Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Researchers propose training-free continual learning for multimodal time series forecasting using outcome-driven prompt evolution with a frozen LLM.
Researchers introduced a success-aware critic learning method to improve reinforcement learning credit assignment for language agents.
Researchers propose SliMOO, an evolutionary search method for LLM depth pruning that optimizes layer selection across multiple tasks.
Research paper examines whether optimized language model agent harnesses elicit existing capabilities or expand new task boundaries.
Researchers propose RANGER, combining language models and reinforcement learning for multi-hop reasoning over knowledge graphs.
A new arXiv preprint evaluates dynamic expert-skipping methods in Mixture-of-Experts models for inference efficiency.
Researchers propose Self-Credit, an RL method using intra-group rescaling to address sparse rewards in long-horizon LLM agents.
New research explores how multi-turn LLM agents can dynamically learn to allocate inference budgets across environment interactions.
A systematic empirical study on arXiv investigates when LLMs improve tabular feature selection using semantic names and statistics.
Researchers propose the Token-Energy Law, a model linking GPU energy per output token to active bytes and realized decode time.
Researchers propose a dynamic dual-feedback conformal inference method for uncertainty quantification in time series forecasting.
Researchers identify selection bias issues in membership inference attacks used for auditing differential privacy guarantees.
Researchers propose new counterfactual metrics for generative limit order book models to accurately simulate market impact.
Researchers propose guided diffusion for private synthetic text generation to address training data privacy risks.
Researchers propose Vector-Borda Empirical-Gap Elimination (VB-EGE) for efficient multi-objective LLM evaluation and Pareto frontier mapping.
Researchers propose an anytime-valid runtime monitor for learned barrier certificates to detect safety deterioration in stochastic systems.
Researchers identify nine fidelity issues in ExeBench, a standard benchmark used for evaluating LLM-based decompiler performance.
Researchers propose formalizing LLM agent collaboration as block optimization with a global coordinator to ensure convergence.
Researchers propose an information decomposition framework to rigorously assess emergence and synergy in language and multi-agent AI systems.
Researchers propose a framework using LLM-driven evolution to optimize semantic identifiers for generative retrieval systems.
A newly announced textbook provides a hands-on introduction to building federated learning systems for privacy-preserving model training.
Researchers propose OPTIMA, a one-shot LLM pruning method using quadratic programming reconstruction to reduce inference costs.
Researchers propose CRC, a selective residual correction framework to prevent overfitting in time-series forecasting models.
Researchers demonstrate that fixed tokens-per-parameter training ratios distort neural scaling law extrapolations for language models.
Researchers introduce STRABLE, a benchmarking suite for tabular machine learning models handling string and numeric data features.
Researchers propose LightSplit, a split learning method using orthogonal projections to balance communication efficiency and privacy.
QwenGyre is a reinforcement learning framework designed to train extreme-long-horizon LLM agents across extensive token rollouts.
Researchers propose Loopy, a low-bit quantization framework designed to reduce memory footprints and inference costs of looped language models.
Researchers propose SiamSA-PPM, a self-supervised framework using Siamese learning and data augmentation to address labeled data scarcity in process prediction.
Researchers propose Magneton, a method for identifying and debugging software-level energy inefficiencies in machine learning systems.