acceptodds
Under review as a conference paper at ICLR 2027

Spec2Twin-Chain: Orchestrating Bi-Level Optimization with LLMs for Blockchain Digital Twin Construction

Abstract

Building a blockchain digital twin largely requires translating domain knowledge and specific system descriptions into a simulator architecture, calibrating its parameters against behavioral evidence, and validating the constructed twin. These steps are commonly performed through application-specific modeling efforts that can be difficult to reuse across systems and downstream decision problems. We consider automating this process through Spec2Twin-Chain, a framework that formulates blockchain digital-twin construction as a bi-level optimization problem. At the upper level, a large language model proposes and revises structurally admissible architectures using system specifications, behavioral evidence, and feedback from evaluated designs. At the lower level, a simulation-based optimizer calibrates the architecture-conditioned parameters under explicit objectives and guardrail constraints. The two levels iterate. The evaluated candidates at lower levels are retained in a global archive and used to guide subsequent proposals at upper levels. We conduct controlled experiments involving twin calibration, feedback-driven recovery, stress analysis, downstream policy optimization, and policy updating. The results demonstrate that the framework can construct behaviorally accurate twins, improve initial designs through iterative feedback, and reuse calibrated twins to support downstream decisions.

Primary area:
generative models
Keywords:
AI agents, bi-level optimization, blockchain, digital twin, large language models (LLMs), simulation-based optimization

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_4f7ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Spec2Twin-Chain, a framework leveraging Large Language Models (LLMs) to construct blockchain digital twins through a two-tier optimization process. At the upper tier, LLM agents propose candidate simulator architectures based on predefined structural specifications and validation constraints. At the lower tier, simulation-based optimization algorithms calibrate architecture-specific parameters to align with reference behavioral evidence. The framework employs an iterative feedback mechanism, utilizing results from previous evaluations to guide subsequent architectural designs. The framework was evaluated primarily using Ethereum Proof-of-Work (PoW) scenario data generated by BlockSim. Furthermore, the experiments encompass congestion analysis, adversarial stress testing, strategy optimization, and adaptive strategy updates. The results demonstrate that the framework can generate structurally compliant simulator configurations, achieve low calibration errors within controlled environments, and support a variety of downstream simulation-based tasks.

Critical strengths

This paper provides a solution to the practical challenge of building and calibrating blockchain digital twins—a process that typically demands extensive manual modeling and domain-specific expertise. Automating parts of this workflow offers significant practical value for blockchain simulation and system analysis.

Critical weaknesses

In this method, a Large Language Model (LLM) designs the structure of a simulator, an off-the-shelf NSGA-II algorithm is used to tune the parameters, and the final results are fed back to the LLM. Each component relies on existing technology; the approach does not incorporate any truly novel algorithms or mechanisms.

The authors use BlockSim to generate reference data, based on which they fit their constructed system. This primary experiment involves only 33 blocks and 4,534 transactions; this is insufficient to demonstrate that the developed digital twin can accurately simulate a real-world blockchain. Furthermore, the study lacks predictive validation for unseen scenarios.

The authors claim that large language models (LLMs) can automatically search for effective architectures, but they do not provide a detailed comparative analysis.

Questions for authors

What is the specific advantage of using an LLM for architecture search?

Could the authors compare Spec2Twin-Chain against random search, manually designed architectures, or conventional optimization methods under the same computational budget?

How much architectural diversity is actually explored by the LLM?

The primary calibration experiments rely on reference data generated by BlockSim. How well does the constructed digital twin model generalize when faced with previously unseen scenarios, independent simulators, or actual blockchain execution traces?

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

Spec2Twin-Chain separates symbolic architecture proposals from numeric simulation optimization. An LLM proposes structures within a declarative admissible space; a lower-level optimizer tunes the permitted variables and returns feasibility, objective, and archive diagnostics. The primary construction study calibrates a proof-of-work Ethereum twin to BlockSim evidence. Additional studies use the twin for congestion and adversarial stress, transaction-policy optimization, and selective policy updating. The paper appropriately distinguishes simulator-grounded calibration from a validated model of a live or post-Merge blockchain.

Critical strengths

S1 A well-defined separation of responsibilities. Schema validation, structural checks, and guardrails prevent the LLM from serving as its own correctness oracle. The explicit architecture-conditioned parameter domain and run-scoped Pareto archive make the workflow understandable and potentially reusable. S2 Transparent diagnostic results. The full KPI table exposes residual timing errors instead of reporting only the aggregate fit. Feedback-recovery and progressive-feedback ablations are provided. The adversarial study also discloses that exact aggregate fit reproduces only two of five strict attack fingerprints. S3 Stronger controls in a downstream study. The locked selective-update comparison uses eight held-out seeds, common random numbers, and baseline rules tuned on separate development seeds. This is more informative than a single illustrative policy trajectory, although it validates a downstream controller rather than the core construction claim.

Critical weaknesses

W1 The contribution of LLM architecture search is not identified. The core calibration experiment lacks a matched-budget comparison against a fixed admissible architecture with the same optimizer, random or enumerative structure search, and a non-adaptive LLM proposal baseline. The recovery ablations remove evidence-guided bounds and vary diagnostic availability, but they do not isolate architectural reasoning from parameter-domain refinement or ordinary optimizer access. The downstream rule baselines do not answer this construction question. It is therefore unclear which part of the achieved fit requires the proposed upper level. W2 Calibration fit is not independent fidelity validation. Section 3.1 fits ten aggregate KPIs from only 33 blocks and 4,534 transactions. Several targets are arithmetically or operationally related, such as counts, transactions per block, throughput, and uncle statistics. The mean relative fit of 0.022484 is thus not ten independent checks of system dynamics; it also averages over a 14.1 percent average-block-time discrepancy. Held-out seeds, longer horizons, new workloads, and distributional or trajectory-level checks are needed to establish predictive behavior beyond the calibration evidence. W3 Most experimental breadth is downstream rather than additional construction evidence. The appendix explicitly clarifies that policy and strategy experiments can reuse a calibrated twin and should be interpreted as DT-enabled optimization. That distinction is correct, but it leaves the core construction claim supported by a narrow instance. Deterministic stress fixtures and zero aggregate fit do not independently validate security behavior, as the two-of-five fingerprint result illustrates. A second genuinely distinct construction request would be more persuasive than additional optimization tasks on the same modeling substrate. W4 Search cost and architectural diversity are insufficiently characterized. Accepted-proposal counts exclude raw request-processing calls and within-candidate repair turns. The reported lower-level evaluation counts are useful but do not reveal total LLM tokens, repair effort, wall-clock cost, invalid proposals, or sensitivity across independent upper-level runs. The paper also needs concrete before-and-after architecture examples to show whether feedback changes modules and interactions or mainly bounds and numeric choices.

Questions for authors

Q1 Under equal simulation and LLM budgets, how does the core calibration compare with fixed-architecture optimization, random admissible architectures, and one-shot LLM construction? Q2 Which structural decisions actually change during the calibration loop? Can you provide a candidate-by-candidate architecture diff and separate structural improvements from parameter-space changes? Q3 When all fitting choices are frozen, how well does the twin predict new BlockSim seeds, workloads, and longer horizons? Please report timing distributions and state-level behavior as well as aggregate KPIs. Q4 What are the total raw model calls, repair turns, tokens, simulation evaluations, and elapsed time per successful construction, and how variable are they over independent runs?

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

The paper proposes Spec2Twin-Chain for constructing blockchain digital twins. An LLM proposes schema-constrained simulator architectures, and a lower-level optimizer calibrates their parameters. Evaluated configurations and diagnostics guide subsequent proposals. Experiments calibrate an Ethereum proof-of-work mempool twin against BlockSim evidence and examine controlled stress analysis, policy optimization, and policy updating.

Critical strengths
  1. Separating architecture proposals from parameter calibration gives the framework a clear division of responsibilities. Schema checks and runtime guardrails constrain candidate execution to establish correctness.
  2. The feedback loop improves calibration under weakened initial information, and the controlled downstream studies illustrate potential reuse of the twin.
Critical weaknesses
  1. The benefit of LLM-driven architectural search is not isolated. The implementation has only one registered interaction template, and the calibration study does not show which substantive structural choices change across proposals. Without fixed-architecture or LLM-free search baselines under matched evidence and simulation budgets, the gains cannot be separated from parameter optimization and evidence-guided bounds. The central architectural-search contribution therefore remains insufficiently validated.
  2. Behavioral fidelity is evaluated too narrowly. Calibration uses only 33 blocks and 4,534 transactions, while the aggregate error of 0.022484 coexists with a 14.14% average block-time error. Averaging several near-exact aggregate KPIs can conceal errors in important temporal dynamics. Held-out trace validation is needed to establish that the calibrated twin predicts behavior beyond the evidence used for fitting.
Questions for authors
  1. Which architectural components or interactions actually change across calibration proposals, and how do those changes affect the optimized fit?
  2. Under identical evidence and simulation budgets, how does the framework compare with a fixed architecture and an LLM-free proposal strategy?
  3. Does the calibrated twin retain its accuracy on unseen traces, particularly for block-time distributions and other temporal dynamics?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.