Machine Learning
Papers on learning algorithms, neural networks, deep learning, and optimization.
cs.LG · 1353 papersElastic Attention Cores for Scalable Vision Transformers
VECA introduces elastic core-periphery attention for Vision Transformers, achieving linear-time complexity and competitive performance with learned core tokens.
Task-Adaptive Embedding Refinement via Test-time LLM Guidance
This paper introduces an LLM-guided query refinement method that adapts embedding models in real-time for challenging zero-shot search and classification tasks.
Learning, Fast and Slow: Towards LLMs That Adapt Continually
Fast-Slow Training enables LLMs to adapt continually with improved efficiency and less forgetting by combining fast context and slow parameter updates.
Beyond GRPO and On-Policy Distillation: An Empirical Sparse-to-Dense Reward Principle for Language-Model Post-Training
A new principle for LM post-training uses sparse rewards for strong teachers and dense distillation for students, outperforming direct sparse RL.
MEME: Multi-entity & Evolving Memory Evaluation
MEME is a new benchmark evaluating LLM agents' multi-entity and evolving memory, revealing severe limitations in dependency reasoning.
Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts
This paper reveals a geometric coupling between SMoE routers and experts, explaining how routers learn effective assignment geometry and proposing a coupling-based router.
KV-Fold: One-Step KV-Cache Recurrence for Long-Context Inference
KV-Fold enables stable, training-free long-context inference by treating the KV-cache as an accumulator, achieving high fidelity and memory efficiency.
Solve the Loop: Attractor Models for Language and Reasoning
Attractor Models introduce a stable, efficient fixed-point refinement method for iterative Transformers, significantly boosting performance in language and reasoning tasks.
High-arity Sample Compression
This paper shows that high-arity sample compression schemes imply high-arity PAC learnability, extending learning theory to product spaces.
Search Your Block Floating Point Scales!
ScaleSearch optimizes Block Floating Point quantization scales by searching for minimal error, significantly improving generative model performance.
Towards Affordable Energy: A Gymnasium Environment for Electric Utility Demand-Response Programs
DR-Gym is a new Gymnasium environment for training RL agents to optimize electric utility demand-response programs, improving grid flexibility and affordability.
A proximal gradient algorithm for composite log-concave sampling
A new proximal gradient algorithm efficiently samples from composite log-concave distributions, matching state-of-the-art for specific cases and extending to broader settings.
Multi-Stream LLMs: Unblocking Language Models with Parallel Streams of Thoughts, Inputs and Outputs
Multi-Stream LLMs introduce parallel computation streams to unblock language models, enabling simultaneous reading, thinking, and acting for improved efficiency.
TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection
TextSeal is a new LLM watermark using dual-key generation and multi-region localization for robust, distortion-free detection and distillation protection.
Enabling AI-Native Mobility in 6G: A Real-World Dataset for Handover, Beam Management, and Timing Advance
This paper introduces a real-world dataset from a commercial 5G network to enable AI-native mobility, focusing on handover and timing advance.
ORCE: Order-Aware Alignment of Verbalized Confidence in Large Language Models
ORCE improves LLM verbalized confidence by decoupling its estimation from answer generation and using rank-based optimization for better calibration.
Environment-Adaptive Preference Optimization for Wildfire Prediction
EAPO is a new framework that uses environment-adaptive preference optimization to improve wildfire prediction, especially for rare events and under distribution shifts.
Learning Minimally Rigid Graphs with High Realization Counts
This paper uses reinforcement learning to discover minimally rigid graphs with record-breaking numbers of realizations, improving bounds for spherical graphs.
ORBIT: Preserving Foundational Language Capabilities in GenRetrieval via Origin-Regulated Merging
ORBIT prevents catastrophic forgetting in GenRetrieval LLMs by regulating weight drift, preserving foundational language capabilities.
Stories in Space: In-Context Learning Trajectories in Conceptual Belief Space
LLMs update beliefs in a low-dimensional conceptual space, showing in-context learning as trajectories through this space, grounded in structured representations.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.