acceptodds
Under review as a conference paper at ICLR 2027

Efficient Calibration via Rejection Sampling

Abstract

Exponential Tilting (ET) is a classical technique for model calibration, alignment and reward maximization. For a given statistic (reward) function , ET transforms a base model distribiution via the multiplicative update rule . Braverman et al. (ICML 2020) showed that fitting to minimize cross-entropy loss against some reference distribution, ensures the tilted model does not lose quality, while maximizing the average value of . Unfortunately, ET has a prohibitive computational cost of computing the normalizing partition function, requiring multiple foward passes to evaluate on the full support of tokens, which renders this technique impreactical in the autoregressive LLM setup. We propose a natural, highly efficient algorithm for implicitly sampling from the tilted distribution via rejection-sampling, which dramatically reduces the query access to the tilted model and avoids computing the normalizing partition function altogether. We demonstrate the efficiency of our technique in theory and practice, by analyzing the expected query count and upper bounding it in terms of the cross-entropy distance to the base model and the “variance" of . We then corroborate our bounds on small-size LLMs (Qwen2.5, Qwen3 and SmolLM3), applying our approach to the substantive case study of entropy calibration of LLMs (Cao et al., NIPS 2025). Our experiments show improvement in entropy miscalibration in base models, while incurring 15–25% additional sampling attempts due to rejections. We expect this overhead to improve when moving to more capable and larger models.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Calibration, Entropy Calibration, Rejection Sampling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_aGZDOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper tackles calibration in Large Language Models. They study a notion of calibration termed ff-calibration, which matches the expectation of a statistic ff induced by the model with the expectation of the same statistic under a reference distribution. The main problem authors tackle is that the sought calibration procedure, i.e.i.e., exponential tilting (ET), requires expensive computations of a partition function. The authors propose an efficient approach to ET based on rejection sampling. Rejection sampling enables to efifciently sample from the tilted distribution and authors additionally propose a methodlogy to estimate the exponential tilt α\alpha. Authors validate they proposal in terms of entropy calibration with Qwen 2.5\texttt{Qwen 2.5}, Qwen 3\texttt{Qwen 3} and SmolLM3-3B\texttt{SmolLM3-3B} with 2 different datasets. The proposal effectively improves entropy calibration of the base models.

Critical strengths

The main empirical strength of the paper is that the proposed method consistently reduces entropy miscalibration relative to the corresponding base models across several model families and datasets.

I also consider the idea behind the methodological contribution appealing and potentially useful. More precisely, the idea to use rejection sampling to enable exact sampling from the exponentially tilted distribution without explicitly computing the partition function is compelling.

Moreover, to the best of my knowledge, the paper is well connected to the relevant literature on ET, ff-moments matching and entropy calibration, and rejection sampling.

Critical weaknesses

The biggest weakness of the paper is clarity. I found the paper difficult to follow, with several important methodological components presented in a fragmented manner. In particular, the procedure used to estimate the tilt parameter α\alpha is scattered across different sections, and the connection between the cross-entropy objective, its derivative, the moment-matching condition, and the practical bisection procedure is not explained clearly in one place.

Moreover, the estimation of the entropy difference is entirely detailed in the Appendix. This is a crucial element of the experimental evaluation and the estimator is sufficiently nontrivial that I believe at least its main idea, properties, or practical implementation should be described in the main text. Considering the paper appears to have available space, I do not see why this was not the case in the first place.

The thoeretical contribution of the paper is also of difficult interpretation. Authors state several lemmas and a theorem, but the surrounding discussion provides relatively little intuition about what the bounds imply, when they are informative, and how they connect to the practical algorithm.

I also have concerns about terminology. In particular, the paper uses the term “local calibration” to refer to the teacher-forced setting. I understand the intended distinction, but “local calibration” is already an established term in the calibration literature, typically referring to calibration notions that incorporate locality or similarity between inputs or their representations [1-6]. I therefore consider this terminology potentially confusing and suggest replacing it.

The second major wekaness of the paper is the experimental evaluation. Authors do not include any baselines other than the base model. This surprises me because the introduction and related-work sections explicitly discuss several alternative approaches and motivate the proposal partly through their limitations. Without such baselines, it is difficult to assess whether the proposed method provides a meaningful practical advantage beyond improving the targeted entropy-calibration objective relative to the unmodified model.

This issue is particularly relevant because authors mention existing methods suffering from a quality-diversity trade-off. However, the experiments do not evaluate this trade-off directly nor do they compare the proposal against the alternative methods that should suffer from it. I would have expected an empirical comparison reporting calibration together with quality and diversity metrics.

Finally, the paper makes several claims that appear stronger than what is supported experimentally. For example, the discussion suggests possible applications to phenomena such as over-empathy and sycophancy, but no experiments are provided for these settings. Similarly, the abstract states that the rejection-sampling overhead is expected to improve for larger and more capable models, while the current experiments only cover a limited set of relatively small models. These claims should either be supported experimentally or stated as directions for future work.

REFERENCES

[1.] Barbera, Cesare, et al. "Multiclass local calibration with the Jensen-Shannon distance." AISTATS2026.

[2] Luo, Rachel, et al. "Local calibration: metrics and recalibration." UAI2022.

[3] Ding, Zhipeng, et al. "Local temperature scaling for probability calibration." ICCV2021.

[4] Xiong, Miao, et al. "Proximity-informed calibration for deep neural networks." NeurIPS2023.

[5] Lin, Zhen et al. "Taking a step back with kcal: Multi-class kernel-based calibration for deep neural networks." ICLR2023.

[6] Valk, Kaspar, and Meelis Kull. "Assuming locally equal calibration errors for non-parametric multiclass calibration." TMLR2023.

Questions for authors

I would ask reviewrs to address my concerns in the weaknesses section, especially regarding the experimental evaluation.

One additional question I have is: although the proposed framework is formulated for generic ff-calibration, what guarantees does improving this moment-matching objective provide with respect to other existing notions of calibration? Does the proposed method improve probabilistic quality beyond the targeted entropy statistic? It would be informative to report additional proper scoring rules, such as NLL and the Brier score, and compare them with those of the base model and the baselines.

4 comments

We thank the reviewer for the careful reading and detailed suggestions. We are glad that the reviewer finds exact sampling without the partition function compelling and the reduction in miscalibration consistent across models and datasets. We addressed each point in the revision, in the order of the review.

1. Estimating α. Section 5.1.1 describes the procedure at a high level, and the new Algorithm 2 (Appendix E) gives the exact algorithm we use. α minimizes the CE ℓ(α) on calibration documents. This objective is convex, with derivative ℓ′(α) = μ_α − μ_P, so minimizing CE and matching the moment are the same problem. We estimate ℓ′(α) from R rejection-sampled tokens per calibration context, without the partition function, and find α by bisection.

2. Entropy estimator. We agree that the estimator is non-trivial, but we kept its derivation in the appendix because it is an evaluation tool rather than part of the method. Neither fitting α nor sampling uses it. It only serves to report the tilted model's entropy in Figures 1–3. We preferred to use the main-text space for the new held-out results (Table 1). To address the concern, Remark 7 now states the estimator's main idea and guarantees in the main text. The derivation and the practical implementation remain in Appendix C.

3. Interpreting the theory. Each result in Section 4 is followed by its interpretation (lines 228–232, 241–247 and 257–269). In short, sampling is exact for any valid bound (Lemma 2), its cost grows with the tilt times the typical value of the statistic (Lemmas 3 and 4), and stopping the CE minimization early gives a smaller, cheaper tilt at a quantified CE loss (Lemma 5, Theorem 6). The revision adds the converse: if the CE can improve only slightly, the fitted tilt must be small and therefore cheap to sample (lines 257–262). Section 4.1 connects the theory to the practical algorithm for entropy, where B = 0 is always a valid bound.

4. Terminology. Thank you for pointing out the clash with the established notion of local calibration in classification, and for the references. We renamed local and global calibration to next-token calibration and full-sequence calibration throughout the paper, where calibration means moment matching calibration (or f-calibration).

5–6. Baselines and the quality–diversity trade-off. We added held-out log-loss (NLL), the diversity measure used by Cao et al., for the base and tilted models (Table 1). For quality, Cao et al. (2025, Appendix E) find that MAUVE is highest when generation entropy roughly matches the model's log-loss, which is the direction in which our tilt moves the model. Temperature and truncation lower entropy at the cost of log-loss, which is well established (Zhang et al. 2020, Cao et al. 2025). Our tilt does not raise log-loss, as Table 1 confirms. The new tilt sweep (Appendix F) traces this trade-off along the tilt itself: held-out log-loss is lowest at the fitted tilt, and larger tilts lower generation entropy further at a growing cost in log-loss.

7. Claims. We removed the sentence on sycophancy and over-empathy, and the statement on larger models is now explicitly a prediction of our bounds, and not an empirical observation only. The prediction is not that larger models are always cheaper. It is a general one: a model that is more capable at a given task leaves less room for CE improvement and has lower average perplexity levels, so it needs a smaller tilt at a lower entropy level, and by Lemma 3 both lower the rejection rate. Table 1 shows this in both families. Within Qwen2.5, the held-out CE gain and the fitted tilt both shrink with model size. Within Qwen3, the 8B model has a smaller CE gain and lower log-loss than the 1.7B model. Qwen3-4B, the one exception to the size trend, has the largest CE gain in its family, hence a larger tilt and more rejections, so it shows the same mechanism at work.

Other notions of calibration, NLL and Brier. We target moment calibration of a chosen statistic, as Braverman et al. and Cao et al. do, and make no claims about notions that condition on the model's confidence. For NLL the method has a guarantee. Since α minimizes CE and α = 0 is a candidate, the tilted model's log-loss on the calibration data cannot exceed the base model's. Table 1 reports held-out NLL, which is lower than the base model's in all eight settings, and, as suggested, the Brier score. The fit does not target the Brier score, which is unchanged within its 95% interval (slightly better for Qwen2.5-7B), so it serves as an independent check.

We believe the revision addresses the presentation and evaluation concerns, and we would be grateful if the reviewer would consider raising their scores.

The review does not point to a specific error in the analysis. If a soundness concern remains, we would be grateful to hear it during the discussion, so that we can address it.

I appreciate the authors rebuttal. I tried to download the new pdf, but the changes are very hard to track. Before I evaluate their response, I would kindly ask authors to highlight the changes (text of a different color, also for captions of tables and images, is a quite standard pratice)

We posted a new revision with the highlighted text (in case you missed the general comment we posted). Thank you for your consideration, we hope this clarifies our paper.

I thank the authors for their detailed rebuttal and substantial revisions. The presentation has improved, and I appreciate the additional experiments, particularly the NLL and Brier score evaluations. Most of my concerns have been addressed.

However, my main reservation regarding the experimental evaluation remains. The absence of direct comparisons with alternative decoding strategies e.g., temperature or min-p, makes it difficult to assess the practical advantages of the proposed method, particularly concerning the quality-diversity trade-off emphasized in the paper. Given the relatively limited methodological novelty, I consider this comparative validation particularly important.

Overall, I acknowledge the improvements and am willing to increase my score from 1 to 2, but I do not believe the remaining experimental evidence is sufficient to support acceptance.

Reviewer_CBYxOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies LLM calibration via exponential tilting, following Braverman et al'20. Previously, this calibration procedure was noted to be theoretically desirable yet not immediately implementable due to the classic bottleneck of having to compute the normalizing factor. This paper addresses the computational limitation in a very simple yet effective way --- via rejection sampling, A bit more precisely, supposing f is the statistic that we are calibrating to, and that one has a global bound (in an appropriate sense) on the statistic f times a tilting parameter, it suffices to resample a word from the original distribution until acceptance with probability proportional to the tilt at f. The paper shows that this sampling approach can then be implemented in the important case of entropy calibration, after coming up with another entropy-specific heuristic to make the acceptance probability evaluation efficient. Empirical evaluations of entropy-recalibration of several small LLMs then demonstrate that the new approach incurs only up to < 30% computational overhead while markedly improving entropy calibration of these LLMs.

Critical strengths

The main strength of this paper is exactly as it declares --- using a simple heuristic approach, it demonstrates how to efficiently perform entropy calibration of LLMs in practice. This is accompanied by mathematical rigor and thoroughness in stating the guarantees and limitations of the heuristics used. For instance, the above-mentioned additional entropy-specific trick on top of rejection sampling is to estimate only the entropy corrections (their difference before and after tilting) --- and Appendix C fairly rigorously and carefully derives how to estimate the covariance integral involved in estimating these corrections. For another example, the caveat that too-high entropy levels/bad tilting can be problematic for the proposed method is stated quite explicitly and is also visible from the mathematical guarantees.

Critical weaknesses

As alluded to above, the paper is very upfront about its limitations, and no "critical" or unaccounted-for weaknesses are to be noted.

That being said, the reason my evaluation is not higher is that the paper in its current form, while clean, is also possibly insufficiently ambitious. As noted, the rejection sampling idea by itself is not really new in the slightly broader reward-driven (rather than calibration-driven) literature/context, and as such one would have hoped to get a deeper dive into its consequences for moment matching, and in particular see it applied more broadly than for a single calibration task (entropy). I agree with the authors that multi-statistic-calibration extensions are logical to push to future work, but applying the rejection sampling method to other single-statistic calibration tasks would fit naturally into this paper rather than aspirational future work.

Moreover, in terms of the pointed-out limitations of the present method (bad tilts or bad statistic distributions), it would have been better if empirical "stress-testing" were provided, so that the reader could concretely observe what amount of empirical deterioration follows in what regimes. In particular, my hunch is that larger LLMs would offer more opportunity for such stress-testing to happen naturally, so I would expand the empirical evaluation accordingly. Furthermore, I may be missing something, but one of the takeaways the paper offers is that for higher-entropy models, it gets harder for the method to perform well --- yet intuitively, wouldn't one somehow be able to obtain "easy gains" on high-entropy settings via a different approach, and then apply rejection sampling?

Questions for authors

As mentioned above, it would be nice, and strengthen the paper, to extend to tasks beyond entropy calibration, as well as add into the mix various larger/newer language models, and (!) stress-test the proposed approach.

3 comments

We thank the reviewer for a thoughtful and generous review, and we are glad that the simplicity of the approach, the rigor of the guarantees and the care in the entropy-estimator appendix came through. Following the review, the revision adds a small-scale stress test (Appendix F), as well as held-out log-loss and Brier score results showing that the tilted models lower entropy without losing predictive quality (Table 1).

Beyond entropy. Nothing in the sampler is specific to entropy. Algorithm 1 needs only a bound B on αf, and α is chosen by CE minimization for any statistic. Rejection sampling is most useful when the statistic depends on the state after a candidate token, since each candidate then needs its own forward pass. We came upon this problem while working on another project, which needs to calibrate a next-token statistic that must be computed for each token separately, and we plan to report it in follow-up work built on this primitive. In this paper we focused on entropy for two reasons. Modern LLMs are well documented to be miscalibrated in it, and Cao et al. (2025) find that scale does not fix this. It is also the case that needs exact rejection sampling most: entropy is a dense look-ahead statistic. Every candidate token has its own value, and computing it takes a forward pass, so the tilt has no closed form and was judged impractical. Closed-form tilts in controlled generation instead rely on sparse or token-level statistics, such as the question-word and repetition indicators of weighted decoding or keyword-based shifts. On the consequences for moment matching itself, the tilt sweep below examines how next-token calibration transfers to full-sequence calibration, and Appendix D relates next-token calibration to reward maximization in a KL trust region.

Stress test. We ran the following experiment (new Appendix F). We fix the tilt at two and four times the fitted α_*, without refitting, on Qwen3-1.7B (WritingPrompts) and Qwen2.5-1.5B (WikiText). For clearer presentation, please check out the new appendix (F).

ModelTiltHeld-out NLLBrierGeneration entropyProposals per token
Qwen3-1.7BBase2.85130.72463.641
α_*2.84900.72473.491.19
2α_*2.84970.72523.311.40
4α_*2.86020.72793.001.90
Qwen2.5-1.5BBase2.12960.59303.151
α_*2.12570.59322.681.27
2α_*2.12900.59482.351.56
4α_*2.15690.60221.762.19

The method degrades only in cost and log-loss. Held-out log-loss is lowest at the fitted α_* and exceeds the base model's only at 4α_*. Meanwhile, generation entropy moves closer to the log-loss, and on Qwen2.5-1.5B 4α_* overshoots it, with entropy falling along the generation. The proposal count grows more slowly than it would at a fixed entropy level. One reason is that stronger tilts move generation into lower-entropy states, where fewer proposals are rejected.

In short, up to twice the fitted tilt, the method costs at most 1.6 proposals per token and no log-loss relative to the base model. At four times the fitted tilt, it costs about 2 proposals per token, and log-loss rises above the base model's.

Larger models. We would have liked to validate our results on larger models, but our compute was limited. Cao et al. show that miscalibration persists at scale, so the problem remains relevant, and our analysis predicts that larger models are cheaper to calibrate, since the cost grows with the tilt (Lemma 3) and a model with lower CE needs a smaller tilt (Lemma 5, Theorem 6). This is supported by the empirical evidence, and allows us to predict that rejection efficiency will improve for larger models.

Cheap gains first, then rejection sampling. Yes. A tilt whose normalizing constant is available from one forward pass, such as temperature, can be built into the proposal. One samples from the temperature-sharpened model and accepts with probability e^(αf). The result is exactly the combined tilt, and the rejection cost applies only to the remaining part. The catch is that the cheap step costs log-loss, which our method is designed to avoid, so both parameters are best fit jointly by CE minimization. Thank you for this question.

We hope the new stress test and these answers show that the theory makes the method's behavior predictable, including its limits, and that this supports raising the score.

Dear Reviewer,

We would appreciate it if you could let us know whether we have addressed your questions and concerns, or if there are any points that still require clarification. If you feel that we have satisfactorily addressed your concerns, we would kindly request you to reconsider your rating.

Thanks so much for your response, to me and to the other reviewers, as well as for your revision --- I have examined these. The revised portions further clarify the exposition (which I think was already mostly good even prior to that). For the empirics, I do appreciate the new stress-testing experiment, which gives a better sense of the method's limits.

I have carefully contemplated potentially raising the score as I had indicated I might, with all the new information in mind, and including my new thoughts and other reviewers' thoughts; ultimately, my honest decision is that I'll keep the numerical assessment, but with the understanding that my opinion of the paper has improved by some amount thanks to this revision/responses. Again, I do think the revisions have strengthened the work, and I appreciate the authors' efforts, but such is my posterior at this moment.

The main reason is that the empirical evaluation does truthfully remain a bit of a blocker for the highest score (while the methodology is reasonably natural/simple, which is good but also it cannot carry the paper "single-handedly"). Several desirable conclusions about the method's performance, such as evaluating on larger models/other metrics of quality/comparing to alternatives like min-p or temperature (which this one was noted by the other reviewers), are currently reached via indirect support by conclusions from existing works. And some initially declared potential future applications, such as to sycophancy etc., are also still out of reach; I do however appreciate the authors' transparent decision to remove some of these mentions from the paper. This is not detrimental per se, I believe (hence why my score/assessment is positive!), and I completely understand constraints such as lack of compute. On balance, though, my initial rating seems like a slightly more precise assessment of the manuscript than a "clear/unconditional accept" evaluation.

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

The paper proposes exact rejection sampling from the one-step-lookahead exponential tilt following Braverman et al. (2020), which avoids computing the partition function. The theory bounds the cost of evaluation, and experiments on Qwen2.5, Qwen3 and SmolLM3 show less entropy drift at about 10–30% extra proposals.

Critical strengths
  1. An exact, cheap sampler. It makes Braverman et al.’s local entropy calibration practical for modern LLMs at about 1.1–1.3× proposals. Braverman et al. had demonstrated it mainly on a 3-layer LSTM trained on Penn Treebank.
  2. Consistent transfer across several models and datasets. Local calibration partly fixes global calibration, apparently by steering trajectories away from high-entropy states. That is a real difference from temperature or truncation.
Critical weaknesses
  1. Thin novelty, overstated theory.
  • The tilt and the CE fitting of α come from Braverman et al.
  • Rejection sampling for exponentially tilted kernels has been done before, e.g. (Madhow et al. 2026) "Learnable Chernoff Baselines for Inference-Time Alignment" with adaptive acceptance probabilities.
  • Lemmas 2–5 are standard results.
  • No plot compares the bound with observed proposal counts.
  1. The experiments don’t test the motivating claims.
  • The only metric is entropy. There is no temperature or min-p baseline at matched entropy, no held-out CE, and no diversity or quality metric.
  • Efficiency is counted only in proposals, with no wall-clock time and no exact top-k tilt baseline.
  • There are no error bars.
  • The claim that efficiency improves with scale is unsupported. Qwen3-4B needs more proposals than 1.7B.
Questions for authors
  1. Is the forward pass of an accepted proposal reused for the next step? What is the wall-clock overhead?
  2. How does the method compare with temperature/min-p at matched entropy, and with exact top-k tilting, on held-out CE and on diversity/quality?
3 comments

We thank the reviewer for a precise review, and we are glad the reviewer sees our sampler as exact and cheap, making Braverman et al.'s calibration practical for modern LLMs. We also appreciate the observation that the tilt steers generation away from high-entropy states, which sets it apart from temperature scaling and truncation. The revision supports this observation in two ways. The tilted models lower entropy without losing held-out log-loss (Table 1), and stronger tilts flatten the growth of entropy along the generation (new tilt sweep, Appendix F).

Q1. Reuse and wall-clock. Yes, the forward pass is reused. Evaluating f(c, w) = H(Q_cw) requires running the model on the context with w appended, and that output is exactly the next-step distribution if w is accepted. Rejected proposals are masked out of the KV cache and periodically evicted (see the supplementary code). Each generated token therefore costs one forward pass per proposal, roughly 10–30% more than standard sampling. Wall-clock time depends on implementation choices, and our research code is not optimized, so we report the implementation-independent number of forward passes (Section 5.1).

Q2. Comparisons. We added held-out log-loss (NLL) and Brier score for the base and tilted models (Table 1). Temperature and min-p lower entropy by raising log-loss, which is well established (Zhang et al. 2020, Cao et al. 2025), while our tilt is chosen by minimizing CE and does not raise it. Top-k tilting is how Cao et al. approximated this same tilt. With k = 32 it was 33 times slower than standard sampling, according to their OpenReview discussion, and it is still only approximate. Our sampler is exact. The new tilt sweep (Appendix F) also shows what happens when our own tilt is pushed to lower entropy. At twice the fitted tilt, held-out log-loss is still below the base model's, and it exceeds it only at four times the fitted tilt.

Novelty. The tilt and the CE-based choice of α are due to Braverman et al., as the paper states. Our contribution is making this tilt usable for modern LLMs, where Braverman et al. could only demonstrate it on a small LSTM and Cao et al. judged it impractical. Madhow et al. (2026) address a different question. For a given tilt, they trade exactness for efficiency, which pays off for large tilts. We focus on choosing the tilt, and the tilts chosen by CE minimization are small (due to small CE improvement), so exact sampling is already cheap. We expanded this comparison in the related work. The techniques in Madhow et al. are complementary to our work, and could be applied to gain even faster sampling.

Theory. The lemmas are deliberately simple. Their role is to make the cost predictable before sampling and to expose the trade-off between the available CE improvement and the tilt. The revision adds the converse direction after Theorem 6: if the CE can improve only slightly, the fitted tilt must be small, ∣α∗∣≤2(ℓ(0)−ℓ(α∗))/Vmin⁡|\alpha_*|\le \sqrt{2(\ell(0)-\ell(\alpha_*))/V_{\min}}, and by Lemma 3 it is cheap to sample. Evaluating the bounds exactly requires the look-ahead statistic on the whole vocabulary at every visited context, so instead of plotting them we test their predictions. The tilt sweep (Appendix F) shows the proposal count growing with the tilt, as Lemma 3 predicts, and Table 1 explains the Qwen3-4B case below.

Model scale. This is a prediction of our bounds, and the revised text says so explicitly. The cost grows with the tilt (Lemma 3, Theorem 6), and a model with little room for CE improvement needs only a small tilt (Lemma 5 and the bound after Theorem 6). The new Table 1 shows that Qwen3-4B fits this picture. Its held-out CE gain is the largest in its family (2.9 × 10⁻³ nats, against 2.3 × 10⁻³ for 1.7B and 1.5 × 10⁻³ for 8B), so it needs a larger tilt (−0.066 versus −0.049 for 1.7B) and more proposals. Within Qwen2.5, the CE gain and the tilt both shrink with size (tilts of −0.084, −0.063 and −0.049 for 1.5B, 3B and 7B).

Error bars. Table 1 reports paired 95% bootstrap intervals over documents for the Brier score, and held-out NLL is lower than the base model's in all eight settings. Each entropy curve averages 1,000 documents (250 per setting in the tilt sweep). Adding bands to Figures 1–3 would require us to rerun the generations, and in our experience bands make these overlaid curves hard to read.

The revision answers the reviewer's requests with evidence: held-out log-loss and Brier scores, a tilt sweep beyond the fitted tilt, and an explanation of the Qwen3-4B case. A tilt that Cao et al. judged impractical now costs 10–30% extra sampling attempts and does not hurt log-loss. We believe this makes a strong case for the paper and would be grateful if the reviewer would consider raising their score.

Dear Reviewer,

We would appreciate it if you could let us know whether we have addressed your questions and concerns, or if there are any points that still require clarification. If you feel that we have satisfactorily addressed your concerns, we would kindly request you to reconsider your rating.

Reviewer_bBVtOpenReview ↗

Substantial rebuttal

I would like to thank the authors for the substantial rebuttal. My concerns were partially resolved and I raise my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.