Pith. sign in

REVIEW 5 major objections 5 minor 42 references

Lookahead resampling with entropy- and power-based rewards steers LLM decoding toward OR formulations whose short simulated continuations are most concentrated, giving reported pass@1 gains not yet separated from added inference budget.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 01:43 UTC pith:OLL5MUSS

load-bearing objection A sensible training-free decoding recipe for OR formulation generation with a correct but modest theory and a genuinely useful diagnostic — but the headline result isn't established until the evaluation gets budget-matched baselines, error bars, defined estimators, and held-out hyperparameters. the 5 major comments →

arxiv 2608.00019 v1 pith:OLL5MUSS submitted 2026-07-01 cs.LG cs.AI

Uncertainty-Aware Simulation-Based Inference for Operations Research with Large Language Models

Liang Guo , Lin Shaochong , Shen Zuo-Jun Max , Zhang Kun This is my paper
classification cs.LG cs.AI
keywords formulationcoherentdemonstratedownstreamframeworkgenerationinferencelanguage
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Large language models write answers one token at a time, choosing each token from local probabilities. For operations-research (OR) tasks — where the output is a mathematical model plus solver code — a locally plausible choice such as omitting an integer constraint can make the whole formulation wrong, and later steps rarely recover. The paper attacks this with a decoding-time search: at each block, it samples N candidate continuations, rolls each forward H tokens M times with the same frozen model, and computes a reward from the rollout cloud. The entropy reward penalizes candidates whose rollouts diverge widely; the power reward favors candidates whose future probability mass concentrates on few strong continuations. Candidates are then importance-resampled using combined local-likelihood and rollout rewards. A theorem shows this resampling converges to the ideal reweighted selection rule as N and M grow.

The experiments use ORLM-LLaMA-3-8B on NL4OPT, MAMO, and IndustryOR. Reported pass@1 correctness rises from 22.7% to 40.3% on hard MAMO instances and from 24.0% to 37.0% on IndustryOR, beating low-temperature sampling. Two caveats limit the claims. First, the method spends roughly 50x more tokens per answer than single-sample baselines, and no best-of-N or self-consistency baseline with the same budget is run, so part of the gain may be just more compute. Second, reward hyperparameters alpha and beta are tuned in a sensitivity analysis on the same benchmarks that are then reported, there are no error bars, and the estimator for the entropy reward from only four rollouts is never specified. The core idea is coherent, but the empirical support is not yet decisive.

Core claim

The abstract claims the framework 'consistently outperforms both standard and low-temperature baselines,' and Section 5 states power reward sampling 'achieves absolute correctness improvements of 17.6% and 9.0% over standard sampling on MAMOComplexLP and IndustryOR,' with entropy reward reaching 37.0% on IndustryOR. If correct: a decoder that samples N block continuations, simulates H-token rollouts M times, scores them with entropy/power rewards, and importance-resamples achieves higher pass@1 formulation correctness than myopic single-sample decoding on NL4OPT, MAMO, and IndustryOR; Theorem 4.1 guarantees the resampling recovers the ideal reward-guided local selector as N, M go to infinity.

Load-bearing premise

The load-bearing premise is the surrogate-reward bridge: that short-horizon concentration of the base model's own rollouts (low conditional entropy, high power moment) is a valid proxy for the probability a candidate prefix leads to a correct final OR formulation, and that this signal appears within the H=8 lookahead tokens used in Table 1. Section 4.1 asserts 'these uncertainties often reveal themselves over a relatively short horizon' and supports it with a single manually clustered example (Figure 4, 16 candidates, 32 full rollouts). If confident-but-wrong paths are also low-entropy, or errors surface only beyond the horizon, the resampling steers toward internally consistent but incorrect formulations, and the reported gains would be attributable to extra compute alone. The paper's own Section 6 concedes it 'measures uncertainty mainly through naive entropy.' This bridges rollout dispersion to terminal correctness — a different statement from the performance claim itself.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes a training-free, inference-time framework for LLM-based OR formulation generation. At each block-level decision point the algorithm samples N candidate blocks from the base model, generates M short H-token rollouts for each, computes a reward based on the base model's own future probability concentration (power reward) or predictive entropy (entropy reward), and then importance-resamples among the candidates. Section 4 formalizes the procedure as an approximation to an ideal reward-guided local selector, with Theorem 4.1 claiming convergence of the resampled distribution to the local reweighted target q_H as N,M grow. Section 5 reports pass@1 success/correctness rates on NL4OPT, MAMO (EasyLP/ComplexLP), and IndustryOR, comparing against standard sampling and low-temperature sampling. The paper also includes a motivating case study, a Pass@k versus CoT-Pass@k analysis, and a sensitivity study of the reward hyperparameters. The abstract claims consistent improvements over both baselines.

Significance. If the surrogate-reward bridge is valid, the framework is an interesting training-free alternative to RLVR-style lookahead: it requires no parameter updates or external reward models, has a clean asymptotic justification in Theorem 4.1, and operates on semantically meaningful blocks. The proof in Appendix A is carefully structured, and the sensitivity and budget experiments are a useful start. The framework is falsifiable and the paper is honest about its limitations (notably the naive entropy measure in Section 6). However, the empirical core has not yet established the headline claim: the proposed method consumes far more tokens per output than the baselines, and no budget-matched competitor is run. The central proxy—short-horizon rollout concentration as a predictor of terminal correctness—rests on one hand-clustered example. These gaps are fixable with additional experiments, but they are load-bearing for the paper's central claim.

major comments (5)
  1. [§5, Table 1; §4.2 token formula] The proposed runs use (N,M,B,H)=(10,4,8,8), giving roughly N T + (N M H / B) T = 50T tokens per output, while the standard and low-temperature baselines use T tokens. No best-of-N, self-consistency, majority-vote, or other budget-matched baseline is reported. The headline gains (17.6% on MAMOComplexLP, 9.0% on IndustryOR for power; 37.0% vs 24.0% on IndustryOR for entropy) may therefore be attributable to additional computation rather than to the lookahead/resampling mechanism. Budget-matched comparisons are needed before the claims 'consistently outperforms' and 'efficient' are supported.
  2. [§4.1, Figure 4; §6] The load-bearing premise is that short-horizon entropy/power concentration of the base model's own rollouts is positively correlated with V, the probability of terminal correctness. The only direct evidence is Figure 4: one MAMOEasyLP variable-definition step with 16 candidates and 32 full rollouts, manually clustered. No systematic correlation analysis is reported for the benchmarks in Table 1, and Section 6 concedes that the framework 'measures uncertainty mainly through naive entropy.' Theorem 4.1 proves convergence to q_H, not that q_H approximates the ideal value-based selector q* ∝ p exp(λV). The paper should provide a direct empirical or theoretical link between φ_H and V.
  3. [§4.2, Algorithm 1, Appendix A] The estimator φ_{M,H} is never defined. Algorithm 1 takes it as input, and the proof of Theorem 4.1 assumes φ_{M,H} converges to φ_H uniformly over a finite A_t, but no formula is given for the Monte Carlo estimate of H_t or of E[p(x>t|x≤t)^α]. Without a definition, the theorem is not checkable and the algorithm is not replicable. Please specify the estimator, including normalization, how the M rollouts of length H are used, and how variable-length or EOS-containing rollouts are handled.
  4. [§5, Figure 6 vs Table 1] The hyperparameters α and β are selected via sensitivity analysis on IndustryOR (Figure 6) and then evaluated on IndustryOR in Table 1, with no held-out protocol described. Because (α,β)=(3,0.2) appears to be chosen on the same benchmark whose results appear in Table 1, the reported IndustryOR gains may partly reflect in-sample selection. A validation-split or nested tuning protocol should be reported. The same concern applies to any hyperparameter or configuration choices made after seeing test-set results.
  5. [§5, Table 1] No variance estimates, confidence intervals, or significance tests are reported for the pass@1 numbers. Several comparisons are very close (e.g., 76.7% vs 76.3% on NL4OPT; 80.2% vs 79.0% on MAMOEasyLP), so the claim that the method 'consistently outperforms' would be more convincing with error bars or a paired significance test across the instances.
minor comments (5)
  1. [§3.2] The diet example is presented as a failure of low-temperature sampling, but the prompt does not state that food quantities must be integer servings. The 'correct model' with the integer restriction therefore appears to rely on an unstated assumption. Please clarify whether the ground-truth solution actually requires integrality, or rewrite the example so that the correct model follows from the stated problem.
  2. [§4.2] The quantity labeled E[token] is deterministic given N,M,B,H,T. Rename it 'token count' or 'total tokens consumed' to avoid confusion with an expectation over random generation.
  3. [Figure 2] The Pass@k curves are plotted without error bars or a description of how k samples are used to estimate the probabilities. Also, the axis labels inconsistently capitalize 'COT' and 'CoT'. Please standardize.
  4. [Figure 6(a)] If Eq. (4.2) is implemented as written, α=0 should make the power reward identically one, so the algorithm should reduce to standard sampling in distribution. The reported α=0 accuracy on IndustryOR (~0.19) differs markedly from the standard sampling correctness in Table 1 (0.24). This discrepancy is unexplained and may indicate an implementation or reporting mismatch that should be checked.
  5. [§4.1, Eq. (4.1)] The notation in φ_entropy uses log p(xt|x<t), which is the local probability of the candidate before appending; after appending, the conditional distribution over future tokens is p(·|x<t,xt). Clarify the notation by writing p(a|x<t) explicitly in Eq. (4.1), as is done in Eq. (4.3).

Circularity Check

1 steps flagged

Mild in-sample hyperparameter tuning on IndustryOR; central derivation is not circular.

specific steps
  1. fitted input called prediction [Section 5, Figure 6 and Table 1]
    "Figure 6 reports the parameter sensitivity of the two proposed reward-based sampling schemes... on IndustryOR. ... Entropy reward sampling (α, β) = (3,0.2) ... IndustryOR ... 37.0%"

    α and β are called 'tuning parameters' in Section 4.1. Figure 6 sweeps them on IndustryOR, and Table 1 then reports IndustryOR as evidence of superiority using the selected (α,β)=(3,0.2). The headline 37.0% IndustryOR gain is therefore an in-sample selected value from the same benchmark used to choose the hyperparameters, not an out-of-sample prediction. This is a mild evaluation-protocol circularity; the formal derivation (Eqs. 4.1–4.3, Theorem 4.1) is independent of this selection.

full rationale

The paper's formal derivation is not circular. Eq. (3.3) is a standard KL-regularized reward-maximization identity; Eqs. (4.1)–(4.2) define rollout-based rewards from the base model's own conditional distributions; Eq. (4.3) and Algorithm 1 define an importance-resampling target. Nothing in these equations defines the reward in terms of the terminal correctness V, so the claimed correlation between low rollout entropy and correctness is an empirical assumption rather than a definitional reduction. Theorem 4.1 is a consistency result for the Monte Carlo approximation of q_H, not a proof that q_H approximates V; it does not reduce to its own input, though its proof assumes (without defining φ_{M,H}) that φ_{M,H}→φ_H. That is an omitted definition/proof, not circularity. The only direct evidence for the entropy-correctness bridge is the single hand-clustered example in Figure 4, and Section 6 concedes the framework 'measures uncertainty mainly through naive entropy'—a validity risk rather than a circular step. There are no load-bearing self-citations: the Karan & Du and Zhao et al. references are external and used only as related ideas. The mild circularity-adjacent problem is that the reward hyperparameters α and β are selected on IndustryOR (Figure 6) and then IndustryOR appears in Table 1 with the selected values, making part of the reported empirical gain in-sample. This does not undermine the derivation itself, so the score is moderate rather than high.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

No invented entities: the paper introduces reward functions, not new causes or mechanisms. The main ledger items are one domain assumption (rollout dispersion proxies terminal correctness), two tuned reward coefficients (alpha, beta), and the N/M/B/H evaluation budget. The theory rests on standard probability. The central cost the reader pays is believing that a frozen model's short self-rollouts carry signal about its own long-term correctness — evidenced by one manually clustered example — while alpha and beta are set on the same benchmarks whose results are then reported.

free parameters (3)
  • alpha (reward shaping exponent) = 3
    Chosen from sensitivity runs on IndustryOR (Fig. 6a-b, alpha in {0,1,3,7,99}); accuracy improves with alpha up to a plateau. Tuned on the evaluation benchmark, not a held-out set.
  • beta (entropy penalty coefficient) = 0.2
    Chosen from Fig. 6c (beta in {0,0.1,0.2,0.4,0.8}); beta=0 is not optimal and large beta hurts, so the reported value is selected to maximize the benchmark result.
  • N, M, B, H (evaluation budget) = 10, 4, 8, 8
    Candidate blocks, rollouts per candidate, block length, and lookahead horizon for Table 1. Fig. 7 shows accuracy varies with N (24% to 33% as N goes 1 to 100), M, and H; chosen by hand. H=8 underlies the short-horizon surrogate assumption.
axioms (5)
  • domain assumption Downstream rollout concentration / low conditional entropy of a candidate prefix is positively correlated with the probability of a correct final OR formulation (Eqs. 4.1-4.2).
    Section 4.1; supported only by Figure 4 (a single MAMO instance with manual clustering). The paper itself notes a confident-wrong path can be low-entropy and counters with the local-likelihood term, but the proxy's validity is assumed, not tested.
  • domain assumption Structural OR errors reveal themselves within the lookahead horizon H (H=8 tokens in Table 1).
    Section 4.1: 'these uncertainties often reveal themselves over a relatively short horizon.' Appendix B gives one example. The short-horizon surrogate is the crux of the efficiency claim and is not systematically validated.
  • domain assumption OR final-answer correctness implies reasoning-trajectory correctness (CoT-Pass@k approximately equals Pass@k), so steering intermediate steps can improve final answers.
    Section 3.2 and Figure 2, shown on only two benchmarks (MAMOEasyLP and AIME2025). Used to justify trajectory-level evaluation; if error recovery were common, myopic sampling would not need fixing.
  • domain assumption The base model p already contains valid OR reasoning paths; inference steering only reallocates probability mass among them.
    Section 3.1 assumes the SFT model 'captures substantial task-relevant knowledge and reasoning patterns'; all reported gains depend on valid paths existing within the support of p.
  • standard math Customary probability machinery: law of large numbers, Hoeffding's inequality, dominated convergence, and finiteness of the candidate set A_t.
    Appendix A proof of Theorem 4.1. The theorem's statement is correct; the only gap is that the estimator phi_M,H for the entropy reward is never defined, so its assumed uniform convergence in probability is unverified.

pith-pipeline@v1.3.0-alltime-deepseek · 19212 in / 23058 out tokens · 203471 ms · 2026-08-04T01:43:10.452105+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Simulation-Based Inference for Operations Research with Large Language Models." pith.science (2026). https://pith.science/paper/OLL5MUSS

@misc{pith2026260800019,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Simulation-Based Inference for Operations Research with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OLL5MUSS}},
  note         = {Machine review of arXiv:2608.00019}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deploying large language models (LLMs) for operations research (OR) tasks remains challenging because correctness depends on a coherent modeling process, not merely a correct final answer. Standard autoregressive generation operates on a myopic policy, which sometimes fails to anticipate whether a partial formulation can be validly extended into a globally consistent optimization model. Consequently, locally plausible steps may propagate into catastrophic downstream formulation or solver code errors. To address this, we propose an uncertainty-aware, training-free inference framework for OR mathematical modeling. Without updating model parameters, our method evaluates intermediate candidate steps using short lookahead simulations to quantify downstream predictive uncertainty or probability concentration. Candidates that demonstrate a higher likelihood of yielding coherent mathematical formulations are then dynamically selected via importance resampling. Empirical evaluations across multiple OR benchmarks (including NL4OPT, MAMO, and IndustryOR) demonstrate that our framework consistently outperforms both standard and low-temperature baselines, establishing an efficient, training-free paradigm for reliable OR formulation generation.

Figures

Figures reproduced from arXiv: 2608.00019 by Liang Guo, Lin Shaochong, Shen Zuo-Jun Max, Zhang Kun.

Figure 1
Figure 1. Figure 1: How an LLM generates text step by step. Starting from the input question (prompt) [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pass@k versus COT-Pass@k can behave differently across domains. The key distinction is whether correct final answers can be produced without a correspondingly correct reasoning process [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The base LLM (top) serves strictly as a frozen probability engine. While standard [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A motivating illustration of the relationship between downstream dispersion and empirical [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the sampling procedure at time step [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity analysis of reward scaling hyperparameters [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pass@1 accuracy on IndustryOR using power reward sampling under various hyperparam [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 10 linked inside Pith

  1. [1]

    2025 , note =

    Aghazadeh and others , title =. 2025 , note =

    2025

  2. [2]

    Proceedings of the 41st International Conference on Machine Learning , series =

    Ali AhmadiTeshnizi and Wenzhi Gao and Madeleine Udell , title =. Proceedings of the 41st International Conference on Machine Learning , series =. 2024 , publisher =

    2024

  3. [3]

    2021 , eprint=

    Evaluating Large Language Models Trained on Code , author=. 2021 , eprint=

    2021

  4. [4]

    2026 , note =

    Zhen Chen and others , title =. 2026 , note =

    2026

  5. [5]

    Proceedings of the 12th International Conference on Learning Representations , year =

    Yung-Sung Chuang and Yujia Xie and Hongyin Luo and Yoon Kim and James Glass and Pengcheng He , title =. Proceedings of the 12th International Conference on Learning Representations , year =

  6. [6]

    Proceedings of the 40th AAAI Conference on Artificial Intelligence , year =

    Zhiyuan Ding and others , title =. Proceedings of the 40th AAAI Conference on Artificial Intelligence , year =

  7. [7]

    Nature , volume=

    Detecting hallucinations in large language models using semantic entropy , author=. Nature , volume=. 2024 , publisher=

    2024

  8. [8]

    Gordy and Sandeep Juneja , Journal =

    Michael B. Gordy and Sandeep Juneja , Journal =. Nested simulation in portfolio risk measurement , Volume =

  9. [9]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    Pith/arXiv arXiv

  10. [10]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Rewarding the unlikely: Lifting grpo beyond distribution sharpening , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  11. [11]

    arXiv preprint arXiv:2503.24290 , year=

    Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model , author=. arXiv preprint arXiv:2503.24290 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2405.13144 , year=

    MAMO: A Mathematical Modeling Benchmark with Solvers , author=. arXiv preprint arXiv:2405.13144 , year=

    Pith/arXiv arXiv

  13. [13]

    Operations Research , year=

    Orlm: A customizable framework in training large models for automated optimization modeling , author=. Operations Research , year=

  14. [14]

    Transactions of the Association for Computational Linguistics , volume =

    Zhengbao Jiang and Jun Araki and Haibo Ding and Graham Neubig , title =. Transactions of the Association for Computational Linguistics , volume =

  15. [15]

    Proceedings of the 13th International Conference on Learning Representations (ICLR) , year=

    LLMOPT: Learning to Define and Solve General Optimization Problems from Scratch , author=. Proceedings of the 13th International Conference on Learning Representations (ICLR) , year=

  16. [16]

    LLM for Large-Scale Optimization Model Auto-Formulation: A Lightweight Few-Shot Learning Approach , author=

  17. [17]

    Proceedings of the 12th International Conference on Learning Representations , year =

    Hunter Lightman and Vineet Kosaraju and Yura Burda and Harri Edwards and Bowen Baker and Teddy Lee and Jan Leike and John Schulman and Ilya Sutskever and Karl Cobbe , title =. Proceedings of the 12th International Conference on Learning Representations , year =

  18. [18]

    Smith and Yejin Choi , title =

    Alisa Liu and Maarten Sap and Ximing Lu and Swabha Swayamdipta and Chandra Bhagavatula and Noah A. Smith and Yejin Choi , title =. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics , year =

  19. [19]

    2022 , note =

    Saurav Kadavath and Tom Conerly and Amanda Askell and Tom Henighan and Dawn Drain and Ethan Perez and Nicholas Schiefer and Zac Hatfield-Dodds and Nova DasSarma and Eli Tran-Johnson and Scott Johnston and Sheer El-Showk and Andy Jones and Nelson Elhage and Tristan Hume and Anna Chen and Yuntao Bai and Sam Bowman and Stanislav Fort and Deep Ganguli and Dan...

    2022

  20. [20]

    Advances in Neural Information Processing Systems , volume =

    Sanyam Kapoor and Nate Gruver and Manley Roberts and Katherine Collins and Arka Pal and Umang Bhatt and Adrian Weller and Samuel Dooley and Micah Goldblum and Andrew Gordon Wilson , title =. Advances in Neural Information Processing Systems , volume =

  21. [21]

    arXiv preprint arXiv:2510.14901 , year=

    Reasoning with sampling: Your base model is smarter than you think , author=. arXiv preprint arXiv:2510.14901 , year=

    Pith/arXiv arXiv

  22. [22]

    Proceedings of the 11th International Conference on Learning Representations , year =

    Lorenz Kuhn and Yarin Gal and Sebastian Farquhar , title =. Proceedings of the 11th International Conference on Learning Representations , year =

  23. [23]

    arXiv preprint arXiv:2306.03081 , year=

    Sequential monte carlo steering of large language models using probabilistic programs , author=. arXiv preprint arXiv:2306.03081 , year=

    Pith/arXiv arXiv

  24. [24]

    Yu and Rindranirina Ramamonjison and Giuseppe Carenini and Zirui Zhou and Yong Zhang , title =

    Mahdi Mostajabdaveh and Timothy T. Yu and Rindranirina Ramamonjison and Giuseppe Carenini and Zirui Zhou and Yong Zhang , title =. Proceedings of the 39th AAAI Conference on Artificial Intelligence , year =

  25. [25]

    Jaakkola and Regina Barzilay , title =

    Victor Quach and Adam Fisch and Tal Schuster and Adam Yala and Jae Ho Sohn and Tommi S. Jaakkola and Regina Barzilay , title =. Proceedings of the 12th International Conference on Learning Representations , year =

  26. [26]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  27. [27]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    Augmenting operations research with auto-formulation of optimization models from problem descriptions , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    2022

  28. [28]

    NeurIPS 2022 competition track , pages=

    Nl4opt competition: Formulating optimization problems based on their natural language descriptions , author=. NeurIPS 2022 competition track , pages=. 2023 , organization=

    2022

  29. [29]

    arXiv preprint arXiv:2506.10947 , year=

    Spurious rewards: Rethinking training signals in rlvr , author=. arXiv preprint arXiv:2506.10947 , year=

    Pith/arXiv arXiv

  30. [30]

    International conference on machine learning , pages=

    Long horizon temperature scaling , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  31. [31]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =

    Zhenning Shi and Yijia Zhu and Yi Xie and Junhan Shi and Guorui Xie and Haotian Zhang and Yong Jiang and Congcong Miao and Qing Li , title =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =

    2025

  32. [32]

    Proceedings of the 11th International Conference on Learning Representations , year =

    Xuezhi Wang and Jason Wei and Dale Schuurmans and Quoc Le and Ed Chi and Sharan Narang and Aakanksha Chowdhery and Denny Zhou , title =. Proceedings of the 11th International Conference on Learning Representations , year =

  33. [33]

    2025 , eprint=

    Large Language Models in Operations Research: Methods, Applications, and Challenges , author=. 2025 , eprint=

    2025

  34. [34]

    2025 , note =

    Longxuan Wei and Yubo Zhang and Junchi Yan , title =. 2025 , note =

    2025

  35. [35]

    arXiv preprint arXiv:2506.14245 , year=

    Reinforcement learning with verifiable rewards implicitly incentivizes correct reasoning in base llms , author=. arXiv preprint arXiv:2506.14245 , year=

    Pith/arXiv arXiv

  36. [36]

    Proceedings of the 12th International Conference on Learning Representations , year =

    Ziyang Xiao and Dongxiang Zhang and Yangjun Wu and Lilin Xu and Yuan Wang and Xiongwei Han and Xiaojin Fu and Tao Zhong and Jia Zeng and Mingli Song and Gang Chen , title =. Proceedings of the 12th International Conference on Learning Representations , year =

  37. [37]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics , year =

    Kevin Yang and Dan Klein , title =. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics , year =

    2021

  38. [38]

    arXiv preprint arXiv:2407.09887 , year=

    OptiBench Meets ReSocratic: Measure and Improve LLMs for Optimization Modeling , author=. arXiv preprint arXiv:2407.09887 , year=

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2504.13837 , year=

    Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model? , author=. arXiv preprint arXiv:2504.13837 , year=

    Pith/arXiv arXiv

  40. [40]

    Bootstrap-based budget allocation for nested simulation , Volume =

    Zhang, Kun and Liu, Guangwu and Wang, Shiyu , Journal =. Bootstrap-based budget allocation for nested simulation , Volume =

  41. [41]

    2024 , note =

    Haiteng Zhao and others , title =. 2024 , note =

    2024

  42. [42]

    arXiv preprint arXiv:2404.17546 , year=

    Probabilistic inference in language models via twisted sequential monte carlo , author=. arXiv preprint arXiv:2404.17546 , year=

    Pith/arXiv arXiv