Dongsheng Hou

Framing context reranking and pruning as a cooperative game to surface high-value evidence with Shapley-style contributions.

TL;DR: From Spatial to Spectral, we introduce a plug-and-play Decompose–Enhance–Reconstruct (DER) operator instantiated in the backbone/neck/head to preserve high-frequency cues for efficient small-object detection across diverse detectors.

BCPPO combines constrained PPO with bootstrapped cost-critic disagreement and a Bachelier-inspired expected-excess penalty, encouraging caution around uncertain cost predictions while preserving standard critic learning and policy-only deployment.

WebCover frames automated web GUI testing as submodular reinforcement learning, rewarding marginal coverage gains and prioritizing untried actions to reduce redundant interactions and improve coverage within fixed testing budgets.

Introducing a controlled evidence-compute protocol for evaluating individual stages in multi-stage RAG cascades. Across multiple rerankers, readers, and datasets, semantic coarse stages improve efficiency and often quality, while downstream reranker value depends strongly on the surrounding composition.

PermVLA treats action-chunk factorization order as a regularizer for vision-language-action learning. Its causally anchored permutation objective creates multiple conditional prediction tasks from each demonstration, improving left-to-right training across LIBERO, LIBERO-Plus, and CALVIN.

Expected-value Factor-Graph Interface (EFGI) is a checked, fail-closed handoff for expected-value gradients with parametric discontinuities. It resolves supported boundaries into an unexpanded fixed-domain expression, reconstructs sparse factors, verifies bucket-elimination replay, and only then applies automatic differentiation—preserving the declared finite quadrature's value and gradient under stated conditions without full-grid tensorization.