Decentralized Finance (DeFi) transaction sequences can obscure a user’s high-level goal behind multi-step smart-contract calls, routing abstractions, and rapidly changing market conditions. Many prior intent-mining pipelines rely on transaction-level semantic labels, which can miss the sequential decision structure of long-horizon strategies. We formulate DeFi intent discovery as sequential reward inference and learn a parametric reward Rθ(s, a) from expert demonstrations via Maximum Entropy Inverse Reinforcement Learning (MaxEnt IRL). To capture short-horizon market trends under volatility and partial observability, we augment the state with a temporal market-gradient term ∇t m, defined as the block-to-block finite difference of market-state variables. We refer to this state design as Chemo-IRL. To enable controlled synthetic evaluation under intent hiding, we introduce Gym-DeFi, a Gymnasium-compatible simulator with a configurable action-obfuscation channel that corrupts observed action identifiers during data generation. We evaluate reward recovery using Reward Recovery Error (RRE) and downstream intent probing using macro-F1 from a fixed post-hoc attribution-based decoder. On this controlled synthetic benchmark under high obfuscation, Chemo-IRL attains the lowest RRE among intent-label-free baselines while remaining competitive on macro-F1 within the same label-free setting. These results should be interpreted as benchmark-level comparisons in a controlled synthetic environment rather than as direct evidence of real-world DeFi deployment performance.
Download Full PDF@inproceedings{aizierjiang26chemoirl,
title={DeFi Intent Discovery via Maximum Entropy Inverse Reinforcement Learning},
author={Aiersilan, Aizierjiang and Yen, Jerome and Wang, Sheng},
booktitle={2026 International Joint Conference on Neural Networks (IJCNN)},
year={2026},
organization={IEEE}
}