- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper examines temperature scaling in posterior-induced flow language models. The authors point out that standard posterior tempering can create a mismatch between the tempered distribution and the underlying sampling dynamics. To address this, they propose Bayes-Consistent Tempering (BCT), which adjusts the intermediate distributions and conditional velocities to account for the temperature change. The paper develops both training-based and training-free approaches for hyperspherical and von Mises–Fisher flows. Experiments on GSM8K and Sudoku show that BCT generally performs better than standard tempering, with the most noticeable improvements at lower temperatures.
Critical strengths
1.The method is well motivated. The authors identify a concrete inconsistency in standard posterior tempering and derive bridge-aware conditional dynamics for two distinct flow constructions. This provides a useful theoretical perspective on temperature control in flow language models.
2.The experimental comparisons are informative. The paper includes matched continuation baselines, training-free variants, two denoiser backbones, and Pass@K evaluations. The improvements on GSM8K are substantial, and the Sudoku results provide additional evidence beyond mathematical reasoning.
Critical weaknesses
1.The main implementation does not fully preserve the claimed Bayesian consistency. Equation (8) gives the exact correction, but Equation (9) introduces a tunable coefficient λ. The main GSM8K results use λ = 0.5 for S-arch and λ = 0.25 for DiT, rather than λ = 1. For these values, the resulting posterior generally no longer corresponds to the tempered bridge whose conditional velocity is used. This weakens the connection between the theoretical motivation and the reported improvements.
2.The empirical scope remains limited. The evaluation focuses on GSM8K and Sudoku, with only one training run per continuation. The reported uncertainty therefore does not capture training variability. The vMF experiments also show a noticeable performance drop for BCT-L at T = 1. These results make the broader robustness claim less convincing.
Questions for authors
1.Can the authors report GSM8K results with the exact Bayes correction (λ = 1) and explain how the selected partial corrections remain consistent with the sampling dynamics?
2.Can the authors separate the effects of tempered-bridge training, corrected conditional velocities, and the tuned posterior correction? How sensitive are the gains to training seeds?
3 comments
Thank you for the Rebuttal
Thank you for the thorough response and the additional results. They address most of my concerns.
Thank you for the follow-up
Thank you for the follow-up.
We are glad that our rebuttal response addresses your concerns: the exact correction, the role of each component, training-seed variability, the vMF case at , and the empirical scope of our claims. Your questions led directly to these experiments and have made the evidence of the paper considerably stronger and clearer.
Thank you again for your careful and constructive engagement through the review.
Response to Reviewer f2Gv
We thank the reviewer for the careful reading. In short: the exact correction (λ=1) already gives the gains in three of four settings, and λ<1 is a principled safeguard against denoiser error; the corrected velocity carries most of the gain, with TFT keeping it under strong tempering; the gains hold across training seeds; and the vMF drop at T=1 reflects the checkpoint, not Proposition 1. All numbers are GSM8K test accuracy (%).
W1 / Q1: Exact correction (λ=1) and the role of λ
Exactness holds only at λ=1, and λ=1 alone already gives the gains in three of four settings. On S-arch, λ=1 outperforms standard tempering at every T<1; on Sudoku and the vMF flow, the held-out selection already chose λ=1. On DiT, λ=1 degrades under strong tempering (3 generation seeds):
Why λ is needed in practice. Proposition 1 assumes an exact denoiser, and no trained denoiser is exact. Its logits are inevitably slightly inconsistent with the analytic term of Eq. (10), whose magnitude grows as T decreases, so the inconsistency is amplified at low temperature. λ<1 damps the amplification by trading it for a known deviation, a slightly sharper bridge qt,Tγ with γ=β−λ(β−1), which vanishes at λ=1. λ is selected once per model on held-out data, never on the test set, and the held-out sweep already anticipated the DiT result. Moreover, most of the gain comes from the corrected velocity, which is exact for every λ (next item).
In the revised PDF: role of λ (
§3.1); λ=1 and the held-out sweep (App. C,Tables 2–3).Q2 (a): Separating TFT, the corrected velocity, and the posterior correction
The corrected velocity carries most of the gain; TFT does not help alone but keeps that gain under strong tempering; the posterior correction adds to it, most at the lowest temperature. The samplers differ only in the three switches shown in the table. At T=0.1 (generation seed 1):
Without TFT (last row), the corrected velocity visits intermediate states the denoiser was not trained on and degrades.
In the revised PDF: all temperatures (
App. K,Table 11).Q2 (b) / W2: Variability across training runs
The gains are robust to training seeds. We re-trained both continuations on S-arch with two more training seeds, keeping λ=0.5 without re-selection. For every training seed, BCT-L outperforms standard tempering at every T<1, and the standard deviation of BCT-L across training seeds (0.3–0.8 points) is far below its gain (6–9 points at T≤0.2). We will report training-seed variability for all settings in the camera-ready version.
In the revised PDF:
App. L,Table 12.W2: The vMF flow at T=1
The drop (21.6% vs. 23.6%) does not contradict Proposition 1: at T=1, the two entries use the same sampler and differ only in the checkpoint. Every BCT correction is the identity at T=1 (Proposition 1(ii)); TFT is trained on tempered bridges over a range of temperatures, the released and Cont. checkpoints on the original bridge only. Continual training itself is not the cause (Cont. matches the released checkpoint), and of the six TFT checkpoints in this work, this is the only one whose T=1 accuracy differs from its matched Cont. checkpoint by more than 0.5 points. It is thus a property of this learned denoiser, outside the idealized setting of Proposition 1. Finally, it has no practical cost: BCT reduces to the original sampler at T=1, so the base checkpoint can be used there.
In the revised PDF:
App. G,Table 8;Proposition 1(§3.2).W2: Scope of the robustness claim
We now state it precisely: at every T<1, BCT-L never falls below standard tempering with the same continual training, in all four settings and, on S-arch, for each of three training seeds. The Limitations now state that Proposition 1 covers the idealized sampler and that robustness to a learned denoiser is an empirical finding, established on two tasks, GSM8K and Sudoku.
In the revised PDF:
§4.2; Limitations (§6).We hope these results address the reviewer's concerns and are happy to clarify further.