Skip to content

Berkeley CS285 L11–18: From Variational Inference and LLM RL to Offline RL

Aug 22, 2026 1 min
TL;DR L11–18 connect control as inference, LLM RL, model-based RL, and offline RL, with HW4 and HW5 providing two compute-intensive implementations.
Table of Contents
  1. L11–14: control as inference
  2. L15–16: learning or using dynamics
  3. L17–18: learning from a fixed dataset
  4. A reduced self-study path
  5. References

🌏 中文版

The official schedule assigns L11–18 to Variational Inference, VI in RL, Control as Inference, LLM RL, two lectures on Model-Based RL, and two on Offline RL. Together they ask what signal an agent can trust when data, models, and objectives are incomplete.

L11–14: control as inference

L11–13 establish latent-variable and variational-inference machinery, then express optimality as a probabilistic event. Reward, trajectory distributions, and entropy enter one language. Section 6 supports the derivation; Section 7 connects IRL and LLM RL.

L14 applies policy optimization to token policies and verifiable rewards. HW4 implements REINFORCE and GRPO on format-copy and harder math tasks.

It is not a laptop-first exercise. The homework compute ledger owns the supporting H100 and required-run details.

L15–16: learning or using dynamics

Model-based RL learns or uses dynamics, then plans or improves a policy through that model. Data reuse is the advantage; compounding model error is the danger. For Section 8, draw three loops—data collection, model learning, and planning or policy learning—and mark every possible distribution shift.

L17–18: learning from a fixed dataset

Offline RL cannot collect corrective interactions. Out-of-distribution actions can therefore receive overestimated values. HW5 implements SAC+BC, IQL, and FQL on OGBench tasks. Its long-running experiments and tuning burden are detailed in the homework compute ledger.

A reduced self-study path

Read the slides and sections first. For HW4, start with the smallest format-copy run. For HW5, choose one task, seed, and baseline, and verify loading, evaluation, and checkpoints before scaling. The goal is understanding failure modes, not reproducing course-only compute support.

See the series overview's access boundary for current course assets and the proper use of historical video.

References