1

We introduce GradLoc, a white-box diagnostic framework for reinforcement learning with verifiable rewards (RLVR) that localizes global gradient spikes to exact culprit tokens via distributed binary search. GradLoc reveals a new failure mode, layerwise gradient heterogeneity, and motivates LayerClip, a layerwise clipping strategy that stabilizes long-horizon RLVR training more effectively than black-box heuristic mitigations.

We study why exploration collapses in reinforcement learning with verifiable rewards (RLVR) and identify low-probability yet valuable tokens, termed reasoning sparks, as a critical driver of sustained exploration. To preserve these tokens during training, we propose Low-probability Regularization (Lp-Reg), which stabilizes long-horizon on-policy training and improves average accuracy on five math benchmarks.

We introduce PaSa, an advanced Paper Search agent powered by large language models. PaSa can autonomously make a series of decisions, including invoking search tools, reading papers, and selecting relevant references, to ultimately obtain comprehensive and accurate results for complex scholarly queries.

We propose an Implicit Regularization Enhancement (IRE) framework to accelerate the discovery of flat solutions in deep learning, thereby improving generalization and convergence.

We propose an robust AI-generated text detector against adversarial perturbations.

We propose an iteratively parallel generation methods for document-level extraction to alleviate the inadequate training of parallel methods.

We propose a novel Siamese Adversarial Network (SAN) architecture for automatic speech recognition, which aims at solving the difficulty of fuzzy audio recognition.