Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

SynapseRoute: An Auto-Route Switching Framework on Dual-State Large Language Model

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A trained router can beat thinking-only accuracy while using 40% fewer tokens.

desk verdict Practical routing for dual-mode LLMs with solid cost savings, but the accuracy gain over thinking mode is not statistically established. read the letter →

arxiv 2507.02822 v1 pith:53SJYYP4 submitted 2025-07-03 cs.CL cs.AIcs.LG

classification cs.CLcs.AIcs.LG
keywords dynamicroutingdual-modeLLMthinkingmodenon-thinkingmedicalquestionansweringinferencecostreductionover-reasoningAITindex
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that a dual-mode LLM's expensive 'thinking' mode is wasted on a large share of medical questions, and that a cheap classifier can decide per question which mode to run. Its central result is a routing framework, SynapseRoute, which on 6,365 medical multiple-choice questions attains accuracy 0.8390 versus 0.8272 for thinking-only mode, while cutting inference time by 36.8% and token use by 39.66%. The reason this works is that roughly 58% of those questions are answered correctly by the non-thinking mode, and the thinking mode can actually hurt on short factual questions by over-reasoning. If correct, the finding means dual-mode LLMs can be made cheaper and more accurate at the same time without deploying separate models.

What carries the argument

The load-bearing component is a Logistic Regression classifier trained on vector embeddings of the question text. Its training labels come from an automated annotation rule: run the dual-mode model in both modes; if only one mode answers correctly, label the question accordingly; if both answer correctly, label it by the faster and cheaper mode; if both fail, discard the question. At inference time this classifier decides whether to activate the thinking mode or the non-thinking mode. The AIT index is the auxiliary evaluation mechanism that turns accuracy, inference time, and token count into one normalized score.

What would settle it

Run the same automated labeling and the same logistic-regression router on a different dual-mode LLM over the same 6,365 questions; if the non-thinking-answerable fraction falls well below 58% or the router's routing accuracy approaches the label-noise floor, the central claim that question complexity rather than model-specific behavior drives the savings would be wrong.

Watch

Extended reading notes

Core claim

The paper claims that in a sample of 6,365 medical multiple-choice questions, about 58% can be answered correctly by the non-thinking mode of a dual-mode LLM, and that a supervised router can learn to identify such questions from text embeddings alone. On the test set, SynapseRoute beats thinking-only accuracy (0.8390 vs. 0.8272) while reducing inference time by 36.8% and token consumption by 39.66%. The accuracy gain is attributed to avoiding over-reasoning on direct, information-poor questions, where the thinking mode tends to bring in irrelevant associations and self-defined terminology that lead to wrong answers. The paper also introduces the Accuracy-Inference-Token (AIT) index, a weighted combination of accuracy, inference time, and token count, as a single scalar for comparing modes under different cost-performance preferences.

Load-bearing premise

The labels that tell the router which questions need thinking are generated by one particular model's behavior, so if that model's mode behavior is not representative of other dual-mode LLMs, the reported gains may not transfer.

Editorial extensions

If this is right

  • Routing roughly 58% of medical questions to the non-thinking mode preserves or improves accuracy while cutting inference time and token cost.
  • Avoiding the thinking mode on short, direct questions removes a source of over-reasoning hallucinations, which is why overall accuracy can rise above thinking-only accuracy.
  • A lightweight logistic regression router suffices for the routing task, outperforming a LoRA-fine-tuned 4-billion-parameter model on routing accuracy while being cheaper to deploy.
  • The AIT index gives deployers a single score for comparing modes under different accuracy-versus-cost weighting schemes.

Reading between the lines

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

  • The same routing idea likely transfers to non-medical dual-mode LLM use, such as customer support or code generation, but the paper only tests medicine; labeling and routing questions from law or finance would show whether the non-thinking share and savings persist.
  • The automatic labeler is model-specific: a different dual-mode LLM may have a different non-thinking-answerable fraction, so a router trained for one base model may need retraining when the base model changes.
  • The paper's failure analysis suggests that a simple text-length or information-density heuristic might capture much of the router's behavior; testing whether a length threshold alone reproduces the accuracy and cost results would isolate what the embedding-based router adds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SynapseRoute, a routing framework that decides, per medical multiple-choice question, whether to invoke the thinking or non-thinking mode of a dual-mode LLM (Qwen3-30B-A3B). Questions are automatically labeled as 'thinking' or 'non-thinking' from the model's own correctness, inference time, and token usage in the two modes, with 'fail' questions excluded. A logistic regression classifier over bge-large embeddings is trained on these labels, and a decision threshold is chosen to maximize F1. On four medical QA datasets, the authors report that the dynamic mode improves accuracy over thinking-only mode (0.8390 vs 0.8272) while reducing inference time by 36.8% and token consumption by 39.66%. The paper also introduces an Accuracy-Inference-Token (AIT) index for jointly evaluating accuracy and cost.

Significance. If the reported accuracy gain is real, the work is practically useful: it shows that a lightweight router can cut inference cost substantially within a single dual-mode model while preserving or slightly improving accuracy. The automated labeling pipeline is cheap and scales to other datasets, and the AIT index is a reasonable way to make cost-accuracy trade-offs explicit. The cost savings are large and robust because the router sends most questions to the cheap mode. However, the central accuracy claim is currently fragile: the gain is small, is not accompanied by confidence intervals or significance tests, and may be affected by test-set threshold selection and by the exclusion of 'fail' questions from the evaluation. The paper's broader claim that the 'thinking'/'non-thinking' distinction is a property of questions rather than of one particular model is not supported by the experiments.

major comments (4)
  1. [Routing Algorithm; Table 1] The paper states that for the logistic regression router, 'we further aimed to maximize the F1 score, selecting an optimal threshold for final class prediction,' but it does not say that this threshold was chosen on a held-out validation set. The same section reports test-set AUC and accuracy, which suggests the threshold may have been tuned on the test set. Since the end-to-end accuracy advantage of SynapseRoute over thinking mode (0.8390 vs 0.8272) is the only non-mechanical part of the central claim, this is load-bearing. On 1,177 binary test questions, the difference corresponds to roughly 14 questions, and the binomial standard error is about 0.011, so the observed gap is within one standard error. Please specify the threshold selection protocol, and report confidence intervals or a significance test for the accuracy comparison.
  2. [Table 7; Results] The end-to-end accuracy comparison appears to be computed only on the 1,177 non-fail test questions, not on the full 1,273-question test set. Table 7 shows that the full test set has 1,273 questions and the binary (non-fail) test set has 1,177, meaning 96 'fail' questions are excluded. The abstract claims that SynapseRoute 'improves overall accuracy,' but on the full test set, where both modes are incorrect on fail questions, the comparison is not reported. The accuracy of thinking mode on the full set may differ from the listed 0.8272. Please clarify the exact evaluation set for Tables 1 and 2, and report the accuracy comparison on the full test set as well.
  3. [Data Labelling; Figure 2b; Abstract] The abstract and introduction state that 'approximately 58% of medical questions can be accurately answered by the non-thinking mode alone,' but the 57.8% figure is the proportion of questions labeled as 'non-thinking' after the labeling rule that breaks ties between modes by efficiency. Under that rule, a question where both modes are correct is labeled 'thinking' if the thinking mode has lower token usage and shorter latency, even though the non-thinking mode also answers it correctly. Therefore the non-thinking label count is a lower bound on the set of questions answerable correctly by non-thinking mode, and the actual percentage could be larger. The paper conflates the router's target label distribution with the model's capability. Please report the actual proportion of questions for which the non-thinking mode produces the correct answer, independent of the efficiency tie-break.
  4. [Data Labelling; Discussion] The automated labeling defines 'thinking' and 'non-thinking' questions by the behavior of a single model, Qwen3-30B-A3B. The 57.8% non-thinking figure and the router's training targets are therefore properties of this model, not intrinsic properties of the questions. The paper claims the framework can be applied broadly to other dual-mode LLMs, but no cross-model validation is provided, and the Discussion's limitations section does not mention this model dependence. At minimum, the manuscript should state this limitation explicitly; ideally, it should include a small experiment with a second dual-mode model to assess label transfer.
minor comments (5)
  1. [Results heading] The subsection heading 'Cross-Dimensional Joint Evaluation Metrics Anlysis' contains a typo; it should be 'Analysis'.
  2. [Figure 1 caption] The caption states '1.5% Accuracy Gain,' but the reported improvement is 0.8390 - 0.8272 = 0.0118, i.e., about 1.18 percentage points, not 1.5%. Please use percentage points or reconcile the numbers.
  3. [Discussion / References] In the Discussion, 'Zhang et al. (Yang et al., 2023)' is inconsistent: the citation label says Zhang et al., but the reference is Yang et al. (2023). Please correct the in-text citation.
  4. [Tables 2 and 6] Table 2 reports an average thinking-mode token size of 789.4933, while Table 6 reports a mean of 799.31 tokens for thinking mode. The difference presumably arises from the different evaluation subsets (test set vs full dataset), but this is not stated; please clarify the populations used in each table.
  5. [Data source description] The paper lists USMLE as a data source, but the Discussion refers to 'MedQA' instead. MedQA and USMLE are sometimes used interchangeably, but the names should be consistent throughout.

Circularity Check

2 steps flagged · score 4.0 of 10

The 58% statistic and the time/token savings are built into the paper's own labeling rule, but the headline accuracy gain is an end-to-end measurement and is not forced by construction.

  1. self definitional [Abstract; Methods > Data Labelling; Exploratory Data Analysis]
    "we reveal that approximately 58% of medical questions can be accurately answered by the non-thinking mode alone... If only the non-thinking mode produces the correct answer, the question is labeled as a non-thinking question... Otherwise, if the non-thinking mode is more efficient in terms of tokens and latency, the question is labeled as a non-thinking question. The results show that non-thinking questions constituted 57.8%."

    The 58% figure is not inferred from an independent measure of question complexity; it is the fraction of examples assigned the paper's own 'non-thinking' label, and that label is defined by whether Qwen3-30B-A3B happens to answer correctly in non-thinking mode (or is cheaper while correct). The 'revelation' that 58% of medical questions can be answered without thinking is therefore a restatement of the labeling rule, not a derived result. Any routing benefit built on this statistic inherits the definitional status.

  2. fitted input called prediction [Results > Comparison of inference time and token size; Methods > Data Labelling]
    "SynapseRoute's average inference time was 10.8 seconds, significantly lower than the thinking mode's 17.1 seconds, representing a 36.8% reduction. In terms of token usage, the averages were 476.4 for SynapseRoute and 789.5 for thinking mode, meaning SynapseRoute reduced token consumption by 39.66%. ... Otherwise, if the non-thinking mode is more efficient in terms of tokens and latency, the question is labeled as a non-thinking question."

    Because the labeling rule defines the 'non-thinking' class as the mode that is cheaper whenever it is also correct, and the router is optimized (F1 threshold) to predict that class, the time/token savings are substantially built into the training labels rather than discovered at evaluation. Table 6's 5-token, 1.25-second non-thinking mode versus 799.31-token, 17.24-second thinking mode makes the reported 36.8% time and 39.66% token reductions a mechanical consequence of routing a large share of questions to the cheap class. The savings are not fully forced because the router is imperfect (accuracy 0.74), so this is partial circularity, not a tautology.

full rationale

The paper's headline accuracy improvement (0.8390 vs 0.8272, Table 1) is an end-to-end measured comparison and is not logically forced by the labeling rule: the logistic-regression router has test accuracy of only 0.74 and AUC 0.82, so routing errors could plausibly depress accuracy, and the reported gain is an empirical outcome. The circularity is confined to (i) the abstract's '58%' claim, which simply restates the paper's own label distribution, and (ii) the time/token savings, which are a mechanical consequence of labeling the cheap non-thinking mode as the preferred class whenever it is correct and routing to that class. Because the central accuracy claim retains independent empirical content, the score is 4 rather than 6+. Concerns about threshold selection on the test set are a validity and reporting issue, not a circularity issue.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central results rest on labels generated by the same Qwen3-30B-A3B model whose modes are being routed, on a sampled multiple-choice subset of four public datasets, with a logistic regression threshold whose selection procedure is not fully specified. No new physical or conceptual entities are introduced; the AIT index is a weighted composite metric with hand-chosen scenario weights.

free parameters (2)
  • Logistic Regression decision threshold = not reported (chosen to maximize F1)
    The routing classifier outputs probabilities; a threshold converts them into thinking/non-thinking predictions. The paper states the threshold was selected to maximize F1 score, but does not say on a held-out validation set, so it may be tuned on the test set.
  • AIT scenario weights (a, b, c) = five hand-set scenarios (e.g., a=0.9, b=0.05, c=0.05)
    These weights define the AIT evaluation scenarios and are not fitted to data. They do not affect the central routing claim, but they determine the reported AIT scores.
assumptions (3)
  • domain assumption Qwen3-30B-A3B's thinking and non-thinking mode outputs are deterministic and representative of the dual-mode LLM class
    All labels and evaluations come from a single model; the framework's benefit is only shown for this model.
  • domain assumption The stratified sample of 6,365 multiple-choice questions from USMLE, MedMCQA, PubMedQA, and CareQA is representative of medical QA tasks
    The paper states stratified sampling preserved original distribution, but coverage of clinical subdomains is acknowledged as limited in the Discussion.
  • domain assumption Accuracy on multiple-choice questions is a sufficient measure of medical answer quality
    All evaluation is accuracy on MCQ; no human or clinical validation of answers is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SynapseRoute: An Auto-Route Switching Framework on Dual-State Large Language Model." pith.science (2026). https://pith.science/paper/53SJYYP4

@misc{pith2026250702822,
  author       = {Pith},
  title        = {Pith review of: SynapseRoute: An Auto-Route Switching Framework on Dual-State Large Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53SJYYP4}},
  note         = {Machine review of arXiv:2507.02822}
}
read the original abstract

With the widespread adoption of large language models (LLMs) in practical applications, selecting an appropriate model requires balancing not only performance but also operational cost. The emergence of reasoning-capable models has further widened the cost gap between "thinking" (high reasoning) and "non-thinking" (fast, low-cost) modes. In this work, we reveal that approximately 58% of medical questions can be accurately answered by the non-thinking mode alone, without requiring the high-cost reasoning process. This highlights a clear dichotomy in problem complexity and suggests that dynamically routing queries to the appropriate mode based on complexity could optimize accuracy, cost-efficiency, and overall user experience. Based on this, we further propose SynapseRoute, a machine learning-based dynamic routing framework that intelligently assigns input queries to either thinking or non-thinking modes. Experimental results on several medical datasets demonstrate that SynapseRoute not only improves overall accuracy (0.8390 vs. 0.8272) compared to the thinking mode alone but also reduces inference time by 36.8% and token consumption by 39.66%. Importantly, qualitative analysis indicates that over-reasoning on simpler queries can lead to unnecessary delays and even decreased accuracy, a pitfall avoided by our adaptive routing. Finally, this work further introduces the Accuracy-Inference-Token (AIT) index to comprehensively evaluate the trade-offs among accuracy, latency, and token cost.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge

    cs.AI 2026-05 unverdicted novelty 6.0 of 10

    RACER routes between reasoning and non-reasoning LLM judges via constrained distributionally robust optimization to achieve better accuracy-cost trade-offs under distribution shift.

  2. AdaThink-Med: Optimizing Inference-Time Compute for Medical Reasoning via Uncertainty Quantification

    cs.CL 2025-09 conditional novelty 6.0 of 10

    An entropy-guided reward that shortens answers to easy, confident questions and extends reasoning on hard ones reduces medical QA tokens by 4.7x to 6.4x with minimal accuracy loss.

Reference graph

Works this paper leans on

32 extracted references · 16 canonical work pages · cited by 2 Pith papers

  1. [1]

    A., Hinjos, D., Bernabeu-Perez, P., Ganzabal, L

    Arias-Duart, A., Martin-Torres, P. A., Hinjos, D., Bernabeu-Perez, P., Ganzabal, L. U., Mallo, M. G., Gururajan, A. K., Lopez-Cuena, E., Alvarez-Napagao, S., and Garcia-Gasulla, D. (2025). Automatic evaluation of healthcare llms beyond question-answering. arXiv preprint arXiv:2502.06666

  2. [2]

    A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M

    Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258

  3. [3]

    Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2024). A survey on evaluation of large language models. ACM transactions on intelligent systems and technology , 15(3):1--45

  4. [4]

    Chen, L., Zaharia, M., and Zou, J. (2023). Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176

  5. [5]

    Firoozi, R., Tucker, J., Tian, S., Majumdar, A., Sun, J., Liu, W., Zhu, Y., Song, S., Kapoor, A., Hausman, K., et al. (2025). Foundation models in robotics: Applications, challenges, and the future. The International Journal of Robotics Research , 44(5):701--739

  6. [6]

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. (2025). Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

  7. [7]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. (2022). Lora: Low-rank adaptation of large language models. ICLR , 1(2):3

  8. [8]

    Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Carney, A., et al. (2024). Openai o1 system card. arXiv preprint arXiv:2412.16720

Show all 32 references
  1. [9]

    Jin, D., Pan, E., Oufattole, N., Weng, W.-H., Fang, H., and Szolovits, P. (2021). What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences , 11(14):6421

  2. [10]

    W., and Lu, X

    Jin, Q., Dhingra, B., Liu, Z., Cohen, W. W., and Lu, X. (2019). Pubmedqa: A dataset for biomedical research question answering. arXiv preprint arXiv:1909.06146

  3. [11]

    B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D

    Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361

  4. [12]

    S., Xu, X., McDuff, D., Lee, H., Ghassemi, M., Breazeal, C., and Park, H

    Kim, Y., Park, C., Jeong, H., Chan, Y. S., Xu, X., McDuff, D., Lee, H., Ghassemi, M., Breazeal, C., and Park, H. W. (2024). Mdagents: An adaptive collaboration of llms for medical decision-making. Advances in Neural Information Processing Systems , 37:79410--79452

  5. [13]

    Kolawole, S., Dennis, D., Talwalkar, A., and Smith, V. (2024). Revisiting cascaded ensembles for efficient inference. arXiv preprint arXiv:2407.02348

  6. [14]

    Marinelli, R., Pichlmeier, J., and Bisztray, T. (2025). Harnessing chain-of-thought metadata for task routing and adversarial prompt detection. arXiv preprint arXiv:2503.21464

  7. [15]

    K., and Sankarasubbu, M

    Pal, A., Umapathi, L. K., and Sankarasubbu, M. (2022). Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. In Conference on health, inference, and learning , pages 248--260. PMLR

  8. [16]

    Y., Darzi, A., Yuan, W., and Topol, E

    Qiu, J., Lam, K., Li, G., Acharya, A., Wong, T. Y., Darzi, A., Yuan, W., and Topol, E. J. (2024a). Llm-based agentic systems in medicine and healthcare. Nature Machine Intelligence , 6(12):1418--1420

  9. [17]

    P.-W., Xiao, B., et al

    Qiu, J., Li, L., Sun, J., Peng, J., Shi, P., Zhang, R., Dong, Y., Lam, K., Lo, F. P.-W., Xiao, B., et al. (2023). Large ai models in health informatics: Applications, challenges, and the future. IEEE Journal of Biomedical and Health Informatics , 27(12):6074--6087

  10. [18]

    Qiu, J., Li, L., Sun, J., Wei, H., Xu, Z., Lam, K., and Yuan, W. (2025). Emerging cyber attack risks of medical ai agents. arXiv preprint arXiv:2504.03759

  11. [19]

    Qiu, J., Yuan, W., and Lam, K. (2024b). The application of multimodal large language models in medicine. The Lancet Regional Health--Western Pacific , 45

  12. [20]

    Strubell, E., Ganesh, A., and McCallum, A. (2020). Energy and policy considerations for modern deep learning research. In Proceedings of the AAAI conference on artificial intelligence , volume 34, pages 13693--13696

  13. [21]

    Su, J., Healey, J., Nakov, P., and Cardie, C. (2025). Between underthinking and overthinking: An empirical study of reasoning length and correctness in llms. arXiv preprint arXiv:2505.00127

  14. [22]

    Sun, J., Curtis, A., You, Y., Xu, Y., Koehle, M., Guibas, L., Chitta, S., Schwager, M., and Li, H. (2024). Hierarchical hybrid learning for long-horizon contact-rich robotic assembly. arXiv preprint arXiv:2409.16451

  15. [23]

    Sun, J., Liu, R., and Zhou, B. (2021). Hiabp: Hierarchical initialized abp for unsupervised representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 35, pages 9747--9755

  16. [24]

    Sun, J., Zheng, C., Xie, E., Liu, Z., Chu, R., Qiu, J., Xu, J., Ding, M., Li, H., Geng, M., et al. (2023). A survey of reasoning with foundation models: Concepts, methodologies, and outlook. ACM Computing Surveys

  17. [25]

    M., Hauth, A., Millican, K., et al

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., Millican, K., et al. (2023). Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

  18. [26]

    J., Ting, D

    Thirunavukarasu, A. J., Ting, D. S. J., Elangovan, K., Gutierrez, L., Tan, T. F., and Ting, D. S. W. (2023). Large language models in medicine. Nature medicine , 29(8):1930--1940

  19. [27]

    Xu, T., Huang, Z., Sun, J., Cheng, S., and Lam, W. (2025). Seqpo-simt: Sequential policy optimization for simultaneous machine translation. arXiv preprint arXiv:2505.20622

  20. [28]

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. (2025). Qwen3 technical report. arXiv preprint arXiv:2505.09388

  21. [29]

    Yang, H., Li, M., Zhou, H., Xiao, Y., Fang, Q., and Zhang, R. (2023). One llm is not enough: Harnessing the power of ensemble learning for medical question answering. medRxiv

  22. [30]

    Ye, Z., Chen, J., Light, J., Wang, Y., Sun, J., Schwager, M., Torr, P., Li, G., Chen, Y., Yang, K., et al. (2024). Reasoning in reasoning: A hierarchical framework for better and faster neural theorem proving. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24

  23. [31]

    Zheng, C., Gao, Y., Shi, H., Xiong, J., Sun, J., Li, J., Huang, M., Ren, X., Ng, M., Jiang, X., et al. (2024). Dape v2: Process attention score as feature map for length extrapolation. arXiv preprint arXiv:2410.04798

  24. [32]

    Zheng, C., Wang, H., Xie, E., Liu, Z., Sun, J., Xin, H., Shen, J., Li, Z., and Li, Y. (2023). Lyra: Orchestrating dual correction in automated theorem proving. arXiv preprint arXiv:2309.15806

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.