acceptodds
Withdrawn from ICLR 2027

Approximating Longer Birth and Death Processes with Large Language Models and the Self Improvement Framework

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Birth and Death Processes are continuous-time Markov chains that transition to either the next or previous state. They model population sizes; state transitions lead to the births and deaths of individuals in a population. We model large,structured birth and death processes on large language models by simulating population data and training large language models to predict the next population. As large language models struggle with large and complicated inputs, we apply previously studied length generalization methods to the problem, including extending the self improvement framework to data from birth and death processes to benchmark performance. To overcome underfitting caused by low difficulty added per self improvement framework round, we parameterize the variance added by each self improvement round through the number of population states added. The Chapman-Kolmogorov equation shows us that the number of population states added each self improvement round influences the variance that the large language models fit to. Finally, we compare applied methods of length generalization in this domain. Our source code can be found on Github at https://anonymous.4open.science/r/self_improving_bd_processes/README.md.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Self Improving Transformers, Length Generalization, Large Language Models, Stochastic Processes

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_mEF6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether self-improvement training can help T5 models extrapolate to longer sequences from simulated birth–death processes. It varies the number of population states added per self-improvement round (KK), and evaluates both output structure and numerical prediction across Verhulst, Poisson, and Ricker processes. The results suggest that self-improvement can help on these controlled tasks, with the effect of KK depending on the process and metric. The question is interesting, but the paper does not yet establish that these findings transfer to general-purpose LLMs or broader long-context forecasting.

Critical strengths
  1. Interesting, underexplored application of self-improvement to structured stochastic processes is a useful test case for length generalization. The process setup gives the authors control over the data-generating mechanism and sequence difficulty.

  2. The study evaluates multiple process families and complementary measurements to distinguish structural validity from numerical prediction quality. It also includes prompting and fine-tuning comparisons rather than relying on a single training setup.

  3. The paper connects the number of states added per round to changes in the variance of the process, using the Chapman–Kolmogorov equation, and tests different KK values. This offers a concrete starting point for understanding why self-improvement may help in some settings.

Critical weaknesses
  1. Limited evidence for broader relevance and unclear positioning against forecasting work. The central experiments use synthetic data and T5-small/T5-base on a narrow family of serialized population sequences. This is not enough to support broad conclusions about contemporary pretrained LLMs, generic long-context OOD generalization, or time-series forecasting. The paper should more clearly position itself relative to work such as Gruver et al. (2023), which studies zero-shot forecasting of numeric series serialized as text. The distinction seems to be that this paper studies self-improvement and length extrapolation on simulated birth–death processes, but it does not establish how much this setup adds beyond prior LLM forecasting results. Nor does it compare against strong time-series foundation models or process-specific statistical baselines, making it difficult to tell whether language-model training is particularly useful here. The application example is also based on simulated extrapolation, rather than a broad evaluation on real population forecasts.

  2. The mechanism and training setup need clearer validation. The motivation for KK is plausible, but increasing KK changes both sequence length and the distribution/variance of prediction targets. The experiments do not clearly isolate which change causes the observed effects. In particular, Section 3.3 uses linear regression and decision trees to argue that larger KK creates a more complex mapping; this is not direct evidence about the LLMs or the self-improvement mechanism. The main text also makes it difficult to reconstruct the full training recipe and comparisons. The authors should clarify model initialization, per-round data construction and labels, optimization, data splits, compute/updates, and variation across random seeds.

  3. Presentation and claims are not yet sufficiently polished or calibrated. Several plots and tables overflow or are difficult to read, and the prose is repetitive in places. The conclusion that length generalization is “a solved problem for some data” seems too strong for the limited synthetic task families and model sizes studied. The broad claim that length generalization remains difficult is supported mainly by an older citation (Anil et al., 2022), while another statement about prior studies lacks a citation. These claims should be qualified and supported with more recent work. The prompting result is also hard to interpret as a forecasting comparison: the domain-informed prompt performs better than “Find the next number,” but this mainly shows the value of supplying task information, not that the model is a strong general-purpose forecaster.

Questions for authors
  1. How does this work extend prior LLM-based numeric forecasting, such as Gruver et al. (2023), beyond applying self-improvement to synthetic birth–death sequences? Can the authors compare with a contemporary pretrained LLM and a strong time-series foundation or process-specific baseline on the same data and protocol, or narrow the claims to the studied T5/synthetic setting? Any claim that TSFMs generally outperform LLMs should be tied to recent, matched benchmarks and qualified by task and evaluation setting.

  2. Could the authors provide a compact, reproducible description of the complete training pipeline, including initialization, tokenization, data splits, how model-generated outputs enter each round, optimization budget, and number of seeds? How are comparisons controlled for training tokens, updates, and compute?

  3. Section 3.3 uses linear regression and decision trees to motivate the increased-complexity hypothesis. Can the authors clarify that these are diagnostic analyses rather than LLM experiments, and provide an ablation that separates the effect of KK on sequence length, target variance, and training distribution?

  4. The domain-informed prompt provides substantially more task information than “Find the next number.” What conclusion should readers draw from this comparison, and can the prompting baseline be made more comparable to the trained models? Please also update/qualify the broad length-generalization claims and repair the overflowing figures and tables.

Reviewer_oq23OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates length generalization in language models for approximating birth-and-death processes. The authors generate synthetic population trajectories using BirDePy and evaluate T5-based models on Verhulst, Poisson, and Ricker processes. Building on an existing self-improvement framework, they introduce a parameter KK that controls the number of population states added during each training round. The authors argue that increasing KK introduces greater prediction complexity and helps alleviate underfitting. Experiments compare prompting, fine-tuning, and iterative self-improvement, with an additional application to bird population prediction.

Critical strengths
  1. The paper explores length generalization in stochastic population processes, extending an existing self-improvement framework to a different task domain.
  2. The experiments consider three types of population processes and multiple values of KK, providing some empirical observations about model behavior at different sequence lengths.
Critical weaknesses
  1. The central theoretical motivation is not sufficiently justified. The authors argue that increasing KK increases the variance and complexity of the prediction target, thereby alleviating underfitting. However, the Chapman-Kolmogorov equation does not establish that increasing prediction variance necessarily reduces model underfitting or improves length generalization. Moreover, increasing KK simultaneously changes the prediction horizon, training sequence length, and target difficulty, making it difficult to identify the actual source of improvement. The paper lacks controlled experiments that isolate these factors.
  2. The evaluation methodology raises concerns about whether the models actually learn the underlying stochastic dynamics. Birth-and-death processes are stochastic, so a generated next state can differ substantially from a sampled reference trajectory while remaining statistically valid. Nevertheless, the paper relies heavily on token accuracy, MSE, MAPE, and output-structure correctness. In particular, the reported near-perfect structural accuracy does not establish accurate modeling of transition distributions. Although Section 4.7 introduces parameter estimation, the analysis is limited and does not provide a systematic evaluation of learned transition probabilities or trajectory distributions. Stronger distributional evaluations and comparisons against classical stochastic-process models are needed to substantiate the main claims.
  3. The paper also suffers from noticeable presentation and formatting issues. Several tables, particularly Tables 1, 2, and 5, are excessively wide or densely formatted, making the numerical results difficult to read. The appendix tables have similar readability problems. Some figures are too small, and the surrounding text layout is crowded. The authors should resize or reorganize the tables, improve figure readability, and carefully check the overall formatting before publication.
Questions for authors
  1. What theoretical or empirical evidence establishes that increasing the variance of the prediction target reduces underfitting? Can the authors isolate the effects of prediction horizon, sequence length, and target variance?

  2. How do the authors verify that the models learn the correct stochastic transition dynamics rather than merely reproducing plausible sequences with the correct output structure? Can they provide distribution-level evaluations against classical stochastic-process baselines?

Reviewer_uVnaOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the length generalization ability of large language models on birth and death processes. It applies a self improvement framework and explores increasing the number of added population states K to mitigate underfitting. Experiments evaluate the effectiveness of the proposed approach.

Critical strengths

The paper investigates a interesting research problem. The authors provide sufficient experimental results.

Critical weaknesses
  1. Serious formatting issues. The paper has substantial formatting issues and would benefit from careful revision before resubmission.
  2. Poor readability. The presentation is difficult to follow. For example, the experimental analyses often lack explicit references to the corresponding figures and tables.
  3. “Indicating that, in some cases, approximating population size with K > 1 is a more complicated function than when K = 1.” This conclusion is confusing. First, according to the results in Table 1, MSE does not increase monotonically with K. Second, given the stochastic nature of the experimental setting, it remains unclear whether the observed results are caused by large random fluctuations or other confounding factors.
  4. “Simulations and intuition show that variance increase with K; there are less trivial results, intuitively, with K > 1. The data is therefore more complex and takes advantage of the high flexibility of a large language model.” This conclusion is also confusing. Higher variance does not necessarily imply that the results are more difficult or the data are more complex. Similarly, more flexible models do not necessarily benefit from higher variance.
Questions for authors

Please reply to the issues raised in the weakness section.

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

Related papers

Discussion (0)

Sign in to comment.