When Do Institutions Beat Intelligence?

Technical Summary: When Do Institutions Beat Intelligence?

Problem Statement
The paper addresses a fundamental disconnect in multi-agent systems (MAS): more capable individual agents do not necessarily form a more capable collective. A system may possess sufficient distributed information to solve a task yet fail because that information is never jointly accessible, public belief admits unsupported or correlated evidence, shared state becomes stale or strategically distorted, or the resulting representation cannot support reliable action. These failures arise not from the reasoning capacity of individual agents, but from the path by which private observations become public reasons for collective action.

The core research question is diagnostic: under what conditions should resources be allocated to strengthen the reasoner (intelligence intervention) versus changing the structure through which agents observe, report, verify, and act (institutional intervention)? The author argues that current workflows often conflate these factors, making it difficult to identify whether performance gains stem from increased reasoning capacity or improved information routing and state construction.

Methodology
The author proposes a controlled experimental framework using "artificial ecologies" designed around four specific loci of collective failure, drawing functional distinctions from group decision-making and distributed cognition research:

1. Access and Routing: Whether distributed knowledge becomes jointly available (motivated by hidden-profile and transactive memory research).
2. Admission and Dependence: Which reports enter public belief and whether repeated reports provide independent support (motivated by social influence and wisdom of crowds research).
3. State Maintenance and Incentives: Whether public state remains current and whether deviations are observable and consequential (motivated by accountability research).
4. Representation and Action: Whether a constructed state remains useful relative to stronger capability and whether the final agent can reliably act on it (motivated by distributed cognition research).

Experimental Design:

• Intervention Types: The study compares intelligence interventions (stronger models, increased inference budget) against institutional interventions (protocols for routing, verification, state updates, and action interfaces).
• Controls: To isolate causal mechanisms, the author employs:
  • Call-matched baselines: Comparing institutional gains against raw reasoning with equivalent computational budgets.
  • Mechanism-breaking controls: Deliberately breaking the institutional signal (e.g., invalidating credentials, removing checkability) to test if gains persist.
  • Substitution tests: Using frontier models or learned components (e.g., rerankers) to see if they can perform the institutional transformation directly.
  • Fixed-evidence interfaces: Holding evidence constant while changing the representation to isolate action interface bottlenecks.
• Models and Tasks: Experiments span multiple model families (Llama, Qwen, Mistral, Claude, Gemini, DeepSeek) and tasks including synthetic logic puzzles, HotpotQA, and MuSiQue retrieval.

Key Contributions

1. Functional Taxonomy of Collective Failure: A classification of artificial collective failures into four categories (access, public belief, state maintenance, action) that separates failures of information flow from failures of reasoning capacity.
2. Mechanism-Breaking Methodology: A rigorous comparison framework that distinguishes institutional repair from additional reasoning by pairing positive interventions with matched reasoning baselines and controls that break the proposed mechanism.
3. Empirical Boundary Conditions: Identification of specific conditions under which institutions provide an advantage over raw intelligence and conditions under which they become redundant or ineffective.

Results
The experiments reveal a consistent boundary for institutional advantage:

• Institutions Help When: They repair failures in how a collective constructs a usable public state.
  • Access: In distributed evidence tasks, a valid routing institution (ensuring disjoint coverage of clues) allowed a weaker model (8B) to outperform a stronger raw model (70B) by +0.58 accuracy, as the stronger model lacked access to the complete evidence set.
  • Admission: In faulty evidence and wrong-consensus scenarios, grounded validation (rejecting unsupported or correlated reports) allowed institutions to achieve near-perfect accuracy where call-matched raw reasoning failed completely or remained low.
  • Incentives: Checkable audits (where violations are observable) significantly improved accuracy and reduced strategic misreporting. However, sanctions without checkability yielded zero gain.
• Institutions Lose Advantage When:
  • Signals are Uninformative/Uncheckable: If the institutional signal is broken (e.g., invalid credentials) or uncheckable (violations cannot be observed), the gain disappears.
  • Substitution by Capability: When a sufficiently capable raw model or a learned component (e.g., a learned reranker) can perform the same transformation (e.g., evidence selection, state maintenance) directly from raw context, the explicit institutional machinery becomes unnecessary. For instance, frontier models like Claude showed zero incremental interaction in tested ecologies, suggesting they could substitute for the institutional rules.
  • Action Interface Bottlenecks: Even with better evidence construction, if the final interface (e.g., graph vs. ordered list) is difficult for the finalizer to execute, gains are erased. Fixed-evidence experiments showed that reformatting the same accepted evidence could reverse performance.

Significance and Claims
The paper claims to recast the choice between intelligence and institutions as a diagnostic problem of where collective reasoning fails. It argues that institutional value is not a permanent property of a workflow but is relative to the capability frontier.

The author concludes with three design principles:

1. Buy Intelligence when capability directly performs the required transformation (substitution).
2. Build an Institution when a structural failure (access, validity, maintenance) survives scaling, the rule observes it reliably, and the resulting state supports action.
3. Redesign or Abandon when the rule is uninformative, redundant, or produces a state the finalizer cannot use.

The work emphasizes that institutional machinery introduces computation and failure modes, while additional reasoning is wasted if the state remains unusable. The primary contribution is not a ranking of workflows, but a method to identify the binding constraint in collective computation.