Skip to content

Epistemic Independence

Epistemic independence is the property that two assessments are generated from separate information, separate reasoning processes, or separate models, so that agreement or disagreement carries independent evidential weight. It is the central architectural requirement for any multi-agent system that claims to provide independent verification or parallel analysis.

Five Dimensions of Independence

The distinction between single-agent simulation and concurrent multi-agent operation is graded across five dimensions, each a continuum:

  1. Context isolation: fully isolated (separate processes, separate contexts) to unified context
  2. Model diversity: different models with different training to same model with different prompts
  3. History divergence: agents with long separate session histories to fresh invocations with no history
  4. Temporal independence: truly concurrent execution to sequential simulation
  5. Information asymmetry: architecturally enforced to none

Where a system falls on each dimension determines which independence properties it can provide. For use cases requiring structured task decomposition, the low-independence end (Pattern 0) is adequate. For use cases requiring independent verification, the high-independence end (Pattern 9 with model diversity) is a correctness requirement, not an optimisation.

Why Single-Agent Simulation Cannot Provide Independence

Two properties of llm-architecture constrain what single-agent simulation achieves:

Unified attention. The transformer self-attention operation computes weights across the full context at every layer. When a single model generates responses for "Agent A" and then "Agent B," information from A's reasoning is visible to B's generation. There is no partition. Prompting cannot override this — it is mathematical.

Shared weights. All characters generated by a single model share trained parameters and distributional biases. Errors are correlated at their source. Stochastic variation from different random seeds is not independence. Chain-of-thought does not partition reasoning within a single context. RAG provides different knowledge access but not different reasoning processes.

Empirical Evidence

The Ultimatum Game Experiment (Sreedhar & Chilton, 2024)

The most direct controlled comparison: single GPT-4 instances simulating both players vs separate GPT-4 instances (one per player) on a five-round Ultimatum Game with personality-differentiated players (greedy and fair).

Structure Human-Consistent Simulations
Multi-agent (GPT-4) 87.5%
Multi-agent (GPT-3.5) 82.5%
Single-LLM (GPT-4) 50.0%
Single-LLM (GPT-3.5) 42.5%

The difference is statistically significant (χ² = 13.091, p < .001). The error analysis reveals the mechanism: in multi-agent systems, 100% of errors were strategy creation errors with zero gameplay mistakes — once agents had separate contexts, they executed their strategies faithfully. In single-LLM systems, 25–39% of errors were gameplay execution failures — the single model couldn't maintain separate player strategies because information leaked across the shared context. This directly confirms the unified attention problem: the proposer's strategy was visible during the receiver's reasoning.

The hardest condition (Greedy-Greedy, requiring differentiated selfish strategies) showed the largest gap: multi-agent achieved 80% vs single-LLM only 10%. The value of architectural independence is greatest precisely when differentiated strategies matter most.

Multi-Agent Debate: Format vs Independence

Du et al. (2023) demonstrated that multi-agent debate (multiple LLM instances debating over rounds) improves reasoning by +14.8 points (arithmetic) and factuality by +7.8 points (biographies) over single agents. However, all agents were instances of the same model (ChatGPT). The improvement comes from structured cross-examination format (catching surface errors) rather than from epistemic independence (providing diverse perspectives).

The critical test: when ChatGPT and Bard (different models) debated together, performance improved further — joint debate solved 17/20 problems vs ChatGPT alone (14) or Bard alone (11). Cross-model debate provides benefit beyond what same-model debate achieves, supporting the case for model diversity.

Subsequent work clarified the limits of same-model debate: - Chan et al. (2024) provided the clearest direct test: when all debate agents received the same role prompt, multi-agent evaluation performed no better than single-agent (53.8% vs 60% with diverse roles). The structure of debate adds no value without genuine perspective diversity. Performance also peaked at 3–4 agents and degraded beyond 2 discussion rounds, suggesting coordination overhead limits. - Wu et al. (2025) showed that same-model debate suppresses independent reasoning under majority pressure - Wynn et al. (2025) found that model diversity is a necessary condition for productive debate - Du et al. (2023) themselves noted that agents are "relatively agreeable" — likely from RLHF training (see sycophancy) — and that "stubborn" prompts produce better debates

Theory of Mind Studies

Studies on LLM theory of mind (Kosinski, 2024; Ullman, 2023) show that the ability to track another entity's belief state separately from one's own is fragile, degrading under minor perturbations.

Measuring Emergent Collective Properties (Riedl, 2025)

Riedl (2025) provides the first information-theoretic framework for measuring whether multi-agent LLM systems exhibit genuine emergent properties. Using partial information decomposition and time-delayed mutual information, the framework distinguishes three levels: (1) no emergence (mere aggregation), (2) identity-linked differentiation without integration (agents specialise but don't complement each other), and (3) true collective coordination (differentiated roles with goal-directed complementarity). Only the third level — achieved by combining distinct personas with explicit theory-of-mind reasoning — produces synergy that amplifies redundancy to improve performance. Both differentiation and integration are required; either alone is insufficient. This provides the measurement tools to verify whether a multi-agent system's independence claims are genuine or illusory.

Empirical limitation. No published study has compared single-agent simulation and concurrent operation on safety-critical monitoring tasks with operational data. The architectural argument is strong; empirical confirmation of its magnitude in safety-critical domains is pending.

Independence as a Continuum

Full model diversity (different base models from different training pipelines) is a hard requirement for: - Independent safety verification where regulatory independence criteria apply - Any function where the multi-agent architecture is credited as a safety barrier

For applications where multi-agent provides operational value but is not credited as a safety function (ambient monitoring, shift handover support, collaborative problem-solving), partial decorrelation from fine-tuned variants is acceptable. The decision criterion: is this function credited in the safety case? If yes, full model diversity required. If no, independence requirement is proportional to consequence of correlated failure.

Where Single-Agent Simulation Is Sufficient

Single-agent simulation delivers real value for: functional decomposition (structured task routing), adversarial review of surface errors where extrinsic feedback is available, narrative and exploratory reasoning, and prototyping multi-agent approaches before building infrastructure. In each case, the improvement comes from structured process rather than independent perspectives. Single-agent simulation is adequate when functional diversity is the primary need. It is inadequate when epistemic-independence, error independence, or temporal parallelism is required.

Relevance to Safety-Critical Systems

  1. Claims of independence must specify the architectural basis. Whether agents use different base models, whether they have separate information access, and whether their reasoning is computationally isolated are verifiable engineering properties.

  2. Same-model "independent" review does not satisfy independence requirements. An adversarial agent on the same base model will tend toward the same conclusions despite its role assignment.

  3. Cross-domain mapping. In aviation, independent flight path verification requires separate computational processes. In medical diagnostics, independent second opinions require clinicians with different training and experience. In oil and gas, independent well control verification requires separate analysis teams. The AI domain needs the same structural separation.

  4. The cost of independence is the cost of model diversity. Maintaining multiple base models from different providers approximately doubles operational complexity. For safety-critical verification roles, this cost is the price of genuine independence.