pith. sign in

arxiv: 2606.03102 · v1 · pith:AY2VTHFUnew · submitted 2026-06-02 · 💻 cs.CL

Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling for Test-Time Scaling

Runpeng Dai , Tong Zheng , Rui Liu , Chengsong Huang , Hongtu Zhu This is my paper

Pith reviewed 2026-06-28 10:22 UTC · model grok-4.3

classification 💻 cs.CL
keywords adaptive samplingtest-time scalingreinforcement learninglarge language modelsMarkov decision processsampling controlleranswer correctnesscomputation cost
0
0 comments X

The pith

A reinforcement learning controller trained on answer statistics decides when to stop sampling from large language models to balance correctness against latency and cost.

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

The paper formulates the decision of when to stop generating additional samples as a Markov decision process. It then trains a small RL agent to choose stop or continue at each step, optimizing for a combination of answer quality and resource use. This approach is designed to be lightweight, requiring only statistics from the final answers rather than internal model states. A reader would care because current test-time methods for improving LLM reasoning are computationally expensive, and this offers a principled way to adapt the number of samples dynamically.

Core claim

By casting adaptive sampling as an MDP and training a lightweight RL controller, the method jointly optimizes for answer correctness, sampling rounds, and total samples, achieving better trade-offs than heuristic baselines such as ASC and ESC. The framework also admits an interpretation as the Lagrangian relaxation of a constrained optimization problem with explicit budget constraints. The controller decides at each round whether to stop or acquire additional samples.

What carries the argument

The Markov decision process formulation of the sampling decision, with the RL-trained lightweight controller that takes statistics of final answers as state and outputs stop or continue actions.

If this is right

  • The controller relies only on final answer statistics and can be trained and deployed on CPU.
  • The method improves trade-offs among answer correctness, sampling rounds, and total samples required compared to strong baselines.
  • The resulting framework can be viewed as the Lagrangian relaxation of a constrained optimization problem with explicit budget constraints.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the controller generalizes across different LLMs without retraining, it could act as a reusable module attached to any base model.
  • The MDP structure could allow reward functions that penalize answer diversity loss in addition to correctness and cost.
  • Extending the state to include partial reasoning traces might increase decision quality at the price of added complexity.

Load-bearing premise

Statistics of final answers alone are sufficient for the controller to make effective stop-or-continue decisions without needing model internals or additional context.

What would settle it

If experiments on reasoning benchmarks show that the RL controller requires more total samples than ASC or ESC to reach equivalent accuracy levels, the claim of improved trade-offs would be falsified.

Figures

Figures reproduced from arXiv: 2606.03102 by Chengsong Huang, Hongtu Zhu, Rui Liu, Runpeng Dai, Tong Zheng.

Figure 1
Figure 1. Figure 1: Overview of the RL-Guided Sampling framework. Top two blocks illustrates mechanism of two adaptive sampling baselines. ASC sequentially samples one response at a time and stops when the posterior probability exceeds a predefined threshold. ESC samples in fixed batches and stops only when intra-batch consistency is achieved. Different from those approaches, RL-Guided Sampling guides the sampling process via… view at source ↗
Figure 3
Figure 3. Figure 3: Correlation between the average total samples [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy scaling behavior of RL-Guided Sam [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy–token scaling curves comparing the SC, ASC, ESC and RL-Guided Sampling. across different models and benchmarks. Results are generated with Qwen3-4B-Instruct on the AIME24 and AIME25 datasets. metrics, [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Correlation between total samples per query [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Accuracy–token scaling curves(right) and Accuracy–sampling scaling curves(left) comparing the SC, ASC, [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
read the original abstract

Test-time scaling improves the reasoning performance of large language models but incurs substantial cost in both total computation and latency. Existing adaptive sampling methods partially mitigate this issue by dynamically deciding when to stop sampling, yet they typically rely on heuristic rules or rely on distribution assumptions. In this work, we formulate adaptive sampling as a Markov decision process (MDP). We train a lightweight sampling controller with reinforcement learning (RL) to jointly balance answer correctness, latency, and computation cost. At each round, the controller decides to stop sampling or to acquire additional samples. Our method is lightweight which only relies on statistics of final answers, and can be trained and deployed on CPU. We further show that the resulting framework admits an interpretation as the Lagrangian relaxation of a constrained optimization problem with explicit budget constraints. Experiments against strong baselines such as ASC and ESC show that our method achieves improved trade-offs among answer correctness, sampling rounds, and total samples required.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript formulates adaptive sampling during test-time scaling of LLMs as a Markov decision process. A lightweight controller is trained via reinforcement learning to decide at each round whether to stop or acquire additional samples; the controller uses only aggregate statistics of the sampled final answers as its state and is trained to jointly optimize answer correctness against latency and total compute. The framework is shown to admit a Lagrangian-relaxation interpretation of a constrained optimization problem with explicit budgets. Experiments against ASC and ESC baselines report improved trade-offs among correctness, number of sampling rounds, and total samples.

Significance. If the empirical gains are robust, the work supplies a model-agnostic, CPU-deployable controller that replaces heuristic stopping rules with a learned policy, while the Lagrangian view supplies a clean optimization-theoretic framing. These features would be useful for practical deployment of test-time scaling under resource constraints.

major comments (2)
  1. [§3 (MDP formulation)] §3 (MDP formulation): the state is restricted to statistics of final answers alone. Because these statistics cannot distinguish high-agreement correct answers from high-agreement incorrect answers, the learned policy has no signal to avoid premature stopping on consistent errors; this directly undermines the central claim that the RL controller produces improved correctness–cost trade-offs.
  2. [§5 (Experiments)] §5 (Experiments): the reward function, the precise definition of the MDP state vector, and the training protocol (including how the controller is optimized and whether it is retrained per LLM or task) are not specified in sufficient detail to allow reproduction or to verify that the reported gains are not artifacts of the particular reward shaping.
minor comments (2)
  1. [§4] The Lagrangian interpretation is presented as an after-the-fact view; making explicit how the dual variables are handled during RL training would strengthen the connection between the two framings.
  2. [§3] Notation for the per-round statistics (e.g., agreement rate, variance) is introduced without a consolidated table; a single table listing all state features would improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and constructive comments. We address the major comments point by point below and will revise the manuscript accordingly where appropriate.

read point-by-point responses

  1. Referee: [§3 (MDP formulation)] §3 (MDP formulation): the state is restricted to statistics of final answers alone. Because these statistics cannot distinguish high-agreement correct answers from high-agreement incorrect answers, the learned policy has no signal to avoid premature stopping on consistent errors; this directly undermines the central claim that the RL controller produces improved correctness–cost trade-offs.

    Authors: We appreciate this observation regarding the limitations of the state representation. The state indeed consists solely of aggregate statistics (e.g., answer frequencies, entropy) without access to ground-truth correctness during inference. However, during RL training, the reward function incorporates a correctness term (based on matching to reference answers in the training data), allowing the policy to learn stopping decisions that correlate with improved expected correctness in the distribution of tasks. The empirical results demonstrate that this leads to better trade-offs compared to baselines, suggesting that the learned policy effectively avoids many consistent error cases through patterns in the statistics. We will add a discussion of this limitation and potential extensions (e.g., incorporating model confidence if available) in the revised manuscript. revision: partial

  2. Referee: [§5 (Experiments)] §5 (Experiments): the reward function, the precise definition of the MDP state vector, and the training protocol (including how the controller is optimized and whether it is retrained per LLM or task) are not specified in sufficient detail to allow reproduction or to verify that the reported gains are not artifacts of the particular reward shaping.

    Authors: We agree that additional details are necessary for reproducibility. In the revised version, we will include the exact formulation of the reward function (including weights for correctness, latency, and compute), the full definition of the state vector components, the training algorithm (e.g., PPO or similar), hyperparameters, and clarification on whether the controller is trained per model/task or in a general manner. This will allow verification of the results. revision: yes

Circularity Check

0 steps flagged

No significant circularity in RL controller training or MDP formulation

full rationale

The paper formulates adaptive sampling as an MDP and trains a lightweight RL controller on final-answer statistics to balance correctness, latency, and cost. This is an empirical learning procedure whose policy is obtained via optimization against an external reward signal, not a closed-form derivation that reduces to its own inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems, no fitted parameters are relabeled as predictions, and no ansatz is smuggled through prior work. The Lagrangian interpretation is presented as an after-the-fact view of the trained policy rather than a definitional equivalence. The method is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, invented entities, or non-standard axioms are described.

axioms (1)
  • domain assumption Adaptive sampling can be formulated as an MDP whose state depends only on statistics of generated answers.
    Directly stated as the starting point of the method.

pith-pipeline@v0.9.1-grok · 5697 in / 1081 out tokens · 37154 ms · 2026-06-28T10:22:32.569108+00:00 · methodology

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

141 extracted references · 74 canonical work pages · 37 internal anchors

  1. [1]

    arXiv preprint arXiv:1911.10422 , year=

    Semeval-2010 task 8: Multi-way classification of semantic relations between pairs of nominals , author=. arXiv preprint arXiv:1911.10422 , year=

    work page arXiv 2010

  2. [2]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

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

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning

    Logic-rl: Unleashing llm reasoning with rule-based reinforcement learning , author=. arXiv preprint arXiv:2502.14768 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    The Twelfth International Conference on Learning Representations , year=

    Let's verify step by step , author=. The Twelfth International Conference on Learning Representations , year=

  5. [5]

    Alphazero-like Tree-Search can Guide Large Language Model Decoding and Training.arXiv preprint arXiv:2309.17179, 2023

    Alphazero-like tree-search can guide large language model decoding and training , author=. arXiv preprint arXiv:2309.17179 , year=

    work page arXiv

  6. [6]

    arXiv preprint arXiv:2410.06508 , year=

    Towards Self-Improvement of LLMs via MCTS: Leveraging Stepwise Knowledge with Curriculum Preference Learning , author=. arXiv preprint arXiv:2410.06508 , year=

    work page arXiv

  7. [7]

    arXiv preprint arXiv:2504.10160 , year=

    MT-R1-Zero: Advancing LLM-based Machine Translation via R1-Zero-like Reinforcement Learning , author=. arXiv preprint arXiv:2504.10160 , year=

    work page arXiv

  8. [8]

    ACM Computing Surveys , volume=

    A comprehensive survey on relation extraction: Recent advances and new frontiers , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    2024

  9. [9]

    Cognitive computation , volume=

    Deep neural approaches to relation triplets extraction: a comprehensive survey , author=. Cognitive computation , volume=. 2021 , publisher=

    2021

  10. [10]

    ACM Computing Surveys , volume=

    A comprehensive survey on automatic knowledge graph construction , author=. ACM Computing Surveys , volume=. 2023 , publisher=

    2023

  11. [11]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Hipporag: Neurobiologically inspired long-term memory for large language models , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  12. [12]

    From Local to Global: A Graph RAG Approach to Query-Focused Summarization

    From local to global: A graph rag approach to query-focused summarization , author=. arXiv preprint arXiv:2404.16130 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2310.01061 , year=

    Reasoning on graphs: Faithful and interpretable large language model reasoning , author=. arXiv preprint arXiv:2310.01061 , year=

    work page arXiv

  14. [14]

    Computers in Biology and Medicine , volume=

    Alzheimer's disease knowledge graph enhances knowledge discovery and disease prediction , author=. Computers in Biology and Medicine , volume=. 2025 , publisher=

    2025

  15. [15]

    Nature Communications , volume=

    Large language model powered knowledge graph construction for mental health exploration , author=. Nature Communications , volume=. 2025 , publisher=

    2025

  16. [16]

    Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages=

    AliMeKG: Domain knowledge graph construction and application in e-commerce , author=. Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages=

  17. [17]

    World Wide Web , volume=

    Llms for knowledge graph construction and reasoning: Recent capabilities and future opportunities , author=. World Wide Web , volume=. 2024 , publisher=

    2024

  18. [18]

    BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension

    Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension , author=. arXiv preprint arXiv:1910.13461 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  19. [19]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  20. [20]

    LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

    LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling , author=. arXiv preprint arXiv:2605.08083 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    Do Not Waste Your Rollouts: Recycling Search Experience for Efficient Test-Time Scaling

    Do Not Waste Your Rollouts: Recycling Search Experience for Efficient Test-Time Scaling , author=. arXiv preprint arXiv:2601.21684 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  23. [23]

    Proceedings of the conference

    Revisiting relation extraction in the era of large language models , author=. Proceedings of the conference. Association for Computational Linguistics. Meeting , volume=

  24. [24]

    arXiv preprint arXiv:2305.01555 , year=

    How to unleash the power of large language models for few-shot relation extraction? , author=. arXiv preprint arXiv:2305.01555 , year=

    work page arXiv

  25. [25]

    arXiv preprint arXiv:2310.07641 , year=

    Evaluating large language models at evaluating instruction following , author=. arXiv preprint arXiv:2310.07641 , year=

    work page arXiv

  26. [26]

    SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training

    Sft memorizes, rl generalizes: A comparative study of foundation model post-training , author=. arXiv preprint arXiv:2501.17161 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    ECAI 2020 , pages=

    Span-based joint entity and relation extraction with transformer pre-training , author=. ECAI 2020 , pages=. 2020 , publisher=

    2020

  28. [28]

    2018 , publisher=

    Reinforcement Learning: An Introduction , author=. 2018 , publisher=

    2018

  29. [29]

    arXiv preprint arXiv:2503.01491 , year=

    What's Behind PPO's Collapse in Long-CoT? Value Optimization Holds the Secret , author=. arXiv preprint arXiv:2503.01491 , year=

    work page arXiv

  30. [30]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    VAPO: Efficient and Reliable Reinforcement Learning for Advanced Reasoning Tasks

    Vapo: Efficient and reliable reinforcement learning for advanced reasoning tasks , author=. arXiv preprint arXiv:2504.05118 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    A Survey on LLM-as-a-Judge

    A survey on llm-as-a-judge , author=. arXiv preprint arXiv:2411.15594 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  34. [34]

    Journal of Artificial Intelligence Research , volume=

    Computational benefits of intermediate rewards for goal-reaching policy learning , author=. Journal of Artificial Intelligence Research , volume=

  35. [35]

    On Time, Within Budget: Constraint-Driven Online Resource Allocation for Agentic Workflows

    On Time, Within Budget: Constraint-Driven Online Resource Allocation for Agentic Workflows , author=. arXiv preprint arXiv:2605.06110 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  36. [36]

    Proceedings of the Sixteenth International Conference on Machine Learning , pages=

    Policy Invariance Under Reward Transformations: Theory and Application to Reward Shaping , author=. Proceedings of the Sixteenth International Conference on Machine Learning , pages=

  37. [37]

    Proceedings of the web conference 2021 , pages=

    A trigger-sense memory flow framework for joint entity and relation extraction , author=. Proceedings of the web conference 2021 , pages=

    2021

  38. [38]

    arXiv preprint arXiv:2309.10105 , year=

    Understanding catastrophic forgetting in language models via implicit inference , author=. arXiv preprint arXiv:2309.10105 , year=

    work page arXiv

  39. [39]

    A survey of large language models , author=

  40. [40]

    Measuring Mathematical Problem Solving With the MATH Dataset

    Measuring mathematical problem solving with the math dataset , author=. arXiv preprint arXiv:2103.03874 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [41]

    2021 , eprint=

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

    2021

  42. [42]

    First Conference on Language Modeling , year=

    Gpqa: A graduate-level google-proof q&a benchmark , author=. First Conference on Language Modeling , year=

  43. [43]

    International Workshop on Semantic Evaluation (SemEval-2018) , pages=

    Semeval-2018 task 7: Semantic relation extraction and classification in scientific papers , author=. International Workshop on Semantic Evaluation (SemEval-2018) , pages=

    2018

  44. [44]

    The contribution of LLMs to relation extraction in the economic field , author=. The Joint Workshop of the 9th Financial Technology and Natural Language Processing (FinNLP), the 6th Financial Narrative Processing (FNP), and the 1st Workshop on Large Language Models for Finance and Legal (LLMFinLegal) , year=

  45. [45]

    arXiv preprint arXiv:2305.02105 , year=

    Gpt-re: In-context learning for relation extraction using large language models , author=. arXiv preprint arXiv:2305.02105 , year=

    work page arXiv

  46. [46]

    arXiv preprint arXiv:2310.12024 , year=

    CORE: A Few-Shot Company Relation Classification Dataset for Robust Domain Adaptation , author=. arXiv preprint arXiv:2310.12024 , year=

    work page arXiv

  47. [47]

    arXiv preprint arXiv:2404.18085 , year=

    CRE-LLM: a domain-specific Chinese relation extraction framework with fine-tuned large language model , author=. arXiv preprint arXiv:2404.18085 , year=

    work page arXiv

  48. [48]

    arXiv preprint arXiv:2505.01077 , year=

    Zero-Shot Document-Level Biomedical Relation Extraction via Scenario-based Prompt Design in Two-Stage with LLM , author=. arXiv preprint arXiv:2505.01077 , year=

    work page arXiv

  49. [49]

    Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Bidirectional recurrent convolutional neural network for relation classification , author=. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  50. [50]

    arXiv preprint arXiv:2503.13939 , year=

    Med-r1: Reinforcement learning for generalizable medical reasoning in vision-language models , author=. arXiv preprint arXiv:2503.13939 , year=

    work page arXiv

  51. [51]

    arXiv preprint arXiv:2505.15817 , year=

    Learning to Reason via Mixture-of-Thought for Logical Reasoning , author=. arXiv preprint arXiv:2505.15817 , year=

    work page arXiv

  52. [52]

    arXiv preprint arXiv:2504.03714 , year=

    Breach in the Shield: Unveiling the Vulnerabilities of Large Language Models , author=. arXiv preprint arXiv:2504.03714 , year=

    work page arXiv

  53. [53]

    Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models

    Stop overthinking: A survey on efficient reasoning for large language models , author=. arXiv preprint arXiv:2503.16419 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  54. [54]

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

    Vision-r1: Incentivizing reasoning capability in multimodal large language models , author=. arXiv preprint arXiv:2503.06749 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  55. [55]

    Instruction-Following Evaluation for Large Language Models

    Instruction-following evaluation for large language models , author=. arXiv preprint arXiv:2311.07911 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  56. [56]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  57. [57]

    Advances in neural information processing systems , volume=

    Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in neural information processing systems , volume=

  58. [58]

    Solving math word problems with process- and outcome-based feedback

    Solving math word problems with process-and outcome-based feedback , author=. arXiv preprint arXiv:2211.14275 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  59. [59]

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning , author=. arXiv preprint arXiv:2503.09516 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  60. [60]

    Biotechnology Advances , pages=

    Advancing microbial production through artificial intelligence-aided biology , author=. Biotechnology Advances , pages=. 2024 , publisher=

    2024

  61. [61]

    Annual Meeting of the Association for Computational Linguistics , year=

    Revisiting the Negative Data of Distantly Supervised Relation Extraction , author=. Annual Meeting of the Association for Computational Linguistics , year=

  62. [62]

    2026 , eprint=

    Semantic-Space Exploration and Exploitation in RLVR for LLM Reasoning , author=. 2026 , eprint=

    2026

  63. [63]

    TabularMath: Understanding Math Reasoning over Tables with Large Language Models

    TabularMath: Understanding Math Reasoning over Tables with Large Language Models , author=. arXiv preprint arXiv:2505.19563 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  64. [64]

    arXiv preprint arXiv:2601.01984 , year=

    Thinking with Blueprints: Assisting Vision-Language Models in Spatial Reasoning via Structured Object Representation , author=. arXiv preprint arXiv:2601.01984 , year=

    work page arXiv

  65. [65]

    Computation , volume=

    The Health-Wealth Gradient in Labor Markets: Integrating Health, Insurance, and Social Metrics to Predict Employment Density , author=. Computation , volume=. 2026 , publisher=

    2026

  66. [66]

    GPT-4 Technical Report

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  67. [67]

    American Invitational Mathematics Examination (

  68. [68]

    2026 , eprint=

    Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs , author=. 2026 , eprint=

    2026

  69. [69]

    GPQA: A Graduate-Level Google-Proof Q&A Benchmark

    Gpqa: A graduate-level google-proof q&a benchmark , author=. arXiv preprint arXiv:2311.12022 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  70. [70]

    Measuring Massive Multitask Language Understanding

    Measuring massive multitask language understanding , author=. arXiv preprint arXiv:2009.03300 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2009

  71. [71]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    Self-consistency improves chain of thought reasoning in language models , author=. arXiv preprint arXiv:2203.11171 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  72. [72]

    arXiv preprint arXiv:2305.11860 , year=

    Let's Sample Step by Step: Adaptive-Consistency for Efficient Reasoning and Coding with LLMs , author=. arXiv preprint arXiv:2305.11860 , year=

    work page arXiv

  73. [73]

    arXiv preprint arXiv:2401.10480 , year=

    Escape sky-high cost: Early-stopping self-consistency for multi-step reasoning , author=. arXiv preprint arXiv:2401.10480 , year=

    work page arXiv

  74. [74]

    Journal of Machine Learning Research , year =

    Antonin Raffin and Ashley Hill and Adam Gleave and Anssi Kanervisto and Maximilian Ernestus and Noah Dormann , title =. Journal of Machine Learning Research , year =

  75. [75]

    2016 , Eprint =

    Greg Brockman and Vicki Cheung and Ludwig Pettersson and Jonas Schneider and John Schulman and Jie Tang and Wojciech Zaremba , Title =. 2016 , Eprint =

    2016

  76. [76]

    arXiv preprint arXiv:2509.07980 , year=

    Parallel-r1: Towards parallel thinking via reinforcement learning , author=. arXiv preprint arXiv:2509.07980 , year=

    work page arXiv

  77. [77]

    arXiv preprint arXiv:2509.04475 , year=

    Parathinker: Native parallel thinking as a new paradigm to scale llm test-time compute , author=. arXiv preprint arXiv:2509.04475 , year=

    work page arXiv

  78. [78]

    2025 , month = jul, day =

    Luong, Thang and Lockhart, Edward , title =. 2025 , month = jul, day =

    2025

  79. [79]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  80. [80]

    arXiv preprint arXiv:2503.00031 , year=

    Efficient test-time scaling via self-calibration , author=. arXiv preprint arXiv:2503.00031 , year=

    work page arXiv

Showing first 80 references.