- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper builds a toy behavioral eval suite that tells pattern-matching "parrots" apart from generalizing intelligence, and uses it to track the intermediate checkpoints of OLMo3 and Apertus. The authors find that models frequently and abruptly switch between pattern-matching and generalization throughout pre-training (mode-hopping), even while loss and standard benchmarks improve smoothly. They argue that this is not caused by eval noise, metric choice, or standard optimization dynamics, and attribute it to capacity allocation between generalizable and shallow circuits. Larger models mitigate mode-hopping but do not eliminate it. The authors also show two applications: selecting intermediate checkpoints that generalize better after post-training, and selecting pre-training data to stabilize generalization dynamics.
Critical strengths
- Extensive experiments. The eval suite covers six behavioral evals plus two fine-tuning-based tests, spanning many datasets. It is applied to dense intermediate checkpoints of two fully open model families at four scales (OLMo3 7B/32B, Apertus 8B/70B), with results averaged over 4 seeds. The authors also carefully rule out several alternative explanations: eval noise (stable curves on 10 standard benchmarks), metric artifacts (soft probabilities and a FLOPs x-axis), and optimization dynamics (single-step local stability tests and checkpoint averaging). Together, these make the core phenomenon convincing.
- Novel perspective. The paper challenges the common assumption that generalization improves steadily during pre-training. It shows that generalization behavior can switch abruptly between modes even after trillions of tokens and far beyond Chinchilla-optimal budgets, while loss and standard benchmarks improve smoothly. This adds a dimension that loss-based scaling analyses and downstream benchmarks do not capture.
- Practical relevance of checkpoint selection (Application 1). The authors show that an intermediate checkpoint selected with the cheap toy suite (4.5T) beats the final pre- and mid-training checkpoints after post-training, both in math SFT → GPQA transfer (36.3% vs. 29.8%) and in robustness to prefilling attacks (53% vs. 21%). This suggests that the common practice of post-training only the final checkpoint may be suboptimal, and that a lightweight behavioral probe could guide checkpoint choice in practice.
Critical weaknesses
- Limited size and diversity of some evals. Several evals have few test examples or little variety. Truthy Answer has fewer than 300 claims per subset. Each Intuitive Answer subset consists of 200 templated variants of a single CRT problem. Multi-hop Persona QA has only 5 test questions per persona. With such narrow item pools, part of the observed oscillation may reflect sensitivity to a few specific items or templates rather than a general shift in generalization behavior. Given the overall scale of the experiments, this is understandable, but the conclusions drawn from these evals should be stated more carefully.
- Section 5.2 is underspecified. The paper only states that subsets are selected "based on whether they enhance pattern-matching or generalization in the original OLMo3 pre-training run", without describing the actual procedure. It is unclear how a subset is labeled as pattern-encouraging or generalization-encouraging (e.g., by the accuracy change across a checkpoint window, or by document-level scoring), which windows are used, and what training hyperparameters are used. If the data are labeled by how target-eval accuracy changed after training on them in the original run, the result is partly circular. In addition, the experiment covers a single eval (answer+1) and only about 6.7B tokens, with no check on other evals or general capabilities. As it stands, this section is a feasibility demonstration rather than a usable data-selection method.
- Mode-hopping may not be a single, global phenomenon. Figure 12 shows low cross-dataset correlations within the same eval (about 0.15–0.50). This means that a checkpoint that generalizes on one dataset often fails on another. This raises the question of whether mode-hopping reflects a global "generalization prior", as the paper frames it, or a collection of largely independent task-specific fluctuations. It also weakens the case for using the suite as a general checkpoint selector.
Questions for authors
- Is the math SFT in Section 5.1 trained long enough? The authors follow the practice of Ren et al. (2026) [1], which shows that reasoning SFT can generalize as well as RL under multi-epoch training. Ren et al. also report a dip-and-recovery pattern: out-of-domain performance first drops early in SFT, then recovers and improves with continued training up to 8 epochs. In Figure 13(a), GPQA accuracy of the 4.9T checkpoint keeps declining over the roughly 200 post-training steps shown, which looks like it could be the "dip" phase. Could the authors report the number of epochs and total training steps, and say whether training was run to convergence? Specifically, if SFT continues further, does the 4.9T checkpoint (or the final pre-/mid-training checkpoints) recover on GPQA, and does the gap with the 4.5T checkpoint remain? If the gap shrinks after full training, the advantage of the selected checkpoint would be a difference in recovery speed rather than in final generalization, which would substantially change the conclusion of Application 1.
[1] Qihan Ren, Peng Wang, Ruikun Cai, Shuai Shao, Dadi Guo, et al. Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability. arXiv preprint arXiv:2604.06628, 2026.
3 comments
The authors have responded to your reviews. Please read their response and reply. Please specify which concerns are resolved and which remain. Please update your review and score if the assessment has changed.
AC
Thank you for the clarifying response and the additional experiment. The authors have addressed most of my concerns, while I believe the test sets like Multi-hop persona QA should be augmented beyond the five provided examples for its credibility. Accordingly, I maintain my weak accept assessment.
We observe mode-hopping across our eval suite, which in total consists of tens of thousands of test examples covering diverse distributions.
Each individual test set has hundreds to 1K examples. This is a common size for test sets. For example, GPQA, a widely adopted benchmark, has ~500 test examples. Multi-hop persona QA is indeed an outlier with only 5 test questions. We sample hundreds of rollouts per question to stabilize the results.
We would also like to clarify that we never optimize our test set against model behaviors to inflate mode-hopping. All test sets are built before testing models on them.
Yes it’s based on the accuracy change across a checkpoint window.
Our toy experiments in Sec 5.2 aims to show that pre-training generalization dynamics can be controlled by controlling data. This experiment is quite expensive, since we need to run continued pre-training on OLMo3 32B (a dense model). We leave the exploration of controlling pre-training dynamics on more diverse, challenging tasks to future work.
There are our training hyperparameters:
The step count may look small because of our large batch size. In preliminary runs with 20 epochs, training converged by epoch 8.