Eishaan/rcdp-dueling-bandits-repro

Reproduction bundle — RCDP-UCB (ICML 2026 #3478)

Independent reproduction of "Robust Linear Dueling Bandits with Post-serving Context under Unknown Delays and Adversarial Corruptions" (Youngmin Oh, ICML 2026 Poster).

• —Paper: arXiv:2605.01752 · OpenReview RaJnDY8aAS
• —Official code: https://github.com/youngmin0oh/rcdp-public (cloned at commit 3ed73c6)
• —Trackio logbook (full writeup): see the linked Space in the collection.

What's here

code/        the official reference implementation + our reproduction harness
  contextual_dueling_bandit.py   env, RCDP-UCB (DuelingGLMLearner), baselines  [official]
  experiments.py                 experiment registry + driver                   [official]
  repro_scaling.py               our harness: figure configs + T/C/delay scaling
  verify_lower_bound.py          our Theorem 5.3 (lower bound) verifier
  PROVENANCE.txt                 upstream commit
outputs/shards/<tag>/            HF-Job outputs (one folder per CPU job shard)
  curves/*.csv                   per-round mean/std regret curves per config
  summary_*.json                 finals, growth exponents, scaling fits
  lower_bound_verification.json  Theorem 5.3 numeric checks

How it was run

All experiments run on CPU (the workload is numpy Sherman–Morrison updates + a tiny 2-layer MLP; the paper's own repo says "no GPU needed"). Local smoke tests + Hugging Face Jobs (cpu-upgrade flavor) for the paper-scale sweep (T=2000, n_runs=10).

Reproduce a single config locally:

pip install numpy pandas matplotlib torch
python code/repro_scaling.py --mode fast --T 2000 --n_runs 10   # linear + scaling
python code/repro_scaling.py --mode custom --exps post-serving --T 2000 --n_runs 10
python code/verify_lower_bound.py

Headline outcome

RCDP-UCB attains the lowest cumulative regret vs 4 baselines (RCDB, ColSTIM, MaxInP, MaxPairUCB) across linear and post-serving settings, with sub-√T empirical growth, and is robust to both stochastic and strategic (adversarial) delay under one fixed provisioning — reproducing the paper's core claims. The adaptive weight ω=min(1,α/‖Δz‖), α=√d/(C+D), the replay-buffer + weighted-MLE algorithm, and the Ω(√(dΛ)) lower-bound mechanism were all verified.