A foundational assumption in machine learning system design is the "wisdom of the crowd." If a single model is prone to occasional errors, combining multiple models through an ensemble, a majority vote, or a peer-review pipeline should theoretically filter out those individual anomalies. For this statistical insurance policy to work, however, the errors must be independent. If ten people independently guess the weight of an ox, their random deviations tend to cancel each other out. But if all ten people are looking at a broken scale that adds fifty pounds, their collective guess will simply codify that systematic error with high confidence.
In modern deep learning, we are increasingly encountering the machine learning equivalent of the broken scale: correlated model failure. Because models frequently share training datasets, architectural paradigms, and inductive biases, they do not fail randomly. Instead, they converge on the exact same incorrect predictions, creating a false consensus that breaks our most reliable aggregation and validation frameworks.
Why Models Fail Together
To understand why ensembles suffer from correlated failures, we must look at where their knowledge originates. Ensembling different model families (such as combining a transformer with a convolutional network) is often used to capture epistemic uncertainty. However, as shown by Kimia Hamidieh et al. in Complementing Self-Consistency with Cross-Model Disagreement for Uncertainty Quantification, this approach fails when models are exposed to the same gaps or errors in a common web-scale dataset. When the underlying data distribution contains a systematic omission or bias, every model trained on it inherits the same blind spot.
This shared foundation leads to what researchers call "herding behavior." For instance, a study evaluating frontier large language models on real-time forecasting, WorldCup Arena: Prospective, Leakage-Free Evaluation of Frontier LLMs on a Live Tournament, revealed that models frequently exhibit a shared bias toward conservative, prototypical outcomes (such as tracking market favorites and under-predicting draws). Because the models share similar inductive biases, they agree with each other far more often than they agree with reality.
The Mathematical Paradox of Majority Voting
The breakdown of independence fundamentally alters the scaling properties of compound AI systems. In an ideal system with independent errors, increasing the number of model calls or votes monotonically increases the accuracy of the aggregate output.
However, Lingjiao Chen et al. demonstrated in Are More LLM Calls All You Need? Towards the Scaling Properties of Compound AI Systems that when a model is systematically biased toward a specific incorrect answer, adding more samples to a majority vote can paradoxically decrease overall performance.
Independent Errors (Ideal) Correlated Errors (Reality)
Model 1: ❌ (Error A) Model 1: ❌ (Error A)
Model 2: ❌ (Error B) Model 2: ❌ (Error A)
Model 3: (Correct) Model 3: (Correct)
-------------------- --------------------
Result: Correct Wins Result: ❌ Error A Wins
When a model's reasoning is fundamentally skewed along a specific incorrect path, repeated sampling or voting merely reinforces the error. As noted by Zhenyu Wu et al. in Get an A in Math: Progressive Rectification Prompting, aggregation techniques fail to improve accuracy on complex tasks because models consistently produce the same systematic logic errors across multiple independent attempts. The most frequent output ceases to be a proxy for correctness; instead, it becomes a proxy for the strength of the shared bias.
The Blind Spot in Uncertainty Quantification
This false consensus poses a severe threat to safety-critical systems that rely on ensembles to measure uncertainty. Standard ensemble-based uncertainty quantification operates on a simple principle: if the constituent models disagree on an input, the system is uncertain and should trigger a safety mechanism or hand off control to a human.
But when models suffer from correlated failures, this signal disappears. Liting Sun et al. explored this in On complementing end-to-end human behavior predictors with planning, demonstrating that deep learning ensembles fail to detect out-of-distribution (OOD) scenarios when individual models share similar inductive biases. Because the models converge on the exact same incorrect hypotheses, the ensemble variance drops to zero.
The system becomes confidently wrong. As described by Dimity Miller et al. in Density-aware NeRF Ensembles: Quantifying Predictive Uncertainty in Neural Radiance Fields, this lack of variance creates a dangerous blind spot where an ensemble reports absolute certainty for what is effectively a total guess, bypassing critical safety boundaries.
Peer-Review Echo Chambers
The vulnerability extends beyond simple voting to multi-agent architectures, where one model generates content and another acts as a critic or verifier.
In Automated Consistency Analysis of LLMs, Aditya Patwardhan et al. documented an "agreement bias" in cross-validation setups. When peer models are used to validate one another on open-ended or situational tasks, they frequently confirm each other's flawed reasoning. This creates an echo chamber that invalidates the cross-validation process.
This false consensus also undermines automated testing frameworks. For example, the whitebox testing framework introduced by Kexin Pei et al. in DeepXplore: Automated Whitebox Testing of Deep Learning Systems relies on differential testing—comparing multiple models to find discrepancies that indicate bugs. The authors explicitly noted that when multiple models in the testing ensemble share the same underlying bias or training data flaws, differential testing fails entirely. The models reach a false consensus, masking systematic failures that affect the entire group.
Breaking the Consensus Bias
Recognizing that pure model ensembles often amplify rather than mitigate bias, recent work has shifted toward injecting true cognitive diversity into the loop.
One promising avenue is the creation of hybrid systems. Axel Abels et al. demonstrated in Wisdom from Diversity: Bias Mitigation Through Hybrid Human-LLM Crowds that simple averaging of model outputs inherently causes bias amplification because the models lack an independent error distribution. By forming hybrid crowds—combining the individual execution speed of LLMs with the distinct cognitive diversity of human annotators—the authors showed it is possible to break the correlated error patterns that plague pure machine ensembles.
Ultimately, solving the problem of correlated model failure requires moving past the assumption that more compute or more models automatically yield truth. Until we explicitly account for shared data lineages and architectural herding, our systems will remain vulnerable to agreeing confidently on the wrong answers.
Key Foundational Work
Linear and Order Statistics Combiners for Pattern Classification (1999) — contributed 2 concept elements to this direction.