Multimodal Evaluator Preference Collapse: Cross-Modal Contagion in Self-Evolving Agents

Abstract: When AI agents use LLMs to evaluate their own outputs in a feedback loop, systematic biases emerge. We show that Evaluator Preference Collapse (EPC) is dramatically amplified in multimodal settings. Using GPT-4o to evaluate DeepSeek-chat across text and visual tasks, we find that a single strategy (step_by_step) absorbs 48.4% of all weight -- 3.2x the collapse observed in text-only self-evaluation -- while three visual-domain strategies receive only 9.1% combined weight. We then demonstrate a novel phenomenon we term cross-modal contagion: evaluator preferences acquired on one modality transfer to and corrupt strategy selection on another. Through a four-phase isolation training paradigm, we measure contagion coefficients and document strategy inversion -- the optimal strategy for a modality reverses after cross-modal exposure. A Phase 3 statistical validation across five evaluator configurations (N=80 total independent repetitions, ~35,000 API calls) with both text-proxy and real-image visual tasks finds: cross-model evaluation produces strong contagion (JSD~0.19-0.34), real-image inputs yield the most directionally consistent signal (mean gamma_{T->V}=1.145, gamma_{V->T}=0.937, 70% T->V, Cohen's d=0.56), and self-evaluation provides near-complete immunity -- 97% of runs (N=30) yield zero contagion (JSD=0.003, d=0.07). Three methodological ablations and multi-executor validation confirm the effect is not a structural artifact. We introduce the contagion matrix indexed by evaluator identity, release the MM-EPC framework, and identify cross-model evaluator architecture as the primary risk factor for preference drift. Code and data: https://github.com/aidless/mm-epc.

• The paper demonstrates that LLM evaluators in multimodal closed-loop systems induce significant preference collapse, with GPT-4o yielding a PCI over three times higher than self-evaluation.
• It introduces the Multimodal Preference Collapse Index (MPCI) and a contagion matrix to quantify cross-modal bias transfer between text and visual modalities.
• Empirical findings highlight the need for evaluator ensemble strategies and modality isolation to prevent silent convergence to evaluator-biased, suboptimal strategies.

Multimodal Evaluator Preference Collapse and Cross-Modal Contagion in Self-Evolving AI Agents

Background and Motivation

The paper "Multimodal Evaluator Preference Collapse: Cross-Modal Contagion in Self-Evolving Agents" (2606.16682) interrogates the dynamics underlying closed-loop agent adaptation where LLMs serve as evaluators for multimodal tasks. The core proposition is that feedback-driven adaptation pipelines, when extended from text-only domains to multimodal settings, are susceptible to acute evaluator-induced bias and cross-modal transfer of strategy preferences. Motivated by prior work documenting Evaluator Preference Collapse (EPC) in text-only self-evaluation scenarios, this research generalizes EPC to multimodal agent systems and rigorously quantifies the magnitude and structure of cross-modal evaluator bias.

Formalization of Multimodal Preference Collapse

EPC is characterized by excessive concentration of agent strategy weights toward a single, evaluator-favored configuration. The paper introduces the Multimodal Preference Collapse Index (MPCI), which aggregates intra-modality concentration (PCI) and inter-modality divergence (CPCI). Using pairwise comparisons and multiplicative update bandit algorithms over strategy sets, the authors systematically probe how evaluation by GPT-4o amplifies strategy collapse relative to self-evaluation (DeepSeek-chat) and a random baseline.

Empirical Findings: Collapse Magnitude and Modality Dependence

The primary empirical result is the observation that GPT-4o, acting as an external cross-model evaluator, produces a PCI of 1.464—over threefold compared to self-evaluation (PCI=0.461) and twice the random baseline (PCI=0.716). Strategy absorption is extreme: step_by_step accounts for 48.4% of all agent weight, and visual-domain strategies obtain only 9.1% combined weight. Visual tasks exhibit heightened collapse relative to text (PCI=1.464 vs. 1.348), confirming modality-specific amplification. Ground-truth PCI calculated for direct task correctness is substantially lower (0.251), providing a quantitative lower bound on unbiased, merit-based concentration.

Cross-Modal Contagion Phenomenon

A novel experimental isolation training procedure uncovers asymmetric cross-modal contagion: strategy preferences acquired in one modality reliably transfer and contaminate behavior in another. The contagion coefficient $\gamma_{A\to B}$, operationalized via normalized L2 distance, captures this effect. GPT-4o evaluations yield $\gamma_{V\to T}=0.847$ and $\gamma_{T\to V}=0.832$, documenting asymmetric transfer and strategy inversion—where text-trained synthesis becomes dominant for visual tasks after cross-modal exposure and vice versa. Robust cross-model statistical validation (Qwen-plus, DashScope) reveals that contagion direction and magnitude are strongly conditional on evaluator identity, with cross-model evaluators displaying strong bidirectional contagion ($\bar{\gamma} \sim 1.0$–$1.2$) and self-evaluation exhibiting near-complete immunity (97% zero-contagion rate).

The Contagion Matrix Framework

The paper introduces the contagion matrix $\Gamma^{(\mathcal{J})}$, indexed by evaluator identity, as a formal construct for capturing the multi-directional bias transfer across $n$ modalities. This matrix quantifies not only direct A→B contagion but also supports potential analysis of transitive interactions as future work extends to higher-order multimodal settings (e.g., including audio).

Implications for Agent Adaptation and Strategy Optimization

Closed-loop adaptation frameworks relying on LLM evaluators are fundamentally vulnerable to evaluator preference drift. The findings reveal that agents may silently converge to strategies optimized for evaluator bias rather than genuine task-relevant competence. Visual strategy blindness is especially problematic: the agent underuses visual-specific strategies because the evaluator (GPT-4o) consistently defaults to step_by_step reasoning, regardless of modality.

Training order is found to impact final strategy distributions for all modalities, engendering path dependence. This implies the necessity for standardized, modality-isolated evaluation protocols and multi-evaluator ensembles to prevent preference contamination.

Statistical Interpretation and Evaluator Dependency

Hierarchical analysis across evaluator types demonstrates that cross-model evaluation is the primary risk factor for cross-modal contagion. Excessive training rounds (e.g., 50+) with certain evaluators (DashScope) induce degenerate single-strategy collapse, closing the preference transfer channel. Statistical significance analyses reveal contagion is symmetrical and evaluator-conditional, with negligible directional asymmetry. The paper emphasizes the criticality of reporting both PCI and $\Gamma$ indexed by evaluator identity in production agent systems.

Practical Recommendations and Broader Impact

The paper advocates for diagnostic deployment of the MM-EPC framework, automated reporting of PCI and contagion metrics, evaluator ensemble strategy for mitigation, and round count regulation to prevent collapse. Silent convergence to evaluator-optimal (but task-suboptimal) strategies poses systemic risk, particularly in safety-critical applications. Diversity and architectural variation in evaluators are essential to maintain robustness and prevent monocultures of bias transfer. Standardized benchmarking suites for preference diversity and collapse dynamics are recommended.

Limitations and Directions for Future Research

Experimental limitations include modality coverage (text and visual only), strategy set imbalance favoring text, absence of native image inputs, and single evaluator type dependence. Extension to audio, balanced strategy sets, real visual inputs with vision APIs, adaptive strategy selection, and multi-evaluator calibration studies are delineated as vital avenues for rigorous future work. Construct, internal, external, and statistical validity threats are candidly discussed.

Conclusion

This research rigorously demonstrates that multimodal agent systems using closed-loop LLM evaluation are susceptible to preference collapse and cross-modal bias transfer, with magnitude and structure strictly dependent on evaluator identity and architecture. Cross-model evaluation amplifies strategy collapse and contagion, while same-model self-evaluation provides substantial immunity. These findings establish the MM-EPC framework and the contagion matrix $\Gamma^{(\mathcal{J})}$ as foundational tools for ongoing benchmarking and system design. Reporting evaluator-specific PCI and $\Gamma$ is recommended for all deployed closed-loop agent systems.

The implications highlight the necessity of cautious evaluator selection, explicit multimodal phase isolation, and systematic bias quantification to prevent silent convergence to evaluator-biased strategies. Future developments in AI agent adaptation must consider these phenomena to ensure robustness across modalities and evaluator architectures.