4/22/2026
SCATR: Simple Calibrated Test-Time Ranking
SCATR method for test-time scaling in LLMs using efficient ranking scorers as alternative to expensive process reward models.
4/22/2026
REZE: Representation Regularization for Domain-adaptive Text Embedding Pre-finetuning
REZE framework for domain-adaptive text embeddings using contrastive pre-finetuning with representation regularization to prevent task-induced bias.
4/22/2026
Cat-DPO: Category-Adaptive Safety Alignment
Category-adaptive safety alignment method for LLMs that addresses safety across different harm categories rather than uniform safety scoring.
4/22/2026
Beyond the Bellman Fixed Point: Geometry and Fast Policy Identification in Value Iteration
Research paper analyzing geometric structure of Q-value iteration for faster policy identification in dynamic programming and Markov decision processes.
4/22/2026
DuQuant++: Fine-grained Rotation Enhances Microscaling FP4 Quantization
Research paper on DuQuant++ fine-grained rotation method improving MXFP4 quantization for efficient LLM inference on NVIDIA hardware.
4/22/2026
LEPO: Latent Reasoning Policy Optimization for Large Language Models
Latent reasoning policy optimization for LLMs with stochastic sampling via Gumbel-Softmax for improved reasoning diversity.
4/22/2026
Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation
Study of diversity collapse in multi-agent LLM systems during open-ended idea generation.
4/22/2026
Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations
Modular representation compression method for adapting LLMs into industrial recommendation systems efficiently.
4/22/2026
MM-JudgeBias: A Benchmark for Evaluating Compositional Biases in MLLM-as-a-Judge
Benchmark evaluating compositional biases in multimodal LLMs used as automatic judges.
4/22/2026
STaD: Scaffolded Task Design for Identifying Compositional Skill Gaps in LLMs
Framework for identifying compositional skill gaps in LLMs through scaffolded task variations.
4/22/2026
Sessa: Selective State Space Attention
Selective State Space Attention architecture combining transformer attention with state-space models for long-context sequences.
4/22/2026
Compile to Compress: Boosting Formal Theorem Provers by Compiler Outputs
Method leveraging compiler outputs to improve formal theorem proving with LLMs at reduced test-time compute.
4/22/2026
Easy Samples Are All You Need: Self-Evolving LLMs via Data-Efficient Reinforcement Learning
Data-efficient RL approach for LLM self-improvement using easy samples and cognitive learning theory.
4/22/2026
Curiosity-Critic: Cumulative Prediction Error Improvement as a Tractable Intrinsic Reward for World Model Training
Intrinsic reward mechanism for world model training based on cumulative prediction error improvement.
4/22/2026
Discrete Tilt Matching
Discrete Tilt Matching method for fine-tuning masked diffusion LLMs using likelihood-free RL objectives.
4/22/2026
Towards Understanding the Robustness of Sparse Autoencoders
Study evaluating robustness of Sparse Autoencoders integrated into LLM inference for interpretability against jailbreak attacks.
4/22/2026
Streaming Structured Inference with Flash-SemiCRF
Technical paper on Semi-Markov CRFs for efficient streaming structured sequence inference without materializing large tensors.
4/22/2026
Efficient Mixture-of-Experts LLM Inference with Apple Silicon NPUs
Research on optimizing Mixture-of-Experts LLM inference on Apple Neural Engine hardware, addressing routing and operator compatibility challenges.
4/22/2026
HELM: Harness-Enhanced Long-horizon Memory for Vision-Language-Action Manipulation
HELM framework for vision-language-action models addressing memory, verification, and recovery gaps in long-horizon robot manipulation tasks.
4/22/2026
Rethinking Dataset Distillation: Hard Truths about Soft Labels
Analysis of dataset distillation showing soft labels mask underperformance versus random baselines; proposes hard-label evaluation for reliable coreset comparison.
4/22/2026
Curvature-Aware PCA with Geodesic Tangent Space Aggregation for Semi-Supervised Learning
Geodesic Tangent Space Aggregation PCA extends PCA to curved manifolds integrating curvature awareness for semi-supervised representation learning.
4/22/2026
One Step Forward and K Steps Back: Better Reasoning with Denoising Recursion Models
Denoising recursion models improve looped transformer reasoning on complex problems by learning long refinement trajectories from noise to structured solutions.
4/22/2026
Task Switching Without Forgetting via Proximal Decoupling
Proximal decoupling approach for continual learning addressing task switching without catastrophic forgetting through explicit separation of learning and retention signals.
4/22/2026
AC-SINDy: Compositional Sparse Identification of Nonlinear Dynamics
AC-SINDy extends sparse identification of nonlinear dynamics using arithmetic circuits for compositional feature construction instead of explicit feature libraries.
4/22/2026
Harmful Intent as a Geometrically Recoverable Feature of LLM Residual Streams
Analysis of harmful intent as geometrically recoverable feature in LLM residual streams across multiple model families and alignment variants.
4/22/2026
Collaborative Contextual Bayesian Optimization
Collaborative contextual Bayesian optimization extending BO to context-specific optimal design problems requiring mapping from context space to solutions.
4/22/2026
Fine-Tuning Small Reasoning Models for Quantum Field Theory
Fine-tuning study of small 7B reasoning models on theoretical physics with analysis of domain-specific reasoning ability development in LLMs.
4/22/2026
TabEmb: Joint Semantic-Structure Embedding for Table Annotation
Joint semantic-structure embedding approach for table annotation capturing column semantics and inter-column relationships for downstream NLP applications.
4/22/2026
FlowForge: A Staged Local Rollout Engine for Flow-Field Prediction
Deep learning surrogate model for CFD flow-field prediction using staged local rollout with shared lightweight predictors instead of global single-pass models.
4/22/2026
Distillation Traps and Guards: A Calibration Knob for LLM Distillability
Analysis of knowledge distillation failures in LLMs identifying distillation traps (tail noise, off-policy instability, teacher-student gap) and calibration methods.
4/22/2026
Self-Improving Tabular Language Models via Iterative Group Alignment
Self-improving language models for tabular data that learn from generated samples via iterative group alignment and RL-based training to improve statistical properties.
4/22/2026
Low-Rank Adaptation for Critic Learning in Off-Policy Reinforcement Learning
LoRA-based structural regularization for critic learning in off-policy RL. Parameter-efficient scaling technique for RL training.
4/22/2026
Decompose, Structure, and Repair: A Neuro-Symbolic Framework for Autoformalization via Operator Trees
Neuro-symbolic framework for autoformalization using operator trees and LLMs. Hierarchical approach to formal math translation.
4/22/2026
Guiding Distribution Matching Distillation with Gradient-Based Reinforcement Learning
RL-guided distribution matching distillation for few-step generation. Research on combining RL with diffusion distillation.
4/22/2026
Accelerating trajectory optimization with Sobolev-trained diffusion policies
Diffusion policy approach using Sobolev training to accelerate trajectory optimization. ML research on policy learning efficiency.
4/22/2026
FedProxy: Federated Fine-Tuning of LLMs via Proxy SLMs and Heterogeneity-Aware Fusion
Federated fine-tuning method for LLMs balancing IP protection, privacy, and performance on heterogeneous data. Novel training approach.
4/22/2026
Local Linearity of LLMs Enables Activation Steering via Model-Based Linear Optimal Control
Research on activation steering for LLM alignment using linear optimal control. Novel inference-time alignment technique with theoretical grounding.
4/22/2026
FG$^2$-GDN: Enhancing Long-Context Gated Delta Networks with Doubly Fine-Grained Control
FG²-GDN enhances gated delta networks with fine-grained channel-wise control and learnable decay rates for long-context linear attention.
4/22/2026
Policy Gradient Primal-Dual Method for Safe Reinforcement Learning from Human Feedback
Safe RLHF framework using constrained MDPs and policy gradient primal-dual methods for safe LLM alignment from human feedback.
4/22/2026
Learning Posterior Predictive Distributions for Node Classification from Synthetic Graph Priors
Uses synthetic graph priors to train GNNs for universal generalization across diverse graphs without retraining per-graph.
4/22/2026
Intentional Updates for Streaming Reinforcement Learning
Proposes intentional updates for stable streaming reinforcement learning by solving for step sizes matching intended outcome changes.
4/22/2026
Robust Continual Unlearning against Knowledge Erosion and Forgetting Reversal
Addresses continual machine unlearning with mechanisms preventing knowledge erosion and forgetting reversal across multiple unlearning rounds.
4/22/2026
LLMs Know They're Wrong and Agree Anyway: The Shared Sycophancy-Lying Circuit
Identifies shared attention heads in LLMs that detect false statements but agree with sycophantic user pressure, enabling behavior control.
4/22/2026
RL-ABC: Reinforcement Learning for Accelerator Beamline Control
RLABC is open-source Python framework automating RL environment creation from Elegant beamline configurations for accelerator optimization.
4/22/2026
Nexusformer: Nonlinear Attention Expansion for Stable and Inheritable Transformer Scaling
Nexusformer replaces linear Q/K/V projections with nonlinear attention expansion enabling stable transformer scaling without retraining from scratch.
4/22/2026
SAW-INT4: System-Aware 4-Bit KV-Cache Quantization for Real-World LLM Serving
SAW-INT4 provides practical 4-bit KV-cache quantization for LLM serving respecting system constraints like paged memory and fused attention.
4/22/2026
LBLLM: Lightweight Binarization of Large Language Models via Three-Stage Distillation
LBLLM framework achieves W(1+1)A4 quantization in LLMs through three-stage distillation for resource-constrained deployment.
4/22/2026
The Logical Expressiveness of Topological Neural Networks
Analyzes logical expressiveness of topological neural networks compared to graph neural networks using Weisfeiler-Leman hierarchy.
4/22/2026
TEMPO: Scaling Test-time Training for Large Reasoning Models
TEMPO scales test-time training for large reasoning models by addressing reward signal drift with external calibration mechanisms.
4/22/2026
RDP LoRA: Geometry-Driven Identification for Parameter-Efficient Adaptation in Large Language Models
RDP LoRA uses geometric analysis of hidden state trajectories to identify optimal layers for parameter-efficient fine-tuning in LLMs, improving upon standard LoRA methods.