ROM: Rating-based Opacity Matching

• ROM is a dynamic mechanism that uses repeated interactions and rating updates to link review effort with future matching outcomes, addressing adverse selection and moral hazard.
• It employs a constant-step rating update and probabilistic matching rule based on rating distributions to incentivize sustained high review quality and convergence to equilibrium.
• Separately, in AirSplat, ROM filters out degraded 3D primitives using local geometry consistency, highlighting distinct implementations sharing the same acronym.

Searching arXiv for the cited papers to ground the article. arXiv search: AirSplat and the peer-review mechanism paper. Rating-based Opacity Matching (ROM) denotes a rating-dependent, opacity-enhanced matching mechanism for repeated peer review in which agents’ current behavior affects future matches and future payoffs through endogenous ratings (Xiao et al., 2014). In a separate usage of the same acronym, AirSplat introduces a component called Rating-based Opacity Matching within a pose-free novel view synthesis framework, where it is described only at a high level as leveraging local 3D geometry consistency knowledge from a sparse-view NVS teacher model to filter out degraded primitives (Bui et al., 26 Mar 2026). The detailed formal specification associated with ROM is the peer-review mechanism developed in Xiao et al., which uses repeated interactions, ratings, and matching rules that depend on those ratings to address adverse selection and moral hazard simultaneously (Xiao et al., 2014).

1. Problem formulation and motivation

Peer review is presented as an effective and scalable method to evaluate the products of a large number of agents when the number of dedicated reviewing experts is limited. The mechanism is formulated for settings such as grading assignments in Massive Open Online Courses and academic paper review. Two difficulties are central: identifying reviewers’ intrinsic capabilities, characterized as adverse selection, and incentivizing reviewers to exert high effort, characterized as moral hazard (Xiao et al., 2014).

The core motivation for ROM is that one-shot matching rules and exogenously fixed matching rules do not link current reviewing behavior with future matches and future payoffs. The mechanism therefore uses ratings to summarize past review quality and designs matching rules that endogenously depend on those ratings. The stated objective is an equilibrium in which agents are incentivized to exert high effort and receive ratings that precisely reflect their review quality (Xiao et al., 2014).

A common misunderstanding is to treat ROM as a purely static assignment scheme. The formal presentation rejects that interpretation: ROM is explicitly repeated, endogenous, and dynamic. Another misunderstanding is to equate opacity with absence of rating information. In the mechanism, the designer broadcasts the rating distribution, but the realized reviewer assignment is randomized in a way that prevents perfect targeting of future matches.

2. Formal model and rating dynamics

The mechanism considers a finite set of agents $\mathcal N=\{1,\dots,N\}$ over time slots $t=0,1,2,\ldots$. In each slot, every agent submits one product for peer review, is assigned by the mechanism to review $M$ other agents’ products, and chooses an unobservable effort $e_i^t\in[0,e_i^{\max}]$. Each agent has private cost, quality, and benefit functions with the following properties:

• Cost: $c_i(e)$ is strictly increasing, strictly convex, with $c_i(0)=0$ and $c_i'(0)=0$.
• Review quality: $q_i(e)$ is strictly increasing, concave, with $q_i(0)=0$.
• Benefit: $b_i(q)$ from receiving a review of quality $t=0,1,2,\ldots$0 is strictly increasing, concave, with $t=0,1,2,\ldots$1.

Agents may differ in these functions, and they may be grouped into types if they share the same $t=0,1,2,\ldots$2 and patience $t=0,1,2,\ldots$3 (Xiao et al., 2014).

The mechanism maintains a nonnegative rating $t=0,1,2,\ldots$4 for each agent. After reviews are completed and each report $t=0,1,2,\ldots$5 is collected, the rating update is given by constant-step exponential smoothing: $t=0,1,2,\ldots$6

By construction,

$t=0,1,2,\ldots$7

for all $t=0,1,2,\ldots$8 (Xiao et al., 2014).

This update rule is central because it couples present reviewing effort to future matching opportunities. The use of a constant step size also creates a tractable trade-off between responsiveness and stability: the exposition states that $t=0,1,2,\ldots$9 should be sufficiently small to ensure convergence, while larger $M$0 accelerates learning but may oscillate.

3. Opacity-enhanced matching rule

ROM orders the distinct ratings in the current profile as

$M$1

and lets agent $M$2 occupy rank $M$3 so that $M$4. The matching probability

$M$5

depends only on $M$6, not on identities (Xiao et al., 2014).

The baseline ROM matching rule is specified as follows. If two agents share the same rating, they are matched one-to-one among themselves. If an agent has a distinct rating and rank $M$7, the agent is matched probabilistically to the two neighboring ranks: $M$8

The role of opacity is explicit: this randomness prevents any agent from perfectly targeting which reviewer she will face, ensuring opacity against strategic rating fluctuations (Xiao et al., 2014). In this sense, ROM is not a deterministic nearest-neighbor assignment rule. The designer broadcasts the rating distribution $M$9, but the realized matching remains probabilistic.

The exposition also describes extended ROM variants with asymmetric neighbor matching or long-range jumps. These modifications are introduced as ways to tune reward and punishment probabilities and thereby alter equilibrium effort and total quality. A plausible implication is that the baseline rule should be viewed as a minimal opacity-preserving construction rather than the only admissible design.

4. Payoffs, conjectural equilibrium, and incentive properties

Each agent obtains benefit from being reviewed and pays cost for reviewing. The expected period payoff is written as expected benefit from being reviewed minus $e_i^t\in[0,e_i^{\max}]$0. Because future matchings depend on future ratings, and future ratings depend on current effort, the mechanism is intrinsically dynamic (Xiao et al., 2014).

To maintain tractability, the exposition introduces a conjectured future value of the form

$e_i^t\in[0,e_i^{\max}]$1

where $e_i^t\in[0,e_i^{\max}]$2 is the expected balance-of-period benefit resulting from updating $e_i^t\in[0,e_i^{\max}]$3 to $e_i^t\in[0,e_i^{\max}]$4. The myopic best response is

$e_i^t\in[0,e_i^{\max}]$5

A Conjectural Equilibrium is defined as a fixed-point triple $e_i^t\in[0,e_i^{\max}]$6 such that each $e_i^t\in[0,e_i^{\max}]$7 is a best response given $e_i^t\in[0,e_i^{\max}]$8, ratings are correct in the sense that $e_i^t\in[0,e_i^{\max}]$9, and conjectures match actual continuation values (Xiao et al., 2014).

The principal theoretical claims are threefold. First, ROM overcomes adverse selection by using ratings to sort reviewers over time and combats moral hazard by making the matching probabilities depend on those same ratings. Second, under any desirable matching rule—defined as one for which expected benefit is strictly increasing and concave in one’s own $c_i(e)$0 and every agent reviews a fixed positive number of products—the best-response dynamics converge to a unique Conjectural Equilibrium provided the rating-update step size $c_i(e)$1 is sufficiently small. Third, at equilibrium each agent chooses strictly positive effort, and the exposition states a first-order condition in which $c_i(e)$2 is balanced against a term involving $c_i(e)$3 and a marginal matching-reward component (Xiao et al., 2014).

The contrast with one-shot matching is sharp. Under one-shot random matching, total review quality $c_i(e)$4. Under ROM, $c_i(e)$5, and the exposition states that performance is often strictly larger than under fixed-matching benchmarks.

5. Algorithmic implementation and illustrative dynamics

The per-period implementation is described procedurally. The mechanism initializes $c_i(e)$6 for all $c_i(e)$7 and sets $c_i(e)$8. At each time step, the designer broadcasts the rating distribution $c_i(e)$9; each agent submits a product; the designer draws reviewers by probabilities $c_i(0)=0$0; each reviewer chooses effort; reviews occur and authors report realized review quality; the designer updates ratings; and each agent updates the belief offset according to

$c_i(0)=0$1

This cycle then repeats (Xiao et al., 2014).

The four-agent example begins with initial ratings $c_i(0)=0$2. Under the baseline rule, all ranks tie and are matched arbitrarily into two pairs with $c_i(0)=0$3 each. Agents then exert effort solving an objective of the form

$c_i(0)=0$4

which yields $c_i(0)=0$5 if $c_i(0)=0$6 is large enough. Ratings update to

$c_i(0)=0$7

after which agents with higher $c_i(0)=0$8 pull ahead, and the next period’s matching probabilistically favors higher-rated agents with better reviewers (Xiao et al., 2014).

As $c_i(0)=0$9, ratings separate in order of true capability $c_i'(0)=0$0, each converging to its quality $c_i'(0)=0$1. The exposition further states that no one can game the system by over-efforting in a single period since the matching is randomized and converges only gradually via $c_i'(0)=0$2.

The implementation notes identify three practical design choices. The step size $c_i'(0)=0$3 should be small enough to ensure convergence, though larger $c_i'(0)=0$4 accelerates learning at the risk of oscillation. Initial ratings should be set at a modest level, because too low a start traps everyone at zero effort. Extended parameters such as an asymmetric reward parameter $c_i'(0)=0$5 or long-range punishment $c_i'(0)=0$6 may be tuned to fine-tune total review quality versus reviewer cost.

6. Relation to AirSplat and terminological ambiguity

AirSplat uses the same acronym, ROM, in a different technical domain. The abstract describes AirSplat as a training framework that adapts the robust geometric priors of 3D Vision Foundation Models into high-fidelity, pose-free novel view synthesis. It introduces two key technical contributions: Self-Consistent Pose Alignment and Rating-based Opacity Matching. In that context, ROM is said to leverage the local 3D geometry consistency knowledge from a sparse-view NVS teacher model to filter out degraded primitives (Bui et al., 26 Mar 2026).

The same record also states that experimental results on large-scale benchmarks demonstrate that AirSplat significantly outperforms state-of-the-art pose-free NVS approaches in reconstruction quality, and that the framework highlights the potential of adapting 3DVFMs to enable simultaneous visual geometry estimation and high-quality view synthesis (Bui et al., 26 Mar 2026). However, no mathematical formulation, training algorithm, implementation detail, or ablation result for the AirSplat version of ROM is specified there.

This terminological overlap is important. In the peer-review literature, ROM is a fully specified repeated endogenous matching mechanism built around ratings, effort incentives, and opacity-enhanced assignment. In AirSplat, the same acronym names a filtering component inside a feed-forward 3D Gaussian Splatting training framework. A plausible implication is that the acronym should not be interpreted as denoting a single domain-independent method; rather, it labels distinct mechanisms whose commonality is only nominal.