UniRL Roadmap Q2 2026

# UniRL Roadmap Q2 2026 *Contributions and feedback are welcome.* Pick up any open item by opening a `[Tracking]`/`[RFC]` issue and linking it back here. UniRL post-trains **diffusion** (image / video), **autoregressive VLM / LLM**, and **unified AR+diffusion** models under one RL loop. Sections follow that loop — **rollout → reward → algorithm → training** — plus the **models** they run on and the **tooling** around them. ## Rollout How samples are generated — engines, multimodal capture, cross-engine parity. *(`unirl/rollout/`)* - **Rollout-engine hardening & expansion** — **P1** **Today:** Five engines, uneven maturity; capture/replay coverage varies by engine. **Goals:** Bring every engine (not just vLLM-Omni) to one capability bar + checksummed weight-sync + parity; first-class multi-turn / tool-use rollout for VLM / LLM via `sglang_llm`. - **Cross-engine refactor / consolidation** — **P1** **Today:** Engines fork below the ABC — duplicate adapters, weight-sync ×4, duplicate SDE kernels. **Goals:** One shared substrate (adapters, weight-receive, SDE kernels) so adding an engine is cheap. - **Rollout-engine conformance matrix** — **P2** **Today:** Only `sigma_verify` is enforced; capability coverage is uneven across engines. **Goals:** A `model × engine × algorithm` capability/smoke matrix with tolerance-based parity tests. ## Reward How samples are scored — the reward-model zoo, the standalone reward service, and the gradient path for REFL. *(`unirl/r...