Stein Variational Distributional Reinforcement Learning: Score-Guided Particles for Multivariate Return Distributions
We propose Stein variational distributional RL (SVDRL).
ProofPaper ↗
Key points
- Decisions that weigh several reward signals against each other require the joint distribution of the return vector.
- Multivariate distributional critics therefore represent each return distribution by M equal-weight particles trained with a discrepancy loss.
- Its critic outputs Stein particles, which Stein variational gradient descent transports along a learned score field of the distributional Bellman target.
- We learn this field by bootstrapped denoising score matching across noise levels and tie it across the Bellman recursion by a consistency loss, which we prove is a non-expansive temporal-difference scheme in score space.
Sources (1)
- [1]Stein Variational Distributional Reinforcement Learning: Score-Guided Particles for Multivariate Return DistributionsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 08:13 AM
We propose Stein variational distributional RL (SVDRL).
Decisions that weigh several reward signals against each other require the joint distribution of the return vector.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 12, 2026Calibrating LLM Explanations to Human Decisions with Reinforcement Learning
- Oct 12, 2026Growing and Transferring Pathways for LLM Reinforcement Learning
- Oct 12, 2026A Survey of Flow Matching in Reinforcement Learning
- Oct 11, 2026ADAPT: Audio-Depth Action Planning with Transition-Aware Rewards for Lights-Out Industrial Navigation
- Oct 11, 2026AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards
- Oct 11, 2026A Sovereign, Indic-First, Reasoning Model