Skip to content
conftrace_
2026 ICML ICML 2026

Position: RL Should Be Used to Adjust Foundation Models, NOT Abused

This position paper argues that reinforcement learning (RL) should be used to adjust foundation models after pretraining and cold-start supervision, not abused as a default recipe for capability creation or early-stage training. We view RL as a high-cost, high-leverage post-training operator that most reliably reallocates probability mass toward behaviors a model can already express, improving correctness, consistency, and constraint satisfaction. This is not an impossibility claim: RL may discover new behaviors when meaningful supervision or scaffolding is unavailable and verification is strong, but under current foundation-model practice it should not be treated as the default path to reasoning capability. Across modalities and domains, we emphasize a recurring regularity: supervision establishes usable reasoning structure, whereas RL mainly sharpens behavior under constraints. We further advocate reward minimalism through auditable, verifiable, and minimally composed rewards, and discuss self-supervised RL only as a boundary case for structured interaction settings. Together, these arguments motivate treating RL as a disciplined adjustment stage with explicit entry criteria and compute-accountable evaluation. Blog: https://aigeeksgroup.github.io/DontAbuseRL/.

More from ICML 2026

All ICML 2026 papers →