acceptodds
Under review as a conference paper at ICLR 2027

Inferred Generative-Process Diversity Predicts Correlated Failure Across Language Models

Abstract

As language models are increasingly deployed in multi-model systems, robust system function depends on constituent models failing independently. Diversity is widely observed to buffer complex adaptive systems against correlated failure, but only when the type of diversity is relevant to the failure mode of concern. Existing assessments of model diversity often focus on semantic differences between outputs. We instead consider inferred generative-process diversity, the difference between inferred processes capable of generating observed outputs. We estimate this using Normalised Compression Distance between model output byte strings residualised against a permutation control. Estimated from responses of 38 language models to open-ended chat prompts, the measure predicts which model pairs choose the same wrong answer on ten disjoint benchmark families, beyond semantic similarity and capability, with a negative partial association on every benchmark (mean , 95% CI ). Semantic distance predicts no reduction in correlated failure once process diversity is held fixed. Five-model ensembles selected to maximise the Vendi score of the measure at fixed semantic distance converge on the same wrong answer less often than capability-matched random ensembles in every semantic band, whereas the reverse selection does not. Because it requires only observed outputs, inferred generative-process diversity provides a practical black-box approach for auditing and constructing model populations where correlated failure creates system-level risk.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, multi-agent systems, algorithmic information theory, normalised compression distance, socio-ecological resilience

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ny31OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes Inferred Generative-Process diversity (IGP), which is measured from open-ended model responses using compression distances and a byte-permutation control. It tests whether this measure relates to shared wrong answers on separate benchmarks. Experimental results show an association beyond the tested semantic baseline and model capability.

Critical strengths
  1. Understanding when models make the same mistakes is important because shared errors undermine systems that rely on multiple models. The paper studies a problem relevant to improving reliability.

  2. The paper measures diversity and shared errors on separate response datasets, which helps test whether the relationship holds across tasks. Controlling for capability and semantic distance also helps assess whether IGP provides additional information.

Critical weaknesses

Majors:

  1. The task-transfer design shows that IGP measured from open-ended responses is related to shared errors on separate question-answering tasks (p.2, L75-081, p.7, L351-362). However, this does not yet demonstrate that selecting a team with high IGP makes a system more reliable, the reported ensemble comparison measures chance-corrected convergent failure (p.5, L246-256, L264-269; p.8, Fig.4). Testing a fixed team with a clear decision rule (such as voting or abstention) on new tasks would strengthen this claim.

  2. The evaluation metric changes between the two-model and five-model settings. Pairwise CWA conditions on cases where at least one model is wrong (p.4, L201-210, Eq.3), whereas five-model CFR conditions on cases where all models are wrong (p.5, L246-256, Eq.5). This difference makes the two settings less directly comparable. Need clarification why different conditioning events are needed and test whether the conclusions hold when both settings use the same conditioning rule.

  3. The reported effects are modest, with a mean partial correlation of −0.216−0.216 after controlling for semantic distance and capability (p.7, L351-355). In the five-model experiment with semantic distance held fixed, only two of the five bands have 95% confidence intervals that exclude zero in the cross-benchmark summary (p.8, L401-414, Fig.4b), encouraging but the size and consistency of the benefit need stronger support.

Minors:

  1. Notation like pp, VpCV_p^C (p.3, L123-127), and the Vendi score (p.5, L260-264), needs more explanation when first introduced in the main text. There could be more cases like this in the paper.

  2. The English writing needs improvement, as parts of the paper are difficult to follow.

Questions for authors
  1. What is the evidence that supports the step from lower shared-error metrics to better performance under a team decision rule (p.5, L246-269, p.8, Fig.4)? A fixed-team test with voting or abstention would be one useful way to examine this.
  2. Why does the conditioning event change between pairwise CWA and five-model CFR, do the findings hold when both settings use the same conditioning rule?
Reviewer_K6K6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether differences in LLM output patterns can predict their tendency to produce the same wrong answers. It proposes "inferred generative-process diversity," measured by how well pairs of responses compress together after removing the component explained by randomly shuffling their bytes -- a control that preserves byte frequencies while destroying sequence order. Using responses from 38 models on 100 open-ended prompts, the paper find that this measure predicts less shared wrong-answer agreement across ten separate benchmark families, even after controlling for model accuracy and semantic similarity between outputs.

Critical strengths

Useful practical question: the paper is motivated by which kinds of model diversity actually predict shared mistakes, making diversity relevant to ensemble design.

Meaningful empirical support: In this paper, diversity is measured on separate prompts from the failure benchmarks, and the association has the same direction across all ten benchmarks.

Critical weaknesses

Practical benefit is narrower than overall reliability: The ensemble analysis measures agreement conditoning on every member being wrong. It does not establish improved ensenble accuracy or safety.

Robustness remains incomplete: Results depend on the statistical controls, use one semantic embedding baseline, and do not repeat the main failure analysis with alternative compressors. (Appendix B.1 and E.1)

Questions for authors
  1. Eq. 3 pools observed error counts across questions, while Eq 4 and 23 construct the chance baseline using aggregate error rates and an unweighted mean of item level collision probabilities. So would the main findings in Figure 3 remain similar if the observed and expected agreement used the same item weighting?

  2. Although diversity and failure are measured on separate prompts, were choices such as: permutation residualisation, compressor selection, and capability controls fixed before looking into benchmark associations?

Reviewer_VfcSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce a black-box method for predicting correlated failures between language models, with the goal of constructing robust model ensembles with lower failure rates. The authors propose a metric for “inferred generative-process diversity” which uses normalised compression distance. The proposed metric outperforms an embedding-based similarity metric at predicting correlated failures. Ensembles constructed using this compression metric converge on the incorrect answer (all models agree), less frequently than ensembles constructed randomly or using embedding diversity.

Critical strengths

The motivation of the paper is well justified. Models increasingly operate in multi-agent environments, and recent events (e.g. the Hugging Face incident) highlight the importance of improving the robustness of these systems.

The authors propose a novel metric for predicting model failures, and show that the method can be used to construct ensembles with less frequent convergent failures

Critical weaknesses

Baselines: The authors only compare the proposed method to a single embedding-based method for predicting correlated failures. The authors should include additional results from alternative embedding models. The authors should also consider including a comparison to one of the methods introduced in omitted related work (See 2).

The authors do not engage with closely related work [1],[2],[3]. [1] proposes an embedding clustering approach for building ensembles. [2] predicts model performance using output distances, and [3] also proposes building ensembles by pruning based on diversity.

Motivation versus scope: While the authors ambitiously motivate their work with recent issues with multi-agent systems, their results only show the benefits of their approach for reducing correlated errors on benchmarks. It is very unclear if improving “inferred generative-process diversity” of model behaviors in multi-turn, long-horizon settings with tool calls reduces the risk of dangerous ensemble behaviors.

Presentation: The language of the paper is difficult to follow. Language like “in terms of how models generate observed outputs rather than what those outputs mean,” is highly ambiguous, and makes it harder to grasp what could otherwise be a much clearer metric and contribution.

—

Cohen et al. (2026). DFPE: A Diverse Fingerprint Ensemble for Enhancing LLM Performance. Findings of EACL.

Yax, Oudeyer, and Palminteri (2025). PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks. ICLR.

Tekin et al. (2024). LLM-TOPLA: Efficient LLM Ensemble by Maximising Diversity. Findings of EMNLP.

Questions for authors

Why do you think “inferred generative-process diversity” of model behaviors in multi-turn, long-horizon settings with tool calls reduces the risk of dangerous ensemble behaviors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.