remdm-planner-minihack

ReMDM Planner — MiniHack artefacts

Anonymous release accompanying Return-Weighted ELBO Fine-Tuning Degrades Masked Diffusion Planners (under double-blind review). It contains the code, the trained checkpoints and the pre-computed results behind every MiniHack number in the paper.

Anonymity notice. This repository is an anonymised mirror prepared for double-blind review. Author names, institutional affiliations, experiment-tracking identifiers and absolute filesystem paths have been removed or replaced with placeholders. Cluster hardware is referred to only as GPU-24GB and GPU-H200.

What is here

src/  configs/  environments/  experiments/  scripts/  tests/
main.py  pyproject.toml  uv.lock  LICENSE  demo_minihack.ipynb

checkpoint_inference.pth   Stripped EMA weights only (~20 MB), what the notebook loads
checkpoints/
  online/Minihack-Online-Diffusion-DAgger-100M/    DAgger planner (the checkpoint every
                                                   ablation fine-tunes)
  offline/Minihack-Offline-Diffusion-BC-100M/      Offline BC planner

ablation_assets/           Flat copy of the assets the notebook reads:
                           9 PNGs + 4 CSVs + results.json

experiments/rl_finetuning/outputs/minihack_ablations/   The published 25-condition suite
  figures/   113 pre-computed PNGs
  tables/    main_results.csv, group_summary.csv, hypothesis_verdict.csv,
             gradient_analysis.csv, significance_test.txt, gdelta.csv, results.tex, ...
  gdelta/    gdelta_{seed0,seed1,seed2,aggregate}.json   <- the return-term measurement
  results.json, diagnosis.md

results/*.csv          Per-analysis summaries (main results, per-env win rates,
                       forgetting, gradient diagnostics, representation drift, verdicts)
results/inference/     Single-seed evaluations of both checkpoints
results/paper_figures/ The 11 manuscript figures as vector PDF

Headline numbers (MiniHack, in-distribution win rate)

Three conditions finish nominally above the checkpoint, but the best-versus-baseline difference has a bootstrap interval of [−2.1, +13.7] points at p = 0.40, so the paper does not read them as improvements.

MiniHack is not a second confirmation, and the paper says so. The effect here is small relative to a seed standard deviation of 6.1 points. It is reported because it is the same suite run identically and its ordering matches Craftax Classic, where the effect is large and unambiguous. The central claim rests on Craftax Classic.

The return term

experiments/rl_finetuning/outputs/minihack_ablations/gdelta/ holds the direct measurement of g_δ at the pretrained checkpoint, over 3 rollout seeds × 8 noise draws on 4,608-window batches. Reproduce it on a CPU with run_ablations.py --measure-gdelta.

MiniHack does not look like Craftax Classic here: ‖g_δ‖/‖∇L_BC‖ = 1.956 ± 0.275 at cosine 0.803 ± 0.039 (imitation against itself gives 0.986 ± 0.006 on this benchmark), against 0.485 ± 0.007 at cosine 0.02 ± 0.05 on Craftax Classic. A shuffled-δ null costs about five sixths of the ratio here and about a quarter there, so the ratio is largely weight dispersion on Craftax Classic and largely not here.

⚠ One caveat the paper records. The Eq.-4 correctness residual is 4.8 × 10⁻⁵ here against 4.8 × 10⁻⁷ on Craftax Classic — two orders looser, and it did not tighten when the measurement went from 3 draws to 8.

The review arms are not in this release

Three questions outside the 25-condition suite were run on both benchmarks — the unweighted control (bc_all), a learning-rate sweep, and advantage_clip at a matched effective step. The Craftax Classic run directories are published, in AnonMLuser/remdm-planner-craftax under experiments/rl_finetuning/outputs/review_*/.

The MiniHack run directories are not part of this release. Their numbers are reported in the paper and quoted in demo_minihack.ipynb, but no artefact in this bundle backs them:

Unlike Craftax Classic, no same-stack anchor was run here, so these are quoted against the published suite rather than a contemporaneous rerun.

Quick start

from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="AnonMLuser/remdm-planner-minihack", local_dir="remdm-planner-minihack")

Then open demo_minihack.ipynb, or see the project documentation below.

Citation

@inproceedings{anonymous2026returnweighted,
  title     = {Return-Weighted ELBO Fine-Tuning Degrades Masked Diffusion Planners},
  author    = {Anonymous},
  booktitle = {Submitted to NeurIPS 2026},
  year      = {2026},
  note      = {Under double-blind review}
}

---

ReMDM Planner for MiniHack

PyTorch implementation of ReMDM (Remasking Discrete Diffusion Model) for action-sequence planning in MiniHack navigation environments. A dual-stream transformer generates 64-step action plans by iteratively denoising masked token sequences, conditioned on a 9x9 local crop and the full 21x79 dungeon map. Trained with DAgger under BFS oracle supervision, from scratch; generalises zero-shot from 4 in-distribution to 3 out-of-distribution environments.

The sibling repository remdm-planner-craftax implements the same method in JAX on Craftax. Both repos share the same CLI, config layout and README structure; commands transfer between them by swapping the repo name and benchmark-specific values.

Method

The planner starts from a fully-masked action sequence and iteratively unmasks tokens over K denoising steps via the ReMDM Algorithm 1 posterior (per-token Bernoulli unmasking), while ReMDM remasking lets committed tokens be re-predicted for plan refinement.

Two independent training pipelines are compared head-to-head, both supervised by the built-in BFS oracle: --mode online runs DAgger from scratch (primary), --mode offline behaviour-clones a pre-collected oracle dataset. Either output is scored with --mode inference. See Architecture and Diffusion for details.

Setup

Prerequisites: Python 3.12+, uv. nle compiles from source on macOS. Linux GPU use needs NVIDIA driver >= 580 for CUDA 13, or >= 525 with --extra cuda12.

# macOS (arm64)
brew install cmake

# Linux (x86_64), if building from source
sudo apt-get install build-essential cmake bison flex libbz2-dev

git clone https://github.com/ANONYMOUS/remdm-planners.git
cd remdm-planner-minihack

# Default. macOS gets the PyPI build (MPS); Linux gets PyPI's CUDA 13.0 build.
# Installs the dev group (pytest) too.
uv sync

# Linux, explicit CUDA 13.2 (driver >= 580)
uv sync --extra cuda13

# Linux, CUDA 12.6 fallback (driver >= 525, or Maxwell/Pascal cards)
uv sync --extra cuda12

Extras: cuda13 and cuda12 are mutually exclusive and Linux-only. Neither is needed on a modern driver — plain uv sync already yields a CUDA 13.0 build on Linux; use cuda12 only if nvidia-smi reports a driver older than 580.

Install path must not contain spaces. MiniHack's mh_patch_nhdat.sh interpolates paths unquoted and fails silently on whitespace, leaving every environment as the same default level with no goal staircase. src/envs/minihack_env.py detects this and substitutes a Python implementation, but a space-free path avoids it entirely.

Repo layout

remdm-planner-minihack/
├── configs/                Experiment configs (defaults.yaml + presets, see Configuration)
├── environments/           Custom .des scenario files (empty; user-supplied)
├── src/                    Model, diffusion, envs, planner pipelines
├── experiments/
│   └── rl_finetuning/      RL fine-tuning ablation suite (run_ablations.py)
├── scripts/                HF upload utilities, DAgger and ablation profilers
├── tests/                  Smoke suite — uv run pytest
├── checkpoints/            Gitignored — offline/, online/ (see Checkpoints)
├── results/                Gitignored, created on demand — inference/ eval JSONs and
│                           paper_figures/ manuscript PDFs, both published (see Checkpoints)
├── demo_minihack.ipynb     Demo notebook
├── main.py                 CLI entry point
└── pyproject.toml          uv project — deps, cuda extra, dev group

Quickstart

Collects a few oracle trajectories, trains under a 5k env-step budget, prints ID evaluation. A few minutes on CPU.

python main.py --mode smoke

Training

Two independent training methods; neither depends on the other. An offline BC checkpoint can warm-start DAgger via --checkpoint, but this was not used for the paper results.

Online DAgger (primary)

python main.py --mode online                                            # full paper recipe (defaults.yaml)
python main.py --mode online --config configs/final_minihack_gpu_24gb.yaml      
python main.py --mode online --override total_timesteps=1000000 --override dagger_lr=0.0001
python main.py --mode online --checkpoint checkpoints/iter600.pth       # resume
python main.py --mode online --checkpoint checkpoints/iter600.pth --no-warm-start

Per iteration: curriculum-sampled model rollouts, BFS oracle labelling on the same seeds, efficiency filtering into the replay buffer, then grad_steps_per_iteration gradient steps — see DAgger training loop.

Offline BC

First collect a dataset, then train on it:

python main.py --mode collect                                    # 5000 eps/env -> data/dataset.pt
python main.py --mode collect --data data/small.pt --override collect_episodes_per_env=2000

python main.py --mode offline --data data/dataset.pt
python main.py --mode offline --data data/dataset.pt --override total_timesteps=500000

# Resume (restores optimizer, scheduler, step counter, W&B run)
python main.py --mode offline --data data/dataset.pt --checkpoint checkpoints/offline_step40000.pth

Gradient steps default to total_timesteps // offline_batch_size; ID + OOD eval runs on the id_eval_every_timesteps / ood_eval_every_timesteps cadence. The offline_*_grad_steps keys override that in grad-step units — see the hazard note under Configuration.

Evaluation from a checkpoint

python main.py --mode inference --checkpoint checkpoints/iter600.pth    # all ID + OOD
python main.py --mode inference --checkpoint wandb:entity/project/checkpoint-iter600:latest

# Specific environments, save JSON
python main.py --mode inference --checkpoint checkpoints/iter600.pth \
    --envs MiniHack-Room-Random-5x5-v0 MiniHack-MazeWalk-45x19-v0 \
    --episodes 100 --output results/inference/eval.json

python main.py --mode inference --checkpoint checkpoints/iter600.pth \
    --des environments/<your_level>.des        # custom .des scenarios (dir ships empty)
python main.py --mode inference --checkpoint checkpoints/iter600.pth --no-ema

--checkpoint accepts a local .pth path or a wandb: artifact reference (wandb:entity/project/name:version). Inference uses EMA weights unless --no-ema is given.

Write eval JSONs into results/inference/ (created for you): scripts/hf_upload.py publishes every JSON it finds there.

Match the config to the checkpoint. The model is built from the config, not the checkpoint, and a mismatch raises at load. Always evaluate with the checkpoint's own config snapshot:

DIR=checkpoints/online/Minihack-Online-Diffusion-DAgger-100M
python main.py --mode inference --config $DIR/config.yaml --checkpoint $DIR/iter563.pth

Baselines and ablations

RL and imitation baselines

Six algorithms: SB3 discrete-action RL (ppo, a2c, dqn, ppo-rnn), Behavioural Cloning (bc) on oracle demos, and a causal Decision Transformer (dt). All share total_timesteps, so numbers are comparable to DAgger and offline BC. Hyperparameters live under the baselines_* config namespace; outputs go to baselines_output_dir.

python main.py --mode baselines --algo ppo                        # any of the six
python main.py --mode baselines --algo dqn --seeds 0 1 2          # explicit seeds
python main.py --mode baselines --algo bc --num-seeds 3           # or a seed count
python main.py --mode baselines --algo ppo --output results/ppo.json

Architecture ablations

# Local-only planner (no global stream, no goal head), trained from scratch
python main.py --mode online --config configs/ablation_local_only.yaml

# Blind-global: zero the global observation of a trained dual-stream model at eval
python main.py --mode inference --checkpoint checkpoints/iter600.pth --blind-global

RL fine-tuning ablation suite

26 registered ablations (same names as in the craftax repo). See experiments/README.md.

python experiments/rl_finetuning/run_ablations.py --list
python experiments/rl_finetuning/run_ablations.py --checkpoint path/to/ckpt.pth --all
python experiments/rl_finetuning/run_ablations.py \
    --checkpoint wandb:entity/project/checkpoint-iter600:latest \
    --ablations baseline_rl kl_penalty --fast

Pass --emit-tex-macros to also write tables/results.tex, one \newcommand per headline quantity, so the manuscript cites generated numbers instead of retyping them. Macros from this repository are prefixed mh and the sibling suite's rw, so both files can be \input together; the name-mangling rule is shared between the two.

--measure-gdelta splits the return-weighted ELBO gradient into its imitation and return terms at one parameter point. No training, no GPU needed; results land in gdelta/ beside the run's results.json — see experiments/README.md.

python experiments/rl_finetuning/run_ablations.py --measure-gdelta --gdelta-seeds 0 1 2 \
    --checkpoint path/to/ckpt.pth \
    --results-path outputs/minihack_ablations/results.json \
    --output-dir outputs/minihack_ablations

Configuration

One YAML config holds the experiment; the CLI holds the run.

Precedence, lowest to highest: configs/defaults.yaml < --config preset < --override and run flags. Exactly two config layers — a preset never inherits from another preset.

• configs/defaults.yaml: the shared final paper recipe, not a cheap baseline. Both machines train exactly this; running with no --config trains it too.
• Config files (configs/*.yaml): deep-merged onto defaults.yaml, so presets contain only their deltas — never re-state a default value.
• Run flags: --seed, --checkpoint, --data, --output, --episodes, --envs, mode switches.
• --override KEY=VALUE (repeatable): keys are validated against defaults.yaml and cast to the key's type, so a typo is an error, not a silent no-op.

Hazard when writing a preset. Four keys silently override an env-step-derived value when non-null, and defaults.yaml now sets all four as part of the recipe: offline_total_grad_steps, offline_eval_every_grad_steps, offline_checkpoint_every_grad_steps, offline_buffer_capacity. A preset that wants its own total_timesteps to govern the offline budget must pin them back to explicit null — omitting them inherits the pins. Left unpinned, smoke.yaml would train 60,000 offline gradient steps instead of 19. tests/test_config.py enforces the pins for every preset that derives its own budget.

Key hyperparameters are documented inline in configs/defaults.yaml; the appendix tabulates them.

Checkpoints

Training writes to a unique run directory under checkpoint_dir (default checkpoints/), named {tag}_{YYYYMMDD}_{HHMMSS}_{hex4}. DAgger saves iter{N}.pth on the checkpoint_every_timesteps cadence; offline BC saves offline_step{N}.pth and offline_final.pth. With use_wandb on they also upload as versioned W&B artifacts (type model). Every checkpoint stores its W&B run ID, so passing it back via --checkpoint resumes the same curve.

checkpoints/ is gitignored. Released weights live on the Hugging Face Hub: AnonMLuser/remdm-planner-minihack

# All checkpoints
uv run hf download AnonMLuser/remdm-planner-minihack --include "checkpoints/**" --local-dir .

# One checkpoint
uv run hf download AnonMLuser/remdm-planner-minihack \
    --include "checkpoints/online/Minihack-*/**" --local-dir .

Keep the --include. The Hub repo carries its own README.md (the generated model card), LICENSE and .gitattributes; dropping the glob and pulling into --local-dir . overwrites this repository's copies of all three. To fetch everything, add --exclude "README.md" "LICENSE" ".gitattributes", or use a separate --local-dir. Publishing is safe either way — hf_upload.py stages LICENSE and the demo README.md from git, not the working tree.

Each released directory ships <step>.pth (full training state), model.safetensors (EMA weights only, no pickle), config.yaml and selection.json. The -100M suffix counts sample-equivalents, not env steps — the runs behind these train 5,650,000 env steps. See Checkpoint format for the .pth schema and programmatic loading.

Historical note: the released DAgger selection.json records "every": null, "configured_max": null and "unit": "dagger_iterations", written by a selection() that read two since-renamed config keys. It is historical and noncanonical and stays as published (author decision 2026-08-17); the checkpoint's own config_<step>.yaml carries the real cadence and budget. Current code records the candidate set in env steps — "every": 940000, "configured_max": 5650000 for the shipped recipe — and raises rather than writing a null.

Publishing to the Hub

scripts/hf_upload.py rediscovers and uploads four things, each keeping its repo-relative path: checkpoints/ (adding a model.safetensors EMA export and selection.json per directory), every experiments/rl_finetuning/outputs/<run>/ holding a results.json (with diagnosis.md, tables/, figures/, gdelta/), the eval JSONs in results/inference/, and the manuscript figure PDFs in results/paper_figures/. It drops W&B and hub config keys, shortens absolute paths and regenerates the model card.

HF_TOKEN=hf_xxx uv run python scripts/hf_upload.py --repo-id <ANON_HF_REPO_ID> \
    --selection-metric "mean ID+OOD win rate" --dry-run

--dry-run prints the staged tree and card without uploading; drop it to upload. --selection-metric records what the best-of-N checkpoints were chosen on. Also --inference-results <FILE|DIR> ... (eval JSONs kept elsewhere), --private, --yes. Publish one model per directory, with a single .pth and config.

The manuscript figures are built by the sibling repo. Each one puts Craftax Classic and MiniHack side by side, so ../remdm-planner-craftax/scripts/paper_figures.py reads both repositories' ablation results.json and neither can build them alone. Copy the PDFs it emits into results/paper_figures/ here; both Hub repos publish the same set, and the upload warns when they are absent rather than passing over them silently.

A hf download --local-dir . overwrites README.md and LICENSE in the working tree. Publishing is unaffected — hf_upload.py stages LICENSE from git cat-file blob HEAD:LICENSE, and hf_upload_demo.py its bundle's README.md, warning if git cannot be consulted — but restore your own files with git checkout -- README.md LICENSE, or avoid the clobber with the download flags above.

Checkpoint discovery expects the released layout, checkpoints/<role>/<name>/*.pth. A training run writes to its own checkpoints/dagger_<timestamp>/, so copy what you mean to release into checkpoints/{offline,online}/<name>/ first, or nothing is staged. checkpoints/hf/ is skipped — that is where a Hub download lands, and publishing from it would nest already-published artefacts under checkpoints/hf/checkpoints/....

Results, citation, licence

Results tables and the full method description are in Return-Weighted ELBO Fine-Tuning Degrades Masked Diffusion Planners (under submission); demo_minihack.ipynb reproduces the headline comparison. Citation to be added on publication. Licence: MIT, see LICENSE.

---

Appendix: benchmark-specific detail

Environments

Architecture

LocalDiffusionPlannerWithGlobal (~5.2M parameters):

Local stream:   9x9 glyphs -> Embedding(6000,64) -> CNN(64->32->64) -> Linear -> 1 token
Global stream:  21x79 glyphs -> Embedding(6000,32) -> CNN(32->32->64) -> Pool(2,4) -> 8 tokens
                Goal head: mean(global) -> MLP -> [B,2] staircase coords (aux loss)
                Gate: sigmoid(learnable scalar, init=-3.0) * global_tokens
Action stream:  Embedding(14, 256) + timestep_emb(100, 256) + position_emb(64, 256)
Transformer:    concat [1 + 8 + 64 = 73 tokens] -> 4-layer encoder (256D, 4 heads, pre-norm)
Output head:    last 64 tokens -> Linear(256, 12) -> action logits

Signature: (local_obs, global_obs, noisy_action_seq, t_discrete) -> {"actions": [B,64,12], "goal_pred": [B,2]}.

LocalDiffusionPlanner (no global stream, no goal head) is the ablation_local_only variant. Supported by --mode offline and --mode online; the experiments/ ablation suite assumes the goal head is present.

Diffusion

• Forward process (MDLM): each action token is independently replaced with MASK (12) with probability 1 - alpha(t), alpha(t) linear or cosine. PAD (13) is never masked.
• Loss: continuous-time MDLM NELBO: per sample w(t) * sum_masked(CE) / L with w(t) = -alpha'(t) / (1 - alpha(t)) clipped to [0, 1000]; optional label_smoothing.
• Greedy sampling: used for DAgger collection. Same MaskGIT loop, argmax decoding, no temperature/top-K/remasking, diffusion_steps_collect steps.

Reverse sampling (ReMDM Algorithm 1), over K steps (default 10). Per step: predict logits, apply temperature and top-p filtering, sample, and record each committed token's decode probability psi; unmask each masked position independently with posterior probability (alpha_s - (1 - sigma) alpha_t) / (1 - alpha_t); remask each committed position with probability sigma from the configured Section-4.1 schedule. A final greedy cleanup commits anything still masked.

Key hyperparameters

configs/defaults.yaml is authoritative and commented inline. Tabulated here are the keys that change a result, carry a hazard, or are named elsewhere in this README.

Model. n_embd 256, n_head 4, n_layer 4, n_global_tokens 8, seq_len 64, dropout 0.0, global_gate_init -3.0 — the shape every released checkpoint carries (see Architecture); a checkpoint restores only against a matching config. Two model keys are result-affecting in their own right:

Diffusion

Training budget (unified). Offline BC, DAgger and the SB3 baselines share one env-step budget. This is the only knob that should change to scale a run.

• Offline BC: gradient steps = total_timesteps // offline_batch_size, and the cosine LR T_max derives from the same quantity, so any run length decays to the 10% floor at its end.
• DAgger: tracks cumulative env.step() calls (model + oracle) and halts at total_timesteps. episodes_per_iteration and grad_steps_per_iteration set the collect/train ratio and must not scale with the budget.
• Caveat, ema_decay: an absolute-update-count constant (half-life ~ 1 / (1 - decay) steps). Shifting total_timesteps by more than ~2x changes the fraction of training the EMA window covers; set a matching decay by hand for very short or long runs.

Offline grad-step pins. These four override the env-step-derived budget whenever non-null, and defaults.yaml sets all four. A preset whose own total_timesteps should govern must pin them back to explicit null — see the hazard note under Configuration.

Training

Collection, evaluation, performance

Worker counts (collect_num_workers, num_collection_workers, both 8) are machine values. The collect_output, use_wandb, wandb_* and offline_log_every keys mirror the run flags under Configuration, the curriculum_* keys the behaviour under DAgger training loop, and the 21 baselines_* keys hold the SB3/BC/DT hyperparameters; all are commented where they are declared.

DAgger training loop

1. Curriculum sampling: pick an environment weighted by difficulty (low win-rate sampled more).
2. Model rollout: EMA model, greedy sampling, replanning every 16 steps with the executed prefix locked, episodes_per_iteration episodes.
3. Oracle rollout: BFS oracle on the same seed.
4. Efficiency filter: add the oracle trajectory if the model failed or took >1.5x the oracle's steps.
5. Budget accounting: env_steps_total += model_steps + oracle_steps; halt at total_timesteps.
6. Training: sample the buffer, run grad_steps_per_iteration steps, update EMA after each.

BFS oracle priority: (1) kick adjacent doors, (2) BFS to staircase, (3) BFS to frontier, (4) BFS to farthest tile, (5) random cardinal.

Reward shaping

Checkpoint format

# DAgger (iter{N}.pth)
{
    "model_state_dict": ..., "ema_state_dict": ...,
    "optimizer_state_dict": ..., "scheduler_state_dict": ...,
    "curriculum_state": {...},
    "iteration": int,
    "env_steps": int,                  # cumulative env.step() calls
    "wandb_run_id": str | None,
    "rng_states": {"torch", "numpy", "python"},
}

Offline BC step-level (offline_step{N}.pth, when checkpoint_every_timesteps > 0) is the same minus curriculum_state, with step for iteration and env_steps = step * offline_batch_size. Its rng_states is required: resume raises without it. The final offline_final.pth carries only model_state_dict, ema_state_dict and wandb_run_id.

Load programmatically

# Inference, from safetensors (already EMA weights)
from safetensors.torch import load_file
from src.config import load_config
from src.models.denoiser import make_model

DIR = "checkpoints/online/Minihack-Online-Diffusion-DAgger-100M"
cfg = load_config(f"{DIR}/config.yaml")
model = make_model(cfg)
model.load_state_dict(load_file(f"{DIR}/model.safetensors"))
model.eval()

From the full .pth instead, to resume or to pick training over EMA weights: torch.load(..., weights_only=False), then model.load_state_dict(ckpt["model_state_dict"]) and, for what evaluation uses by default, ModelEMA(model, decay=cfg.ema_decay) with load_state_dict(ckpt["ema_state_dict"]) and apply_to(model).

W&B artifacts and run resumption

W&B model artifacts contain the .pth and a config.yaml snapshot; the reference format is wandb:entity/project/artifact-name:version, version being latest, v0, v1. Resuming reads the run ID out of the checkpoint and passes it to wandb.init(resume="must"), so curves continue with no gaps.

# Automatic. A checkpoint predating the feature needs the ID passing by hand:
python main.py --mode online --checkpoint old.pth --override wandb_resume_id=abc123xyz

W&B metric namespaces

Declared in src/planners/logging.py; the key lists there are authoritative.

DAgger and offline BC both emit to eval_id/ and eval_ood/, through the same Evaluator and EMA-weight path.

Performance tuning

use_amp (default true) puts torch.amp.autocast("cuda") + GradScaler in both trainers — roughly 2x on gradient steps, with loss and win rates statistically equivalent to FP32, and a no-op on CPU. torch_compile (default true) shows no measured gain beyond AMP. num_collection_workers affects only the threaded CPU fallback: collection auto-selects GPU-batched (CUDA, episodes_per_iteration > 1) > threaded CPU > sequential.

Profile with python scripts/profile_dagger.py [--override key=value ...].

Testing

uv run pytest            # the default suite
uv run pytest -m slow    # slow entry points only (BC + PPO baselines)

A CPU-only suite, 17 modules. Tiny synthetic data and a shrunken model throughout — no real checkpoints, datasets or network calls, and nothing written outside tmp_path. conftest.py forces CPU and disables W&B; slow marks the multi-second CLI smokes and is deselected by default. For a quality signal, use --mode smoke.

Implementation notes

• MDLM loss returns 0.0 (not NaN) when no masked positions exist. NELBO-weighted per MDLM eq (10).
• PAD tokens are never masked and are excluded from the loss.
• EMA updates after every gradient step, not per iteration. DataCollector syncs EMA weights before each rollout.
• Curriculum starts from a 50/50 prior per environment and buckets the rolling win-rate: [0, 0.15) -> 0.2, [0.15, 0.85) -> 1.0, [0.85, 1.0] -> 0.1.
• Replay buffer pins offline data at the front; only online samples are FIFO-evicted. Returns None when empty.
• Global gate starts at sigmoid(-3.0) ~ 0.047, nearly closed, so the global stream cannot destabilise early training.
• DAgger warm-start: iteration 0 seeds the buffer with 3 oracle trajectories per ID environment (12 total).