- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper studies how much of an LLM's epistemic uncertainty can be captured by deep ensembles as models scale. Treating epistemic uncertainty as the model's reducible loss, the paper splits it into a part that is visible through disagreement among ensemble members and an invisible part consisting of errors shared by all members, including model misspecification.
The main claim is that ensemble disagreement shrinks only slowly with model size and is largely insensitive to more training data or longer context, whereas the invisible part shrinks faster. As a result, ensembles reveal a growing share of epistemic uncertainty at larger scales, and an ensemble of smaller models can become competitive with a single larger model of the same total size. The paper also shows that ensembles whose members share part of their training history (late branching, LoRA) lose much of their diversity in a predictable way.
These claims are supported empirically by fitting scaling laws to ensembles built from public model families and to a newly trained family of byte-level models, with extrapolations to larger scales that depend on an assumed irreducible loss.
Critical strengths
-
Scale and care of the empirical study. Multi-seed pretraining experiments are expensive and rarely available, and the paper assembles a broad set of ensemble comparisons across public model families together with a newly trained model family that allows controlled measurements (e.g., separating the effects of model size, training data, and context length) that public checkpoints do not support. The analyses are accompanied by extensive sensitivity checks and a transparent discussion of limitations. This matters to my assessment because, even where the interpretation is debatable, the measurements themselves are a useful reference for the community.
-
The late-branching and LoRA results. The finding that ensembles sharing training history lose much of their diversity, in a predictable way, is measured in a controlled setting and does not depend on the unknown irreducible loss. It is also directly relevant to practice, since cheap ensembles such as LoRA ensembles and trajectory-based Bayesian approximations are commonly used for LLM uncertainty. I consider this the most robust and actionable contribution of the paper.
Critical weaknesses
-
The central conclusion depends on an unknown quantity. The growing visible share is computed relative to the irreducible loss, which is unknown and approximated by the best available model's loss. The trend weakens as this value is lowered and reverses below a threshold that the paper cannot rule out, so the headline result rests on an unverifiable assumption.
-
Large-scale evidence comes mostly from pairs of different models. True seed ensembles reach only about 1B parameters; beyond that, the "ensembles" pair different models (e.g., Qwen2.5 and Qwen3), whose disagreement reflects differences in data and training recipes rather than epistemic uncertainty. Yet these pairs drive key results, such as the highest projected visible share and the main example of an ensemble beating a larger model with fewer parameters, which are then extrapolated far beyond the measured range.
-
Mismatch between motivation and results. The introduction motivates the work with applications that need uncertainty at individual inputs, such as active learning and deferral. However, the paper's notion of epistemic uncertainty is defined only on average over contexts, and the main results concern averaged quantities. Pointwise uncertainty is addressed only briefly, using a reference-model proxy with modest predictive accuracy, so it is unclear what the scaling results imply for the applications that motivate the paper.
-
Limited novelty. The loss decomposition, the decrease of ensemble disagreement with scale, uncertainty scaling laws, and ensembles outperforming larger models have all been reported in prior work, as the authors acknowledge. The contribution is therefore mainly an empirical characterization at LLM scale. Relevant work suggesting that larger single models can match the benefits of ensembles (e.g., [1]) is not discussed.
[1] Abe, Taiga, et al. "Deep ensembles work, but are they necessary?." Advances in neural information processing systems 35 (2022): 33646-33660.
Questions for authors
-
Key results use pairs of different models trained on different data and recipes (e.g., Qwen2.5 + Qwen3). Is there precedent for treating their disagreement as epistemic uncertainty, and how would the main results change with seed ensembles only?
-
The growing visible share requires the irreducible loss to exceed a threshold. What evidence supports this beyond the best achieved model loss?
-
Does a growing average visible share imply better per-input uncertainty estimates at scale?
-
The fitted projections show no crossover for Pythia or Qwen3 up to 1T. How is this consistent with the claim that ensembling becomes increasingly competitive at scale?
4 comments
Part 2
Weakness 3 and Question 3: Mismatch between motivation and results... Pointwise uncertainty is addressed only briefly, using a reference-model proxy with modest predictive accuracy, so it is unclear what the scaling results imply for the applications that motivate the paper.
A: To make the direct connection between the average trend indicated by scaling laws and individual context-based uncertainties clearer, we added a set of plots (top row of Fig. 6) that show the visible fraction of uncertainty when using a reference model. This quantity is the exact analogue of what the scaling law predicts. In this experiment, the visible fraction is calculated by dividing the ensemble uncertainty/spread by the approximate reducible loss estimated by the KL to the reference model. We see a similar trend, but shifted upward compared to Fig. 5b because the reference models lower-bound the irreducible loss.
Q: Limited novelty. ... The contribution is therefore mainly an empirical characterization at LLM scale. Relevant work suggesting that larger single models can match the benefits of ensembles (e.g., [1]) is not discussed
A: Thank you for the pointer; we have added a brief discussion of Abe et al. (2022) in Section 3.4.
What is novel compared to the prior work, including the one pointed out by the reviewer, is that we characterize the benefits of ensembling both in achieving better accuracy and in uncertainty estimation, through power laws that relate them to model size and branching strategy. As a result, our conclusions can be the opposite of what is reported in such prior works depending on these parameters. For example, the paper above suggests that a single model of appropriate size can match the performance of an ensemble, while we draw the opposite conclusion with a more accurate statement: assuming our scaling laws, if we match the ensemble and the single model (either by total parameters and total compute, or in terms of compute-optimality and total compute), an ensemble always outperforms a single model beyond some model scale. This cross-over is now clearly characterized solely based on the two scaling laws in our claim of Section 3.4. The same is true for our arguments about uncertainty: prior work, such as Kirsch et al. (2024), reports that the ensemble diversity collapses, which in their terms questions the effectiveness of ensembling at scale. Using our scaling laws, we show the opposite: that despite the shrinkage of diversity, due to the nature of power laws, ensembling becomes more effective for uncertainty estimation at scale.
Question 1: Is there precedent for treating [the] disagreement [in heterogeneous ensembles] as epistemic uncertainty?
A: Yes, this is a common practice in various data modalities. We can provide a list of citations if required. Given the relationship between ensembling and Bayesian model averaging, ensembling different architectures is analogous to extending the model class in the Bayesian setting.
The remaining questions are answered in response to the weaknesses above.
Thank you for the response and the revised manuscript. The separation of seed and heterogeneous ensembles, App. E.1, and Fig. 6(f–i) and 16 address my Q3 and much of W1. I have a few remaining questions.
-
Fig. 14 shows that the trend reverses at . At what value of does it reverse, and how close is this to the lower end of the retained range? The trend also looks fairly flat near the lower end, so it would be good to mention this in the main text.
-
The 6.9B Pythia pair is excluded from the primary fits (App. B.2), so the seed-only fits seem to go up to about 1B rather than 7B. Could you clarify this? References for my Q1 would also be helpful.
-
The crossover sizes in App. G.2 are far beyond current model scales, and the measured seed ensembles in Fig. 7 still underperform the larger model. I think this should be stated in the main text, and the abstract and the claim of refuting Abe et al. [1] should be toned down accordingly.
-
In Table 23, the matched-cost comparison is essentially a tie, as App. G.3 also notes, and the fixed-data reference is not annealed. How does this support that the pair beats a larger model in both scenarios?
References
[1] Abe, Taiga, et al. "Deep ensembles work, but are they necessary?" NeurIPS, 2022.
We appreciate your engagement!
We identified a recent release of the Delphi seed ensemble family of models with up to 9.6B parameters, and we added its results to the paper. It supports the general identified trends and further addresses your concern about larger seed ensembles.
Q: At what value [of the floor] does [the trend] reverse, and how close is this to the lower end of the retained range? The trend also looks fairly flat near the lower end, so it would be good to mention this in the main text.
A: The table below shows the reversal threshold for different families. The uncertainty range for the floor is [0.26019,0.33373], so the reversal thresholds do not overlap with our uncertainty for the floor. We have added this table to the same appendix section. Please note that Figure 5 shows the full range with the uncertainty band, so readers can see the lower end of the slopes; however, we have also added a sentence about this in the main text.
| Family | Reversal threshold |
|---|---|
| Character | -0.35597 |
| Pythia | 0.18996 |
| DataDecide C4 | 0.20088 |
| DataDecide Dolma | 0.17154 |
| Dense Qwen2.5/Qwen3 pairs | 0.14397 |
| Delphi | 0.20384 |
| Pythia, if we include 6.9B model | 0.14014 |
Q: The 6.9B Pythia pair is excluded from the primary fits (App. B.2), so the seed-only fits seem to go up to about 1B rather than 7B. Could you clarify this?
A: Pythia 6.9B is a true seed pair; however, it is from an older Pythia release, so we excluded it from the family fit. We still measure its loss and spread, and it does support our claims about visible fraction scaling.
Q: The crossover sizes in App. G.2 are far beyond current model scales, and the measured seed ensembles in Fig. 7 still underperform the larger model. I think this should be stated in the main text, and the abstract and the claim of refuting Abe et al. [1] should be toned down accordingly. In Table 23, the matched-cost comparison is essentially a tie, as App. G.3 also notes, and the fixed-data reference is not annealed. How does this support that the pair beats a larger model in both scenarios?
A: We adjusted the phrasings the reviewer referenced to make them more accurate and added the recommended statements, except for the abstract, which simply states the increased competitiveness of ensembles with scale. We note that the crossover range for the two families discussed starts around 20B to 3T model sizes, which are realistic for today's large open-weight models, but the upper end of the interval includes much larger models.
We also added specialized validation sets (i.e., more OOD datasets). Since the spread increases (ensemble members disagree more), the advantage of ensembling in lowering the loss also increases for these datasets, and here we see many cases of crossover among seed ensembles; see Figure 18.
Q: References for my Q1 would also be helpful.
A: Please find below some examples from different areas:
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. CVPR.
Zaidi, S., Zela, A., Elsken, T., Holmes, C., Hutter, F., & Teh, Y. W. (2021). Neural Ensemble Search for Uncertainty Estimation and Dataset Shift. NeurIPS.
Antorán, J., Allingham, J. U., & Hernández-Lobato, J. M. (2020). Depth Uncertainty in Neural Networks. NeurIPS.
Wang, X., Kondratyuk, D., Christiansen, E., Kitani, K. M., Alon, Y., & Eban, E. (2022). Wisdom of Committees: An Overlooked Approach to Faster and More Accurate Models. ICLR.
Gontijo-Lopes, R., Dauphin, Y., & Cubuk, E. D. (2022). No One Representation to Rule Them All: Overlapping Features of Training Methods. ICLR.
Simonyan, K., & Zisserman, A. (2015). Very Deep Convolutional Networks for Large-Scale Image Recognition. ICLR.
Shwartz-Ziv, R., & Armon, A. (2022). Tabular Data: Deep Learning Is Not All You Need. Information Fusion, 81, 84–90.
Caruana, R., Niculescu-Mizil, A., Crew, G., & Ksikes, A. (2004). Ensemble Selection from Libraries of Models. ICML.
Erickson, N., Mueller, J., Shirkov, A., Zhang, H., Larroy, P., Li, M., & Smola, A. (2020). AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data. arXiv:2003.06505.
Jiang, D., Ren, X., & Lin, B. Y. (2023). LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion. ACL.
Chen, L., Zaharia, M., & Zou, J. (2023). FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176.
Wang, J., Wang, J., Athiwaratkun, B., Zhang, C., & Zou, J. (2025). Mixture-of-Agents Enhances Large Language Model Capabilities. ICLR.
Part 1.
We thank the reviewer for their valuable feedback. Below, we believe we have addressed all of their questions and concerns expressed as weaknesses. We would be happy to address any remaining issues during the discussion period.
Weakness 1 and Question 2: The central conclusion depends on an unknown quantity... The trend weakens as this value is lowered and reverses below a threshold that the paper cannot rule out.
A: First, please note that our scaling law for the spread and the late branching law are independent of the irreducible loss. Only the results of Figures 5 and 7 use this estimate. Concretely, what depends on the noise floor is the exponent in the Chinchilla scaling law, which is well-studied and generally believed to be around .3 but may vary for specific datasets and for smaller studies. Our conclusions hold as long as this number is larger than the small exponents we observe for the spread ≈[.11−.16]. Therefore, our trend is already supported by estimates from the existing LLM scaling law literature.
Independently of existing studies, we get a lower bound for the irreducible loss using a nominal 95% F-based profile confidence interval, and furthermore we now report the same monotonically increasing trend of the visible fraction using a reference model; see Fig. 6 (top row).
Finally, we have visualized the irreducible loss floor and the model-size scaling of the loss. Qualitative inspection of Fig. 2(b) suggests a conservative estimate, and this same range is used in drawing the uncertainty bands for Figures 5 and 7.
Weakness 2 and Questions 2 and 4: Large-scale evidence comes mostly from pairs of different models... Yet these pairs drive key results, such as the highest projected visible share and the main example of an ensemble beating a larger model...
A: First, please note that most of the theory claims and many of the empirical results are true for heterogeneous ensembles. In particular, the theory of Section 2.2 applies to any ensemble; the only exception is the justification for normalization by S/(S−1), which can be dropped without affecting our conclusions. What differs when going from a seed ensemble to an arbitrary ensemble is that the mean member loss Lˉ does not reflect individual members any longer, and therefore the gain due to ensemble spread may not improve over the best ensemble member in terms of (reducible) loss.
Second, we have now clearly separated the results for seed ensembles from other types in all reported experiments and, more importantly, all conclusions can be supported even if we confined the results to seed ensembles. Our seed ensembles range in size from 0.8 M to 7B, approximately four orders of magnitude, and the trends claimed by the paper persist across this range. We now provide examples of seed ensembles beating a larger model in two regimes (parameter-matched and compute-optimal) in Appendix G.3 (Table 23). We also clearly show the theoretical prediction of how the gap should close between ensemble and individual models vs empirical observations for seed ensembles in both of these scenarios; see the updated Figure 7.