JustGRPO
Summary
JustGRPO is the reinforcement-learning method introduced by The Flexibility Trap. It treats a bidirectional masked diffusion language model as an autoregressive policy only during RL rollouts and GRPO optimization: observed prefix tokens are followed by a fully masked suffix, and the next-position logits define an exact left-to-right policy factorization.
The model architecture remains non-causal. After training, the resulting checkpoints can still use parallel diffusion samplers at inference.
Role In The Wiki
Use this page as the object card for the method, codebase, full checkpoints, and LoRA adapters. The source page carries the detailed evidence, social-source audit, limitations, and time-series/world-model transfer boundary.
The key design pattern is different commitment schedules for training and inference:
- training confronts high-entropy logical forks sequentially so RL can explore and assign credit;
- inference can denoise several positions in parallel;
- the best rollout order need not be the best deployment execution order.
Evidence
Official Artifacts
- Project: The Flexibility Trap
- Code: LeapLabTHU/JustGRPO — MIT-licensed training and evaluation implementation.
- Full GSM8K checkpoint: LLaDA-Instruct-JustGRPO-GSM8K
- Full MATH-500 checkpoint: LLaDA-Instruct-JustGRPO-Math500
- Full code checkpoint: LLaDA-Instruct-JustGRPO-Code
- GSM8K LoRA adapter: LLaDA-Instruct-JustGRPO-GSM8K-LoRA
- MATH-500 LoRA adapter: LLaDA-Instruct-JustGRPO-Math500-LoRA
- Code LoRA adapter: LLaDA-Instruct-JustGRPO-Code-LoRA
The full-checkpoint model cards identify LLaDA 8B Instruct as the base model and use Hugging Face custom_code. The adapter pages existed but lacked model cards at the 2026-07-17 audit; use the official code README for their training configuration.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in flexibility-trap-2026 rather than duplicating verdict rows here. At the entity level, JustGRPO is adjacent to dynamic compute, trajectory generation, and training-policy design: it suggests exposing decision-critical branches sequentially during training while retaining parallel future-block generation at inference. It is not direct evidence for multivariate time series, event streams, control inputs, interventions, or action-conditioned world models.