Reproduce “CODA: Difficulty-Aware Compute Allocation for Adaptive Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Siye Wu, Jian Xie, Yikai Zhang, Yanghua Xiao
CODA trains reasoning models to spend inference computation according to problem difficulty rather than applying a uniform token budget or length penalty. Using rollout-group success rates to gate length shaping, it cuts average generation by 16% for an 8B model while slightly improving average accuracy, with especially large savings on easy tasks and preserved reasoning depth on difficult mathematics problems.
Key insights
- Inference tokens have different marginal value across problems: easy instances reach an accuracy plateau quickly, while hard instances can continue benefiting from additional search and self-verification.
- Adaptive allocation is distinct from globally shortening outputs; CODA compresses reasoning primarily on easy instances and retains or slightly expands computation on difficult ones.
- Correctness gating is necessary to avoid length-seeking behavior, and moderate penalty strengths provide better efficiency without suppressing useful exploration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.