pith. machine review for the scientific record. sign in

arxiv: 2604.21018 · v1 · submitted 2026-04-22 · 💻 cs.AI

Recognition: unknown

Adaptive Test-Time Compute Allocation with Evolving In-Context Demonstrations

Bowen Zuo , Dongruo Zhou , Yinglun Zhu
Authors on Pith no claims yet

Pith reviewed 2026-05-09 23:52 UTC · model grok-4.3

classification 💻 cs.AI
keywords test-time computeadaptive inferencein-context demonstrationssemantic similarityreasoning benchmarkscompute allocationlarge language models
0
0 comments X

The pith

A test-time method builds a pool of successful responses from easy queries then uses semantic similarity to evolve in-context examples for harder ones.

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

The paper introduces a framework that splits test-time effort into a warm-up phase and an adaptive phase. In the warm-up, the model solves the easier items in a test set and keeps those question-response pairs. In the adaptive phase it spends remaining compute only on the still-unresolved items, each time conditioning the generation on a small set of successful responses drawn from semantically similar past queries instead of drawing from a static distribution. This joint change in where compute is spent and how each answer is generated produces higher accuracy than fixed-allocation baselines while using less total inference compute across math, coding, and reasoning benchmarks. A reader would care because the approach turns the test set itself into a growing source of useful demonstrations without requiring extra training or external data.

Core claim

The central claim is that jointly adapting compute allocation and generation distributions at test time yields better performance with lower total cost: a warm-up phase first identifies easy queries and assembles a pool of their successful responses; an adaptive phase then concentrates further samples on unresolved queries while reshaping each generation by conditioning on successful responses from semantically related queries in the pool.

What carries the argument

Evolving in-context demonstrations, which select successful responses from the test-set pool by semantic similarity to condition generation for each new unresolved query.

If this is right

  • Accuracy rises on math, coding, and reasoning benchmarks relative to static test-time methods.
  • Total number of model calls needed to reach a target performance level drops.
  • Compute is automatically redirected away from queries that have already been solved correctly.
  • Each generation for a difficult query is conditioned on a changing, query-specific set of prior successes rather than a fixed prompt.

Where Pith is reading between the lines

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

  • The same warm-up-plus-similarity idea could be applied to other modalities or to open-ended generation tasks where success is harder to verify automatically.
  • If semantic similarity proves reliable, future systems might maintain a running pool across multiple related tasks rather than resetting per benchmark.
  • The approach implicitly treats the test set as a source of training signal at inference time, which may change how benchmarks are constructed or protected.

Load-bearing premise

That a warm-up run on the test set itself can identify enough easy queries and assemble a pool of successful responses whose semantic similarity to harder queries is sufficient to improve generation without introducing bias or leakage.

What would settle it

An experiment in which the warm-up phase yields too few successful examples or in which semantic matching produces no accuracy gain over repeated sampling from the base distribution would falsify the claimed benefit.

Figures

Figures reproduced from arXiv: 2604.21018 by Bowen Zuo, Dongruo Zhou, Yinglun Zhu.

Figure 1
Figure 1. Figure 1: Average coverage (%) across 4 model variants, where each result is averaged over 5 benchmarks [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Coverage comparison on MATH-500. 4.2.1 Warm-up In the warm-up stage, we allocate a small, fixed amount of compute to each query. For each x ∈ S, we generate a fixed number of responses g0(x) = {y1, . . . , yK}, yi ∼ p(· | x). Each response is evaluated by a reward oracle r(x, y) usually with a range [0, 1]. If any response satisfies r(x, y) ≥ γ, then the question is marked as solved and removed from furthe… view at source ↗
Figure 3
Figure 3. Figure 3: Coverage comparison on LiveCodeBench. 40 60 80 100 Total Output Tokens (e4) 45 50 55 60 65 70 Coverage (%) Gemini (non-thinking) 40 60 80 Total Output Tokens (e4) 45 50 55 60 65 70 Gemini (thinking) 20 30 40 Total Output Tokens (e4) 45 50 55 60 GPT-4.1 Mini 4 6 8 10 12 Total Output Tokens (e4) 45 50 55 60 65 GPT-5 Nano Unif Elim Ours (Rand) Ours (Adpt) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Coverage comparison on MinervaMath. released between 10/05/2024 and 01/04/2025, following the evaluation protocol reported for Gemini 2.5 (Team, 2025). In addition to these benchmarks, we construct a custom evaluation dataset using Reasoning Gym (Stojanovski et al., 2025). This dataset consists of 400 math-oriented reasoning problems sampled at the hard difficulty level. Details of the dataset construction… view at source ↗
Figure 5
Figure 5. Figure 5: Coverage comparison on GPQA-Diamond. 100 200 300 Total Output Tokens (e4) 45 50 55 60 65 70 Coverage (%) Gemini (non-thinking) 60 80 100 120 Total Output Tokens (e4) 58 60 62 64 66 68 Gemini (thinking) 40 60 80 Total Output Tokens (e4) 35 40 45 50 55 60 GPT-4.1 Mini 2 3 4 5 Total Output Tokens (e4) 72 74 76 78 80 GPT-5 Nano Unif Elim Ours (Rand) Ours (Adpt) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Coverage comparison on Reasoning Gym. new responses are generated only for test questions that remain unsolved. By default, we use a total of R = 4 rounds and Rwarm = 1. For the number of neighboring shots in adaptive rounds, we apply P = 3 for all our settings. For Unif, Elim, and the warm-up round, we apply zero-shot COT prompt (Kojima et al., 2022). We include all details in Appendix B. For unif, we mat… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation studies across fixed demonstrations, temperature tuning, and selection strategies. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Final round coverage comparison Numerical results. We present numerical results for the approach Ours (Adpt) compared with all the other variants. We compare the last round coverage gain for Ours (Adpt) over the other baselines. In general, our method is able to achieve a good performance gain. C Token efficiency analysis We attribute the token efficiency of our approach to two complementary factors: (i) t… view at source ↗
read the original abstract

While scaling test-time compute can substantially improve model performance, existing approaches either rely on static compute allocation or sample from fixed generation distributions. In this work, we introduce a test-time compute allocation framework that jointly adapts where computation is spent and how generation is performed. Our method begins with a warm-up phase that identifies easy queries and assembles an initial pool of question-response pairs from the test set itself. An adaptive phase then concentrates further computation on unresolved queries while reshaping their generation distributions through evolving in-context demonstrations -- conditioning each generation on successful responses from semantically related queries rather than resampling from a fixed distribution. Experiments across math, coding, and reasoning benchmarks demonstrate that our approach consistently outperforms existing baselines while consuming substantially less inference-time compute.

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

3 major / 1 minor

Summary. The manuscript introduces a test-time compute allocation framework for language models. It begins with a warm-up phase that identifies easy queries and assembles a pool of question-response pairs drawn from the test set itself. An adaptive phase then focuses additional computation on unresolved queries while conditioning generation on semantically similar successful responses from the pool as evolving in-context demonstrations, rather than resampling from a fixed distribution. The authors claim this yields consistent outperformance over existing baselines across math, coding, and reasoning benchmarks while consuming substantially less inference-time compute.

Significance. If the empirical results hold after addressing leakage and compute-accounting concerns, the work could provide a practical advance in efficient test-time scaling by jointly adapting allocation and generation distributions via cross-query semantic conditioning. This would differentiate it from static or fixed-distribution baselines and offer a falsifiable path to better performance-compute tradeoffs.

major comments (3)
  1. [Abstract] Abstract: the central claim of consistent outperformance and substantially lower inference-time compute is stated without any quantitative results, baseline definitions, success criteria for 'unresolved queries,' or explicit controls for test-set leakage, rendering the empirical contribution unevaluable from the provided description.
  2. [Method] Method description (warm-up and adaptive phases): assembling the demonstration pool directly from the test set and then evaluating on the same set creates a circularity risk in which performance gains may partly reflect reuse of test information rather than generalization. No mechanism is described for preventing cross-query information flow or for streaming per-query operation without batch test-set access.
  3. [Experiments] Experiments section: the headline efficiency claim requires that the computational cost of the warm-up phase be included in the total reported inference-time compute and still show net savings versus baselines; the manuscript does not clarify whether this accounting is performed.
minor comments (1)
  1. [Abstract] The abstract introduces 'evolving in-context demonstrations' without a concise definition or example of how the pool is updated across queries; a short clarifying sentence would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. We have revised the manuscript to address the concerns about the abstract, potential test-set circularity, and compute accounting. Our responses to each major comment are provided below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of consistent outperformance and substantially lower inference-time compute is stated without any quantitative results, baseline definitions, success criteria for 'unresolved queries,' or explicit controls for test-set leakage, rendering the empirical contribution unevaluable from the provided description.

    Authors: We agree that the abstract would benefit from greater specificity. In the revised version we have added concise quantitative highlights (performance deltas and compute ratios versus the primary baselines), explicit definitions of the baselines used, the criterion for marking a query as unresolved, and a brief statement on leakage controls. These additions make the central claims directly evaluable while remaining within length limits. revision: yes

  2. Referee: [Method] Method description (warm-up and adaptive phases): assembling the demonstration pool directly from the test set and then evaluating on the same set creates a circularity risk in which performance gains may partly reflect reuse of test information rather than generalization. No mechanism is described for preventing cross-query information flow or for streaming per-query operation without batch test-set access.

    Authors: The concern is valid and we have clarified the design. The pool contains only model-generated responses on queries the model itself solved during warm-up; no ground-truth labels are injected. To address streaming and cross-query flow, the revised method section now specifies a sequential, per-query protocol: each new query is processed using only the demonstration pool accumulated from prior queries, with successful generations added on the fly. We have added an explicit streaming algorithm and a limitations paragraph acknowledging that the current implementation assumes access to the full test distribution for the warm-up ordering, which may not hold in strictly online settings. revision: partial

  3. Referee: [Experiments] Experiments section: the headline efficiency claim requires that the computational cost of the warm-up phase be included in the total reported inference-time compute and still show net savings versus baselines; the manuscript does not clarify whether this accounting is performed.

    Authors: We agree that total inference cost must encompass the warm-up phase. The revised experiments section now states explicitly that all reported FLOPs and token counts include warm-up computation, provides a per-phase breakdown, and confirms that the net savings versus static baselines remain positive after this inclusion. A new table column reports the warm-up overhead as a fraction of total cost. revision: yes

Circularity Check

1 steps flagged

Warm-up pool assembly from test set makes performance gains dependent on evaluation data access

specific steps
  1. fitted input called prediction [Abstract]
    "Our method begins with a warm-up phase that identifies easy queries and assembles an initial pool of question-response pairs from the test set itself. An adaptive phase then concentrates further computation on unresolved queries while reshaping their generation distributions through evolving in-context demonstrations -- conditioning each generation on successful responses from semantically related queries rather than resampling from a fixed distribution."

    The 'prediction' of improved outputs on unresolved test queries is performed by conditioning on a pool of responses that was itself constructed by running the model on the full test set. The outperformance and compute savings therefore reduce to the input of having batch access to the evaluation data for demonstration assembly, rather than arising from a test-time procedure that operates without such access.

full rationale

The paper's central empirical claim of consistent outperformance at lower inference compute rests on a method whose adaptive phase explicitly conditions generations on successful responses harvested from the same test set during an initial warm-up. This step is load-bearing for the headline result because the reshaping of generation distributions for unresolved queries is achieved by direct reuse of test-set information rather than an independent mechanism. No equations or self-citations are visible in the provided text to create additional circularity, but the described procedure reduces the reported gains to a form of in-evaluation data reuse. The derivation chain is therefore partially circular at the level of the performance claim itself.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so the full ledger cannot be extracted. The approach appears to rest on standard LLM assumptions (semantic similarity can be measured reliably, success on one query transfers to semantically related queries) plus the domain assumption that the test set may be used to build the demonstration pool at inference time.

pith-pipeline@v0.9.0 · 5415 in / 1252 out tokens · 39516 ms · 2026-05-09T23:52:32.316782+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Active Testing of Large Language Models via Approximate Neyman Allocation

    cs.AI 2026-05 unverdicted novelty 6.0

    Active testing via surrogate semantic entropy stratification and approximate Neyman allocation reduces MSE by up to 28% versus uniform sampling and saves about 23% of the labeling budget on language and multimodal benchmarks.

Reference graph

Works this paper leans on

54 extracted references · 23 canonical work pages · cited by 1 Pith paper · 7 internal anchors

  1. [1]

    Self-improving llm agents at test-time, 2025

    Emre Can Acikgoz, Cheng Qian, Heng Ji, Dilek Hakkani-Tür, and Gokhan Tur. Self-improving llm agents at test-time, 2025. URL https://arxiv.org/abs/2510.07841

    work page arXiv 2025

  2. [2]

    Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, John D Co-Reyes, Eric Chu, Feryal Behbahani, Aleksandra Faust, and Hugo Larochelle

    Rishabh Agarwal, Avi Singh, Lei M Zhang, Bernd Bohnet, Luis Rosias, Stephanie C.Y. Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, John D Co-Reyes, Eric Chu, Feryal Behbahani, Aleksandra Faust, and Hugo Larochelle. Many-shot in-context learning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://op...

    2024

  3. [3]

    Gormley, and Graham Neubig

    Amanda Bertsch, Maor Ivgi, Emily Xiao, Uri Alon, Jonathan Berant, Matthew R. Gormley, and Graham Neubig. In-context learning with long-context models: An in-depth exploration. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors, Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Huma...

    work page doi:10.18653/v1/2025.naacl-long.605 2025

  4. [4]

    Large Language Monkeys: Scaling Inference Compute with Repeated Sampling

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V. Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling, 2024. URL https://arxiv.org/abs/2407.21787

    work page internal anchor Pith review arXiv 2024

  5. [5]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gr...

    1901

  6. [6]

    Pure exploration in multi-armed bandits problems

    S \'e bastien Bubeck, R \'e mi Munos, and Gilles Stoltz. Pure exploration in multi-armed bandits problems. In International conference on Algorithmic learning theory, pages 23--37. Springer, 2009

    2009

  7. [7]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sashank Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bra...

    2023

  8. [8]

    Training Verifiers to Solve Math Word Problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021. URL https://arxiv.org/abs/2110.14168

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  9. [9]

    Learning how hard to think: Input-adaptive allocation of LM computation

    Mehul Damani, Idan Shenfeld, Andi Peng, Andreea Bobu, and Jacob Andreas. Learning how hard to think: Input-adaptive allocation of LM computation. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  10. [10]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai D...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  11. [11]

    Huerta, and Hao Peng

    Yufeng Du, Minyang Tian, Srikanth Ronanki, Subendhu Rongali, Sravan Babu Bodapati, Aram Galstyan, Azton Wells, Roy Schwartz, Eliu A Huerta, and Hao Peng. Context length alone hurts LLM performance despite perfect retrieval. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors, Findings of the Association for Computati...

    work page doi:10.18653/v1/2025.findings-emnlp.1264 2025

  12. [12]

    Rae, and Laurent Sifre

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Oriol Vinyals, Jack W. Rae, and Laurent Sifre...

    2022

  13. [13]

    Livecodebench: Holistic and contamination free evaluation of large language models for code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=chfJJYC3iL

    2025

  14. [14]

    Best-arm identification algorithms for multi-armed bandits in the fixed confidence setting

    Kevin Jamieson and Robert Nowak. Best-arm identification algorithms for multi-armed bandits in the fixed confidence setting. In 2014 48th annual conference on information sciences and systems (CISS), pages 1--6. IEEE, 2014

    2014

  15. [15]

    Why Language Models Hallucinate

    Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. Why language models hallucinate, 2025. URL https://arxiv.org/abs/2509.04664

    work page internal anchor Pith review arXiv 2025

  16. [16]

    Scaling Laws for Neural Language Models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. URL https://arxiv.org/abs/2001.08361

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  17. [17]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA, 2022. Curran Associates Inc. ISBN 9781713871088

    2022

  18. [18]

    Smith, and Hannaneh Hajishirzi

    Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. R eward B ench: Evaluating reward models for language modeling, April 2025. URL https://aclanthology.org/2025.findings-naacl.96/

    2025

  19. [19]

    Solving quantitative reasoning problems with language models

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. Solving quantitative reasoning problems with language models. In Proceedings of the 36th International Conference on Neural Information Proces...

    2022

  20. [20]

    Let's verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=v8L0pN6EOi

    2024

  21. [21]

    Cmcts: A constrained monte carlo tree search framework for mathematical reasoning in large language model, December 2025

    Qingwen Lin, Boyan Xu, Guimin Hu, Zijian Li, Zhifeng Hao, Keli Zhang, and Ruichu Cai. Cmcts: A constrained monte carlo tree search framework for mathematical reasoning in large language model, December 2025. ISSN 0924-669X. URL https://doi.org/10.1007/s10489-025-07044-6

    work page doi:10.1007/s10489-025-07044-6 2025

  22. [22]

    11 Wendy Johnson and Thomas J Bouchard Jr

    Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt- 3 ?, 2021. URL https://arxiv.org/abs/2101.06804

    work page arXiv 2021

  23. [23]

    Can 1b LLM surpass 405b LLM ? rethinking compute-optimal test-time scaling

    Runze Liu, Junqi Gao, Jian Zhao, Kaiyan Zhang, Xiu Li, Biqing Qi, Wanli Ouyang, and Bowen Zhou. Can 1b LLM surpass 405b LLM ? rethinking compute-optimal test-time scaling. In Workshop on Reasoning and Planning for Large Language Models, 2025. URL https://openreview.net/forum?id=CvjX9Lhpze

    2025

  24. [24]

    An optimal algorithm for the thresholding bandit problem, 2016

    Andrea Locatelli, Maurilio Gutzeit, and Alexandra Carpentier. An optimal algorithm for the thresholding bandit problem, 2016. URL https://arxiv.org/abs/1605.08671

    work page arXiv 2016

  25. [25]

    Self-refine: Iterative refinement with self-feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback. In Thirty-seventh Conference on Neural Informat...

    2023

  26. [26]

    Adaptive inference-time compute: Llms can predict if they can do better, even mid-generation, 2024

    Rohin Manvi, Anikait Singh, and Stefano Ermon. Adaptive inference-time compute: Llms can predict if they can do better, even mid-generation, 2024. URL https://arxiv.org/abs/2410.02725

    work page arXiv 2024

  27. [27]

    Few-shot fine-tuning vs

    Marius Mosbach, Tiago Pimentel, Shauli Ravfogel, Dietrich Klakow, and Yanai Elazar. Few-shot fine-tuning vs. in-context learning: A fair comparison and evaluation. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Findings of the Association for Computational Linguistics: ACL 2023, pages 12284--12314, Toronto, Canada, July 2023. Association...

    work page doi:10.18653/v1/2023.findings-acl.779 2023

  28. [28]

    s1: Simple test-time scaling

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Cand \`e s, and Tatsunori B Hashimoto. s1: Simple test-time scaling. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 20286--20332, 2025

    2025

  29. [29]

    Learning to reason with llms, 2024

    OpenAI. Learning to reason with llms, 2024. URL https://openai.com/index/learning-to-reason-with-llms/. Accessed 12-Sep-2024

    2024

  30. [30]

    In-context learning with iterative demonstration selection

    Chengwei Qin, Aston Zhang, Chen Chen, Anirudh Dagar, and Wenming Ye. In-context learning with iterative demonstration selection. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7441--7455, 2024

    2024

  31. [31]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First conference on language modeling, 2024

    2024

  32. [32]

    Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning

    Charlie Victor Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling LLM test-time compute optimally can be more effective than scaling parameters for reasoning. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=4FWAwZtd2n

    2025

  33. [33]

    Reasoning gym: Reasoning environments for reinforcement learning with verifiable rewards

    Zafir Stojanovski, Oliver Stanley, Joe Sharratt, Richard Jones, Abdulhakeem Adefioye, Jean Kaddour, and Andreas K \"o pf. Reasoning gym: Reasoning environments for reinforcement learning with verifiable rewards. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2025. URL https://openreview.net/fo...

    2025

  34. [34]

    Fast best-of-n decoding via speculative rejection

    Hanshi Sun, Momin Haider, Ruiqi Zhang, Huitao Yang, Jiahao Qiu, Ming Yin, Mengdi Wang, Peter Bartlett, and Andrea Zanette. Fast best-of-n decoding via speculative rejection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=348hfcprUs

    2024

  35. [35]

    Adaptive rectification sampling for test-time compute scaling, 2025

    Zhendong Tan, Xingjun Zhang, Chaoyi Hu, Yancheng Pan, and Shaoxun Wang. Adaptive rectification sampling for test-time compute scaling, 2025. URL https://arxiv.org/abs/2504.01317

    work page arXiv 2025

  36. [36]

    Multilingual LLM s are better cross-lingual in-context learners with alignment

    Eshaan Tanwar, Subhabrata Dutta, Manish Borthakur, and Tanmoy Chakraborty. Multilingual LLM s are better cross-lingual in-context learners with alignment. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6292--6307, Toront...

    work page doi:10.18653/v1/2023.acl-long.346 2023

  37. [37]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Google Gemini Team. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities, 2025. URL https://arxiv.org/abs/2507.06261

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [38]

    Jens Tuyls, Dylan J Foster, Akshay Krishnamurthy, and Jordan T. Ash. Representation-based exploration for language models: From test-time to post-training. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=S4PCF1YxoR

    2026

  39. [39]

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

    Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. Solving math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  40. [40]

    Think deep, think fast: Investigating efficiency of verifier-free inference-time-scaling methods, 2025 a

    Junlin Wang, Shang Zhu, Jon Saad-Falcon, Ben Athiwaratkun, Qingyang Wu, Jue Wang, Shuaiwen Leon Song, Ce Zhang, Bhuwan Dhingra, and James Zou. Think deep, think fast: Investigating efficiency of verifier-free inference-time-scaling methods, 2025 a . URL https://arxiv.org/abs/2504.14047

    work page arXiv 2025

  41. [41]

    Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning

    Xinglin Wang, Shaoxiong Feng, Yiwei Li, Peiwen Yuan, Yueqi Zhang, Chuyi Tan, Boyuan Pan, Yao Hu, and Kan Li. Make every penny count: Difficulty-adaptive self-consistency for cost-efficient reasoning. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors, Findings of the Association for Computational Linguistics: NAACL 2025, pages 6919--6932, Albuquerque, Ne...

    work page doi:10.18653/v1/2025.findings-naacl.383 2025

  42. [42]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=1PL1NIMMrw

    2023

  43. [43]

    Chi, Quoc V Le, and Denny Zhou

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?i...

    2022

  44. [44]

    On the role of temperature sampling in test-time scaling, 2025

    Yuheng Wu, Azalia Mirhoseini, and Thierry Tambe. On the role of temperature sampling in test-time scaling, 2025. URL https://arxiv.org/abs/2510.02611

    work page arXiv 2025

  45. [45]

    Rethinking the unsolvable: When in-context search meets test-time scaling, 2025

    Fanzeng Xia, Yidong Luo, Tinko Sebastian Bartels, Yaqi Xu, and Tongxin Li. Rethinking the unsolvable: When in-context search meets test-time scaling, 2025. URL https://arxiv.org/abs/2505.22290

    work page arXiv 2025

  46. [46]

    \ k\ NN prompting: Beyond-context learning with calibration-free nearest neighbor inference

    Benfeng Xu, Quan Wang, Zhendong Mao, Yajuan Lyu, Qiaoqiao She, and Yongdong Zhang. \ k\ NN prompting: Beyond-context learning with calibration-free nearest neighbor inference. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=fe2S7736sNS

    2023

  47. [47]

    & Kankanhalli, M

    Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models, 2025. URL https://arxiv.org/abs/2401.11817

    work page arXiv 2025

  48. [48]

    Griffiths, Yuan Cao, and Karthik R Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik R Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=5Xc1ecxO1h

    2023

  49. [49]

    Making retrieval-augmented language models robust to irrelevant context

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. Making retrieval-augmented language models robust to irrelevant context. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=ZS4m74kZpH

    2024

  50. [50]

    The lessons of developing process reward models in mathematical reasoning

    Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors, Findings of the Association for Computational Linguistics: ACL 2025, pa...

    work page doi:10.18653/v1/2025.findings-acl.547 2025

  51. [51]

    Robust outlier arm identification

    Yinglun Zhu, Sumeet Katariya, and Robert Nowak. Robust outlier arm identification. In International Conference on Machine Learning, pages 11566--11575. PMLR, 2020

    2020

  52. [52]

    Pure exploration in kernel and neural bandits

    Yinglun Zhu, Dongruo Zhou, Ruoxi Jiang, Quanquan Gu, Rebecca Willett, and Robert Nowak. Pure exploration in kernel and neural bandits. Advances in neural information processing systems, 34: 0 11618--11630, 2021

    2021

  53. [53]

    Near instance optimal model selection for pure exploration linear bandits

    Yinglun Zhu, Julian Katz-Samuels, and Robert Nowak. Near instance optimal model selection for pure exploration linear bandits. In International Conference on Artificial Intelligence and Statistics, pages 6735--6769. PMLR, 2022

    2022

  54. [54]

    Strategic scaling of test-time compute: A bandit learning approach

    Bowen Zuo and Yinglun Zhu. Strategic scaling of test-time compute: A bandit learning approach. In The Fourteenth International Conference on Learning Representations, 2026

    2026