A 4B-parameter model trained on expert-curated rubrics came within 0.5 percentage points of a baseline set by a model roughly 60 times larger.
The paper, published on arXiv, introduces ComplexConstraints, a benchmark and training suite built around granular, human-authored evaluation criteria rather than the programmatic checklist tests that dominate current benchmarks. The benchmark includes a public set of 75 prompts with 1,559 rubric criteria and a disjoint 1,000-prompt training set, each prompt paired with 10 to 40 atomic criteria. The authors argue that standard programmatic benchmarks miss the semantic and context-dependent behavior that real instruction following actually demands — and that rubric-based rewards fix that gap during reinforcement learning.
According to the arXiv paper, training the 4B model on ComplexConstraints raised its mean criterion pass rate by 15.5 percentage points on a held-out split, with gains transferring to benchmarks the model never saw: +8.4 pp on AdvancedIF and +10.1 pp on MultiChallenge. A parallel experiment in CoreCraft, a stateful agentic environment, showed similar transfer (+4.5 pp on BFCL, +7.4 pp on tau^2-Bench). The common thread is that better reward signals — not bigger models — drove the improvements.
The efficiency story here matters more than the benchmark numbers. Labs racing to scale parameter counts have largely ignored how much headroom sits inside the reward function itself. If rubric-based RL consistently extracts this much performance from small models, the calculus on expensive pre-training runs gets harder to justify — though one paper's results rarely survive contact with production workloads.