Controlling Exploration–Exploitation in GFlowNets via Markov Chain Perspectives
Lin Chen, Samuel Drapeau, Fanghao Shao, Xuekai Zhu, Bo Xue, Yunchong Song, Mathieu Laurière, Zhouhan Lin · 2026
Read Controlling Exploration–Exploitation in GFlowNets via Markov Chain Perspectives by Lin Chen, Samuel Drapeau, Fanghao Shao, Xuekai Zhu, Bo Xue, Yunchong Song, Mathieu Laurière, Zhouhan Lin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-01749.txt.
Description
Lin Chen and colleagues give GFlowNet training exploration exploitation control, constrained by fixed equal mixing of forward and backward policies. Linking GFlowNets to Markov chains reveals a reversibility equivalence, explaining the constraint and adapting Markov properties. Their alpha GFlowNets vary mixing through a tunable parameter, controlling exploration and discovery, while hybrids with fewer full layers and far more linear layers cannot solve it. Across set, sequence and molecule benchmarks they beat earlier GFlowNet objectives, finding up to ten times as many modes, the first provable separation from full attention.
Catalog record
| CREATOR | Lin Chen, Samuel Drapeau, Fanghao Shao, Xuekai Zhu, Bo Xue, Yunchong Song, Mathieu Laurière, Zhouhan Lin |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 99076 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602.01749