Paper: arXiv 2511.22101

Authors: Zhaofeng Zhang

Abstract

The report goes through the main steps of replicating and improving the article “Adaptive Liquidity Provision in Uniswap V3 with Deep Reinforcement Learning.” The replication part includes how to obtain data from the Uniswap Subgraph, details of the implementation, and comments on the results. After the replication, I propose a new structure based on the original model, which combines Mamba with DDQN and a new reward function. In this new structure, I clean the data again and introduce two new baselines for comparison. As a result, although the model has not yet been applied to all datasets, it shows stronger theoretical support than the original model and performs better in some tests.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced reinforcement learning architectures (Dueling DDQN, Mamba) and complex mathematical formulations for Uniswap V3 modeling, indicating high math complexity. It is heavily data-driven with detailed backtest implementations, specific datasets from the Uniswap Subgraph, and comparative performance metrics, demonstrating high empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal: Replicate & Improve<br>Adaptive Liquidity Provision in Uniswap V3"] --> B["Data Acquisition & Processing"]
  
  subgraph B
      direction LR
      B1["Uniswap V3 Subgraph"] --> B2["Data Cleaning & Normalization"]
  end

  B --> C{"Process Branch"}

  C --> D["Original Model Replication"]
  D --> D1["Implementation: DDQN"]
  D1 --> D2["Results Analysis"]

  C --> E["Enhanced Model Proposal"]
  E --> E1["New Architecture: Mamba + DDQN"]
  E1 --> E2["New Reward Function"]
  E2 --> E3["New Baselines for Comparison"]
  E3 --> E4["Theoretical Support & Performance Tests"]

  D2 --> F{"Key Findings & Outcomes"}
  E4 --> F

  F --> G["Stronger theoretical support than original"]
  F --> H["Better performance in test scenarios"]
  F --> I["Data limitations noted for full deployment"]
  
  style A fill:#f9f,stroke:#333,stroke-width:2px
  style F fill:#ccf,stroke:#333,stroke-width:2px

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a new extended structure named Mamba‑DDQN that combines the Mamba state‑space model with a dueling DDQN and a new reward function.

Source sentence
we propose a new structure based on the original model, which combines Mamba with DDQN and a new reward function.

Assumes 3

  • A flat gas fee is assumed in the experiments. execution
    Source sentence
    there is an assumption about the ”flat” gas fee in the paper
  • The price process for the dynamic programming baseline is assumed to follow a geometric Brownian motion in continuous time. model
    Source sentence
    Based on the assumption of [3], the price should be a GBM in continuous time
  • Each time tick appears only once after data cleaning; rows with NaN values are removed. data
    Source sentence
    In the data cleaning, each time tick only contains once, and more raw data will be removed. The rows containing a NaN value were also removed.

Data

Backtest sample: 2021-08 – 2023-01 · ETH/USDC-0.3% and ETH/USDT-0.3% pools · crypto · daily

  • ETH/USDC-0.3% pool · The Graph crypto-onchain public
    2021/08/02 to 2023/01/25 · ETH/USDC-0.3% pool
    Source sentence
    ... will be collected and used. The dataset used for replication was collected from The Graph
  • ETH/USDT-0.3% pool · The Graph crypto-onchain public
    2021/08/02 to 2023/01/25 · ETH/USDT-0.3% pool
    Source sentence
    ... will be collected and used. The dataset used for replication was collected from The Graph

Costs

Modeled Costs are part of the reported headline results.

Trading fee is modeled as fee = { δ 1−δL(√p′ −√p) ... } and gas fee is a flat fee per reallocation (gas_t = self.gas_flat if action != 0 else 0).

Source sentence
Based on the rules of Uniswap V3, the trading fee can be represented in this way: fee = { δ 1−δL(√p′ −√p) ... }

Finds 3

  • In replication, the M1 baseline achieved the highest PnL and LVR across all periods for both pools. performance
    Source sentence
    Both the PnL of M1 perform best of all in all the periods and the LVR of M1 is always the highest in all the periods.
  • The extended M‑DDQN model outperforms the original Dueling‑DDQN without hedging in all sets and periods, and often beats the hedging version. performance
    Source sentence
    M‑DDQN performs better than the Dueling‑DDQN without hedging in all the sets and periods, but only better than Dueling‑DDQN with hedging in a few of the sets and periods.
  • The new model shows stronger theoretical support than the original and performs better in some tests. theoretical
    Source sentence
    it shows stronger theoretical support than the original model and performs better in some tests.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 16 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (9)

  1. Adaptive liquidity provision in uniswap v3 with deep reinforcement learning (2023)
  2. Dueling network architectures for deep reinforcement learning (2016)
  3. Uniswap v3: Concentrated Liquidity (2021)
  4. Strategic liquidity provision in uniswap v3 (2021)
  5. Uniswap liquidity provision: An online learning approach (2023)
  6. Decentralized finance and automated market making: Predictable loss and optimal liquidity provision (2024)
  7. Playing atari with deep reinforcement learning (2013)
  8. Mamba: Linear-time sequence modeling with selective state spaces
and 1 more
  1. Quantformer: from attention to profit with a quantitative transformer trading strategy — → 2404.00424

1 of 9 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.