BLAST: Blockchain-based LLM-powered Agentic Spectrum Trading

Abstract: The management of radio frequency spectrum is undergoing a paradigm shift from static, centralized command-and-control models to dynamic, market-driven approaches. However, the realization of Dynamic Spectrum Management has been hindered by the lack of an automated, trustworthy, and intelligent coordination infrastructure that can operate without a central authority while preserving participant privacy. In this paper, we introduce BLAST (Blockchain-based LLM-powered Agentic Spectrum Trading), a comprehensive framework that integrates LLM Agents with a permissioned blockchain infrastructure to create a fully autonomous, private, and secure spectrum trading ecosystem. We propose a novel agent architecture that implements the Cognitive Radio cycle through a sequential decision pipeline (perceive, plan, act) enabling agents to reason strategically about economic value and market dynamics. We evaluate the framework through three distinct market mechanisms: Direct Sale, First-Price Sealed-Bid, and Second-Price (Vickrey) Sealed-Bid auctions. Experimental results demonstrate that the Second-Price (Vickrey) auction is the optimal choice for maximizing social welfare and allocative efficiency, capturing up to 71% of the theoretical surplus by incentivizing truthful bidding. We also compare the proposed model against a baseline non-LLM heuristic agentic model and show that utilizing LLM agents yields significant improvements in market competition, reduced wealth and asset concentration, and increased system welfare. Furthermore, we validate the system's privacy preservation, confirming that sensitive bid values remain isolated in private data collections while only cryptographic hashes are committed to the public ledger.

• The paper presents BLAST, an end-to-end system that combines Gemini-powered LLM agents with Hyperledger Fabric, commit-reveal privacy, and smart-contract auctions for decentralized spectrum trading.
• The evaluation finds that second-price auctions deliver the highest heterogeneous-market surplus at $3,050 and 64.9% Shapley efficiency, although they increase ownership concentration with an HHI of 0.507.
• The paper shows that LLM agents outperform fixed heuristics in complex second-price auctions but underperform in some direct-sale and first-price settings, while overall efficiency remains below 75% of the theoretical benchmark.

Motivation and problem statement

BLAST addresses the gap between the theoretical promise of dynamic spectrum management (DSM) and the absence of an automated, trustworthy coordination infrastructure that operates without a central authority while preserving participant privacy. The authors argue that centralized database-driven regimes (TV White Space databases, CBRS SAS) suffer from three structural deficiencies: single points of failure and poor scalability, trust and transparency problems arising from entrusting sensitive operational data to a broker, and the absence of native incentive compatibility for primary users to share spectrum. Prior learning-based approaches to dynamic spectrum access, particularly DRL and MARL, are characterized as effective for low-level numerical optimization but lacking the high-level strategic reasoning, policy comprehension, and market interpretation needed for economic participation; centralized MARL optimizers are additionally constrained by network size and can produce incoherent outputs when processing large global state vectors.

The paper's central design hypothesis is that LLM agents, embedded in a structured game-theoretic decision pipeline and executing against a permissioned blockchain marketplace, can reconcile bounded-rationality LLM behavior with mechanism-design guarantees. The key theoretical anchor is Vickrey's result that truthful bidding is weakly dominant in second-price sealed-bid auctions, which the authors exploit to make allocative efficiency robust to agent irrationality: even if the LLM reasons imperfectly, the dominant strategy ensures the highest-valuation agent wins.

Game-theoretic formulation

Spectrum access rights are tokenized as time-limited digital assets $\tau = \{f_c, B, \Delta t, \mathcal{L}\}$ traded via smart contracts acting as decentralized auctioneers. Operator valuation follows a Shannon-capacity form with a monetization coefficient $\alpha_i$, capturing spectral efficiency as a function of SINR. A notable analytical claim is that economic parameters (monetization coefficient, cost of spectrum, quantity acquired) dominate physical-layer parameters (power, SINR) in determining utility, which motivates shifting optimization effort from channel control to market strategy.

Three mechanisms are analyzed:

• Second-price (Vickrey) auction: truthful bidding is weakly dominant; the paper derives $PoA_{SP} = 1$, i.e., zero efficiency loss at equilibrium.
• First-price sealed-bid: requires bid shading; under uniform independent private values the symmetric BNE is $b^*(v_i) = \frac{N-1}{N}v_i$. Because valuations are unknown and possibly non-uniform in practice, agents augment this with empirical probability maximization over blockchain-recorded winning bids, solving $\arg\max_b (v_i - b)\hat{F}(b)$ where $\hat{F}$ is an empirical CDF of past winning prices.
• Direct sale: modeled as intertemporal price skimming with reserve price decay $r_t = \max(v_s,\ r_{t-1}(1-\delta))$ subject to a break-even floor, approximating perfect price discrimination over time.

The analysis concedes that first-price auctions have strictly greater than 1 Price of Anarchy under asymmetric beliefs, and that direct sale can incur infinite PoA when rigid pricing blocks mutually beneficial trades ($v_s < v_i < r_t$), with dynamic pricing mitigating but not eliminating this inefficiency at the cost of temporal latency.

System architecture

The framework has two layers. The blockchain layer uses Hyperledger Fabric v2, selected for permissioned identity (MSP/X.509), channels and private data collections, deterministic finality (Raft, eliminating fork-induced double-spending of spectrum tokens), execute-order-validate throughput, and general-purpose chaincode. Sealed-bid privacy is enforced through implicit private data collections combined with a commit-reveal protocol: bidders commit $Hash(salt + value)$ on the public ledger during bidding and reveal after close, with the chaincode verifying hash consistency. Endorsement policies validate outcomes among required stakeholders.

The agent layer implements a sequential pipeline of sub-agents explicitly mapped onto Mitola's cognitive radio cycle:

• Analyst sub-agent (perception/orientation): ingests transaction history, computes win rates, realized profit per win, classifies market structure, and flags demand-fading trends with a formal risk assessment.
• Planner sub-agent (decision/reasoning): applies chain-of-thought reasoning to select Buy/Sell/Idle intent and mechanism-specific game-theoretic logic — empirical-Nash bid shading for first-price, near-truthful bidding for second-price, skimming/acquisition rules for direct sale.
• Action executor sub-agent (actuation): validates feasibility (ownership, budget) before invoking tool calls such as start_auction, place_bid, buy_now, a separation-of-concerns measure intended to prevent hallucinated invalid transactions.

Agents are built on Google's Agent Development Kit calling Gemini 2.5 Flash with thinking mode enabled in the planner, with internal (inter-sub-agent handoff) and external (blockchain state) memory, utility-maximizing personas, and function-calling tools.

Experimental evaluation

Simulations use 4 agents (1 seller, 3 buyers), 25 spectrum tokens of 10 MHz blocks at 3.5 GHz, over 100 ticks, under heterogeneous buyer utilities (10/15/20 USD/MHz) and homogeneous utilities (20 USD/MHz each). Metrics include social welfare, Shapley-value efficiency ratio $\eta = \text{surplus}/\Phi$, Gini coefficient, and HHI.

Key results for Scenario 1 (heterogeneous):

Mechanism	Trades	Avg $\alpha_i$0)	Shapley eff.	Gini	HHI
Direct sale	25	7.50	2775	53.4%	0.104
First-price	33	11.06	2600	48.1%	0.191
Second-price	25	12.80	3050	64.9%	0.205

For Scenario 2 (homogeneous):

Mechanism	Trades	Avg $\alpha_i$1)	Shapley eff.	Gini	HHI
Direct sale	25	7.50	4050	75.0%	0.108
First-price	25	16.58	3525	75.0%	0.205
Second-price	23	20.00	4050	71.1%	0.290

The headline claim is that the second-price auction maximizes social welfare and captures up to roughly 71% of the theoretical Shapley surplus by incentivizing truthful bidding. In Scenario 1 it achieves the highest surplus ($3,050) and efficiency (64.9%), because truthful bidding allocates tokens to the highest-valuation buyers, whereas first-price bid shading occasionally produces misallocations. However, the results also expose a tension the authors acknowledge directly: the second-price format yields greater balance dispersion (Gini 0.205) and the highest ownership concentration (HHI 0.507), as the most liquid buyer amasses spectrum — a "winner-takes-all" dynamic. In Scenario 2, homogeneous valuations cause all truthful bids to coincide, eliminating buyer profit in second-price auctions; efficiency there drops slightly below direct sale and first-price (71% vs. 75%). Notably, no mechanism exceeds ~75% of the Shapley benchmark in any configuration, which the authors interpret as persistent frictions even in the best-performing mechanism rather than a fully efficient outcome.

Comparison against non-LLM heuristic baseline

A heuristic baseline implementing fixed game-theoretic rules (symmetric-Nash shading capped between half valuation and budget, 15%/5% seller markups, 1.1× reserves) was run on identical scenarios. Results are mixed and the paper reports them candidly:

• In Scenario 1, the baseline outperforms the LLM in direct sale (57.1% vs. 53.4% efficiency) and first-price auctions (56.6% vs. 48.1%), but the LLM wins decisively in second-price auctions: 64.9% efficiency / $\alpha_i$22,550.
• LLM agents are far more active in auctions (84 transactions in first-price, 54 in second-price, versus 25 and 22 for the baseline), attributed to price-discovery behavior, though some transactions are exploratory and yield no sale.
• In Scenario 2, both models reach parity in direct sale and first-price (75.0%); the baseline edges the LLM in second-price (75.0%, $\alpha_i$34,050), with the LLM generating 103 hyper-active transactions.

The authors' interpretation is that LLM agents excel specifically in complex, truthful mechanisms where aggressive bidding carries less risk, while hand-tuned heuristics remain competitive or superior in stable environments and simple mechanisms. They also note that constructing the baseline required multiple simplifying decisions that the LLM handles implicitly, which supports the flexibility argument for LLM agents, though this is a qualitative rather than quantitative advantage.

Privacy verification

The Hyperledger Fabric deployment was verified such that during the bidding phase the world state contains only bid hashes, with actual bid values confined to the private data collections of participating peers. This confirms commercial confidentiality of sealed bids, which the authors identify as a prerequisite for industry adoption. The full Vickrey auction lifecycle (creation, commit, close, reveal, settlement, token transfer) executes end-to-end in chaincode.

Limitations and open questions

Several limitations are stated or evident. The evaluation scale is small — four agents, one seller, 25 tokens, 100 ticks — so claims about scalability to realistic operator populations and high-frequency micro-transactions remain untested despite the throughput arguments made for Fabric's execute-order-validate design. Efficiency never exceeds ~75% of the Shapley ideal in any run, and the paper does not fully decompose the sources of this residual gap (LLM reasoning errors, mechanism frictions, or simulation artifacts). The second-price mechanism's tendency toward ownership concentration (HHI up to 0.507, above the 0.25 high-concentration threshold the paper itself cites) is reported but not resolved; whether repeated trading de-concentrates markets over longer horizons is left open. The comparison against the baseline shows the LLM underperforming in two of six mechanism-scenario pairings, indicating the benefit of LLM agency is conditional on mechanism complexity rather than universal. Valuations are static per scenario; adaptation to time-varying demand, interference dynamics, and adversarial agents (e.g., collusion or false-name bidding) is not evaluated. The authors propose two concrete extensions: benchmarking different LLM backbones for rational reasoning in dynamic spectrum sharing, and extending mechanisms to combinatorial auctions over bundles of frequency, time, and location.

Conclusion

BLAST demonstrates an end-to-end integration of LLM-driven cognitive radio agents with a permissioned blockchain marketplace, showing that autonomous strategic spectrum allocation is feasible with verifiable privacy guarantees. Its strongest empirical finding is the superiority of the Vickrey mechanism under LLM agency in heterogeneous markets, consistent with its dominance-strategy theory, while its most instructive negative results — sub-heuristic performance in some settings and residual inefficiency across all mechanisms — delineate precisely where LLM-based market participation currently falls short of classical rational-agent assumptions.