acceptodds
Under review as a conference paper at ICLR 2027

Topology and Shared Error Exposure in Large Language Model Lineages

Abstract

As the population of large language models (LLMs) rapidly grows, understanding how shared training sources shape errors across the LLM ecosystem and how developers' training choices create these exposures becomes an important system level question. We develop a lineage representation that maps directed training relationships into error loadings and decomposes model error covariance into three components arising from common tasks, shared base families, and idiosyncratic errors propagated through the lineage. This decomposition shows when concentrated source exposure sustains aggregate error variance as the model population expands. At fixed individual error distributions, different patterns of shared exposure can yield different aggregate errors. We illustrate the mechanism through three canonical lineage topologies. To examine how individual choices shape concentration, we study a one-generation teacher-choice model and derive an explicit condition under which every developer prefers the same teacher, while a diversified teacher assignment reduces aggregate error.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Lineage Topology, Algorithmic Monoculture, Error Propagation, Error Covariance, Endogenous Teacher Choice, Ecosystem Risk, Knowledge Distillation, Model Lineages

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_A5ihOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how shared training sources creates dependence among model errors. It represents model relationships as a DAG and decomposes error covariance into task, family, and propagated model-specific components. The decomposition allows characterizing when population growth diversifies propagated errors. Authors also study the difference between the individual objective and the population objective. Experiments are conducted on synthetic regression with ridge learners and neural networks to illustrate these effects.

Critical strengths

The framework gives an interpretable explanation of why adding more models don't necessarily diversify their errors. The root, star, and chain examples make the role of shared ancestry clear. The teacher-choice result clarifies the trade-off between individual accuracy and shared errors.

Critical weaknesses

The LLM problem was used as a motivation, but the authors conducted no actual experiments on LLMs or actual LLM lineages. I agree that synthetic regression experiments in sec 5 do provide useful illustrations, and I agree the theory paper doesn't need a complete LLM benchmark, but it would be beneficial to demonstrate how the mathematical model applies to e.g. a specific LLM training setting.

Questions for authors
  1. Can the authors give one concrete LLM application example and explain what the teacher and student errors represent, why the averaged-prediction loss is useful in that context, and how to interpret the error-transmission assumption there?
  2. Which conclusions are from existing network-aggregation results [1,2,3]? What new insight is from the lineage and teacher-choice formulation? A more explicit comparison would help assess the contributions.

References

[1] Seifert, Bastian, Chris Wendler, and Markus Püschel. "Causal Fourier analysis on directed acyclic graphs and posets." IEEE Transactions on Signal Processing 71 (2023): 3805-3820.

[2] Acemoglu, Daron, et al. "The network origins of aggregate fluctuations." Econometrica 80.5 (2012): 1977-2016.

[3] Golub, Benjamin, and Matthew O. Jackson. "Naive learning in social networks and the wisdom of crowds." American Economic Journal: Microeconomics 2.1 (2010): 112-149.

1 comment

Thank you for your thoughtful questions. We clarify the applications and the relationship to the three references below.

Q1. Concrete applications and the interpretation of our framework

We consider two applications. First, in LLM distillation, student models learn to evaluate answers using scores provided by a teacher. Teacher and student errors are deviations from a common reference score, such as a human evaluation target. When several student scores are averaged to assess an answer, averaged-prediction MSE directly measures the accuracy of that assessment. If students inherit the same scoring bias, adding more scoring models only provide little improvement. In this setting, the error-transmission represents a student's centered error as a weighted combination of inherited teacher errors and a local error component. The transmission coefficients describe how strongly teacher errors carry into student predictions. In appendix B.1, we derive this relationship from a squared-loss training objective combining local supervision and teacher predictions.

Second, consider several banks choosing the same well-documented teacher LLM to reduce development costs and support model validation and compliance reviews. Each bank uses the teacher's risk assessments to train its own student model to estimate corporate default probabilities. Teacher and student prediction errors are the differences between predicted probabilities and subsequent default outcomes. While learning a teacher's analytical capabilities, students can also inherit its tendency to underestimate risk for certain firms. Sharing a high-quality teacher can therefore improve each bank's model accuracy while making the banks more likely to misjudge the same risks. The MSE of the average forecast measures whether these errors cancel across banks or remain in the collective assessment.

This has practical financial significance. The Bank of England[1] warns that common weaknesses in widely used models could cause multiple institutions to misestimate risks simultaneously, leading to credit mispricing and misallocation. This motivates our study of how shared errors accumulate through training relationships and how individually reasonable teacher choices can create population-level risks.

Q2. Relationship to existing results and the new insight

[2] provides a source-and-path representation on DAGs, which is one mathematical foundation of our framework. [3] studies how production networks transmit shocks and generate aggregate fluctuations. [4] characterizes how the distribution of influence in social learning determines whether individual noise disappears as the population grows. Our path expansion and error-diversification conditions build on these principles. We further explain how the training topology emerges endogenously from developers' teacher choices. When developers choose teachers to improve their own models' accuracy, they may converge on the same teacher or on teachers with shared ancestors. These individually optimal choices can concentrate inherited errors and worsen population performance. Our contribution therefore links how training relationships form to how errors spread through them and whether those errors diminish as the population grows. For example, theorem 7 establishes exact conditions in a finite population under which every developer strictly prefers the same teacher, yet every population-optimal assignment uses both teachers. This moves beyond studying error aggregation on a given network.

We also connect teacher selection to the origins of teacher-error overlap. Rather than taking this overlap as given, we express it through shared ancestors and transmission paths. This reveals when choosing different teachers actually diversifies the underlying error sources and when common ancestry preserves shared errors despite population growth. Appendix B.5 to B.6 further extends the teacher-choice analysis to multiple teachers, independent training, and compute costs.

For completeness, our General Response also reports additional experiments on transmission accuracy, population mechanisms, and larger teacher–student populations.

Reference

[1]. Bank of England. Financial Stability in Focus: Artificial Intelligence in the Financial System. April 2025.

[2] Seifert, Bastian, Chris Wendler, and Markus Püschel. Causal Fourier analysis on directed acyclic graphs and posets. IEEE Transactions on Signal Processing 71 (2023): 3805-3820.

[3] Acemoglu, Daron, et al. The network origins of aggregate fluctuations. Econometrica 80.5 (2012): 1977-2016.

[4] Golub, Benjamin, and Matthew O. Jackson. Naive learning in social networks and the wisdom of crowds. American Economic Journal: Microeconomics 2.1 (2010): 112-149.

Reviewer_dVvnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines how lineage topology and shared training sources affect aggregate error in model populations. The authors show that increasing the number of models does not necessarily reduce the mean squared error of their average prediction, because the models may inherit errors from common sources. Through mathematical analysis, they characterize when propagated model-specific error variance persists and establish conditions under which diversified teacher assignments achieve lower aggregate MSE than having all students choose the same individually optimal teacher. Synthetic regression experiments with ridge and neural learners illustrate these mechanisms.

Critical strengths

This paper uses a mathematical model to show that errors inherited from a common teacher can persist under aggregation as the model population grows. This helps clarify how shared training sources affect aggregate error in model populations. Besides, it provides explicit conditions under which assigning some students to an individually suboptimal teacher with different inherited error patterns can reduce aggregate mean squared error. This clarifies the trade-off between individual accuracy and diversity in teacher selection.

Critical weaknesses
  1. The validation is limited, and the reliability of the propagation approximation remains unclear. All experiments use synthetic regression tasks with ridge and small neural learners. The mean relative RMSE for perturbation-response prediction reaches 0.714 on the nonlinear task, raising questions about whether the conclusions remain reliable under such large deviations from the affine error recursion. The teacher-choice experiment is also restricted to two teachers and four students. These results illustrate the mechanisms but leave their robustness and relevance to real LLM populations unresolved.
  2. The practical implications are not sufficiently established. The analysis focuses on the MSE of averaged model predictions, whereas LLMs often serve users independently or are combined through voting or routing. The paper does not clearly explain how its conclusions extend to shared failures in these settings. The absence of comparisons with existing teacher-selection or diversity-aware methods further limits the evidence for practical benefits.
Questions for authors

Q1: On the nonlinear task, the relative RMSE for perturbation-response prediction reaches 0.714. With an approximation error this large, can the theory still predict which lineage topology has lower aggregate MSE and whether diversified teacher assignments are beneficial? Q2: The paper mainly analyzes averaged model predictions. Which conclusions still hold when models serve different users independently or are used through voting or routing? Q3: Have the authors conducted experiments with real language models or with larger numbers of teachers and students?

3 comments

Thank you for your questions. We sincerely appreciate your engagement with our work and we respond to each below.

Q1. Accuracy of the propagation approximation and reliability of the population mechanisms

In our General Response, we report the additional propagation, topology, and teacher-choice experiments, showing that the core population mechanisms hold under both the original setting and settings with substantially improved response prediction.

Q2. Independent deployment, voting, and routing

The lineage framework describes joint errors across models. The deployment rule determines how these errors affect performance. For independent deployment, models with shared training sources can still make similar mistakes on common evaluation tasks, even when serving different users. The MSE of the average prediction contains both individual error terms and cross-model error terms. Holding individual MSEs fixed, changes in aggregate MSE directly reflect changes in the cross-model terms. Aggregate MSE therefore measures whether errors cancel across models or remain shared across the population.

For voting, shared errors matter because models may vote for the same wrong answer. But the voting error rate also depends on the voting rule, so translating our results to voting requires additional specifications.

For routing, let's consider a system that selects one model with fixed probabilities, independently of the input. Then its expected loss is the weighted average of the individual model losses. Training lineage still matters because it shapes the errors inherited by each selected model.

Q3. Real models and larger experiments

Existing studies already provide empirical evidence of shared LLM errors and their downstream effects[1], and infer model relationships from output similarity[2]. These inferred relationships themselves require validation, especially when training histories are undocumented. This makes controlled experiments particularly valuable. By specifying the training relationships and changing them while holding other conditions fixed, so we can directly test how lineage generates shared errors and affects population performance.

Our controlled experiments isolate this mechanism by changing training topology or teacher assignment while holding other experimental conditions fixed. Since comparisons between existing LLMs combine differences in training data, architecture, optimization, and shared origins, therefore, controlled experiments separate the effect of training relationships from these other differences and directly test its consequences for population error. We also expand the teacher-choice experiment from two teachers and four students to three teachers and eight students, as reported in our General Response.

References

[1]. Elliot Kim, Avi Garg, Kenny Peng, and Nikhil Garg. Correlated Errors in Large Language Models. ICML, 2025.

[2]. Nicolas Yax, Pierre-Yves Oudeyer, and Stefano Palminteri. PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting Their Performances in Benchmarks. ICLR, 2025.

Thank you for the additional validation and clarifying. Q1: I still have reservations about validation on real LLMs, the connection between aggregate MSE and practical applications, and the contribution beyond prior work. Q2: I think the connection between aggregate MSE and performance under independent deployment, voting, or routing is still unclear. Q3: The validation remains limited to small synthetic regression settings, and the cited LLM studies do not directly validate the proposed framework. Its relevance to real LLM lineages therefore remains an open concern for me.

Thank you for clarifying your concerns. We add a teacher choice experiment on BoolQ and evaluate the same students under prediction averaging, independent deployment, voting, and random routing. The code is in our anonymous repository.

Q1. When is aggregate MSE useful, and what does the paper add to prior work?

On BoolQ, each model estimates the probability that the answer is yes. If a system averages these estimates, aggregate MSE directly measures the mean squared error of its final predictions against the observed answers. Prior work assumes an exogenous network, but we ask how the training network forms endogenously. Theorem 7 gives conditions under which every developer prefers the same teacher, yet assigning some students to a different teacher lowers aggregate MSE. Our BoolQ experiment illustrates this tradeoff with OLMo and Qwen.

Q2. What changes under independent deployment, voting, or routing?

We fix two assignments before evaluating them on 1,600 new BoolQ questions. In the first, all four students learn from OLMo. In the second, one designated student learns from Qwen.

Prediction averaging: The mixed assignment reduces the MSE of the four students’ average prediction by 0.70%.

Independent deployment: When models serve users separately, a practical question is how often a user gets the wrong answer regardless of which model they use. Under the mixed assignment, the number of questions that all four students answer incorrectly falls from 460 to 404, i.e. it's from 28.75% to 25.25%.

Majority voting: When all four students give the wrong answer, one student can switch to the correct answer without changing the majority vote. Voting accuracy rises from 66.03% to 66.34%.

Random routing: If the system selects one student uniformly at random, independently of the question, its expected MSE is the average of the four individual MSEs. Routing MSE is nearly the same under both assignments as we expect.

Q3. What evidence comes directly from real LLMs?

We extend the two-teacher, four-student experiment in Table 5 to BoolQ using OLMo and Qwen as the teachers. On the new questions, all four students have lower individual MSE when they learn from OLMo. Yet assigning the designated student to Qwen increases that student's own MSE while reducing the MSE of the four students' average prediction by 0.70%.

Reviewer_W3VfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper uses a weighted DAG to model how students inherit their teachers' errors. Sharing a teacher can make students share its errors, limiting the benefit of averaging their predictions. The authors show that having every student choose the best teacher can give worse ensemble performance than using different teachers. Synthetic experiments with ridge regression and small neural networks illustrate this tradeoff.

Critical strengths
  • The idea of relating training lineages to shared errors is interesting. The distinction between individual teacher quality and population performance provides a useful way to reason about teacher concentration.
  • The equations and derivations I checked appear correct.
  • The controlled experiments clearly illustrate the divergence between individual and population objectives. The authors report results across eight seeds with confidence intervals.
Critical weaknesses
  • The connection to actual neural distillation is a bit weak. The central error recursion assumes affine transmission along teacher-student relationships. This holds exactly in the constructed ridge-regression setting, but the approximation is already quite inaccurate for the small neural learners. At 600 updates, the relative response RMSE is 0.42 on linear tasks and 0.714 on nonlinear tasks (Appendix C.2). The qualitative diversification results are consistent with the theory, but these substantial discrepancies provide limited support for the propagation model itself. It is unclear how well the assumed transmission mechanism captures actual LLM distillation.

  • The theoretical machinery is heavy relative to the additional insight it establishes. The conflict between individual teacher quality and aggregate error can already be understood through a simple bias-variance-covariance tradeoff, as illustrated by the authors' own two-teacher example in Appendix B.5. The lineage analysis gives structural conditions for persistent aggregate error, but it is unclear how much additional predictive understanding this provides about distillation, particularly given the weak neural approximation. A more focused model capturing the central tradeoff, with stronger validation of its assumptions, could make the contribution more convincing.

The idea is interesting, but the weak validation of the error propagation model and unclear value of its full generality limit the contribution. I therefore recommend a weak reject.

Questions for authors

I understood the authors' intent and the manuscript. I do not have specific questions that need to be addressed.

1 comment

Thank you for checking the correctness of the derivations. We appreciate your careful evaluation.

  1. Connection between distillation and error transmission

We address the neural approximation errors in our General Response, where the additional experiments show that our core mechanisms hold under both the original and improved training settings.

The connection between distillation training and error transmission is developed in the appendix to keep the main text concise. In appendix B.1, we start from a squared-loss objective combining local supervision with teacher predictions and derives the transmission strength and local error exactly. Furthermore, in appendix B.3, we examine how increasing the distillation weight strengthens the inheritance of teacher errors while reducing the student's own local error. It shows that accounting for this local-error scaling preserves the conclusions that stars retain shared errors and chains progressively diversify propagated errors.

  1. For the second point, the two-teacher example can show choosing the individually more accurate teacher worsen population performance. But only the full lineage framework explains how the overlap between teachers' errors arises and whether their shared exposure disappears as the population grows. This especially matters for practical teacher selection: in financial applications, banks using different models can still inherit errors from the same upstream teachers, leading them to misjudge the same risks and amplify the financial crisis.

Furthermore, a standard bias-variance-covariance comparison takes the teachers' joint error term QpqQ_{pq} as given. Our framework relates it to their exposure to common ancestors: Qpq=b(p)b(q)+a2∑jΦpjΦqjsj2.Q_{pq}=b(p)b(q)+a^2\sum_j\Phi_{pj}\Phi_{qj}s_j^2. Here, ΦpjΦqj\Phi_{pj}\Phi_{qj} measures the two teachers' shared exposure to ancestor jj, and each lineage loading collects transmission along the paths from that ancestor. This identifies which ancestors contribute to the overlap and how their errors reach both teachers. It therefore explains whether choosing different teachers actually diversifies the underlying error sources.

Our framework also follows successive generations in which teachers are themselves trained from earlier models. It characterizes whether adding models disperses exposure to individual error sources or leaves enough concentration for aggregate error to persist. Holding local error sources fixed, it further derives how changing training relationships changes shared errors. Therefore, only the full framework can explain how training relationships shape both the benefits of population growth and the choice of teachers.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response for supplementary validation and shared concerns

Our first two points address W3Vf's and dVvn's concerns about the response RMSE of 0.714. And the third point addresses dVvn's question for experiments with more teachers and students.

First, 0.714 is the highest mean response error among our neural training comparisons at 600 updates. Standard improvements to neural network training can substantially reduce this value. To isolate the optimizer's effect, we keep the original network architecture and replace only Adam with standard gradient descent. This reduces RMSE on the nonlinear task from 0.714 to 0.177, a 75.2% reduction, with improvement in all eight seeds. Even keeping Adam, using a wider network with parameter regularization reduces the nonlinear task error to 0.160, a 77.6% reduction. On the linear task, the error also falls from 0.420 to 0.088, a 79.1% reduction. We originally reported this value in the appendix because our paper's core mechanisms hold even in this worst performing baseline setting with a response error of 0.714. These mechanisms continue to appear after training is improved, which is our next second point.

Second, we highlight our study focuses on differences in population error caused by topology and the divergence between individual and population objectives. On the nonlinear task with the original network, aggregate MSE averaged over eight seeds satisfies MSEroots<MSEchain<MSEstarMSE_{roots} < MSE_{chain} < MSE_{star} This shows systematic differences in population error across training topologies. Compared with independent roots, the star topology, where students share the same teacher, reduces average individual MSE by 25.8% but increases aggregate MSE by 50.3%. Individual error falls while aggregate error rises in all eight seeds.

Furthermore, we test the same mechanisms with the regularized network whose relative response RMSE falls to 0.160. On the nonlinear task, sharing a teacher reduces average individual MSE by 30.8% but increases aggregate MSE by 52.4%. Thus, both the differences across topologies and the individual-population tradeoff persist, from the original setting with larger response error to the setting with substantially improved transmission prediction.

  1. In response to the experimental scale, the divergence between individual and population-optimal teacher choices also holds at a larger scale, and the gains from population-based assignment generally increase. We expand the experiment from two teachers and four students to three teachers and eight students. Result shows same mechanism persists at the larger scale: each student individually prefers the same teacher, while the population-optimal assignment introduces other teachers and achieves lower test error. Compared with the original scale, the test gains from population-based assignment are generally larger, and positive gains extend to a wider range of teacher quality gaps.

In summary, the original worst 0.714 baseline can still exhibits our core mechanisms. Standard training adjustments substantially improve transmission prediction while preserving these mechanisms. The larger experiment further supports the divergence between individual and population-optimal teacher choices. These additional results are reported in the added Supplementary validation subsection of the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.