PaperlayerView exact revision on arXiv

Source-linked AI summary

Hallucination Cascade: Analyzing Error Propagation in Multi-Agent LLM Systems

Saeid Jamshidi, Arghavan Moradi Dakhel, Kawser Wazed Nafi, Foutse Khomh

arXiv:2606.07937v1cs.CR

TL;DR

LLMs can produce fluent but unsupported claims, while existing evaluations largely treat hallucination as a static property of isolated outputs. This paper tracks claim-level factual inconsistencies through multi-agent cascades and finds that sequential refinement reduces hallucination but also lowers factual accuracy and response quality. The resulting framework characterizes hallucination as a dynamic signal that propagates, attenuates, and transforms across agents.

  • Problem

    Existing hallucination research largely evaluates isolated outputs, leaving dynamic error evolution across multi-agent interactions insufficiently modeled.

    Source: E. Domain-Specific Hallucination… · p. 2

  • Method

    The method decomposes responses into atomic claims, grounds them with reference facts and semantic validation, and tracks aggregated scores through sequential agent cascades.

    Source: §I · p. 2, D. Cascade System · pp. 4–5

  • Results

    Hallucination decreases across cascade stages, but refinement is accompanied by declines in factual accuracy and response quality.

    Source: §VI · pp. 17–18, §VII · p. 19

  • Takeaways & Limitations

    Hallucination should be evaluated as a dynamic propagation signal alongside factual preservation, semantic richness, and response utility.

    Source: §VI · p. 18, §VII · p. 19

    The approach depends on accurate claim extraction, complete reference knowledge bases, and learned representations that may introduce estimation bias.

    Source: A. Limitations · p. 18

Abstract

from arXiv · show

Large Language Models (LLMs) generate fluent text but remain vulnerable to hallucinations, producing unsupported, inconsistent, and factually incorrect claims. Most prior work treats hallucination as a static property of isolated outputs. In multi-agent LLM systems, however, responses are exchanged across agents, revised through sequential stages, and reused as context for later reasoning. Hallucination, therefore, becomes a dynamic process shaped by interaction history, cascade depth, and model heterogeneity. This paper analyzes hallucination dynamics in multi-agent LLM cascades by tracking claim-level factual inconsistencies across sequential agent interactions. We conduct 500 cascade experiments across 10 knowledge domains using GPT-5.3, DeepSeek-V3, and LLaMA-3-70B-Instruct, yielding 1,250 evaluated responses. Results show that deeper cascades reduce the normalized hallucination score from 0.422 at the first agent to 0.272 at the final agent in 3-agent chains, with an amplification factor of 0.644, indicating net attenuation. This reduction is accompanied by a decline in factual accuracy from 0.789 to 0.769, revealing a trade-off between hallucination suppression and factual preservation. Transition-level analysis shows that each agent-to-agent refinement reduces hallucination by an average of 0.072, with small but consistent losses in factual consistency and response quality. Model-level results reveal reliability-efficiency trade-offs: LLaMA-3-70B-Instruct achieves the lowest hallucination score, whereas GPT-5.3 provides faster generation with a higher hallucination rate. Domain-level analysis shows that hallucination varies with topic complexity, with lower scores in well-grounded scientific domains and higher scores in more abstract domains.

I. INTRODUCTION

The paper reframes hallucination from a static property of isolated outputs to a dynamic process that propagates across multi-agent interactions. It combines claim-level estimation with cascade analysis to characterize how errors evolve and affect reliability.

I. INTRODUCTION

III. METHODOLOGY

I. INTRODUCTION

A. Hallucination in LLMs

II. RELATED WORK

A. Claim Extraction and Representation

  • Claim extraction uses semantic segmentation followed by deterministic refinement that removes duplicates, normalizes coreference, completes clauses, and filters non-factual fragments.Source: A. Claim Extraction and Representation · p. 3

B. Overview of the Approach

The approach decomposes responses into verifiable claims, grounds them against domain references, and aggregates their scores across sequential agents. This converts hallucination from an isolated response property into a trackable cascade trajectory.

B. Overview of the Approach

C. Formal Problem Definition

  • The estimator combines semantic similarity, grounding uncertainty, rule-based verification, and LLM-based judgment to identify unsupported or inconsistent claims.Source: C. Formal Problem Definition · p. 3
  • Response-level hallucination is defined as the mean claim-level score, providing a normalized measure of factual inconsistency for a response and domain.Source: C. Formal Problem Definition · p. 3

D. Cascade System

The cascade system models hallucination as information and factual errors propagate through sequential agents, while combining grounding-based and LLM-based assessment. It also defines parallel and specialist architectures that use model diversity and reliability-aware aggregation.

Sequential cascade

  • Each agent receives the original prompt and accumulated prior outputs, enabling refinement while allowing earlier factual errors to affect later generations.The sequential formulation supports attenuation, preservation, transformation, amplification, and recovery of hallucinated content.Source: D. Cascade System · pp. 4, 6

Adaptive hallucination estimation

  • H(t, k) = w(t, k) · Hrule(t, k) + (1 −w(t, k)) · HLLM(t, k) combines rule-based grounding with LLM-based semantic assessment.The adaptive weight favors grounding when reference support is reliable and the LLM judge when evidence is weak or incomplete.Source: D. Cascade System · pp. 4–5

Claim-level estimation

  • Claim-level rule-based verification uses reference knowledge and hallucination patterns to estimate factual support and penalize fabricated entities, unsupported relations, temporal inconsistencies, and contradictions.The estimator is bounded, adaptive to grounding reliability, and less dependent on either component alone.Source: D. Cascade System · p. 5

LLM-based semantic judgment

  • The LLM judge classifies each claim across factual correctness, unsupported inference, fabrication, logical contradiction, and contextual inconsistency, then maps the probabilities to a scalar hallucination score.Calibration aligns predicted probabilities with empirical annotations and keeps scores comparable across cascade settings.Source: D. Cascade System · p. 5

Parallel and specialist architectures

  • Parallel aggregation removes dependence on previous outputs and selects responses using semantic agreement and factual reliability, while high disagreement can reveal useful diversity and low disagreement can indicate correlated failures.A positive diversity gain indicates hallucination reduction beyond individual model performance.Source: D. Cascade System · p. 6
  • Reliability-aware specialist aggregation weights heterogeneous outputs toward low-hallucination specialists and supports diagnosis of specialist, coordinator, cross-specialist, and downstream errors.A coordinator prioritizes factual consistency, semantic coherence, and preservation of verified information.Source: D. Cascade System · pp. 6–7

E. Hallucination Cascade Dynamics

Hallucination dynamics are represented as a temporal stochastic process in which prior hallucination states, prompt uncertainty, and generation noise shape later outputs. The formulation supports trajectory analysis, propagation measurement, stability assessment, and long-chain behavior analysis.

Dynamic formulation

  • The hallucination sequence {h_i} models how errors evolve across agents, with propagation coefficients, prompt uncertainty, and stochastic variability determining each step.The first-order dependency conditions the current hallucination score on the immediately preceding score.Source: E. Hallucination Cascade Dynamics · p. 7

Stability and information preservation

  • Stability requires past hallucination states to decay rather than accumulate, with first-order dynamics attenuating when the coefficient magnitude is below one.Even stable dynamics may preserve hallucination when the original prompt is ambiguous and lacks factual context.Source: E. Hallucination Cascade Dynamics · p. 7

Algorithmic analysis

  • The cascade algorithm generates an initial response, iteratively aggregates prior context, estimates each hallucination score, and returns the trajectory with propagation and amplification metrics.It also computes cumulative deviation from the initial response and handles zero initial hallucination separately.Source: E. Hallucination Cascade Dynamics · p. 8

Stability and information preservation

  • Repeated refinement can reduce hallucination while also compressing, weakening, or transforming factual information, enabling analysis of semantic drift and long-chain failure cases.The formulation therefore treats hallucination as an information-dissipation factor rather than only an output-level score.Source: E. Hallucination Cascade Dynamics · p. 8

Propagation metrics

  • Negative ρ_i indicates attenuation, values near zero indicate preservation, and positive ρ_i indicates amplification in the cascade.Γ(L) measures total deviation, while A_L summarizes global attenuation or amplification.Source: E. Hallucination Cascade Dynamics · pp. 7–8

IV. EXPERIMENTAL SETUP

The experimental setup evaluates hallucination as a dynamic cascade phenomenon, covering propagation, attenuation, recovery, semantic drift, long-chain stability, and failure dynamics. It links the proposed formulation to measurable cascade behavior.

Experimental scope

  • The experiments evaluate how hallucination evolves across sequential agents and how cascade depth affects factual reliability and model heterogeneity affects correction behavior.The setup explicitly targets propagation, attenuation, recovery, semantic drift, long-chain stability, and failure dynamics.Source: §IV · p. 8

A. Multi-Agent Experimental Design for Hallucination Analysis

The benchmark studies sequential multi-agent hallucination across three models, ten knowledge domains, and 500 cascade runs, producing 1,250 evaluated responses. Outputs are assessed at claim level for hallucination and related reliability, quality, efficiency, and propagation measures.

Multi-agent benchmark design

  • The benchmark uses GPT-5.3, DeepSeek-V3, and LLaMA-3-70B-Instruct in sequential cascades where downstream agents receive the prompt and all preceding outputs.The design makes hallucination observable as a trajectory rather than a single response-level score.Source: A. Multi-Agent Experimental Design for… · p. 8
  • 500 cascade runs comprise 250 two-agent chains and 250 three-agent chains, yielding 1,250 independently evaluated agent-level responses.The experiments span ten named knowledge domains, with 50 baseline runs per domain.Source: A. Multi-Agent Experimental Design for… · p. 8

Evaluation measures

  • Responses are decomposed into atomic claims and scored through rule-based grounding and LLM-based semantic judgment on normalized hallucination scores in [0, 1].The evaluation also measures factual accuracy, response quality, semantic drift, cascade risk, response time, token usage, inference cost, and propagation metrics.Source: A. Multi-Agent Experimental Design for… · p. 8
  • The study additionally analyzes cross-model, long-chain, injected-hallucination, error-recovery, and failure-case cascades to examine convergence, saturation, rebound, information loss, correction, overcorrection, and unresolved hallucination.All experiments use consistent prompting and inference settings.Source: A. Multi-Agent Experimental Design for… · p. 8

B. Cascade Configurations

The study uses multiple cascade configurations to isolate how interaction structure affects hallucination propagation, recovery, and failure.

B. Cascade Configurations

  • Short sequential cascades provide the baseline for attenuation and amplification analysis.Cross-model cascades test heterogeneous model sequences, while long-chain cascades assess effects of added depth.Source: B. Cascade Configurations · p. 8
  • Injected-hallucination cascades trace controlled factual errors across agents.Error-recovery and failure-case analyses connect aggregate propagation metrics with claim-level outcomes.Source: B. Cascade Configurations · p. 8

C. Evaluation Metrics

The evaluation measures response reliability, efficiency, and hallucination propagation using claim-level and trajectory-based metrics.

C. Evaluation Metrics

  • Response-level assessment covers hallucination, factual accuracy, response quality, semantic drift, and cascade risk.These metrics assess factual correctness, utility, meaning shift, and reliability degradation.Source: C. Evaluation Metrics · p. 9
  • Efficiency is measured through response time, token usage, and inference cost.The experimental configuration and cascade designs provide the context for these evaluations.Source: C. Evaluation Metrics · p. 9, Table I · p. 9, Table II · p. 9
  • Propagation analysis tracks factual inconsistency across agents with trajectory, deviation, amplification, recovery, and failure measures.The relative propagation factor ρ_i, cumulative deviation Γ(L), and amplification factor A_L quantify local, cumulative, and global change, respectively.Source: C. Evaluation Metrics · p. 9

D. Baselines

The experimental setup compares the proposed multi-agent cascade with controlled baselines and frames the study around cascade dynamics, depth, model heterogeneity, and empirical findings.

D. Baselines

  • Four baselines compare direct generation, reasoning-based prompting, retrieval-grounded generation, and same-model iterative revision under matched evaluation conditions.Using the same prompts, knowledge domains, and metrics isolates the effect of multi-agent interaction.Source: D. Baselines · p. 9

E. Research Questions

  • RQ1 examines whether hallucination accumulates, attenuates, recovers, or fluctuates across sequential cascades.The analysis uses trajectories, propagation rates, cumulative deviation, amplification, recovery, injected errors, and failure cases.Source: E. Research Questions · p. 9
  • RQ2 evaluates whether deeper cascades improve mitigation while introducing saturation, semantic drift, factual decay, information loss, or quality degradation.Short and long-chain cascades provide the comparison.Source: E. Research Questions · p. 9
  • RQ3 compares GPT-5.3, DeepSeek-V3, and LLaMA-3-70B-Instruct across homogeneous and heterogeneous cascades.The comparison covers reliability, attenuation, recovery, semantic drift, quality, latency, token usage, and inference cost.Source: E. Research Questions · p. 10

V. EXPERIMENTAL RESULTS AND ANALYSIS

  • The results section reports empirical findings from the evaluation.The supplied passage identifies the section's role but does not specify individual results.Source: §V · p. 10

A. Overall System and Model Performance

Across models, cascade configurations, and domains, the system exhibits measurable hallucination with structured reliability–efficiency trade-offs. Deeper refinement generally suppresses hallucination and drift, but can reduce factual preservation and response quality.

A. Overall System and Model Performance

B. Chain Dynamics, Stability, and Propagation

C. Drift, Retention, and Risk Behavior

D. Ablation Study

  • Removing claim decomposition increases hallucination from 0.272 to 0.338 and semantic drift from 0.257 to 0.312, while removing the model-based estimator also substantially increases hallucination.The full method achieves the lowest hallucination, semantic drift, and cascade risk.Source: D. Ablation Study · pp. 12–13, Figure 4 · p. 13, Figure 5 · p. 13

E. Claim-Level Transition and Recovery Analysis

Claim-level analysis shows that sequential refinement usually corrects or weakens hallucinated claims, while topic complexity and system-level trade-offs shape the reliability of the final cascade.

E. Claim-Level Transition and Recovery Analysis

  • Negative claim-level changes indicate correction or weakening, while positive changes indicate hallucination amplification.Claims without semantic matches are deleted, and factual claims becoming unsupported, distorted, or unnecessarily removed are classified as overcorrected.Source: E. Claim-Level Transition and Recovery… · p. 13
  • Amplified hallucinations account for only 7.3% of transitions, while preserved hallucinations decrease from 24.7% in the first transition to 17.9% in the second.Deleted and overcorrected claims account for 15.2% of transitions overall, indicating that lower hallucination reflects both correction and weakening as well as some deletion and overcorrection.Source: E. Claim-Level Transition and Recovery… · p. 13

F. Topic-Level Variability

  • Photosynthesis and DNA have the lowest hallucination scores, while Black Holes and Roman Empire have the highest.The reported scores are 0.265489 for Photosynthesis, 0.266954 for DNA, 0.486860 for Black Holes, and 0.473012 for Roman Empire.Source: F. Topic-Level Variability · pp. 13–14
  • Topic difficulty affects hallucination, factual accuracy, semantic drift, and cascade risk across knowledge domains.Photosynthesis has factual accuracy of 0.905600, while Black Holes has 0.588800; Quantum Computing has the highest drift at 0.309514, and DNA the lowest at 0.195635.Source: F. Topic-Level Variability · p. 14

G. Hallucination-Quality Trade-off and Pareto Analysis

  • Higher hallucination generally coincides with lower response quality and higher cascade risk, producing a narrow Pareto-efficient region.The results frame mitigation as a balance among factual reliability, semantic utility, and system stability rather than optimization of one metric.Source: G. Hallucination-Quality Trade-off and… · pp. 14–15, Figure 9 · p. 16

H. Comparison with Existing Approaches

  • The proposed method extends single-stage detection by tracking how claims are preserved, corrected, transformed, and amplified across interacting agents.This propagation-aware view complements uncertainty estimation, external verification, metamorphic testing, and other existing approaches.Source: H. Comparison with Existing Approaches · pp. 15–16

I. Model-Order Analysis in Cross-Model Cascades

Model order materially changes cascade reliability and efficiency: placing LLaMA-3-70B-Instruct last minimizes hallucination, while GPT-5.3-last cascades are faster and preserve more accuracy and quality. Statistical results and baseline comparisons reinforce that hallucination attenuation comes with information-preservation trade-offs.

I. Model-Order Analysis in Cross-Model Cascades

  • LLaMA-3-70B-Instruct as the final agent yields the lowest hallucination scores, 0.263 and 0.271, while GPT-5.3-last cascades score 0.326 and 0.338.GPT-5.3-last configurations retain higher factual accuracy and response quality, establishing a reliability-efficiency trade-off.Source: §I · pp. 16–17, Table XVII · p. 18

J. Statistical Significance Analysis

  • The mean transition-level hallucination change is −0.072489 (p < 0.001), while factual accuracy and response quality decline slightly.Chain-3 significantly improves final hallucination and cascade risk over Chain-2 but introduces greater factual decay.Source: J. Statistical Significance Analysis · p. 17, Table XVIII · p. 19

K. Comparison with Baselines

  • The proposed cascade lowers hallucination from 0.417 for Single-Agent generation to 0.272, the lowest among the compared methods.RAG achieves the highest factual accuracy, 0.815, and response quality, 0.781, while CoT improves accuracy and quality but retains hallucination of 0.398.Source: K. Comparison with Baselines · p. 17, Table XIX · p. 19

VI. DISCUSSION

  • Sequential refinement suppresses hallucination but can omit valid details, simplify nuanced claims, and reduce response utility through semantic drift.Reliable cascades therefore require balancing correction, drift control, and information retention.Source: §VI · pp. 17–18

A. Limitations

  • The evaluation is limited by claim-extraction errors, dependence on reference-knowledge completeness, estimator bias, restricted models and topics, and absent adversarial testing.The paper calls for broader validation across domains and architectures, including prompt-injection scenarios.Source: A. Limitations · p. 18

B. Future Work

  • Future work proposes adaptive routing, adversarial evaluation, energy-aware analysis, learned propagation parameters, and real-world domain validation.These directions target hallucination control, robustness, sustainability, predictive stability modeling, and practical utility.Source: B. Future Work · pp. 18–19

VII. CONCLUSION

  • The paper models hallucination as a stochastic dynamical process in which claim-level errors propagate, attenuate, and transform across agents.Its conclusion links structured correction with measurable losses in factual accuracy and response quality.Source: §VII · p. 19

Related papers

Loading 2606.07937v1…