Control-R: Towards controllable test-time scaling

Di Zhang; Dongzhan Zhou; Haonan He; Jianbo Wu; Jiatong Li; Jingdi Lei; Junxian Li; Peng Ye; Shufei Zhang; Wanli Ouyang

REVIEW 5 major objections 5 minor 13 references

Control-R: Towards controllable test-time scaling

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Control scores in the prompt let a 32B model tune its reasoning depth

desk verdict Competitive 32B accuracy with SteerLM-style control fields, but the controllability claim rests on accuracy alone — no direct evidence that reasoning effort changes. read the letter →

arxiv 2506.00189 v1 pith:XQVJUKKU submitted 2025-05-30 cs.AI cs.CL

Di Zhang , Weida Wang , Junxian Li , Xunzhi Wang , Jiatong Li , Jianbo Wu , Jingdi Lei , Haonan He

show 5 more authors

Peng Ye Shufei Zhang Wanli Ouyang Yuqiang Li Dongzhan Zhou

This is my paper

classification cs.AIcs.CL

keywords reasoningcontrolfieldscontrollabletest-timescalinglongchain-of-thoughtconditionaldistillationfinetuningunderthinkingoverthinkingtreesearchmathbenchmarks

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that long chain-of-thought reasoning can be made controllable at test time by conditioning the model on a short string of eleven integer scores, which the authors call Reasoning Control Fields. The scores regulate how deep and broad the search should be, how aggressively errors should be detected and corrected, how often strategies should switch, and how the process should score on correctness, efficiency, completeness, coherence, knowledge accuracy, and clarity. If correct, a user could dial reasoning effort up or down per query instead of accepting whatever depth the model chooses, directly addressing the twin failures of underthinking and overthinking. To support this, the paper builds a curated dataset, proposes a conditional distillation fine-tuning objective, and reports that a 32B model trained this way outperforms comparable 32B reasoning models on AIME2024 and MATH500.

What carries the argument

The Reasoning Control Field (RCF) string is a prompt suffix such as "<control> search_depth: 8; search_breadth: 7; error_detection: 8; error_correction: 7; strategy_switching: 6; correctness: 9; efficiency: 7; completeness: 8; coherence: 8; knowledge_accuracy: 9; clarity_of_steps: 8 <control/>", prefixed by a line break. Each of the eleven integer scores, ranging 0 to 9, is intended as a regulator for a specific aspect of the tree-search process, and Conditional Distillation Finetuning forces the model's next-step probabilities to depend on the entire string, so changing any score should change how the model searches.

What would settle it

Run Control-R-32B on a fixed question set with all control scores set to 0 versus 9 and measure actual token counts, number of distinct strategies, and backtracking events in the sampled traces; if the all-9 condition does not produce longer or structurally deeper reasoning while improving accuracy, the central controllability claim is not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that long CoT reasoning can be viewed as an in-context tree search and that a fixed textual vector of eleven control scores can modulate that search. It formalizes generation as $P(R|q,C)$, where $C$ is the control string, and trains with the conditional distillation objective $\mathcal{L}_{CDF}=\mathbb{E}[-\sum_i \log P(r_i|r_{<i},q,C;\theta)]$. The resulting Control-R-32B scores 70.0% on AIME2024 and 93.2% on MATH500, which the paper reports as state-of-the-art among 32B models, and 61.1% on GPQA-Diamond and 55.0% on AIME2025 Part I. The ablation shows the control string is load-bearing: removing it collapses AIME2024 accuracy from 70.0% to 6.7% and MATH500 from 93.2% to 3.2%, while setting all scores to 0 or 5 gives intermediate results.

Load-bearing premise

The paper assumes that accuracy differences under different control scores prove the model adjusts its reasoning effort, but it never measures reasoning length, step count, or strategy use, and the control labels were assigned by an automated annotator to the same traces used for training, making the supervision signal partly circular.

Editorial extensions

If this is right

If the control fields truly modulate reasoning effort, users can trade accuracy against compute per query by adjusting the eleven scores rather than switching to a different model.
The control string is not decorative: removing it degrades accuracy dramatically, so deployment of Control-R-32B must always supply the field.
The conditional objective applies to any next-token language model, suggesting the same recipe could be transferred to other base models and other long-reasoning tasks.
Task sensitivity varies: AIME2024 moves substantially with control settings while MATH500 moves only slightly, so the optimal control configuration is task-dependent rather than universal.
Framing long CoT as tree search with execution-control and process-quality scores gives an interpretable vocabulary for reasoning behavior, which could be reused for analysis and monitoring.

Reading between the lines

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

A natural extension the paper does not build is a difficulty-estimating router that chooses control scores automatically per query, turning the all-9 setting into one point on a cost-accuracy frontier rather than a fixed recommendation.
Because the control labels were produced by an automated annotator reading the same traces used for training, the eleven scores may reflect that annotator's stylistic preferences more than objective reasoning properties; an independent re-annotation study would clarify this.
The tree-search framing suggests a direct test the paper leaves implicit: token counts, branch counts, and backtracking events should move monotonically with stated search_depth and related scores, which would turn the accuracy-based proxy into a mechanism-level verification.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper takes SteerLM's attribute-conditioned generation, ports it to long CoT reasoning with 11 hand-defined control fields, and trains a 32B model that hits competitive numbers on AIME2024 and MATH500. That application is new, and the Control-R-4K dataset, with its per-trace control-field annotations, could be useful to people working on reasoning distillation. The method is simple and clearly described, and the ablation shows the control string matters: without it, accuracy collapses, and raising all fields from 0 to 9 improves AIME2024 from 63.3% to 70.0%. That is real evidence of some sensitivity, and I don't want to minimize it.

The soft spot is the central claim. The paper says RCFs let users dial reasoning effort, but it never measures reasoning effort. No token lengths, step counts, backtracking frequency, or strategy composition under different control values. Table 3 only shows accuracy, and accuracy is a weak proxy — especially when the test-time prompt sets correctness and knowledge_accuracy to 9, which leaks the outcome. The presence/absence gap (6.7% vs 70%) also suggests the model has learned a format dependency rather than a graded dial; the graded response between 0 and 9 is modest, and on MATH500 it's almost flat. The labeling loop is also self-referential: ChatGPT-4o judges traces from DeepSeek-R1/QwQ, and the model trains on those judgments with no external validation that the scores track real process attributes.

Two smaller issues: MATH500 is sampled from MATH, which appears in the training data, so the headline 93.2% is in-distribution. And the AIME2025 results use Pass@1 with n=4 while other tables use n=1, so the comparison is apples-to-oranges. No artifacts are released, which makes the dataset hard to reuse or verify.

Overall, this is a serious paper about a real problem, but the load-bearing claim — controllable test-time scaling — is not yet supported. The fix is straightforward: measure the reasoning process under different control settings, drop or ablate the outcome-leaking fields, and report error bars. The authors even acknowledge the calibration issue in their limitations, so the gap is recognized rather than hidden.

I'd send this to peer review. The idea is timely, the dataset has value, and the concerns are addressable with additional experiments. But the reviewers should push hard on the controllability evidence before this is accepted.

Referee Report

5 major / 5 minor

Summary. The paper proposes Reasoning Control Fields (RCFs), an 11-dimension set of integer-valued control signals appended to prompts, together with a Conditional Distillation Finetuning (CDF) procedure and a new dataset, Control-R-4K, to train a 32B model (Control-R-32B) that can adjust long chain-of-thought reasoning effort at test time. Control-R-32B is evaluated on AIME2024, MATH500, GPQA-Diamond, and AIME2025 Part I, reporting 70.0%, 93.2%, 61.1%, and 55.0%, respectively. The central claim is that RCFs enable 'controllable Long CoT reasoning' and state-of-the-art performance at the 32B scale.

Significance. If the central claim were substantiated, the paper would offer a useful, lightweight mechanism for trading reasoning depth against compute in long-CoT models, and the Control-R-4K dataset with per-trace attribute annotations would be a reusable resource for conditional reasoning distillation. The paper also gives a clear probabilistic framing of conditional generation over reasoning traces. However, the evidence presented does not currently demonstrate that the control fields modulate reasoning behavior: accuracy differences are small, unreplicated, and confounded by outcome-related prompt fields, and no direct measurement of reasoning effort under differing control values is reported. The MATH500 result is also likely in-distribution because the training data includes a random sample of MATH, of which MATH500 is a subset.

major comments (5)

[§3.3, Table 3] The central claim that RCFs enable controllable test-time scaling is not directly supported by the reported experiments. Table 3 reports only accuracy under all-fields-equal-to-0/5/9; it never reports token length, number of reasoning steps, backtracking frequency, strategy-switch counts, or any other process-level measure under different control values. Table 2 reports token lengths and 'Wait' occurrences only by correctness outcome, not by control-field setting. Thus the paper shows that a prompt string affects accuracy, not that it controls the reasoning process.
[§3.3, Table 3] The sensitivity of accuracy to control-field values is weak and likely within noise. On AIME2024, all-0 and all-5 both give 63.3% and all-9 gives 70.0%; on MATH500 the numbers are 91.4%, 91.8%, and 93.2%. The paper uses Pass@1 with n=1 and provides no error bars or repeated-sampling statistics, so the 6.7- and 1.8-point differences do not establish fine-grained control. This is especially important because the test-time prompt sets 'correctness' and 'knowledge_accuracy' to 9, so part of the accuracy gain may come from directly prompting the model to be correct rather than from modulating search effort.
[§2.2, Appendix B.3] The RCF labels are produced by ChatGPT-4o annotating the same long-CoT traces that are later used as training targets, with no inter-annotator agreement, no validation against externally defined reasoning-effort measures, and no evidence that the scores correlate with actual process attributes. This makes the control semantics circular: the model learns to imitate the annotator's score-to-trace mapping rather than to implement an independently verified notion of search depth, breadth, or error correction.
[Appendix B.1] The claim of state-of-the-art MATH500 performance is undermined by training-data contamination: the main subset of Control-R-4K includes a random sample of problems from the MATH dataset, and MATH500 is a subset of MATH. Thus the 93.2% figure is in-distribution for Control-R-32B, while the comparison baselines were not trained on this subset. The paper should report results on a held-out, non-overlapping benchmark and should state explicitly which AIME years and MATH splits were excluded from training.
[§3.1, pass@k definition] The paper defines pass@k but reports all main results as Pass@1 with n=1. This means each benchmark number is a single deterministic or single-sample run, so the reported differences in the ablation study have no sampling uncertainty. The authors should either report multiple seeds with means and confidence intervals or clearly state that the numbers are single-run point estimates and avoid drawing conclusions from small gaps.

minor comments (5)

[Abstract] The first sentence has a grammatical error: 'This paper target in addressing' should be 'This paper targets the challenges of...'.
[Figure 1] The label 'Cotrol-R' in Figure 1 is a typo for 'Control-R'.
[Table 3 caption] The caption reads 'Ablation study on the affection of reasoning control fields'; 'affection' should be 'effect'.
[Table A2 caption and §3.1] Table A2 reports the metric as Pass@1 (n=4), while §3.1 defines the metric as Pass@1 and Table 1 uses n=1. This inconsistency should be reconciled.
[References] The references list Lightman et al. 2023a and 2023b as separate entries for the same paper; this should be a single citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: control fields are external annotations, CDF is conditional LM training, and benchmark outcomes are external accuracy measurements.

full rationale

No load-bearing circular step is present. The RCFs are external annotations of reasoning traces produced by ChatGPT-4o, and the CDF objective (Eq. 6) is standard conditional language-model training on (q, C, R) triples; the reported controllability evidence is benchmark accuracy on external sets (AIME2024, MATH500, GPQA-Diamond, AIME2025) rather than a re-prediction of the control labels. The only self-citation (Zhang et al. 2024 for self-refine) is a supporting citation for a known strategy and is not used to license the central claim. The concern that controllability is inferred from accuracy rather than from direct measurements of reasoning depth or breadth, and that the all-9 test prompt includes outcome-leaking fields such as correctness, is an experimental-validity limitation (partially acknowledged in the Limitations section), not a circular derivation: no equation or fitted parameter reduces the benchmark result to the annotation input. Hence score 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The central method relies on hand-designed control dimensions, a manual test-time operating point (all scores 9), and the assumption that LLM-generated annotations are valid. The conceptual tree-search framing is an invented analogy without a corresponding algorithmic implementation.

free parameters (3)

Test-time control field configuration (all set to 9) = 9 across 11 fields
The main experiments set every control field to maximum; this is a manually chosen operating point, not a learned value.
Control field definitions (11 dimensions and scales)
The decomposition of reasoning into 11 attributes is hand-designed; the 0-9 scales are arbitrary and set by the annotation prompt.
Sampling temperature for response generation = 1.0
Chosen in Appendix B.2 to increase diversity; affects training data distribution.

assumptions (5)

domain assumption Long CoT reasoning can be represented as an in-context tree search, with each reasoning step a node.
Section 2.1 introduces this analogy to justify the RCF dimensions.
domain assumption The 11 control attributes completely capture execution control and process quality of reasoning.
Section 2.1.1 and 2.1.2 define them as the full set of control fields.
domain assumption ChatGPT-4o annotations of reasoning traces are accurate and consistent enough to serve as training labels.
Section 2.2 and Appendix B.3 describe the annotation pipeline with temperature 0.
domain assumption Fine-tuning on Control-R-4K transfers to AIME2024, MATH500, GPQA-Diamond, and AIME2025.
This is the standard transfer assumption underlying the evaluation in Section 3.2.
domain assumption PRM800K grading tool correctly determines answer equivalence.
Used for grading in Section 3.1.

invented entities (2)

Reasoning Control Fields (RCF)
purpose: A prompt-injected set of 11 integer scores intended to control reasoning effort at test time.
RCFs are a new representational construct. No external evidence links the scores to actual reasoning behavior beyond the paper's own accuracy results.
In-Context Tree Search (ICTS) analogy
purpose: A conceptual model of L-CoT reasoning as a search tree to motivate the RCF dimensions.
It is a metaphor; no tree search is actually performed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Control-R: Towards controllable test-time scaling." pith.science (2026). https://pith.science/paper/XQVJUKKU

@misc{pith2026250600189,
  author       = {Pith},
  title        = {Pith review of: Control-R: Towards controllable test-time scaling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XQVJUKKU}},
  note         = {Machine review of arXiv:2506.00189}
}

read the original abstract

This paper target in addressing the challenges of underthinking and overthinking in long chain-of-thought (CoT) reasoning for Large Reasoning Models (LRMs) by introducing Reasoning Control Fields (RCF)--a novel test-time approach that injects structured control signals to guide reasoning from a tree search perspective. RCF enables models to adjust reasoning effort according to given control conditions when solving complex tasks. Additionally, we present the Control-R-4K dataset, which consists of challenging problems annotated with detailed reasoning processes and corresponding control fields. To further enhance reasoning control, we propose a Conditional Distillation Finetuning (CDF) method, which trains model--particularly Control-R-32B--to effectively adjust reasoning effort during test time. Experimental results on benchmarks such as AIME2024 and MATH500 demonstrate that our approach achieves state-of-the-art performance at the 32B scale while enabling a controllable Long CoT reasoning process (L-CoT). Overall, this work introduces an effective paradigm for controllable test-time scaling reasoning.

Figures

Figures reproduced from arXiv: 2506.00189 by the authors.

**Figure 1.** The main idea of Control-R. thinking. On the other hand, there is still a chance that LRMs engage and stuck in redundant or excessive reasoning, also known as overthinking (Kumar et al., 2025; Wang et al., 2025; Anderson et al., 2025). As shown in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** For the explanation of 11 Reasoning Control Fields. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The overall pipeline of Control-R. processes align with both the query requirements and the predefined reasoning attributes. For details of the prompting strategies used to obtain these annotations, refer to Appendix A. Additionally, the dataset employed in the annotation process, Control-R-4K, is described in Appendix B. 2.3 Conditional Distillation Fine-tuning To effectively teach the model to reason in the Long C… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Radar chart of experiments results on cutting [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 6 canonical work pages

[1]

Execution Control, 5 sub-items: a) Search Depth (search_depth) - 0: Extremely shallow, stops after one or two steps, lacks depth of thought - 9: Extremely deep, includes multi-layer reasoning or step-by-step reﬁnement, or demonstrates advanced strategies such as divide-and- conquer b) Search Breadth (search_breadth) - 0: Almost no branching, follows a sin...

work page
[2]

Quality Evaluation, 6 sub-items: a) Final Correctness (correctness) - 0: Answer is clearly incorrect - 9: Conclusion aligns perfectly with facts or the standard answer b) Reasoning Focus/Search Efficiency (efficiency) - 0: Excessive redundant branches, many ineffective attempts - 9: Achieves high-quality search with minimal resources/steps c) Search Compl...

work page
[3]

arXiv preprint arXiv:2406.16061

Port: Preference optimization on reasoning traces. arXiv preprint arXiv:2406.16061. Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xi- angru Peng, and Jiaya Jia. 2024. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. arXiv preprint arXiv:2406.18629. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, T...

arXiv 2024
[4]

arXiv preprint arXiv:2404.19733

Iterative reasoning preference optimization. arXiv preprint arXiv:2404.19733. Bhrij Patel, Souradip Chakraborty, Wesley A Sut- tle, Mengdi Wang, Amrit Singh Bedi, and Di- nesh Manocha. 2024. Aime: Ai system optimiza- tion via multiple llm evaluators. arXiv preprint arXiv:2410.03131. Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Z...

arXiv 2024
[6]

In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume1: Long Papers), pages 2609–2634

Plan-and-solve prompting: Improving zero- shot chain-of-thought reasoning by large language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume1: Long Papers), pages 2609–2634. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024a. Large search model: Redefining sear...

arXiv 2025
[7]

analysis

Large language models are better reasoners with self-verification. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2550–2575. Yuxi Xie, Anirudh Goyal, Wenyue Zheng, Min-Yen Kan, Timothy P Lillicrap, Kenji Kawaguchi, and Michael Shieh. 2024. Monte carlo tree search boosts reasoning via iterative preference learning. arXiv pr...

arXiv 2023
[8]

Execution Control

for optimization, fine-tuning based on Qwen2ForCausalLM from the Huggingface Trans- formers library 1. The training is distributed across four nodes with a total of 32 A100 GPUs on a slurm cluster (eight GPUs per node) and utilizes the Deep- Speed ZeRO-3 (Rajbhandari et al., 2020) optimiza- tion strategy to minimize memory consumption and enhance computat...

work page 2020
[11]

analysis

Your output must strictly follow the JSON format without any extra text or explanations. The root structure is as follows: { "analysis": { "execution_control_scores": { "search_depth": <int>, "search_breadth": <int>, "error_detection": <int>, "error_correction": <int>, "strategy_switching": <int> }, "quality_evaluation_scores": { "correctness": <int>, "ef...

work page

Show all 13 references

[12]

Each score must be an integer between 0 and 9
[13]

===================================================================== 【Task Instructions】 - I will provide a piece of text to be annotated

The justiﬁcation ﬁeld should contain 1 to 5 brief sentences explaining the reasoning behind your scores. ===================================================================== 【Task Instructions】 - I will provide a piece of text to be annotated. - Carefully read it and assign s...

2025
[2021]

arXiv preprint arXiv:2112.08633

Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633. Anikait Singh, Kushal Arora, Sedrick Keh, Jean Mercat, Tatsunori Hashimoto, Chelsea Finn, and Aviral Ku- mar. Improving the efficiency of test-time search in llms with backtracking. SRI_Lab....

2025 arXiv
[2023]

See https://vicuna

Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. See https://vicuna. lmsys. org (accessed 14 April 2023), 2(3):6. Tri Dao. 2023. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691. DeepSee...

2023 arXiv
[2024]

arXiv preprint arXiv:2412.16720

Openai o1 system card. arXiv preprint arXiv:2412.16720. Dhiraj Kalamkar, Dheevatsa Mudigere, Naveen Mellem- pudi, Dipankar Das, Kunal Banerjee, Sasikanth Avan- cha, Dharma Teja V ooturi, Nataraj Jammalamadaka, Jianyu Huang, Hector Yuen, et al. 2019. A study of bfloat16 for dee...

2019 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Execution Control, 5 sub-items: a) Search Depth (search_depth) - 0: Extremely shallow, stops after one or two steps, lacks depth of thought - 9: Extremely deep, includes multi-layer reasoning or step-by-step reﬁnement, or demonstrates advanced strategies such as divide-and- conquer b) Search Breadth (search_breadth) - 0: Almost no branching, follows a sin...

work page

[2] [2]

Quality Evaluation, 6 sub-items: a) Final Correctness (correctness) - 0: Answer is clearly incorrect - 9: Conclusion aligns perfectly with facts or the standard answer b) Reasoning Focus/Search Efficiency (efficiency) - 0: Excessive redundant branches, many ineffective attempts - 9: Achieves high-quality search with minimal resources/steps c) Search Compl...

work page

[3] [3]

arXiv preprint arXiv:2406.16061

Port: Preference optimization on reasoning traces. arXiv preprint arXiv:2406.16061. Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xi- angru Peng, and Jiaya Jia. 2024. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. arXiv preprint arXiv:2406.18629. Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, T...

arXiv 2024

[4] [4]

arXiv preprint arXiv:2404.19733

Iterative reasoning preference optimization. arXiv preprint arXiv:2404.19733. Bhrij Patel, Souradip Chakraborty, Wesley A Sut- tle, Mengdi Wang, Amrit Singh Bedi, and Di- nesh Manocha. 2024. Aime: Ai system optimiza- tion via multiple llm evaluators. arXiv preprint arXiv:2410.03131. Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Z...

arXiv 2024

[5] [6]

In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume1: Long Papers), pages 2609–2634

Plan-and-solve prompting: Improving zero- shot chain-of-thought reasoning by large language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume1: Long Papers), pages 2609–2634. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024a. Large search model: Redefining sear...

arXiv 2025

[6] [7]

analysis

Large language models are better reasoners with self-verification. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2550–2575. Yuxi Xie, Anirudh Goyal, Wenyue Zheng, Min-Yen Kan, Timothy P Lillicrap, Kenji Kawaguchi, and Michael Shieh. 2024. Monte carlo tree search boosts reasoning via iterative preference learning. arXiv pr...

arXiv 2023

[7] [8]

Execution Control

for optimization, fine-tuning based on Qwen2ForCausalLM from the Huggingface Trans- formers library 1. The training is distributed across four nodes with a total of 32 A100 GPUs on a slurm cluster (eight GPUs per node) and utilizes the Deep- Speed ZeRO-3 (Rajbhandari et al., 2020) optimiza- tion strategy to minimize memory consumption and enhance computat...

work page 2020

[8] [11]

analysis

Your output must strictly follow the JSON format without any extra text or explanations. The root structure is as follows: { "analysis": { "execution_control_scores": { "search_depth": <int>, "search_breadth": <int>, "error_detection": <int>, "error_correction": <int>, "strategy_switching": <int> }, "quality_evaluation_scores": { "correctness": <int>, "ef...

work page