Active working paper

When Does Information Sharing Improve Decentralized Discovery?

Identifies when pooled error contraction beats lost independent rescue, then separates the selected positive result from its equilibrium-selection boundary.

Active working paper28 pages

Download PDF

Recommended citation

Nakajima, Yohei (2026). When Does Information Sharing Improve Decentralized Discovery? Aggregation, Independent Rescue, and Equilibrium Selection. arXiv:2609.01814 [cs.AI].

BibTeX

@misc{Nakajima2026InformationSharingFrontier,
  author       = {Yohei Nakajima},
  title        = {When Does Information Sharing Improve Decentralized Discovery? Aggregation, Independent Rescue, and Equilibrium Selection},
  year         = {2026},
  howpublished = {arXiv preprint},
  eprint       = {2609.01814},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi          = {10.48550/arXiv.2609.01814},
  url          = {https://arxiv.org/abs/2609.01814},
  note         = {v1, primary category cs.AI, cross-list cs.GT; DOI registration pending at observation; not peer reviewed}
}

Source and provenance

Read the build source · Inspect validation and provenance

Technical details

scientific: valid-with-scope · editorial: active · publication: submitted · lifecycle: active-working-paper · DOI 10.48550/arXiv.2609.01814 · DOI registration pending at observation · arXiv 2609.01814 [cs.AI, cs.GT] · arXiv license: arXiv perpetual non-exclusive distribution license · arXiv submitted · not peer reviewed

The download is copied only after its SHA-256 matches the committed validation record.

SHA-256 8d116b86cdbbc6cda66d65ac72077d29c15b05b86f4dc862ee8ec8e69d8f4ac0

Translation layer

Related formulations

These formulations translate a canonical object without changing its scientific scope.

OR/MS

independent rescue

residual independent coverage outside the pooled block

registered one-hit rescue protocol

Economics / AGT

independent rescue

outside option value of unpooled search attempts

mechanical coverage term not a payoff theorem

Multi-agent AI

independent rescue

agents kept on independent task attempts can recover pooled misses

unevaluated software-agent analogy

Practitioner

independent rescue

keep some independent attempts available to catch what the shared answer misses

model-based design language

OR/MS

action-budget discovery profile

cardinality-constrained posterior coverage frontier

declared finite channel and top-L planner

Economics / AGT

action-budget discovery profile

value of an experiment under a portfolio-action decision problem

specialized discovery payoff and centralized authority

Multi-agent AI

action-budget discovery profile

best exact task coverage available with L coordinated agent actions

theoretical oracle baseline not observed agents

Practitioner

action-budget discovery profile

how much the group can cover with one two or more distinct attempts

model frontier under centralized coordination

OR/MS

recovery budget

minimum resource budget meeting a benchmark value

baseline and authority must be named

Economics / AGT

recovery budget

smallest action capacity restoring a declared outside option

centralized comparator unless otherwise stated

Multi-agent AI

recovery budget

number of distinct coordinated agent actions needed to match a private baseline

exact task baseline not evaluated agent evidence

OR/MS

Information Sharing Frontier

error-contraction versus residual-coverage frontier

registered channels protocols and selection

Economics / AGT

Information Sharing Frontier

aggregation benefit net of decentralized action compression

selected positive result is not every-equilibrium

Multi-agent AI

Information Sharing Frontier

shared context helps only if better joint inference beats lost task diversity

theoretical mapping without model evaluation

Practitioner

Information Sharing Frontier

share when the map improves faster than the portfolio narrows

model-derived formulation

OR/MS

centralized recovery

centralized portfolio recovery benchmark

authority and comparison baseline are explicit

Economics / AGT

centralized recovery

full-information team benchmark with action assignment

comparator not incentive-compatible mechanism

Multi-agent AI

centralized recovery

oracle task assignment from pooled evidence

theoretical upper comparator

Practitioner

centralized recovery

what a coordinator could recover by assigning different attempts

centralized model benchmark

OR/MS

equilibrium-selection boundary

implementation performance depends on the selected stable outcome

declared binary equal-share game

Economics / AGT

equilibrium-selection boundary

selected-equilibrium result with an adverse correspondence witness

no refinement or learning selection claimed

Multi-agent AI

equilibrium-selection boundary

one coordination policy can look good while other compatible policies fail

theoretical policy-class analogy

Practitioner

equilibrium-selection boundary

a good outcome is not enough if the rules also permit bad ones

explanation of a model boundary

Open the complete contextual concordance