HINT-WAM: Hindsight-Informed Training for World Action Models
TL;DR HINT-WAM post-trains a world action model's action expert by self-distillation, where the same model, also shown the demonstration's future video, supervises the student on the denoising states its own sampler visits. In simulation on LIBERO-90 it raises success from 78.9% (SFT) to 85.2% and forgets less of the original LIBERO suites.
What they do
Supervised fine-tuning (SFT) of a world action model trains its flow-matching action expert only on noised demonstration actions, never on the states its own sampler produces, and one observation can fit several demonstrated continuations. HINT-WAM post-trains Fast-WAM’s action expert by self-distillation: the same model, given hindsight (the demonstration’s future video), acts as teacher. All experiments are in simulation (LIBERO-90, RoboTwin 2.0).
How it works
- Stage 0: SFT of the 1.02B-parameter action expert; the 5B video expert stays frozen throughout.
- Stage A, hindsight teacher: a small encoder pools future video latents into 8 tokens, read through gated cross-attention adapters (about 13M parameters). Only these are trained, so the teacher’s edge comes from hindsight alone.
- Stage B, distillation: the student runs its own 10-step denoising on demonstration observations; at visited states its implied actions are matched to one teacher sample with the energy score, which matches distributions instead of averaging modes, plus a behavior-cloning anchor. No rewards or simulator rollouts.
Results
- LIBERO-90 (50 trials per task): SFT 78.9%, full SFT 81.2%, HINT-WAM 85.2%. The same extra steps without the teacher drop to 71.6%; a teacher reading another episode’s future reaches 80.3%.
- Objectives: velocity matching 79.0%, MMD 80.6%, action MSE 82.9%, energy score 85.2%. Without the anchor, distillation falls to 74.4%.
- RoboTwin 2.0 hard subset: 87.0% clean / 83.7% randomized, vs 85.4% / 82.1% for the BC control and 86.5% / 83.4% for the released checkpoint.
- Without replay, the original LIBERO suites drop from 96.6% to 44.1% (SFT: 37.5%). Training costs 1.28x SFT and 0.56x full SFT.
Limitations
Simulation only. Observations still come from demonstrations, so compounding errors across control steps are not addressed. RoboTwin gains over the released checkpoint are not significant. 22 of 90 LIBERO-90 tasks get worse, mostly pick-and-place into a basket or caddy.
Videos
Figures



Abstract
World action models (WAMs) have shown promising generalization and data efficiency in robot manipulation by coupling action generation with predictions of future visual dynamics. Translating these capabilities into reliable task execution, however, remains a challenge for existing post-training methods. Offline supervised fine-tuning provides dense supervision on expert trajectories, but its flow-matching action expert is trained only on noised demonstration actions and never on the intermediate states produced by its own sampler; task-specific adaptation can also compromise pretrained capabilities. On-policy reinforcement learning trains on the policy's own behavior, yet sparse task rewards make exploration and credit assignment difficult, while dense reward shaping still provides scalar evaluations rather than direct action targets. We introduce \textbf{HINT-WAM}, the first on-policy self-distillation framework that supervises the action expert on its own sampling trajectories using hindsight, the future video of the demonstration, as privileged information available to the WAM itself. For each demonstration observation, the student generates its action through its own denoising trajectory, and at the states along this trajectory the same model acts as a teacher conditioned additionally on this hindsight, providing a conditional action distribution as a dense training target, while the student sees only deployment-time inputs and needs no environmental rewards or simulator interaction. To transfer this supervision to a flow-matching action expert, HINT-WAM distills the teacher into the student with an energy score, a distribution-level objective that matches the teacher's action distribution from the student's own samples rather than regressing its velocity field. The project web is available in https://tuandattt.github.io/hintwam/.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
