💻 computer science

When Does Hallucination Propagate? Reconciling Six Formalizations of Error Spread in Multi-Agent LLM Systems

This paper reconciles contradictory findings on hallucination propagation in multi-agent LLM systems by demonstrating that six distinct formalizations are projections of a single process governed by error correlation and independent verification channels, revealing that propagation depends on structural metrics rather than topology alone and that increasing initial agent diversity can paradoxically amplify false claims.

Original author: Zhang Shaoxuan

Published 2026-11-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original author: Zhang Shaoxuan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of artificial intelligence, engineers are increasingly building teams of digital assistants rather than relying on a single, powerful brain. The logic is intuitive: if one large language model makes a mistake, a second model can spot it, and a third can confirm the correction. This approach, known as a multi-agent system, is being deployed to review medical triage, analyze financial contracts, and screen scientific literature, where a single factual error can be costly. The hope is that by having these agents check one another, the group becomes more reliable than any individual member. However, a growing body of research suggests that this teamwork sometimes backfires. Instead of catching errors, the agents can accidentally reinforce them, causing a false claim to spread through the group like a rumor, gaining weight and confidence with every repetition.

This paradox is the focus of a new systematic review by Shaoxuan Zhang, a researcher at the Harbin Institute of Technology. The paper tackles a confusing situation where two recent studies, both tracking the same type of error through a chain of agents, reached opposite conclusions. One study found that errors were getting smaller and easier to spot as they moved down the line. The other found that errors were becoming harder to detect and were persisting longer. Both studies were correct, but they were measuring different things. One was counting how many false claims existed, while the other was measuring how easy it was to find them. The review reconciles these findings by showing that a false claim can become rarer in number while simultaneously becoming more invisible to the system.

The core of Zhang's work is a framework that explains exactly when a false claim will grow and when it will fade. The author identifies two main structural factors that decide the outcome. The first is how much the agents rely on the same information. If all the agents are reading the same conversation history or using the same training data, they tend to make the same mistakes. When they check each other, they are not offering independent opinions; they are simply echoing the same error. The second factor is whether there is a source of truth that exists outside the group. If the agents only talk to each other, they are trapped in a loop. But if one agent has access to an external tool, a database, or a rule-based check that the others do not use, the system can break the loop and correct the error.

The review analyzed forty-two papers published in the last year to see how researchers have been studying this problem. It found that the field is currently using six different mathematical languages to describe the same process, making it difficult to compare results. Some researchers measure how fast a claim spreads, while others measure how much the agents agree with each other. Zhang maps these different approaches onto a single timeline of how a claim travels. The journey starts with the generation of a claim, moves through transmission between agents, and can end in amplification where the error grows, or in recovery where it is caught.

A striking pattern emerged from the data. Every study that included an independent verification channel—something outside the group of agents to check the facts—reported that errors were reduced. Conversely, the few studies that reported errors getting worse did not have such an external check. This suggests that the key to stopping hallucinations is not just adding more agents or making them talk more, but ensuring that at least one part of the system has access to evidence that the others cannot generate on their own. The review also highlights a hidden danger: a system can appear to be improving because it is discarding false claims, but it might be discarding correct answers at the same time. If the group agrees on a wrong answer and ignores the one agent who is right, the system has failed even if the number of "hallucinations" seems to drop.

The paper concludes that the current design of many multi-agent systems is missing a crucial component: a dedicated recovery mechanism. Prevention, which tries to stop errors from happening in the first place, is not enough because errors will inevitably slip through. Once a false claim enters the system, it needs a way to be removed. The review argues that simply voting on the final answer is not a solution, because a vote cannot create new information; it can only choose between the options already present. If the correct answer is not in the pool of ideas generated by the agents, no amount of voting will find it. The only way to lower the error rate is to introduce independent evidence that changes the pool of options itself.

This work provides a clear set of questions for engineers building these systems. They must decide what counts as a carrier of a claim, whether their verification methods are truly independent, and which type of error they are actually measuring. The findings suggest that the promise of multi-agent systems is real, but only if the design accounts for the fact that agents can easily reinforce each other's mistakes. Without an external check or a way to recover from errors that have already spread, the system may end up amplifying the very mistakes it was built to solve. The path forward is not just about making the agents smarter, but about structuring their interactions so that truth has a way to enter the conversation from the outside.

Technical Summary: When Does Hallucination Propagate?

Problem Statement

Multi-agent Large Language Model (LLM) systems are often deployed under the assumption that agents checking one another reduce hallucinations, based on the intuition that independent samples yield independent errors. However, recent empirical studies present contradictory findings: some report that hallucination rates attenuate (decrease) as claims move through a pipeline, while others report amplification (increase). The paper argues that these contradictions arise because the literature measures different quantities along different axes without a unified frame. Specifically, one study may measure the prevalence of unsupported text (extensive channel), while another measures the detectability of that text (intensive channel). A claim can simultaneously become rarer (lower prevalence) and harder to detect (lower detectability).

Furthermore, the field lacks a unified vocabulary to reconcile six distinct formalisms proposed in the past year, which model error spread as Markov processes, network dynamics, stochastic cascades, context divergences, confidence pooling, or causal influence tensors. Existing surveys organize by agent workflow, system capability stages, or application settings, but none take the "travelling claim" as the unit of analysis.

Methodology

The paper employs a systematic review of a purposively sampled corpus of 42 papers (32 included studies) drawn from arXiv records from late 2022 to late 2026. The sampling was quota-based, targeting specific stages of error spread (amplification, consolidation, transmission, attribution, recovery) and formalizations.

Coding Protocol:
The author coded the corpus against a new framework defined by:

Unit of Analysis: The claim (proposition), not the agent or system.
Structural Variables: The correlation of agent errors and the independence of verification channels.
Outcome Channels: Distinguishing between False Positive (FP) amplification (false claims gaining support) and False Negative (FN) truth suppression (correct claims being discarded).
Verification Definition: A strict definition of "independent evidence" where the verifier's inputs must be disjoint from the claim's ancestry (provenance). Heterogeneous models reviewing the same context do not count as independent.

The study also audited the reporting practices of the field, checking for the presence of propagation-level metrics, independent verification channels, and recoverable out-degree distributions in communication graphs.

Key Contributions

1. A Conditional Account of Error Spread

The paper proposes that whether a false claim grows is determined by two structural variables:

Correlation Pressure: The degree to which agents share priors, prompts, context, or conversation history, leading to correlated errors.
Independent Verification: The presence of a channel supplying evidence disjoint from the agent set's ancestry (e.g., external tools, retrieval from non-agent corpora).

Proposition 1: Aggregation reduces error rates only if (i) errors are not strongly positively correlated AND (ii) at least one independent verification channel exists. If errors are correlated, the effective number of voters collapses, and without independent evidence, the system cannot self-correct.

2. Reconciliation of Six Formalisms

The author demonstrates that the six existing formalisms (Markov processes, network processes, stochastic cascades, etc.) are not contradictory but are projections of the same underlying $(k, e)$ process onto different observables:

$k$ : Corroboration count (support from other agents).
$e$ : Independent evidence count.
Each formalism measures a different aspect (e.g., escape probability, reproduction number, divergence score) but fails to separate the FP and FN channels or account for the independence of evidence.

3. The Threshold and Topology

The paper defines a branching ratio $R = f \cdot s$ , where $f$ is the fan-out and $s$ is the per-hop survival probability.

Proposition 2b: In a linear chain topology, $f=1$ , making $R < 1$ by construction. Thus, chains cannot amplify the number of carriers (extensive channel); any attenuation observed is a topological consequence, not necessarily evidence of self-correction.
However, chains can amplify undetectability (intensive channel) via semantic drift ( $\mu$ ) and provenance laundering ( $\lambda$ ).
Conjecture 1: The disagreement in the literature regarding whether topology or model scale dominates may be a function of where specific systems sit relative to the $R=1$ threshold.

4. Governance and Interventions

The paper maps interventions to parameters and outcome channels:

Aggregation Ceiling (Proposition 3): No aggregation rule (voting, consensus) can remove aggregation loss ( $\Lambda$ ) uniformly or create information not present in the candidate pool. Aggregation-side interventions are zero-sum between FP and FN channels.
Independent Channels: Only adding an independent verification channel can lower both FP and FN rates simultaneously by altering the pool of candidates rather than just reweighting them.
Recovery as a First-Class Component: The paper argues for treating recovery (detection, containment, rollback) as distinct from prevention. Due to the "ratchet" effect in persistent memory (where writes are cheap and irreversible, but retractions are expensive), systems without rollback mechanisms accumulate durable errors.

Results from the Coded Corpus

Reporting Gaps: Of 32 included studies, only 10 report any propagation-level metric, 9 report an independent verification channel, and only 7 report a graph structure allowing out-degree recovery.
The Co-occurrence Pattern: A critical finding is the absence of co-occurrence between independent verification and amplification. All 9 studies with an independent channel reported attenuation. Both studies reporting amplification lacked an independent channel.
Metric Scarcity: The effective reproduction number ( $R_{eff}$ ) and aggregation loss ( $\Lambda$ ) are reported in exactly one paper each, highlighting a lack of commensurable quantities across the field.
Diversity Paradox: Contrary to design intuition, increasing the initial diversity of agent answers can increase false adoption (conformity) because disagreement provides a majority with a target to converge upon.

Significance and Claims

The paper claims to provide the first framework that reconciles contradictory findings in multi-agent LLM reliability by distinguishing between the extensive channel (carrier count) and the intensive channel (detectability/support strength).

Its primary significance lies in:

Resolving Disagreement: Showing that "amplification" and "attenuation" are not mutually exclusive but describe different projections of the same process.
Defining Independence: Establishing a strict provenance-based definition of independent verification that reclassifies many "verification" procedures in the literature as mere coordination protocols.
Identifying the Aggregation Limit: Proving that voting and consensus mechanisms cannot fix errors if the underlying candidate pool lacks correct answers or independent evidence.
Setting an Agenda: Proposing a reporting standard (carrier set, aggregation loss, out-degree distribution, comparator) to make future studies comparable and suggesting specific experiments to separate causal mechanisms from research artifacts.

The paper concludes modestly, noting that while the mechanisms are documented, the quantitative parameters (like the false-adoption coefficient $g$ and truth-abandonment coefficient $a$ ) remain unestimated in the current literature, and the framework serves primarily as a scheme for organizing and comparing evidence rather than a fully calibrated model.