acceptodds
Under review as a conference paper at ICLR 2027

From Anonymous Wallets to Behavioral Agents: A Calibrated Policy Test Bed for DAO Governance

Abstract

Unlike traditional corporate hierarchies, Decentralized Autonomous Organizations (DAOs) manage a multi-billion-dollar economy by utilizing token-voting mechanisms that aim to distribute governance power. While DAOs envision democratic governance, their complex economic dynamics can produce unforeseen issues, such as token concentration and runaway inflation. Navigating this complexity requires scalable simulation testbeds that construct heterogeneous behavioral agents from large volumes of on-chain transaction data. We propose DAO-SIM, the first realistic data-driven, LLM-based token-flow simulator that uses inverse behavioral profiling through event-chain clustering and two-stage LLM interpretation to reconstruct heterogeneous participant profiles under the anonymity of DAOs. Then a hierarchical agent architecture separates probabilistic decisions by profile-level super agents from independent execution by wallet-level child agents, enabling flexible simulation scaling-up while reducing redundant LLM inference. Built on over 3 million on-chain transactions collected from Uniswap, DAO-SIM simulates decision-making for an initial population of 5,000 agents and up to 10,000 at peak. Compared to existing agent-based baselines, our simulation aligns much more closely with observed empirical data and faithfully reproduces 3 key stylized facts of token economics: whale effect, path dependence, and the-rich-get-richer effect. Leveraging this validated fidelity, we deploy DAO-SIM as a robust testbed for counterfactual policy experiments. Our findings reveal: modest benefits from perfectly equal initial allocations, substantial inequality reduction from equal inflation issuance, and – contrary to prior studies – little redistributive effect from staking. Consequently, we recommend combining fairer initial allocations with equally issued inflation to improve DAO governance.

Primary area:
generative models
Keywords:
Large Language Models, Social Simulation, Decentralized Autonomous Organizations, Token Holding Inequality, Counterfactual Policy Evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8tVqOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes DAO-SIM, an LLM-based simulator that simulates governance-token flows and tests policies for reducing voting-power concentration. It derives participant profiles from UNI wallets, assigns them to LLM super agents, and uses 5000 to 10000 child agents to execute trading decisions at observed prices.

Calibrated on about 3 million Uniswap records from September 2020 to September 2021, DAO-SIM reproduces the whale effect, path dependence, and rich-get-richer dynamics.

Policy tests show that equal initial allocation modestly reduces concentration, equal inflation substantially reduces it, proportional inflation does not, and staking provides little redistribution.

Critical strengths
  1. The data grounding is well motivated. Using public wallet-level on-chain records for inverse behavioral profiling is a good fit for this domain. This also means that the participant profiles are inferred from actual transactions rather than written as artificial personas.
  2. The validation is strong. The authors compare the simulation with real data using both aggregate measures and a wallet-level accumulation test.
  3. The architecture is a practical scaling pattern. Separating profile-level LLM reasoning from wallet-level stochastic execution is a reasonable way to keep heterogeneity while controlling inference cost.
  4. Prompts, evidence rules, calibration fits, and the action vocabulary are documented in detail.
Critical weaknesses
  1. Super agents only produce cluster-level transition probabilities, which could be estimated directly from the data. Without a non-LLM baseline, the gains over the ABM may come from profiling and calibration rather than LLM reasoning.

  2. DAO-SIM use empirical heavy-tailed balances and allows entry and exit, where the ABM uses prescribed endowments and a fixed population. The results from ABM with matched initialization and population dynamics is needed.

  3. Profiling, calibration, and validation use the same DAO and period. A temporal holdout or second DAO is needed to for generalization.

  4. The prompts heavily constrain behavior. They specify numeric anchors, wealth-based priors, adjustment rules, and probability caps. Without prompt ablations or another LLM, it is unclear how much behavior comes from model reasoning.

  5. Some policy results may follow directly from their design. Equal rewards automatically reduce inequality, while proportional staking rewards do not redistribute tokens. Unequal new entrants may also explain why concentration returns. Controls without trading or new entrants would show whether agent behavior contributes to these results.

Questions for authors
  1. Could you compare DAO-SIM with a non-LLM baseline where super-agent probabilities are replaced by empirical per-cluster, per-wealth-segment rates estimated from the same data? Does the LLM version still outperform it on Tables 3 and 4 and Figure 6?

  2. How does the ABM perform when given the same empirical balance and USD-capacity distributions and the same entry and exit dynamics as DAO-SIM? Which of the reported fidelity gaps remain after this matching?

  3. Can you report validation on a held-out period or on a second DAO?

  4. How do the results change if the prompt constraints (numeric anchors, wealth-based priors, adjustment rules, and probability caps) are removed or varied, or if a different LLM is used?

  5. What are the policy outcomes when (a) the policy is applied without trading, (b) there are no new entrants, (c) agent behavior is fixed at default probabilities? Under equal inflation and equal initial allocation, what fraction of the Gini change remains attributable to agent responses?

Reviewer_yyj8OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DAO-SIM, an LLM-based simulation framework for DAO governance. It combines behavioral profiling from on-chain transactions with a hierarchical agent architecture to simulate large-scale token flows. Experiments on Uniswap data show improved agreement with historical token concentration and accumulation patterns compared to a traditional ABM. The framework is further used to evaluate initial allocation, inflation, and staking policies.

Critical strengths
  1. The paper addresses a meaningful application of LLM agents to DAO governance, using large-scale real-world on-chain transaction data.
  2. The hierarchical agent architecture provides a practical approach to scaling simulations while reducing redundant LLM inference.
Critical weaknesses
  1. The necessity of LLMs is not established. DAO-SIM introduces empirical calibration, behavioral profiling, and hierarchical simulation simultaneously. Without a controlled non-LLM baseline using the same framework, the improvements over ABM cannot be attributed to LLM-based decision-making.

  2. The counterfactual policy conclusions lack sufficient validation. Reproducing historical statistics does not establish predictive validity under unseen policies. Moreover, the simulator embeds wealth-dependent behavioral assumptions in its prompts and uses exogenous historical prices. The reported policy effects may therefore reflect modeling assumptions rather than realistic behavioral responses.

Questions for authors
  1. Can the authors show that LLM-based decisions outperform statistical or rule-based alternatives when all other simulation components are held constant?

  2. What evidence demonstrates that the predicted effects of inflation and staking are not primarily driven by assumptions embedded in the prompts and simulation design?

Reviewer_UdRuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors use Uniswap transaction data to learn a diverse set of market participants, where each is instantiated by an LLM (more precisely, an LLM that outputs probability parameters for individual participant, who is represented by a simpler function). This agent-based model (ABM) is shown to match three aggregate quantities of interest (e.g. wealth distribution) better than a simpler traditional ABM.

Critical strengths

The application of ABM to understanding DAO governance is interesting and novel. Using LLMs to extract idiosyncratic behaviors from raw blockchain is a promising direction, since the transaction addresses are known and we can accurately track flows, unlike traditional securities transaction data (e.g. equities and futures).

Critical weaknesses

Although the problem setup is broadly interesting, it's hard to gauge how difficult the modeling problem is, due to the weak and under-specified ABM baseline, and the lack of ablation studies. The authors' implementation of ABM uses data to infer individual-level statistics, and then evaluates it on aggregate-level statistics. This could be an easy task, if the two sets of statistics are highly related. Thus stronger baselines and ablations could help clarify whether the authors' proposed ABM is modeling the system well. One example of an ablation is an ABM with participants who act probabilistically based on past price moves, where those probabilities are directly estimated from the same transaction data you use to learn the behaviors.

Questions for authors

Please specify the exact approach used to calibrate the baseline ABM. What are the parameters, what is the optimization objective, and how did you optimize it?

How do you evaluate whether the probabilities outputted by your LLM agents are sensible?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.