debajyotidasgupta/repro-contextual-rollout-bandits-for-reinforcement-learning-with-verifiable-rewards-artifacts

Reproduction: Contextual Rollout Bandits for RLVR (ICML 2026, #985)

Independent reproduction of "Contextual Rollout Bandits for Reinforcement Learning with Verifiable Rewards" (Lu, Wang, Chai, Yin, Lin, Chen, Luo, Zhuang, Ban, Wang) — OpenReview weMYE1B16x, arXiv 2602.08499.

Part of the Hugging Face × AlphaXiv ICML-2026 reproduction challenge. Official code: github.com/lxd99/CBS_public (verl 0.5.x fork).

What CBS is

The paper reframes rollout scheduling in RLVR as a contextual bandit: each rollout is an arm with a 10-dim feature vector; its reward is the induced policy-performance gain between consecutive optimization steps. A small neural scheduler (CBS = global reuse over a replay buffer; CBS* = intra-group top-p% selection) uses ε-greedy to pick high-value rollouts, cutting noise and training time.

What reproduces (and what does not)

Full replication needs the released verl recipes trained for 200–500 steps each across 54 cells (Qwen3-1.7B/4B/8B × {GRPO, DAPO, GSPO} × {base, CBS, CBS*}) on an 8-GPU node — many GPU-days, far beyond this budget, and HF Jobs was unavailable (402, no credits). We instead reproduce the method, theory, and efficiency mechanism:

Claim	What we did	Verdict
1 · 10-dim arm	re-implement encoder, check vs Table 1 + official aux_info_names	SUPPORTED (exact)
2 · sub-linear regret (Thm 5.4)	neural ε-greedy bandit sim; fit exponent; vary reward geometry	SUPPORTED (α≈0.63)
3 · buffer bound (Thm 5.3)	φ(M) monotone; decompose Fig.11 inverted-U	SUPPORTED
4 · 6 benchmarks × 3 methods	verify all datasets on Hub; load MATH-500	SUPPORTED (setup)
5 · +5/+5.8/+2% & −50.4%	recompute from Table 2/Fig.2b; GPU time-reduction mechanism	PARTIAL
6 · overhead < 4%	measure faithful residual-MLP scheduler cost	SUPPORTED

Layout

src/cbs/         features.py (10-dim encoder) · scheduler.py (MLP + official residual-MLP,
                 ε-greedy, EMA reward) · buffer.py (FIFO) · bandit.py (synthetic env)
experiments/     claim{1,3,4,6}_*.py, claim2_regret.py, claim5_arithmetic.py,
                 gpu_grpo_timing.py (Claim 5B/6 on a real GPU), make_plots.py
outputs/         per-claim CSV/JSON results + figures/
make_poster.py   build the reproduction poster (poster_embed.html)
build_logbook.py populate the Trackio logbook

Rerun

pip install torch numpy scipy plotly datasets transformers
for c in claim1_features claim2_regret claim3_buffer claim4_setup claim5_arithmetic claim6_overhead; do
    python experiments/$c.py
done
python experiments/make_plots.py
# GPU mechanism (Claim 5B/6):
python experiments/gpu_grpo_timing.py --model Qwen/Qwen3-0.6B-Base --n_prompts 32 --group 8

Compute: local CPU/MPS + one Vast.ai RTX 3090 (~$0.12/hr). Total ~1–2 h, < $1.