OPEN LINK ↗
// 1h agoRESEARCH PAPER
Celeris Shares Outcome-Routed Diffusion RL Paper
Celeris researchers propose a likelihood-free post-training pipeline for diffusion language models that routes rollout groups by verifier outcomes, combining ratio-free reinforcement learning with selective teacher distillation. The paper reports an 18.31-point tool-calling improvement over a distilled checkpoint while preserving broader capabilities with anchored KL supervision.
// ANALYSIS
Celeris addresses a genuine weakness in applying standard policy-ratio RL to bidirectional denoising models, but the results suggest targeted gains still require careful capability preservation.
- –Mixed-success rollout groups receive group-centred REINFORCE updates without unreliable policy likelihood ratios
- –All-failed student groups can receive dense teacher supervision when the teacher succeeds, while low-signal groups are skipped
- –Routing improved direct-answer accuracy by 13.21 points but slightly reduced reasoning-trace accuracy
- –A KL anchor recovered trace performance, showing the method needs explicit safeguards against reward overfitting
- –The tool-calling gains replicated across seeds, while the math result remained seed-dependent
// TAGS
celerisllmfine-tuningtrainingreasoningresearch
DISCOVERED
1h ago
2026-10-16
PUBLISHED
2h ago
2026-10-15
RELEVANCE
9/ 10
AUTHOR
Celeris_ai