Suppose many autonomous agents, such as vehicles, drones, or radios, each run a local detector that reports how unusual its current input looks. The detector might flag a sensor fault, weather it was not trained on, or a faint signal. A shift that affects a region is often too weak for any single agent to flag. Pooling the agents' evidence can reveal it. A well-studied way to pool evidence is to convert each detector's score into a p-value: the fraction of clean reference inputs that looked at least as unusual. P-values from different detectors can then be combined into one test with a guaranteed false-alarm rate.
The question here is which agents should pool with which. Wireless research answers this with coalition formation, where agents repeatedly merge into groups or split apart until the grouping is stable. That work has always assumed fusion rules whose false alarms grow as groups get larger, and that growth is what caps group size. We argue that calibrated p-value fusion removes this cap. What then limits useful group size is different and statistical: weak or redundant members dilute the combined evidence, and the dependence among members has to be estimated from a finite pool of shared reference data. We propose a coalition game whose utility is each agent's statistical detection power. We derive the group-size cap it implies and test whether stable coalitions beat both "everyone pools with everyone" and "pool with the nearest few."
Core Claim
Existing work on coalitional spectrum sensing, Coalitional Games for Distributed Collaborative Spectrum Sensing in Cognitive Radio Networks, proves a maximum useful coalition size, which it calls $M_{\max}$. The proof rests on each extra radio adding its own false-alarm probability, which a barrier penalty then punishes. That mechanism belongs to the rule "declare detection if any member detects." It does not carry over to p-value fusion. There, one combined test is run at a fixed level $\alpha$, the chosen false-alarm rate, so the coalition's false-alarm rate stays at $\alpha$ whatever its size, provided the dependence between members is accounted for.
Meta-analytic out-of-distribution detection supports the opposite intuition, "more is better." Combine and Conquer fuses 14 detectors with Fisher's method, which sums $-2\ln p$ across tests. It uses Brown's correction to account for correlation between the detectors. So is there any cap left on pooling?
Our claim is that there is a cap, with two sources. The first is dilution: members that the event does not affect contribute pure noise to the combined statistic. The second is dependence-estimation cost: the covariance among $k$ members must be estimated from $m$ jointly observed clean frames, and this estimate gets worse as $k$ grows relative to $m$. Finite calibration data is already known to cause volatile true-positive rates in single-model settings, as shown by A Closer Look at Generalized BH Algorithm for Out-of-Distribution Detection. Here the same limitation shows up as a coordination cost between agents.
Both effects can be computed locally, and both differ from agent to agent. That makes them a natural utility for a merge-and-split process governed by the Pareto order, in which a merge or split happens only if nobody loses and someone gains.
Mechanism
Each agent $j$ holds a pretrained detector and a private set of clean calibration scores. At test time it turns its score into an empirical p-value $p_j$ and then into a normal-scale value $z_j = \Phi^{-1}(1-p_j)$, where $\Phi$ is the standard normal distribution function.
For a coalition $S$ of size $k$, let $\Sigma_S$ be the correlation matrix of the members' $z$ values on clean data. This is estimated from $m$ frames that all members observed at the same time. Let $\delta_{S,i}$ be the expected upward shift in each member's $z$ for an event centered on agent $i$. We model this shift with a prior that decays with distance and with how much two agents' sensors see the same area. The best weighted combination then gives agent $i$ approximate power
$$\pi_i(S) \approx \Phi\!\left(\sqrt{\delta_{S,i}^{\top}\Sigma_S^{-1}\delta_{S,i}} - z_{\alpha}\right),$$
where $z_\alpha$ is the normal critical value at level $\alpha$. The expression follows from the Gaussian form of a weighted Stouffer test, which sums normal-scale values with weights.
The quadratic term shows dilution directly. A member with near-zero $\delta$ adds nothing to the numerator. A member that is highly correlated with existing members adds little once $\Sigma_S^{-1}$ discounts it.
With plug-in $\Sigma_S$ estimated from finite $m$, the realized false-alarm rate drifts above $\alpha$. Correcting for this means raising the threshold by an amount that grows with $k/m$. This is the usual behaviour of estimating a $k$-dimensional covariance from $m$ samples, and deriving the exact correction is part of the work.
Agent $i$'s utility is $\pi_i(S)$ under this corrected threshold, minus a latency cost for negotiation rounds. Merges and splits follow the Pareto rule. The stable partitions this process converges to are the $\mathbb{D}_{hp}$-stable partitions of Distributed Coalition Formation Games for Secure Wireless Transmission, which we reuse without change.
One design rule is essential for validity. Coalitions are chosen only from calibration data and past windows, never from the current test p-values. Otherwise the choice of group would itself inflate false alarms.
Bandwidth is deliberately not the binding constraint here. At test time a member sends a single scalar, unlike the feature maps exchanged in Where2comm or OPV2V. The real coordination costs are three:
- the one-time exchange of calibration scores when coalitions merge;
- the need for jointly observed calibration frames to estimate $\Sigma_S$;
- negotiation latency when agents move.
Research Program
The theory work derives a closed-form or tightly bounded $M_{\max}$ under p-value fusion as a function of three quantities: how quickly the event's signal decays across agents, the average correlation between members, and the joint calibration size $m$. We will state precisely when the grand coalition becomes stable, meaning everyone in one group. We expect this when signals are spread widely and $m$ is large. This is exactly where "more is better" should hold, and it would reconcile the two literatures instead of declaring one of them wrong.
We will also prove that restricting coalition choice to calibration data preserves the level-$\alpha$ guarantee. We then quantify the false-alarm inflation when Brown's correction or Fisher's method is applied with an estimated $\Sigma_S$.
The empirical work tests the game in three settings that range from fully controlled to realistic.
Baselines
No cooperation. Each agent tests alone at level $\alpha$ using its own calibration set. Every other result is measured against this reference.
Classical coalition game. The coalitional sensing game of Coalitional Games for Distributed Collaborative Spectrum Sensing in Cognitive Radio Networks, run with its original fusion rule and false-alarm barrier penalty. This is the strongest existing method on this exact question. It reports up to an 86.6% reduction in average missing probability relative to non-cooperative sensing, at false-alarm threshold $\alpha=0.1$. We will reimplement its simulation using the same setup described in that paper: a 3 km by 3 km area, Rayleigh fading, and path-loss exponent 3. We will match the realized false-alarm rate across methods. Beating it means lower missing probability at equal realized false-alarm rate across network sizes up to 30 or more radios.
Broadcast-all pooling. Every agent pools the p-values of all neighbors in radio range. We test two combination rules:
- the Brown-corrected Fisher combination of Combine and Conquer;
- the Cauchy combination test evaluated by Enhancing Out-of-Distribution Detection with Multitesting-based Layer-wise Feature Fusion, which stays valid under arbitrary dependence.
No reported result was found on Lacuna for distributed broadcast-all p-value pooling on any benchmark below. The closest centralized reference points are Combine and Conquer's 89.8% average AUROC on ResNet-50 ImageNet and MLOD-Fisher's CIFAR-10 FPR95 of 7.47% with KNN scores.
Fixed-$K$ neighbors. Each agent pools with its $K$ nearest agents. $K$ is tuned on the same validation scenarios as our method's decay prior, with the same tuning budget.
Centralized oracle partition. The best partition found by exhaustive search. This is only feasible for small networks, since the centralized problem is reported intractable beyond 7 radios. It bounds how much the distributed game leaves on the table.
Ablations. Three ablations isolate the key ingredients:
- the same merge-and-split process with the classical false-alarm-penalty utility swapped in for our power utility;
- our utility with plug-in covariance and no $k/m$ correction, to show whether ignoring dependence-estimation cost inflates realized false alarms;
- coalitions selected using current test p-values, to show the validity failure that the design rule prevents.
All methods use identical detectors, calibration pools, and mobility traces.
Why not just…?
Why not just broadcast your p-value to everyone in range and use the Cauchy combination? This is the real threat. Messages are scalars, and Cauchy combination is dominated by the smallest p-values, so it resists dilution better than Fisher's method. If localized-event power under broadcast-all Cauchy matches stable coalitions at equal realized false-alarm rate, the game adds nothing, and we will report that. The experiments settle it by sweeping how far the event's signal spreads across agents, from one or two agents to the whole network. Our derivation predicts a gap when signals are moderately spread and weak per agent. In that regime Cauchy's reliance on the smallest p-value wastes evidence and Fisher's method dilutes it, while the chosen coalition does neither. If no such regime appears, the contribution reduces to the $M_{\max}$ theory alone.
Why not just pool with the nearest $K$ agents? OPV2V reports that detection gains saturate around 4 cooperating vehicles, which makes a small fixed $K$ tempting. Where2comm argues, however, that physical closeness is a poor proxy for how informative a partner is. Nearby agents with overlapping views are highly correlated, and $\Sigma_S^{-1}$ discounts them. The tuned fixed-$K$ baseline, run under mobility traces, measures this directly. We expect the game to win mainly when agent density and sensor overlap vary across the scene, so that no single $K$ suits every region.
Why not just learn a combiner, such as attention or logistic regression over neighbors' scores, and calibrate it on held-out clean data? A learned combiner that is calibrated afterwards also controls false alarms. However, it needs labeled examples of shifts to learn from, and it must be retrained whenever the neighborhood changes. The target setting is unknown shifts in a changing topology, so both requirements are serious. We will give it an advantage by training on some corruption types and testing on held-out types. Beating it on held-out shifts is the relevant bar.
Proposed Experiments
Cognitive-radio replication. The first experiment reproduces the classical collaborative-sensing simulation and swaps only the fusion rule and the utility. It directly tests whether p-value fusion removes the false-alarm-driven $M_{\max}$. It then checks whether our predicted cap, driven by dilution and calibration, matches the empirical distribution of coalition sizes as $m$ varies.
Centralized bridge. The second experiment treats each of the 14 detectors in Combine and Conquer, and each network layer in MLOD, as an "agent." Each carries a compute cost. We ask how much of the reported 89.8% AUROC and 7.47% FPR95 a Pareto-stable subset retains, and at what fraction of the detectors or layers. Both papers flag the cost of running every detector or layer as an open problem, so this experiment answers it directly.
Multi-agent driving. The third experiment runs on OPV2V, V2X-Sim, and DAIR-V2X. Each agent computes a per-frame shift score from its detector's bird's-eye-view features and calibrates it on clean frames. We then inject localized corruptions of controlled severity and spatial spread, such as fog-like point dropout, sensor noise, and miscalibration, into a subset of agents. The metrics are:
- true-positive rate at realized false-alarm rate $\alpha$;
- detection delay in frames;
- realized false-alarm rate on clean scenes, where the proposal fails if it exceeds $\alpha$ beyond sampling error;
- calibration bytes exchanged;
- merge-and-split rounds under mobility.
Decisive checks. Three results decide the proposal:
- realized false-alarm rate stays at $\alpha$ for every coalition size when the $k/m$ correction is used, and inflates without it;
- stable coalitions beat broadcast-all Cauchy on moderately spread weak shifts;
- the observed coalition sizes track the derived $M_{\max}$ as joint calibration size and event spread vary.
References
- Coalitional Games for Distributed Collaborative Spectrum Sensing in Cognitive Radio Networks
- Combine and Conquer: A Meta-Analysis on Data Shift and Out-of-Distribution Detection
- A Closer Look at Generalized BH Algorithm for Out-of-Distribution Detection
- Distributed Coalition Formation Games for Secure Wireless Transmission
- Where2comm: Communication-Efficient Collaborative Perception via Spatial Confidence Maps
- OPV2V: An Open Benchmark Dataset and Fusion Pipeline for Perception with Vehicle-to-Vehicle Communication
- Enhancing Out-of-Distribution Detection with Multitesting-based Layer-wise Feature Fusion