mavrl 0.1

Unified Multi-modal Feedback using Amortized Variational Inference

MAVRL — Multi-feedback Amortized Variational Reward Learning

Code release for the ICML 2026 camera-ready of "MAVRL: Learning Reward Functions from Multiple Feedback Types with Amortized Variational Inference". Implements an amortized variational reward learner that combines preference, demonstration, rating and stop feedback in a single ELBO.

Repository layout

• mavrl/ — the algorithm: encoders, feedback likelihoods, losses, retraining utilities. import mavrl.
• mavrl_experiments/ — experiment infrastructure: Optuna search, file queues, table printers, CLI. Depends on mavrl.
• scripts/ — entrypoints. reproduce_paper.sh, reproduce_run.py, per-figure/table renderers, and SLURM launchers for re-running on a cluster.
• results/ — the canonical artifact bundle: model checkpoints, hyperparameter logs, MCMC results, normalization constants, rendered tables, paper source, published figures.
• FeedbackInformativeness/ — the Julia MCMC pipeline for the baseline-comparison.
• expert_policies/ — pre-trained external policies used as Boltzmann-rational simulators (inputs, not our outputs; see the README inside).
• tests/ — pytest suite for the algorithm + feedback models.

Top-level entry-point scripts (train.py, transfer.py, evaluate_reward_model.py) live at the repo root.

Installation

conda env create -f environment.yml
conda activate mavrl-env
pip install -e .

Reproducing the paper

All tables and figures regenerate from results/. No Optuna journals, cluster access, or retraining required.

conda activate mavrl-env
bash scripts/reproduce_paper.sh

Each step is skipped if its output exists; pass --force to rebuild everything.

Paper element	Renderer	Output	Runtime
tab:fixed_allocation_combined_paper	python -m mavrl_experiments.fixed_allocation_table --camera-ready-root results	results/tables/fixed_allocation.tex	~5 s
tab:equal_budget_combined	python -m mavrl_experiments.equal_budget_table --camera-ready-root results	results/tables/equal_budget.tex	~5 s
tab:misspec-excerpt, tab:app-full-misspec	python scripts/summarize_misspec.py --tex …	results/tables/misspec.tex	~1 s
tab:baselines (Post-Hoc Avg row)	python scripts/baseline_comparison_table.py	stdout (manual splice)	~1 s
tab:baselines (MCMC row)	inspect results/mcmc/*_combined_*.json	manual	—
fig:robustness-ood	python scripts/plot_combined_transfer.py	results/figures/transfer_combined.{png,pdf}	~10 s
fig:appendix-transfer-*	python scripts/plot_individual_transfer.py	results/figures/transfer_*_full.{png,pdf}	~30 s
fig:qualitative-comparison-app-grid-trap	python scripts/visualize_checkpoint_comparison.py --config results/qualitative_configs/grid_trap.json …	results/figures/appendix_grids/final-fig_grid-trap.png	~5 s
fig:qualitative-comparison-app-grid-{cliff,sparse}	committed PNG (regeneration deferred to a follow-up checkpoint bundle)	—	—

Reproducing a single training run

Each cell under results/models/{equal_budget,fixed_allocation}/<env>/<subset>/ ships per-seed checkpoints plus a hparams.json with the full merged training config. To re-train one seed:

# Grid env (~15 s):
python scripts/reproduce_run.py \
    --cell results/models/fixed_allocation/grid_trap/pdrs --seed 0

# Control env (~15 min):
python scripts/reproduce_run.py \
    --cell results/models/fixed_allocation/cartpole_v1/pdrs --seed 0

# Lunar Lander (~15 min):
python scripts/reproduce_run.py \
    --cell results/models/fixed_allocation/lunar_lander_v3/pdrs --seed 0

The script prints the produced final eval metric next to the published _meta.per_seed_values[seed] so you can eyeball the match. Swap --cell and --seed to cover any other cell.

Bit-exact reproduction vs. from-scratch retraining

Loading a published r_model_seed{i}.pt and re-evaluating it reproduces the per-seed metric bit-identically on any machine — that's the canonical reproducibility path.

Re-training the same config from scratch (reproduce_run.py) produces metrics in the same distribution but may bounce between basins per-seed. This is the well-known floor of PyTorch non-determinism across hardware/BLAS backends (Linux x86 + MKL vs macOS arm64 + Accelerate). Deterministic algorithms only fix intra-platform non-determinism, not cross-platform. The aggregate means in the paper tables average this out; individual seeds bounce.

Re-running the full pipeline on a SLURM cluster

For readers who want to redo the experiments end-to-end (not just verify the published numbers), the cluster-side entrypoints are:

• scripts/launch_fixed_allocation_table.sh — 66 Optuna studies for the fixed-allocation table (uses configs/optuna/<env>_fixed_paper.py).
• scripts/launch_equal_budget_table.sh — 66 Optuna studies for the equal-budget appendix table.
• scripts/launch_transfer_fixalloc.sh — transfer/perturbation experiments (consumes the published encoders from results/models/fixed_allocation/).
• scripts/launch_misspec.sh — misspecification robustness sweeps.
• scripts/pregenerate_datasets.py — populates the dataset cache before launching the Optuna sweeps.
• FeedbackInformativeness/scripts/submit_grid_mcmc_euler.sh — Julia MCMC baseline.

All launchers read --help-style docstrings at the top of each file. They default to writing into $SCRATCH/mavrl/optuna_studies/.

Testing

pytest tests/

Citation

@inproceedings{baur2026mavrl,
  title={MAVRL: Learning Reward Functions from Multiple Feedback Types with Amortized Variational Inference},
  author={Baur, Rapha\"el and Metz, Yannick and Gkoulta, Maria and El-Assady, Mennatallah and Ramponi, Giorgia and Kleine Buening, Thomas},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026},
}