Decentralised Preference Discovery Protocol

• Decentralised preference discovery protocols are distributed mechanisms that infer and aggregate agents' private preferences using asynchronous, scalable, and sybil-resistant methods.
• They are applied in voting, matching markets, and service ranking, enabling robust consensus and stable matchings even under adversarial or partial information conditions.
• Key advancements include modular aggregation rules, convergence in O(n) time, and resilience to strategic manipulation, ensuring consistent performance across diverse application domains.

A decentralised preference discovery protocol is a distributed mechanism for aggregating and inferring individual agents’ preferences in a network without relying on a central authority. These protocols address fundamental problems in computational social choice, multi-agent coordination, market matching, and peer-to-peer service discovery, enabling robust consensus or optimal selection even under adversarial conditions, privacy constraints, and limited communication. Several families of protocols—gossip-based voting, matching-market learning, signal propagation in reputation graphs, and leaderless consensus—embody this theme, each specializing for distinct application domains and theoretical criteria.

1. Formal Foundations and Problem Models

Decentralised preference discovery is characterized by a set of autonomous agents, each holding a subjective, private preference over a space of alternatives (candidates, services, partners, trajectories, etc.). The agents form a network—often asynchronous and permissionless—seeking to jointly determine a collective outcome or ranking. Core models include:

• Voting/choice: Each agent $i$ has a strict ranking $R_i$ or full cardinal utility $O_i$ over a finite candidate set. The task is to aggregate to a unique winner or full ranking associated with some aggregation rule $F$ (Kotsialou, 20 Dec 2025).
• Matching markets: Agents partition into proposers and acceptors, each with unknown value functions; the collective goal is a stable matching achieving maximum welfare, subject to individual exploration and learning (Shah et al., 2024).
• Reputation and service discovery: Agents transact with services (and possibly one another), producing an interaction graph; the emergent task is to rank or select services with highest collectively endorsed preference/reputation (Shi et al., 31 Oct 2025).

The protocols are constructed to operate under adversarial communication, partial visibility (no agent sees the full profile), and asynchronicity: all consistency, liveness, and sybil-resistance guarantees must hold without central orchestration.

2. Gossip-Based Iterative Aggregation: The Snowveil Framework

Snowveil (Kotsialou, 20 Dec 2025) exemplifies gossip-based decentralised preference discovery tailored for large electorates:

• Protocol design: Every agent (voter) repeatedly samples $k$ peers’ current “locked” choices, aggregates these via a deterministic social-choice function (notably, the Constrained Hybrid Borda—CHB), then probabilistically updates or “locks” onto a candidate. The process iterates in $m-1$ rounds to yield a full ranking (single-winner election at each round, with winner removed and reset).
• Formal steps: Each agent holds a dynamic state $s_i(t)\in\{\bot, p_1,\dots,p_m\}$ (unlocked or locked to a candidate). Upon activation, the local “VoterUpdate” procedure samples peers, counts repeated wins, and locks if a threshold is met.
• Aggregation rule: The CHB rule combines Borda count (consensus) and plurality filters (broad support, first-place majorities) with tunable parameters $(\alpha, \beta, \lambda)$ enforcing popularity and consensus thresholds at every sampling stage.
• Key guarantees: With only partial, randomly sampled information and asynchronous updates, the protocol almost surely converges in finite time to a unique winner corresponding to the canonical aggregation over the full (unknown) profile. Full rankings are constructed by iterative winner selection.
• Scalability: Empirical and theoretical analysis indicates $O(n)$ expected time to convergence, robust to parameter variation and resistant to local coalition manipulation. Monotonicity, uniqueness, and positive responsiveness are rigorously satisfied.

3. Decentralised Learning in Matching Markets

In matching domains where preferences are initially unknown even to agents themselves, decentralised preference discovery harnesses stochastic learning and local signal-based updates (Shah et al., 2024):

• Market structure: A bipartite set of proposers and acceptors with latent cardinal utilities; each proposer must learn their preferences by interacting (trial-propose) in discrete rounds.
• Local protocol: Proposers maintain minimal internal state—mood $(C,W,D)$, baseline action, observed utility—and select actions via structured exploration (controlled by parameter $\epsilon$). Acceptors deterministically accept their most-preferred proposal each round.
• State evolution: Mood and baseline are updated upon payoff observation using strictly decreasing maps $F,G$ and randomized Bernoulli experiments determining promotion/retention of a new baseline.
• Markov convergence: The full system forms an irreducible, regularly perturbed Markov chain. As $\epsilon\to 0$, the unique stochastic stationary distribution concentrates on the proposer-optimal stable matching $\mu^*$, with arbitrarily high probability.
• Decentralisation: The only communication is inherent in propose/reject messages; no coordination, voting, or message passing between proposers is required. Computational and memory complexity per round are $O(1)$. Communication-bounded environments thus admit stable and welfare-maximising outcomes despite total lack of global preference or utility visibility.

4. Information Propagation Protocols for Sybil-Resistant Service Discovery

In decentralised service ranking, preference discovery is achieved by propagating reputation scores over payment/endorsement graphs, ensuring resistance to sybil attacks and bias (Shi et al., 31 Oct 2025):

• Graph construction: Agents (nodes) interact via payment transactions (edges), each with timestamp and value. An initial seed vector marks trusted nodes with positive reputation; others begin at zero.
• Flow accumulation and normalisation: For each ordered $(j\to i)$ payment, flows are aggregated with temporal decay and logarithmic value weighting, then column-normalised to form a stochastic propagation matrix $W$.
• Iterative propagation: A PageRank-style iteration with damping parameter $\alpha$ converges to a fixed point $r^*$, which reflects reputation preference; only agents with trust-seeded predecessors can accumulate score.
• Sybil-resistance: Nodes with zero-seed ancestry cannot accrue reputation, making the system robust to spamming by low-reputation actors. Only endorsements by reputable agents can promote service ranking.
• Preference-aware querying: To answer natural-language discovery queries, semantic embeddings of service profiles are fused with TraceRank scores via multiplicative combination; the result filters both “semantic fit” and network-preferred/endorsed services.
• Fully decentralised operation: All data is globally replicated (public ledger), computation is local at each agent, and there is no central server or coordination point.

5. Protocols for Emergent Multi-Agent Preference Coordination

Leaderless, reference-free Model Predictive Control (MPC) protocols operationalize decentralised preference discovery in continuous control domains (Wartnaby et al., 2019):

• Preference as trajectory: Each agent continuously computes a “desired” trajectory (its unconstrained optimum absent peers), broadcasts it, and receives others’ desired/planned paths.
• Constraint negotiation: The actual “planned” trajectory is computed accounting for high-weight penalties on planned trajectory overlap (hard safety) and soft penalties on desired trajectory overlap (preference accommodation).
• Preference communication: Broadcasting the “desired” trajectory allows each agent to signal its intent and level of urgency (derived from the relative cost of deviation), which peer agents incorporate (scaled) as soft constraints in their own planning.
• Emergent cooperation: Without negotiation or voting, iterated MPC steps resolve conflicts—agents incrementally yield to each others’ high-urgency wishes while always satisfying their own safety and constraint priorities. The process gives rise to robust, leaderless coordination with linear computational cost and scalability to large swarms.

6. Axiomatic and Performance Guarantees

The efficacy and fairness of decentralised preference discovery protocols hinge on both formal social choice axioms and system-theoretic convergence properties:

• Determinism and uniqueness: Expressive aggregation rules (e.g., CHB) are constructed to yield unique winners on any profile in $O(km)$ time (Kotsialou, 20 Dec 2025).
• Responsiveness and monotonicity: Minimal improvements in one candidate’s Borda score can flip outcomes; larger coalitions are required to mount strategic attacks ($\Omega(n)$ for $n$ voters).
• Strict submartingale convergence: Preference propagation Markov chains possess strictly positive drift—potential functions increase each non-terminal step, with almost sure absorption into consensus states (Kotsialou, 20 Dec 2025).
• Sybil and manipulation resistance: Properly seeded flow propagation guarantees that reputation cannot be sybil-inflated; preference manipulation requires non-trivial collusion whose cost scales linearly.
• Complexity and scalability: All examined protocols, whether voting, learning, or ranking, achieve $O(1)$ or $O(n)$ per-agent computation, reach convergence in $O(n)$ steps, and are robust to high asynchrony and partial knowledge.

7. Extensions and Practical Considerations

These protocols admit extensive generalization and adaptation:

• Aggregation rule modularity: Snowveil’s protocol generalizes to any rule satisfying positive responsiveness and uniqueness—robust to customisation for different fairness or consensus notions, including multi-winner or approval styles (Kotsialou, 20 Dec 2025).
• Contextual preferences: In service discovery, multiple contexts or utility classes can be supported by separate seed vectors and propagation matrices (Shi et al., 31 Oct 2025).
• Continuous and dynamic environments: Preference discovery remains stable in high-churn multi-agent systems, with incremental convergence, on-the-fly agent addition/removal, and compatible with privacy-respecting, censored communications channels.
• Accelerated learning: Matching-market protocols may incorporate empirical payoffs and UCB-style confidence bounds to speed up discovery at increased storage cost (Shah et al., 2024).
• Cross-domain application: Elements of preference discovery (partial information, decentralisation, sybil-resistance) are increasingly relevant for federated learning, DAO governance, and resilient peer-to-peer marketplaces.

Decentralised preference discovery protocols thus provide formal, scalable, and robust methods for leaderless consensus and preference aggregation in the presence of partial information, adversaries, and asynchronous operation, as demonstrated in computational social choice (Kotsialou, 20 Dec 2025), multi-agent markets (Shah et al., 2024), sybil-resistant reputation networks (Shi et al., 31 Oct 2025), and distributed continuous control (Wartnaby et al., 2019).