Contribution
A training objective, independent noise per token over a causal model, that turns the choice between next-token prediction and full-sequence diffusion into a choice made at sampling time. Also sampling schemes that use this, including zig-zag schedules and Monte Carlo Guidance.
Key insight
Noise level is a continuous mask. Teacher forcing trains “clean past, predict next”. Full-sequence diffusion trains “equally noisy everything”. Training on random per-token noise covers both and everything in between, so the sampler can hold the past slightly noisy (for robustness), keep the future noisier than the present (for causal uncertainty), or guide the whole horizon.
Assumptions
- A causal architecture with a latent state. The implementation is a convolutional RNN
- Standard DDPM hyperparameters
- Small, continuous-token domains: 2D mazes, Minecraft and DMLab video, one robot task, time series
Key results
- Video (Figure 3): stable to 1,000 frames where baselines with the same RNN diverge. Qualitative only
- Maze2D (Table 1): single-task average 141.7 against Diffuser 119.5 (with a PD controller) and 8.7 (without). 129.7 without MCG
- Robot: 80% success, 76% under occlusion, against 48% for a next-frame diffusion baseline
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | The objective lower-bounds the likelihood of all training subsequences | strong | proof (appendix, not read) |
| C2 | Conditioning on slightly noised history stabilizes long autoregressive rollouts of continuous tokens | weak | qualitative video only |
| C3 | Per-token noise enables better guided planning than full-sequence diffusion | moderate | D4RL mazes, with ablation of MCG |
| C4 | The generated actions are causally consistent with the generated states | moderate | Diffuser collapses when executing its own actions and DF does not |
Concepts
- Teacher forcing — training on the ground-truth history
- Zig-zag schedule — denoise the near future ahead of the far future
- Monte Carlo Guidance — average guidance gradients over several sampled futures
Connections
It unifies next-token prediction and full-sequence diffusion (video diffusion, Diffuser for planning, diffusion policy for control). SIREN (LIT-551) shares an author, Sitzmann, but no method.
Recommendations
- R1 — For autoregressive generation of continuous tokens, train with per-token noise and condition on slightly noised history. Filed as SOTA-333
Bearing on the record
- SOTA-333 is new
Limitations
- The headline video result is unquantified
- RNN only, small scale. Transformer and scaling are named as future work
- Few tasks in each domain
Open questions
- How much of the rollout stability survives in large transformer video models, and how it compares with other anti-drift methods