Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Optimizing Temperature for Language Models with Multi-Sample Inference

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A bend in a model's log-entropy curve pinpoints the near-optimal sampling temperature.

desk verdict TURN is a practical, label-free temperature rule that works well on the tested 13-model/2-task grid, but the entropy-spike mechanism needs a top-K sensitivity check before I'd fully trust it. read the letter →

arxiv 2502.05234 v2 pith:ASXEDYXC submitted 2025-02-07 cs.LG cs.AIcs.CL

Weihua Du , Yiming Yang , Sean Welleck This is my paper · ORCID
classification cs.LGcs.AIcs.CL
keywords temperatureselectionmulti-sampleinferencemajorityvotingbest-of-Ntoken-levelentropyturningpointlargelanguagemodelsself-consistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that for a given language model and task, the temperature that maximizes accuracy under multi-sample aggregation, whether majority voting or best-of-N sampling, can be read off a label-free signal: the point where the logarithm of the average token-level entropy of generated samples switches from concave to convex. The authors call this the entropy turning point (EntP) and build an algorithm, TURN, that selects the temperature at EntP, plus a small correction of +0.1 when the aggregation is best-of-N. They report that accuracy at the EntP temperature tracks the best grid-search accuracy with correlations of 0.9998 (MATH) and 0.9913 (MBPP), and that TURN lands inside the epsilon-optimal range in 12 of 13 and 11 of 13 model settings, with average performance drops of 0.32% and 0.59%. The paper also offers a stochastic process model with proper and improper tokens to explain why the entropy spike coincides with sample-quality collapse. If the claim holds, practitioners can tune inference temperature without a labeled validation set.

What carries the argument

The entropy turning point (EntP), defined on the curve $H(T)$, the average token-level entropy of model-generated samples at temperature $T$, as the smallest temperature at which the second derivative of $\log H(T)$ changes from negative to positive. The turning point localizes the sudden rise in entropy that the paper identifies with sample-quality collapse. The selection rule is: use the EntP temperature for majority voting, and EntP + 0.1 for best-of-N; entropy is estimated from randomly chosen problems and the top 1000 tokens per decoding step. A supplementary stochastic process model, in which the error rate rises when an improper token is sampled and falls when a proper token is sampled, reproduces the concave-to-convex shape and shows the improper-token fraction rising quickly past EntP.

What would settle it

Estimate the entropy curve for the same models on MATH and MBPP with top-K set to 500, 2000, and the full vocabulary, and record where the concavity-to-convexity transition falls. If the turning point moves by more than the temperature grid interval of 0.1 as K changes, the claimed signal is an artifact of truncation; if it stays fixed, the signal is a property of the model's full distribution. A second check is to find a task where sample quality does not collapse after the entropy spike, such as a constrained decoding task with a closed form, in which case TURN would systematically predict temperatures that are too low.

Watch

Extended reading notes

Core claim

The central discovery is empirical: across thirteen models spanning pretrained, instruction-tuned, and task-finetuned checkpoints, the temperature at which the log token-level entropy curve becomes convex matches the temperature range where multi-sample aggregation accuracy peaks. This holds for majority voting on MATH and best-of-N on MBPP, and the match is quantitative, with accuracy at the predicted temperature nearly identical to peak accuracy. The paper further shows that the optimal temperature is not universal: models whose training data are closer to the task favor higher temperatures, and the paper's token-level entropy distance correlates negatively with the midpoint of the optimal temperature range, with coefficients of -0.895 on MATH and -0.777 on MBPP. The entropy spike itself is shown to reflect a shift in the sampled distribution rather than mere softmax flattening, because constraining the model to greedy-decoded outputs at the same temperature does not reproduce the spike.

Load-bearing premise

The load-bearing premise is that the concavity-to-convexity transition in the log of average token-level entropy, estimated from a small sample of problems and only the top 1000 tokens per step, is a stable signal that the generated samples are about to collapse in quality, and that this collapse point sits at or just below the accuracy peak.

Editorial extensions

If this is right

  • Temperature can be set per model and per task without validation labels, replacing a single fixed default across all settings.
  • The same EntP rule transfers across aggregation strategies with only a 0.1 shift for best-of-N, so the method may generalize to other selection-based aggregators.
  • Entropy estimation is inexpensive: 40 samples per temperature give low prediction variance and a negligible performance drop, according to the paper's variance analysis.
  • The low-temperature entropy distance predicts which models prefer warm sampling and which prefer cool sampling, offering a principled rule for choosing temperature ranges by model-task proximity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The entropy spike may be a general detector of distributional shift or quality collapse for other decoding hyperparameters such as top-p, top-k, or min-p, since the stochastic process model identifies error-rate amplification as the underlying mechanism.
  • EntP might coincide with the temperature at which the model's predictive uncertainty becomes poorly calibrated; if so, TURN would double as a calibration-free uncertainty threshold.
  • A testable extension is to apply TURN to weighted majority voting or reward-model reranking beyond best-of-N; the paper's aggregation correction was measured on one dataset, and the same measurement procedure would presumably yield the needed offset for other aggregators.
  • The top-1000-token truncation is a potential confound; measuring EntP at several values of K would settle whether the turning point is a property of the full next-token distribution or an artifact of the truncated estimate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper studies temperature selection for multi-sample inference in large language models. It reports that the optimal temperature varies across models and aggregation strategies, proposes a label-free heuristic called TURN that selects the temperature at which the log token-level entropy curve changes from concave to convex (the entropy turning point, EntP), and adds a small aggregation-dependent offset for best-of-N. The method is evaluated with majority voting on MATH and best-of-N on MBPP across 13 models, with reported hit rates of 12/13 and 11/13, average performance drops of 0.32% and 0.59%, and high correlations with grid-search peak accuracy. A stochastic process model is introduced to provide an interpretable mechanism for the entropy spike.

Significance. If the empirical result holds, the paper makes a practically useful contribution: it offers a cheap, label-free way to set a hyperparameter that is usually fixed or tuned with validation data, and it does so across more model families than most temperature studies. The code release, the entropy-based estimator, and the toy model are useful assets. I do not see label circularity in the EntP rule, because Algorithm 1 does not use accuracy labels to select the turning point, and the best-of-N offset is fit on MATH and transferred to MBPP. The main unresolved concerns are statistical and methodological: the entropy estimator is truncated without sensitivity analysis, and the headline performance numbers are point estimates without confidence intervals. The skeptic's concern about top-K truncation lands and is not resolved by the manuscript.

major comments (4)
  1. [§4.1, Eq. (4), Algorithm 1] The EntP signal is computed from entropy truncated to the top-K=1000 tokens, but the paper reports no sensitivity analysis over K and does not state whether the top-K probabilities are renormalized before computing H. Because the tail of the next-token distribution carries increasing mass as temperature rises, a spurious inflection can appear at the point where mass begins to leave the top-1000 window; the greedy-output control in Fig. 4(a) does not isolate this truncation effect, and the stochastic-process model in §4.3 does not model truncation. Since TURN's entire prediction is the first grid point at which the second derivative of log H changes sign, the central result depends on the stability of this estimator. Please report full-vocabulary entropy curves, or at least several values of K, for the models in Table 1, state the normalization convention, and show that EntP moves by less than one grid step under these changes.
  2. [§5.4, Fig. 1(b), Table 1] The accuracy-at-EntP versus peak-accuracy correlations (0.9998 and 0.9913) are reported without confidence intervals and are not by themselves strong evidence for near-optimality, because accuracy at any temperature is largely determined by model quality, so a high correlation with peak accuracy would be expected even for a non-adaptive temperature. The load-bearing numbers are the hit rates, temperature gaps, and performance drops in Table 1, but these are also reported without uncertainty. Accuracy is computed on 200 (MATH) or 100 (MBPP) problems, the averages are driven by many 0% drops, and the claimed average drops of 0.32% and 0.59% are small relative to the expected finite-test-set noise. Please provide bootstrap confidence intervals for HR, TG, PD, and the correlation coefficients, and report the median and maximum drop in addition to the average.
  3. [Algorithm 1, line 14] The EntP detection rule 'find j* = arg min_j (d2ℓ/dt2 > 0)' is not a well-defined optimization problem on a finite grid. The paper should specify the finite-difference formula used to approximate the discrete second derivative, whether the positive sign must persist over a window of grid points, and how ties are broken. Because a one-step change in j* changes the selected temperature by 0.1 and directly changes the hit-rate and performance-drop results, the exact detection rule is load-bearing and must be reproducible from the text.
  4. [Appendix C, Table 4] The best-of-N offset β=0.1 is estimated as the average difference between optimal-range midpoints for best-of-N and majority voting across 13 models on MATH, and then applied to MBPP. This is a legitimate held-out protocol, but the per-model differences in Table 4 range from -0.1 to 0.4, and only the mean is reported. Please report the standard error of this estimate and a small sensitivity analysis over β (for example, 0, 0.05, 0.1, 0.15) for the MBPP hit rate and performance drop, so that the transfer of the offset is itself quantified.
minor comments (5)
  1. [§4.1, Fig. 4(a)] Please specify how the dotted 'fixed (greedy) generation' curve is computed: which temperature is used to evaluate the entropy of greedy outputs, and whether the same prompts and decoding settings are used as for the solid curve. Without this detail, the distribution-shift interpretation is incomplete.
  2. [§5.1, Algorithm 1] The temperature interval t=0.1 is stated, but t_max is not specified for the MATH and MBPP experiments; please state the exact grid range used in each evaluation.
  3. [Eq. (4)] The displayed formula for H(p) has typographical errors in the probability notation, and the text refers to 'Fig. 4(1)' where it should refer to 'Fig. 4(a)'.
  4. [§4.3] The stochastic-process model is presented as an explanation, but it is not fitted to the real entropy curves and its hyperparameters are set by hand. The text says the model 'proving that token-level entropy is a good indicator of sample quality'; 'proving' is too strong, and 'consistent with' would be more appropriate.
  5. [§5.4, Table 3] The variance analysis is reported for one model (Llama-3.1-8B-Instruct on MATH); please state whether the sample-size stability in Table 3 holds for other models and for the MBPP setting, or report at least one additional model.

Circularity Check

1 steps flagged · score 2.0 of 10

EntP temperature prediction is not circular; the only circular element is a minor, non-central stochastic-process 'proof' that encodes quality collapse in its definition of improper tokens.

  1. other [Section 4.3, Appendix B.1.2–B.1.7 and B.3, Figure 5]
    "All tokens are labeled proper or improper irrelevant to contexts, and the number of improper tokens (N1) is much larger than that of proper tokens (N0)... Improper tokens have uniformly low logits L1. Then, the initial error rate pinit is determined as the probability of selecting an improper token based on the logits and temperature... Figure 5(b) shows ... the percentage of improper tokens quickly increases after EntP... The behavior of the stochastic process model is consistent with our observations of language models, proving that token-level entropy is a good indicator of sample quality."

    In Appendix B.1.6, improper tokens receive probability xt/N1 and proper tokens receive 1−xt split by softmax, so the model's 'percentage of improper tokens' is exactly the constructed mass xt. Because xt is labeled as the probability of producing an improper (i.e., incorrect) token, an entropy spike produced by growth in xt is definitionally a quality drop. The model therefore does not independently prove that EntP marks quality collapse; the conclusion is encoded in the proper/improper labeling and in the large N1/low-logit L1 construction. This step is explanatory and is not used to compute TURN's label-free temperature predictions, so it is a minor, non-central circularity.

full rationale

The central TURN derivation (Algorithm 1) computes token-level entropy from model samples without accuracy labels and selects the concavity-to-convexity turning point of log H(T); the reported correlations and hit rates in Table 1 are empirical validations against grid-search accuracy, not fitted outputs. The best-of-N adaptation factor β=0.1 is a label-derived constant (Appendix C) transferred from MATH to MBPP; this is a disclosed calibration that qualifies the 'label-free' wording, but it does not force the per-model MBPP predictions because the EntP term is label-free and varies per model. Self-citations (e.g., Sun et al. 2024, Welleck et al. 2024, Toshniwal et al. 2024) appear as background, models, or software and are not load-bearing for the EntP claim. The top-K=1000 entropy truncation is a robustness risk rather than a circularity. The only circular element found is the Section 4.3/Appendix B stochastic-process 'proof', which builds quality collapse into the definition of improper tokens; it is minor and does not undermine the empirical temperature-selection results.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central prediction method relies on one label-derived constant (beta) and several hand-chosen analysis constants (epsilon, grid step, top-K, sample sizes). The entropy-to-quality-collapse link is an assumption supported by a toy model. The stochastic process model adds many free parameters but is explanatory rather than load-bearing for the TURN predictions.

free parameters (6)
  • Aggregation adaptation factor beta = 0.1 for best-of-N, 0 for majority voting
    Appendix C: computed as the average difference (0.092) between the midpoints of best-of-N and majority-voting optimal temperature ranges across 13 models on MATH; applied unchanged to MBPP.
  • Epsilon for optimal temperature range = 0.02
    Section 3.1: accuracy within epsilon of the peak defines the optimal range; used to compute hit rate, temperature gap, and performance drop.
  • Temperature grid interval t = 0.1
    Section 4.2 and Algorithm 1: candidate temperatures are evaluated every 0.1, so EntP precision and predicted temperatures are quantized to this grid.
  • Top-K for entropy estimation = 1000
    Section 4.1: entropy is computed only over the top 1000 tokens per step; this truncation could damp or shift the entropy spike.
  • Sample size for metric evaluation = 128
    Section 5.2: hit rate, temperature gap, and performance drop are computed under a large sample size of 128, chosen because optimal ranges change slowly beyond 32 samples.
  • Entropy estimation sample budget N = 8 times dataset size in main runs; 40 shown sufficient
    Section 5.4 and Table 3: variance of predicted temperature is low with N=40, but the main experiments use a larger budget.
assumptions (5)
  • domain assumption Accuracy-temperature curves are single-peaked, so the epsilon-optimal range is a contiguous interval around one peak.
    Section 3.1: 'the accuracy curve follows a single-peak pattern'; this assumption justifies the interval-based hit rate and the claim that EntP near the peak is near-optimal.
  • ad hoc to paper The entropy spike indicates sample quality collapse.
    Section 4.1: 'we infer that the sudden rise in the entropy curve implies a substantial drop in sample quality'; this is inferred from a controlled comparison and a toy model, not from direct token-correctness measurements.
  • ad hoc to paper The stochastic process model of proper/improper tokens with error-rate reinforcement captures real language model behavior.
    Appendix B: a Bernoulli error process with update rule x_{t+1}=1-(1-x_t)^alpha uses hand-set hyperparameters (N0=10, N1=30000, L0=0, L1=-10, sigma0=1) to reproduce the entropy curve shape; it is not validated against model internals.
  • standard math The softmax-with-temperature token distribution (with a regularization function F) is the correct generation model.
    Equation (2) defines the sampling distribution; this is the standard autoregressive LM sampling setup used throughout.
  • domain assumption Average token-level entropy estimated from randomly selected problems and top-K truncation approximates the true entropy curve.
    Section 4.2 and Table 3: variance is measured only for Llama3.1-8B-Instruct on MATH; the stability of EntP across other models and tasks is assumed.
invented entities (2)
  • Entropy Turning Point (EntP) independent evidence
    purpose: The temperature where the log token-level entropy curve transitions from concave to convex; used as the predicted optimal temperature for multi-sample aggregation.
    EntP is directly observable in entropy curves (Figures 1, 4, 14-15), but its claimed link to optimal accuracy is an empirical correlation rather than an independently verified mechanism.
  • Noise tolerance rate alpha
    purpose: Controls how quickly the error rate escalates after improper tokens in the toy stochastic process model.
    Hand-set in Appendix B across values 1.5-3.0; there is no external measurement or falsifiable prediction tied to this parameter.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimizing Temperature for Language Models with Multi-Sample Inference." pith.science (2026). https://pith.science/paper/ASXEDYXC

@misc{pith2026250205234,
  author       = {Pith},
  title        = {Pith review of: Optimizing Temperature for Language Models with Multi-Sample Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ASXEDYXC}},
  note         = {Machine review of arXiv:2502.05234}
}
read the original abstract

Multi-sample aggregation strategies, such as majority voting and best-of-N sampling, are widely used in contemporary large language models (LLMs) to enhance predictive accuracy across various tasks. A key challenge in this process is temperature selection, which significantly impacts model performance. Existing approaches either rely on a fixed default temperature or require labeled validation data for tuning, which are often scarce and difficult to obtain. This paper addresses the challenge of automatically identifying the (near)-optimal temperature for different LLMs using multi-sample aggregation strategies, without relying on task-specific validation data. We provide a comprehensive analysis of temperature's role in performance optimization, considering variations in model architectures, datasets, task types, model sizes, and predictive accuracy. Furthermore, we propose a novel entropy-based metric for automated temperature optimization, which consistently outperforms fixed-temperature baselines. Additionally, we incorporate a stochastic process model to enhance interpretability, offering deeper insights into the relationship between temperature and model performance.

Figures

Figures reproduced from arXiv: 2502.05234 by the authors.

Figure 1
Figure 1. (a) The entropy turning point (EntP) (green star) is defined as the temperature point where the log-scale of the token￾level entropy of generation (red line) shifts from concave to convex, implying the sudden spike in the entropy curve (blue line). (b) The accuracy tested at EntP is highly correlated with the best accuracy from grid search over temperatures on the MATH dataset. reasoning (Ahn et al., 2024; Sun et al… view at source ↗
Figure 2
Figure 2. (a) Accuracy Heatmap. Performance of Mistral-7B-Instruct-v0.3 under majority voting across different temperatures. The best temperature for each sampling size is highlighted in bold white, and the optimal temperature range is shaded white. The green line shows the temperature predicted by our method. (b) Midpoint of Optimal Temperature Range vs. Number of Samples. The optimal temperature range varies by model; those… view at source ↗
Figure 3
Figure 3. Plot of midpoints of optimal temperature ranges (x-axis, sample size 128) vs. distances between models and tasks (y-axis). A strong negative correlation is observed on the MATH and MBPP datasets, with correlation coefficients of -0.895 and -0.777. target task and its corresponding optimal temperature. Our key insight is that token-level entropy is a proxy of distance from a task, which motivates our entropy-based me… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Entropy Curve Characteristics. (a) The token-level entropy H (solid blue line) increases slowly at lower temperatures and then jumps sharply at a critical turning point. In contrast, the entropy for a fixed (greedy) generation stays low (dotted blue line). log(H) (red …
Figure 5
Figure 5. Figure 5: Stochastic Process Model. We run our process model in the setting: N0 = 10, N1 = 30000, L0 = 0, σ0 = 1, L1 = −10, and α = 2. (a) The entropy curve is similar to that of the real language model: flat at first, and then sharply increases. (b) We calculate the relation be…
Figure 6
Figure 6. Figure 6: The temperature-entropy curves [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The temperature-log entropy curves [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 9
Figure 9. Figure 9: The Temperature-Improper Token (%) curves [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: The temperature-entropy curves [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: The temperature-log entropy curves. In the case of majority voting, the turning point on the entropy curve aligns with the optimal temperature, so we set its adaptation to 0. For best-of-N, we computed an adaptation on MATH and then tested it on MBPP to confirm genera…
Figure 12
Figure 12. Figure 12: The accuracy heatmap for all tested models on the MATH dataset. The green line is our predicted temperature. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: The accuracy heatmap for all tested models on the MBPP dataset. The green line is our predicted temperature. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: The entropy curves and turning points of language models when testing on the MATH dataset. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: The entropy curves and turning points of language models when testing on the MBPP dataset. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. When Life Gives You Samples: The Benefits of Scaling up Inference Compute for Multilingual LLMs

    cs.CL 2025-06 conditional novelty 6.0 of 10

    Hedged sampling, checklist-based one-pass selection (CHOPS), and cross-lingual MBR (X-MBR) improve multilingual LLM output quality when scaling from one to five samples.

  2. The Paradox of Stochasticity: Limited Creativity and Computational Decoupling in Temperature-Varied LLM Outputs of Structured Fictional Data

    cs.LG 2025-02 conditional novelty 4.0 of 10

    In three LLMs generating fictional names and birthdates, model choice dominates processing time and default name archetypes persist across temperature, while rare names appear mainly at mid-range temperatures.

Reference graph

Works this paper leans on

30 extracted references · 2 canonical work pages · cited by 2 Pith papers

  1. [2]

    Program synthesis with large language models

    Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al. Program synthesis with large language models. arXiv preprint arXiv:2108.07732,

    arXiv

  2. [5]

    Adaptive Decoding via Latent Preference Optimization

    Dhuliawala, S., Kulikov, I., Yu, P., Celikyilmaz, A., Weston, J., Sukhbaatar, S., and Lanchantin, J. Adaptive decod- ing via latent preference optimization. arXiv preprint arXiv:2411.09661,

    work page Pith review arXiv

  3. [6]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783,

    arXiv

  4. [7]

    Deepseek-coder: When the large language model meets programming–the rise of code intelligence.arXiv preprint arXiv:2401.14196,

    Guo, D., Zhu, Q., Yang, D., Xie, Z., Dong, K., Zhang, W., Chen, G., Bi, X., Wu, Y ., Li, Y ., et al. Deepseek-coder: When the large language model meets programming–the rise of code intelligence.arXiv preprint arXiv:2401.14196,

    arXiv

  5. [8]

    Measuring math- ematical problem solving with the math dataset

    Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring math- ematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874,

    arXiv

  6. [9]

    The curious case of neural text degeneration

    Holtzman, A., Buys, J., Du, L., Forbes, M., and Choi, Y . The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751,

    arXiv 1904

  7. [11]

    Openai o1 system card

    Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Car- ney, A., et al. Openai o1 system card. arXiv preprint arXiv:2412.16720,

    arXiv

  8. [12]

    Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D

    Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. d. l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al. Mistral 7b. arXiv preprint arXiv:2310.06825,

    arXiv

Show all 30 references
  1. [14]

    and Sarawagi, S

    Kumar, A. and Sarawagi, S. Calibration of encoder decoder models for neural machine translation. arXiv preprint arXiv:1903.00802,

    1903 arXiv

  2. [15]

    Dynamic stochastic decoding strategy for open-domain dialogue generation

    Li, Y ., Mi, F., Li, Y ., Wang, Y ., Sun, B., Feng, S., and Li, K. Dynamic stochastic decoding strategy for open-domain dialogue generation. arXiv preprint arXiv:2406.07850,

    arXiv

  3. [16]

    Lean-star: Learn- ing to interleave thinking and proving

    Lin, H., Sun, Z., Yang, Y ., and Welleck, S. Lean-star: Learn- ing to interleave thinking and proving. arXiv preprint arXiv:2407.10040,

    arXiv

  4. [17]

    B., Rus, D., Gan, C., and Matusik, W

    Ma, P., Wang, T.-H., Guo, M., Sun, Z., Tenenbaum, J. B., Rus, D., Gan, C., and Matusik, W. Llm and simulation as bilevel optimizers: A new paradigm to advance physical scientific discovery. arXiv preprint arXiv:2405.09783,

    arXiv

  5. [18]

    Turning up the heat: Min-p sampling for creative and coherent llm outputs

    Nguyen, M., Baker, A., Neo, C., Roush, A., Kirsch, A., and Shwartz-Ziv, R. Turning up the heat: Min-p sampling for creative and coherent llm outputs. arXiv preprint arXiv:2407.01082,

  6. [19]

    and Guven, E

    Renze, M. and Guven, E. The effect of sampling tempera- ture on problem solving in large language models. arXiv preprint arXiv:2402.05201,

    arXiv

  7. [20]

    E., Adi, Y ., Liu, J., Sauvestre, R., Remez, T., et al

    Roziere, B., Gehring, J., Gloeckle, F., Sootla, S., Gat, I., Tan, X. E., Adi, Y ., Liu, J., Sauvestre, R., Remez, T., et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950,

    arXiv

  8. [21]

    Deepseekmath: Push- ing the limits of mathematical reasoning in open language models

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y ., Wu, Y ., et al. Deepseekmath: Push- ing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300,

    arXiv

  9. [22]

    Easy-to-hard generalization: Scalable alignment beyond human supervision

    Sun, Z., Yu, L., Shen, Y ., Liu, W., Yang, Y ., Welleck, S., and Gan, C. Easy-to-hard generalization: Scalable alignment beyond human supervision. arXiv preprint arXiv:2403.09472,

    arXiv

  10. [23]

    Openmathinstruct-2: Accelerating ai for math with massive open-source in- struction data

    Toshniwal, S., Du, W., Moshkov, I., Kisacanin, B., Ayrapetyan, A., and Gitman, I. Openmathinstruct-2: Accelerating ai for math with massive open-source in- struction data. arXiv preprint arXiv:2410.01560,

    arXiv

  11. [24]

    Planning in natural language improves llm search for code generation

    Wang, E., Cassano, F., Wu, C., Bai, Y ., Song, W., Nath, V ., Han, Z., Hendryx, S., Yue, S., and Zhang, H. Planning in natural language improves llm search for code generation. arXiv preprint arXiv:2409.03733, 2024a. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., ...

    arXiv

  12. [25]

    From de- coding to meta-generation: Inference-time algorithms for large language models

    Welleck, S., Bertsch, A., Finlayson, M., Schoelkopf, H., Xie, A., Neubig, G., Kulikov, I., and Harchaoui, Z. From de- coding to meta-generation: Inference-time algorithms for large language models. arXiv preprint arXiv:2406.16838,

    arXiv

  13. [26]

    Infer- ence scaling laws: An empirical analysis of compute- optimal inference for problem-solving with language models

    Wu, Y ., Sun, Z., Li, S., Welleck, S., and Yang, Y . Infer- ence scaling laws: An empirical analysis of compute- optimal inference for problem-solving with language models. arXiv preprint arXiv:2408.00724,

    arXiv

  14. [27]

    S., Lee, Y ., Mitchell, E., and Finn, C

    Xie, J., Chen, A. S., Lee, Y ., Mitchell, E., and Finn, C. Calibrating language models with adaptive temperature scaling. arXiv preprint arXiv:2409.19817,

    arXiv

  15. [28]

    B., Shu, T., and Gan, C

    10 Optimizing Temperature for Language Models with Multi-Sample Inference Zhang, H., Du, W., Shan, J., Zhou, Q., Du, Y ., Tenenbaum, J. B., Shu, T., and Gan, C. Building cooperative embodied agents modularly with large language models. arXiv preprint arXiv:2307.02485,

    arXiv

  16. [29]

    Y ., and Li, L

    Zhang, K., Zhou, S., Wang, D., Wang, W. Y ., and Li, L. Scaling llm inference with optimized sample compute allocation. arXiv preprint arXiv:2410.22480, 2024a. Zhang, S., Bao, Y ., and Huang, S. Edt: Improving large language models’ generation by entropy-based dynamic temperat...

    arXiv

  17. [30]

    No additional sampling constraints are imposed for the MBPP dataset. A.3. Metric Calculation To compute the majority vote results for the MATH dataset, we consider two samples to have the same answer if they match after normalization. For the pass@K metric, we follow the defin...

    2021

  18. [2019]

    Hui, B., Yang, J., Cui, Z., Yang, J., Liu, D., Zhang, L., Liu, T., Zhang, J., Yu, B., Lu, K., et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186,

    arXiv

  19. [2021]

    D., McAleer, S., Jiang, A

    Azerbayev, Z., Schoelkopf, H., Paster, K., Santos, M. D., McAleer, S., Jiang, A. Q., Deng, J., Biderman, S., and Welleck, S. Llemma: An open language model for math- ematics. arXiv preprint arXiv:2310.10631,

    arXiv

  20. [2022]

    Chen, M., Tworek, J., Jun, H., Yuan, Q., Pinto, H. P. D. O., Kaplan, J., Edwards, H., Burda, Y ., Joseph, N., Brockman, G., et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374,

    arXiv

  21. [2023]

    L., and Rafiei, D

    Kamalloo, E., Dziri, N., Clarke, C. L., and Rafiei, D. Eval- uating open-domain question answering in the era of large language models. arXiv preprint arXiv:2305.06984,

    arXiv

  22. [2024]

    co/blog/lorinma/yi-coder

    URL https://huggingface. co/blog/lorinma/yi-coder. Ahn, J., Verma, R., Lou, R., Liu, D., Zhang, R., and Yin, W. Large language models for mathematical reasoning: Pro- gresses and challenges. arXiv preprint arXiv:2402.00157,

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.