4/21/2026
FUSE: Ensembling Verifiers with Zero Labeled Data
Unsupervised method for improving LLM verification by ensembling verifiers without labeled data, useful for training and deployment.
4/21/2026
GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling
Gumbel-Softmax based scalar quantization technique for LLM weight compression achieving sub-4-bit accuracy without vector quantization overhead.
4/21/2026
ConforNets: Latents-Based Conformational Control in OpenFold3
Method for controlling protein conformational states in OpenFold3 via latent variable manipulation to capture biologically relevant alternate conformations.
4/21/2026
Benchmarking System Dynamics AI Assistants: Cloud Versus Local LLMs on CLD Extraction and Discussion
Benchmark comparing cloud-based and open-source LLMs on causal loop diagram extraction and system dynamics discussion tasks.
4/21/2026
Back into Plato's Cave: Examining Cross-modal Representational Convergence at Scale
Empirical study examining whether neural networks trained on different modalities converge to shared representations, challenging the Platonic Representation Hypothesis.
4/21/2026
MathNet: a Global Multimodal Benchmark for Mathematical Reasoning and Retrieval
Large-scale multimodal multilingual benchmark of Olympiad-level math problems for evaluating reasoning in LLMs and embedding-based retrieval systems.
4/21/2026
RAYEN: Imposition of Hard Convex Constraints on Neural Networks
Framework for enforcing hard convex constraints on neural network outputs during training and inference for robotics applications.
4/21/2026
Neural Network-Based Score Estimation in Diffusion Models: Optimization and Generalization
Theoretical analysis of score estimation optimization and generalization in diffusion models trained via gradient descent on neural networks.
4/21/2026
Condense, Don't Just Prune: Enhancing Efficiency and Performance in MoE Layer Pruning
Method for pruning Mixture-of-Experts layers in LLMs to reduce memory requirements while maintaining performance through condensation rather than removal.
4/21/2026
Improving Dynamic Object Interactions in Text-to-Video Generation with AI Feedback
Technique using AI feedback to improve text-to-video model generation, specifically addressing dynamic object interactions and physics violations.
4/21/2026
Test-Time Alignment via Hypothesis Reweighting
Method for real-time LLM personalization via hypothesis reweighting that adapts model outputs to individual user preferences with minimal labeled examples.
4/21/2026
Rethinking Uncertainty Estimation in LLMs: A Principled Single-Sequence Measure
Research on efficient uncertainty estimation in LLMs using single-sequence measures instead of computationally expensive multi-sequence methods for trustworthiness evaluation.
4/21/2026
On the Interpolation Effect of Score Smoothing in Diffusion Models
Theoretical study of score smoothing in diffusion models enabling interpolation and generation of new data.
4/21/2026
Judge a Book by its Cover: Investigating Multi-Modal LLMs for Multi-Page Handwritten Document Transcription
Evaluation of multi-modal LLM prompting strategies for zero-shot handwritten document transcription without fine-tuning.
4/21/2026
Two-Stage Regularization-Based Structured Pruning for LLMs
Two-stage pruning method for reducing LLM parameters with regularization to preserve knowledge during deployment.
4/21/2026
PiCa: Parameter-Efficient Fine-Tuning with Column Space Projection
Parameter-efficient fine-tuning method using column space projection as alternative to LoRA for foundation models.
4/21/2026
Saddle-To-Saddle Dynamics in Deep ReLU Networks: Low-Rank Bias in the First Saddle Escape
Theoretical analysis of gradient descent dynamics in deep ReLU networks escaping saddle points with low-rank bias.
4/21/2026
Vision Language Models are Biased
Analysis of bias in vision-language models on objective visual tasks like counting and identification.
4/21/2026
LIFT the Veil for the Truth: Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning
Parameter-efficient fine-tuning method using rank reduction for LLM reasoning tasks with reduced computational cost.
4/21/2026
Rethinking Post-Unlearning Behavior of Large Vision-Language Models
Study of vision-language model privacy concerns and degenerate outputs after machine unlearning.
4/21/2026
End-to-End Optimization of LLM-Driven Multi-Agent Search Systems via Heterogeneous-Group-Based Reinforcement Learning
Framework for optimizing multi-agent search systems coordinating LLM agents with tools using reinforcement learning.
4/21/2026
Textual Bayes: Quantifying Prompt Uncertainty in LLM-Based Systems
Method for quantifying uncertainty in LLM-based systems and prompt sensitivity without access to model internals.
4/21/2026
Time Series Forecasting as Reasoning: A Slow-Thinking Approach with Reinforced LLMs
LLM-based time series forecasting using reinforcement learning and reasoning approach instead of fast pattern extraction.
4/21/2026
"Faithful to What?" On the Limits of Fidelity-Based Explanations
Research on limitations of fidelity-based explanations in XAI, introducing linearity score to measure linear decodability of neural network behavior.
4/21/2026
From 2:4 to 8:16 sparsity patterns in LLMs for Outliers and Weights with Variance Correction
Demonstrates 8:16 semi-structured sparsity for LLM compression, outperforming N:M sparsification methods while handling outlier weights.
4/21/2026
EvoCoT: Overcoming the Exploration Bottleneck in Reinforcement Learning
Introduces EvoCoT to address exploration bottleneck in RL with verifiable rewards for LLM post-training, improving reasoning on hard problems without teacher models.
4/21/2026
Beyond Memorization: Extending Reasoning Depth with Recurrence, Memory and Test-Time Compute Scaling
Studies multi-step reasoning in LLMs using cellular automata framework to understand how models learn and execute reasoning beyond memorization.
4/21/2026
A Ridge Too Far: Correcting Over-Shrinkage via Negative Regularization
Negative regularization technique addressing over-shrinkage in small-data regression through controlled anti-shrinkage.
4/21/2026
Bi-LoRA: Efficient Sharpness-Aware Minimization for Fine-Tuning Large-Scale Models
Bi-LoRA integrates Sharpness-Aware Minimization with LoRA for efficient fine-tuning of large models with improved generalization.
4/21/2026
RefineStat: Efficient Exploration for Probabilistic Program Synthesis
RefineStat method for probabilistic program synthesis combining small language models with statistical exploration under domain constraints.
4/21/2026
Low-rank Orthogonalization for Large-scale Matrix Optimization with Applications to Foundation Model Training
Low-rank orthogonalization technique for matrix optimization improving foundation model training efficiency.
4/21/2026
PiERN: Token-Level Routing for Integrating High-Precision Computation and Reasoning
PiERN architecture enables token-level routing to integrate high-precision numerical computation within LLM reasoning.
4/21/2026
Flow marching for a generative PDE foundation model
Flow Marching algorithm combining neural operators with flow matching for generative PDE foundation models.
4/21/2026
Central Limit Theorems for Asynchronous Averaged Q-Learning
Theoretical analysis establishing central limit theorems for asynchronous Q-learning with convergence rate bounds.
4/21/2026
CaTS-Bench: Can Language Models Describe Time Series?
CaTS-Bench benchmark for evaluating language models on time series captioning and temporal reasoning across 11 domains.
4/21/2026
Semantic-Space Exploration and Exploitation in RLVR for LLM Reasoning
Framework for LLM reasoning via reinforcement learning that optimizes semantic space exploration rather than token-level proxies.
4/21/2026
Downgrade to Upgrade: Optimizer Simplification Enhances Robustness in LLM Unlearning
Method for improving robustness of LLM unlearning by simplifying optimizers, addressing fragility of privacy/safety interventions.
4/21/2026
OptunaHub: A Platform for Black-Box Optimization
OptunaHub platform for distributed black-box optimization components with unified interface, supporting AutoML and hyperparameter tuning.
4/21/2026
On the Convergence and Size Transferability of Continuous-depth Graph Neural Networks
Theoretical convergence analysis of Graph Neural Differential Equations combining GNNs with continuous-depth architectures.
4/21/2026
RACE Attention: A Strictly Linear-Time Attention Layer for Training on Outrageously Large Contexts
RACE Attention proposes linear-time attention layer to handle contexts exceeding 4M tokens, addressing quadratic complexity of softmax attention.
4/21/2026
Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention
Mechanistic analysis of training failures with low-precision transformers and Flash Attention, revealing root causes of loss explosion.
4/21/2026
The Potential of Second-Order Optimization for LLMs: A Study with Full Gauss-Newton
Research on full Gauss-Newton preconditioning for LLM training, establishing practical iteration complexity bounds up to 150M parameters.
4/21/2026
PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling
PaTaRM: Reward modeling approach bridging pairwise and pointwise preference signals for improved RLHF alignment of LLMs with human preferences.
4/21/2026
Finding Culture-Sensitive Neurons in Vision-Language Models
Study identifying culture-sensitive neurons in vision-language models that preferentially activate for culturally grounded inputs and visual reasoning.
4/21/2026
FlexiCache: Leveraging Temporal Stability of Attention Heads for Efficient KV Cache Management
FlexiCache: KV cache management technique exploiting temporal stability of attention heads to reduce memory overhead in long-context LLM serving.
4/21/2026
In Situ Training of Implicit Neural Compressors for Scientific Simulations via Sketch-Based Regularization
In situ training protocol for implicit neural compressors using sketch-based regularization to prevent catastrophic forgetting in scientific simulations.
4/21/2026
ConMeZO: Adaptive Descent-Direction Sampling for Gradient-Free Finetuning of Large Language Models
ConMeZO: Zeroth-order optimizer for gradient-free LLM finetuning using adaptive descent-direction sampling, reducing memory overhead and convergence time.
4/21/2026
Preparation of Fractal-Inspired Computational Architectures for Automated Neural Design Exploration
FractalNet: Framework generating 1,200+ neural network variants systematically through permutations for automated architecture search and LLM analysis.
4/21/2026
Empowering Multi-Turn Tool-Integrated Agentic Reasoning with Group Turn Policy Optimization
Group Turn Policy Optimization: RL training method for multi-turn tool-integrated reasoning in LLMs with fine-grained turn-level rewards improving agentic performance.
4/21/2026
DeepThinkVLA: Enhancing Reasoning Capability of Vision-Language-Action Models
DeepThinkVLA: Framework identifying necessary conditions for chain-of-thought reasoning to improve vision-language-action robot models through systematic analysis.