3/6/2026
OPPO: Accelerating PPO-based RLHF via Pipeline Overlap
Proposes OPPO, a pipeline overlap technique to accelerate PPO-based RLHF training for aligning large language models with human preferences.
3/6/2026
Breaking and Fixing Defenses Against Control-Flow Hijacking in Multi-Agent Systems
Demonstrates control-flow hijacking attacks in multi-agent systems and evaluates LlamaFirewall defense mechanisms for securing inter-agent communications.
3/6/2026
LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery
LLEMA framework coupling LLM scientific knowledge with evolutionary algorithms and memory for multi-objective materials discovery.
3/6/2026
Revisiting Multimodal KV Cache Compression: A Frequency-Domain-Guided Outlier-KV-Aware Approach
Frequency-domain-guided compression method for multimodal LLM KV cache using outlier-awareness, compatible with FlashAttention.
3/6/2026
Guided Flow Policy: Learning from High-Value Actions in Offline Reinforcement Learning
Guided Flow Policy method for offline RL that distinguishes high-value actions through weighted behavior cloning with flow-matching policies.
3/6/2026
Bootstrapped Mixed Rewards for RL Post-Training: Injecting Canonical Action Order
RL post-training technique using mixed rewards and canonical action ordering to improve transformer performance on structured problem solving.
3/6/2026
Sparse Attention Post-Training for Mechanistic Interpretability
Post-training method that sparsifies transformer attention to ~0.4% of edges while maintaining performance for mechanistic interpretability studies.
3/6/2026
Revolutionizing Mixed Precision Quantization: Towards Training-free Automatic Proxy Discovery via Large Language Models
Uses LLMs to automatically discover proxies for mixed-precision quantization without manual design, reducing training-free DNN optimization costs.
3/6/2026
RePo: Language Models with Context Re-Positioning
Proposes RePo, a method to improve in-context learning in LLMs by dynamically repositioning context based on Cognitive Load Theory principles.
3/6/2026
Controlled LLM Training on Spectral Sphere
Research on optimization strategies for scaling LLMs, introducing Spectral Sphere method to address weight drift limitations in μP and Muon optimizers.
3/6/2026
ButterflyMoE: Sub-Linear Ternary Experts via Structured Butterfly Orbits
ButterflyMoE compresses MoE expert parameters to sub-linear scaling via structured butterfly orbits, enabling deployment on edge devices.
3/6/2026
Yuan3.0 Ultra: A Trillion-Parameter Enterprise-Oriented MoE LLM
Open-source trillion-parameter MoE LLM with 68.8B activated parameters using Layer-Adaptive Expert Pruning for enterprise and general-purpose tasks.
3/6/2026
Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models
On-policy self-distillation framework where student LLMs generate their own trajectories while receiving dense token-level supervision from teacher models.
3/6/2026
MiTA Attention: Efficient Fast-Weight Scaling via a Mixture of Top-k Activations
MiTA Attention mechanism using mixture of top-k activations for efficient fast-weight scaling in transformers with extremely long sequences.
3/6/2026
Adaptive Rollout Allocation for Online Reinforcement Learning with Verifiable Rewards
VIP allocation strategy optimizing rollout sampling efficiency in reinforcement learning with verifiable rewards, improving upon fixed GRPO allocation.
3/6/2026
Position: Beyond Model-Centric Prediction -- Agentic Time Series Forecasting
Position paper proposing agentic time series forecasting with iterative refinement, reasoning-driven inference, and continual adaptation beyond static prediction.
3/6/2026
On the Non-Identifiability of Steering Vectors in Large Language Models
Analysis proving steering vectors in LLMs are fundamentally non-identifiable due to large equivalence classes, challenging behavioral control interpretability.
3/6/2026
Beyond the Unit Hypersphere: Embedding Magnitude in Contrastive Learning
Framework showing embedding magnitude should be learned independently in contrastive learning, benefiting retrieval and RAG applications.
3/6/2026
Empirical Stability Analysis of Kolmogorov-Arnold Networks in Hard-Constrained Recurrent Physics-Informed Discovery
Study evaluating Kolmogorov-Arnold Networks in physics-informed neural network architectures for discovering unknown terms in oscillatory systems.
3/6/2026
Learn from Your Mistakes: Self-Correcting Masked Diffusion Models
Framework enabling masked diffusion models to correct their own tokens iteratively, reducing error accumulation in parallel generation.
3/6/2026
QTabGAN: A Hybrid Quantum-Classical GAN for Tabular Data Synthesis
Hybrid quantum-classical GAN framework for synthesizing realistic tabular data with heterogeneous features and privacy constraints.
3/6/2026
Out-of-Support Generalisation via Weight-Space Sequence Modelling
Weight-space sequence modelling technique to improve neural network extrapolation on out-of-support datapoints beyond training distribution.
3/6/2026
Curriculum Learning for Efficient Chain-of-Thought Distillation via Structure-Aware Masking and GRPO
Three-stage curriculum learning framework for distilling chain-of-thought reasoning from large LLMs into compact student models while preserving interpretability.
3/6/2026
cc-Shapley: Measuring Multivariate Feature Importance Needs Causal Context
cc-Shapley method for measuring multivariate feature importance in ML models using causal context.
3/6/2026
Zatom-1: A Multimodal Flow Foundation Model for 3D Molecules and Materials
Zatom-1 open-source foundation model for unified generative and predictive learning on 3D molecules and materials.
3/6/2026
Learn Hard Problems During RL with Reference Guided Fine-tuning
Reference-guided fine-tuning approach for reinforcement learning on mathematical reasoning tasks with sparse rewards.
3/6/2026
AOI: Turning Failed Trajectories into Training Signals for Autonomous Cloud Diagnosis
AOI framework for training LLM agents to diagnose cloud infrastructure failures using failed trajectories as learning signals, addressing safety and data constraints.
3/6/2026
RADAR: Learning to Route with Asymmetry-aware DistAnce Representations
Neural solver for vehicle routing problems using distance representation learning for asymmetric real-world scenarios.
3/6/2026
Why Are Linear RNNs More Parallelizable?
Analysis of why linear RNNs achieve transformer-level parallelizability compared to nonlinear RNNs for language modeling.
3/6/2026
Distilling Privileged Information for Dubins Traveling Salesman Problems with Neighborhoods
Learning approach for Dubins Traveling Salesman with neighborhoods using knowledge distillation from expert trajectories.
3/6/2026
Sink equilibria and the attractors of learning in games
Theoretical research on game theory attractors and replicator dynamics, characterizing learning equilibria.
3/6/2026
Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-Tuning and Can Be Mitigated by Machine Unlearning
Research paper on safety mirage in vision language models: spurious correlations in safety fine-tuning and mitigation via machine unlearning.
3/6/2026
Assessing the Impact of Code Changes on the Fault Localizability of Large Language Models
Evaluates LLM fault localization capabilities on code changes, assessing semantic program reasoning beyond syntax and lexical features.
3/6/2026
Ice Cream Doesn't Cause Drowning: Benchmarking LLMs Against Statistical Pitfalls in Causal Inference
Benchmark evaluating LLMs on rigorous causal inference tasks involving statistical pitfalls relevant to medicine, economics, and public policy.
3/6/2026
ShIOEnv: A Command Evaluation Environment for Grammar-Constrained Synthesis and Execution Behavior Modeling
ShIOEnv: Gymnasium environment for grammar-constrained synthesis and modeling of command-line interface behavior with shell input-output data.
3/6/2026
Highly Efficient and Effective LLMs with Multi-Boolean Architectures
Novel framework using multi-kernel Boolean parameters for weight binarization in LLMs to improve efficiency without full-precision latent weights.
3/6/2026
SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models
SealQA benchmark for evaluating search-augmented LLMs on fact-seeking questions with conflicting or noisy web search results.
3/6/2026
EDINET-Bench: Evaluating LLMs on Complex Financial Tasks using Japanese Financial Statements
EDINET-Bench: New benchmark dataset evaluating LLMs on complex financial document analysis tasks using Japanese financial statements.
3/6/2026
MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model Pretraining
MuRating framework transfers English data-quality signals to score documents in 17 languages for multilingual LLM pretraining.
3/6/2026
Elucidating the Design Space of Arbitrary-Noise-Based Diffusion Models
Research generalizes EDM to arbitrary-noise diffusion models, analyzing design space beyond Gaussian noise for image restoration tasks.
3/6/2026
Structured quantum learning via em algorithm for Boltzmann machines
Quantum EM algorithm for training quantum Boltzmann machines that circumvents barren plateau problem in quantum machine learning.
3/6/2026
Dropping Just a Handful of Preferences Can Change Top Large Language Model Rankings
Study shows LLM ranking systems are sensitive to small changes in preference data, with top model rankings changeable by dropping few preferences.
3/6/2026
How Quantization Shapes Bias in Large Language Models
Comprehensive evaluation of how weight and activation quantization affects model bias across stereotypes, fairness, toxicity, and sentiment.
3/6/2026
New Insights into Optimal Alignment of Acoustic and Linguistic Representations for Knowledge Transfer in ASR
Research on optimal alignment of acoustic and linguistic representations in pre-trained models for automatic speech recognition.
3/6/2026
BabyHuBERT: Multilingual Self-Supervised Learning for Segmenting Speakers in Child-Centered Long-Form Recordings
BabyHuBERT self-supervised speech model trained on 13,000 hours multilingual child-centered recordings for speaker segmentation.
3/6/2026
Diffusion-Based Impedance Learning for Contact-Rich Manipulation Tasks
Framework combining diffusion models with impedance control for robot learning in contact-rich manipulation tasks.
3/6/2026
Noise-to-Notes: Diffusion-based Generation and Refinement for Automatic Drum Transcription
Noise-to-Notes reformulates automatic drum transcription as conditional generative task using diffusion modeling.
3/6/2026
BridgeDrive: Diffusion Bridge Policy for Closed-Loop Trajectory Planning in Autonomous Driving
BridgeDrive applies diffusion-based planning with expert behavior anchors for closed-loop autonomous driving trajectory planning.
3/6/2026
BeyondBench: Contamination-Resistant Evaluation of Reasoning in Language Models
BeyondBench framework uses algorithmic problem generation for contamination-resistant evaluation of reasoning in language models.
3/6/2026
Hyperspherical Latents Improve Continuous-Token Autoregressive Generation
SphereAR addresses variance collapse in continuous-token autoregressive image generation by constraining latents to hypersphere.