ai-sherpa/raco-repro-bundle

RACO reproduction bundle (arXiv 2602.02495)

Independent reproduction of "Reward-free Alignment for Conflicting Objectives" (RACO) for the ICML-2026 Agent Reproducibility Challenge. No official code exists; everything here is implemented from the paper (Algorithm 1, Appendix B.1).

Layout

src/moo.py               CAGrad, CAGrad-Clip (RACO), MGDA, weighted-sum + Appendix-B.1 closed form
src/synthetic.py         high-dim conflicting-quadratics problem (Exp B/C)
src/toy_cagrad.py        the canonical non-convex CAGrad toy (Exp A)
tests/test_moo.py        correctness tests (closed form vs solver, clip binding, ...)
run_tier1.py             Tier 1 driver -> outputs/tier1/{results.json, *.png}
tier2/repro_beavertails.py   Tier 2 UV script: multi-objective DPO on BeaverTails (runs on HF Jobs)
outputs/tier1/           Tier 1 results + figures
outputs/tier2/           Tier 2 results pulled back from the HF Job

Reproduce

Tier 1 (CPU, deterministic, ~3 min):

pip install -r requirements.txt
python -m pytest tests/test_moo.py -v      # algorithm correctness
python run_tier1.py                         # -> outputs/tier1/results.json + figures

Tier 2 (GPU, on Hugging Face Jobs):

hf jobs uv run --flavor a100-large --timeout 3h -d --secrets HF_TOKEN \
  -e MODEL_ID=Qwen/Qwen3-1.7B -e STEPS=180 -e N_TRAIN=1500 -e N_EVAL=200 \
  -e BATCH=4 -e MAXLEN=384 -e CRADIUS=0.5 -e LORA_R=16 \
  -e WEIGHTS=0.1,0.3,0.5,0.7,0.9 -e METHODS=dpo_lw,cagrad,raco \
  -e OUT_REPO=<user>/raco-beavertails-results \
  tier2/repro_beavertails.py

The same script runs on CPU with a tiny model for smoke-testing (set MODEL_ID to a tiny model and small STEPS/N_TRAIN).

Method (one line)

RACO = vanilla CAGrad on per-objective DPO gradients + the single new step p̃ = min(p*, w) (clip correction weights to user weights). See src/moo.py::raco_update.