/9/2026
Synthetic Monitoring Environments for Reinforcement Learning
Synthetic Monitoring Environments (SMEs) benchmark suite for reinforcement learning with configurable tasks and known optimal policies for white-box agent diagnostics.
3/9/2026
Learning to Solve Orienteering Problem with Time Windows and Variable Profits
Learning-based approach (DeCoST) for solving orienteering problems with time windows and variable profits using decoupled discrete-continuous optimization.
3/9/2026
Agentic retrieval-augmented reasoning reshapes collective reliability under model variability in radiology question answering
Studies how agentic retrieval-augmented reasoning pipelines impact reliability in clinical radiology QA under model variability and heterogeneous LLM deployments.
3/9/2026
Stem: Rethinking Causal Information Flow in Sparse Attention
Stem rethinks sparse attention mechanisms from information flow perspective to reduce quadratic computational complexity during LLM pre-filling for long contexts.
3/9/2026
From Entropy to Calibrated Uncertainty: Training Language Models to Reason About Uncertainty
Three-stage pipeline to post-train LLMs for efficient calibrated uncertainty estimation, improving reliability in high-stakes decision-making applications.
3/9/2026
Adaptive Lipschitz-Free Conditional Gradient Methods for Stochastic Composite Nonconvex Optimization
ALFCG: adaptive projection-free optimization framework for stochastic composite nonconvex minimization without global smoothness constants or line search.
3/9/2026
Adapter-Augmented Bandits for Online Multi-Constrained Multi-Modal Inference Scheduling
Adaptive bandit-based scheduling framework for multi-modal LLM inference under heterogeneous budgets, handling variable modality composition and latency constraints.
3/9/2026
NOBLE: Accelerating Transformers with Nonlinear Low-Rank Branches
NOBLE adds nonlinear low-rank branches to transformer linear layers for pretraining from scratch, achieving parameter efficiency different from LoRA/PEFT adapters.
3/9/2026
COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamics
Proposes COLD-Steer, training-free framework for steering LLM activations via in-context one-step learning dynamics without retraining.
3/9/2026
Causal Interpretation of Neural Network Computations with Contribution Decomposition
Introduces CODEC, sparse autoencoder method for causal interpretation of neural network computations via contribution decomposition.
3/9/2026
A recipe for scalable attention-based MLIPs: unlocking long-range accuracy with all-to-all node attention
Presents AllScAIP, attention-based machine learning interatomic potential using all-to-all node attention for long-range interactions.
3/9/2026
Information-Theoretic Privacy Control for Sequential Multi-Agent LLM Systems
Studies privacy preservation in sequential multi-agent LLM systems through information-theoretic controls against inference attacks.
3/9/2026
RoboLayout: Differentiable 3D Scene Generation for Embodied Agents
Introduces RoboLayout for generating differentiable 3D scene layouts from language instructions feasible for embodied agent interaction.
3/9/2026
Attention Meets Reachability: Structural Equivalence and Efficiency in Grammar-Constrained LLM Decoding
Analyzes grammar-constrained LLM decoding as coupling between autoregressive distribution and reachability oracle over context-free grammars.