ExpDis publishes a paper, Apache-2.0 code and checkpoints, reporting gains over DAPO at equal wall-clock budget.
ExpDis separates exploration from optimization in RLVR
2026-10-08Frontier watchTRIAL
Exploration-Distillation trains novelty-seeking explorer policies, filters their correct trajectories and distills them into a separate student that optimizes without the novelty bonus. The authors report gains over DAPO at the same wall-clock budget across seven math benchmarks and two model families, with more diverse correct solutions.
Official · ReferenceDecoupling Exploration from Optimization in RLVROfficial · ReferenceExploration-Distillation code
Availability: Apache-2.0 code and checkpoints are public
What this means for you
For teams training reasoning models, the method offers a way to scale exploration without letting sparse verifiable rewards damage the production policy. The public code and checkpoints make a small independent replication possible, but the reported gains remain author evidence.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | novelty-bonused explorers -> correctness and quality filtering -> distillation into a separate student |
|---|---|
| Reported Evaluation | seven mathematical reasoning benchmarks and two model families; DAPO baseline at equal wall-clock budget |
| Availability | Apache-2.0 code and checkpoints are public |
| Evidence Limit | author-reported preprint; no independent replication found |
Technical impact 4
Engineering value 4
Adoption signal 1
Maturity 2
Verification cost 3
Risk 3