Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Enhancing Reasoning Capabilities of Small Language Models with Blueprints and Prompt Template Search

T0 review · 3 major / 4 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 A framework that has a large LLM write structured reasoning blueprints for small language models, then searches the prompt template per model and task, claims consistent accuracy gains on math, coding, and logic benchmarks without any…

desk verdict A useful but statistically underpowered prompting study for small models; the template-search contribution is confounded, and the paper needs an ablation and uncertainty estimates before the headline claims hold. read the letter →

arxiv 2506.08669 v1 pith:SE43IUS5 submitted 2025-06-10 cs.LG cs.AI

Dongge Han , Menglin Xia , Daniel Madrigal Diaz , Samuel Kessler , Ankur Mallick , Xuchao Zhang , Mirian Del Carmen Hipolito Garcia , Jin Xu
show 2 more authors
Victor Rühle Saravan Rajmohan
This is my paper
classification cs.LGcs.AI
keywords smalllanguagemodelsreasoningblueprintsprompttemplatesearchchain-of-thoughtoptimizationin-contextlearningautomatic
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

Small language models reason worse than large ones and are easily thrown off by small changes in prompt wording or layout. This paper argues that both problems can be mitigated without retraining: a large LLM writes a reusable blueprint, a structured step-by-step reasoning guide, for each task category and each small model, and a search procedure picks the best prompt template (which components to include and in what order) for that specific model and task. The paper claims that this combination consistently beats chain-of-thought and standard prompt-optimization baselines on math (GSM8K), coding (MBPP), and logic (BBH) across three models. If true, it offers a cheap, deployment-friendly way to get more reasoning out of models that can run on-device or in resource-constrained settings.

What carries the argument

The blueprint carries the argument: a reusable, LLM-generated, step-by-step reasoning guide for a class of problems, containing high-level reasoning steps and optionally a worked example of how to apply them. It converts abstract chain-of-thought patterns, which small models struggle to extract from in-context examples, into explicit instructions the model can follow. The second mechanism is the prompt template search, a successive-halving search over 32 combinations of template parameters (number of in-context examples from 0 to 3, task-description/examples order, blueprint inclusion, and chain-of-thought inclusion) that tailors the prompt layout to each model and task. The paper also documents that different small models prefer very different blueprint styles, for instance bullet points help GPT4o-mini most while being among the worst styles for Phi3-mini and Mistral-7B, which motivates the per-model style selection.

What would settle it

Re-run the full pipeline on the same tasks with many different random draws of the 50 training examples, tracking the test accuracy of the chosen blueprint and template across draws; if the selected configuration's test accuracy varies widely or frequently falls back to the chain-of-thought baseline level when the training examples change, the gains are a selection artifact rather than a stable property of blueprints. A cheaper check is to evaluate the losing blueprint styles on the held-out test sets: if a non-winning style matches or beats the winning style's test accuracy, the style-selection step is not actually finding the best blueprint.

Watch

Extended reading notes

Core claim

The central claim is that the limited reasoning capacity of small language models and their sensitivity to prompt variations can be offset at inference time by supplying an explicit, high-level problem-solving plan that is tailored to the task category and to the specific small model. For each task category and model, the framework has a large LLM generate candidate blueprints in twelve distinct styles, selects the style that performs best on a handful of training examples, refines the chosen blueprint with Automatic Prompt Optimization driven by the small model's own errors, and then searches 32 prompt templates via successive halving, varying the number of in-context examples, the ordering of task description and examples, and whether to include the blueprint and a chain-of-thought trigger. The experimental result is that the full pipeline is the best-performing variant in 5 of 9 model/dataset combinations and near-best in the rest, with blueprint-only variants already beating chain-of-thought and APO baselines in nearly every setting, sometimes by large margins such as +20% for Mistral-7B on MBPP.

Load-bearing premise

The load-bearing assumption is that a blueprint and prompt template chosen using a small number of training examples (10 per style, 5 per template per successive-halving round, and 25 or 20 for APO evaluation) remain near-optimal on the held-out test sets of 200 to 300 examples per task; if the search overfits to those few examples, the reported accuracy gains would be inflated.

Editorial extensions

If this is right

  • Blueprint-only prompting, without APO refinement or template search, already beats chain-of-thought (1-shot and 3-shot) and APO task-description baselines on nearly every model/dataset combination, so most of the gain comes from the blueprint itself.
  • Adding the searched prompt template yields the best overall results, winning outright in 5 of 9 combinations and staying near-optimal elsewhere, consistent with the paper's finding that no single template ordering is best across models and tasks.
  • The gains come with no parameter change and no training: the pipeline costs roughly a few hundred small-model calls to select the blueprint and template, then reuses one fixed prompt artifact at inference.
  • Blueprint style preferences differ sharply by model, so per-model style selection is a necessary component; a style that is best for one model can be among the worst for another.
  • APO refinement adds only small and mixed gains over the unrefined blueprint, indicating that blueprint generation and style selection carry most of the improvement.
  • The approach is training-free and deployment-friendly, targeting on-device or resource-constrained environments where the same weights are used with only a per-task text artifact added to the prompt.

Reading between the lines

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

  • If a static step list can move small-model accuracy this much, part of what scaling buys in large models may be the ability to discover such problem-solving structure on the fly; a testable extension is to bake blueprint-style guidance into small-model training data or fine-tuning as a cheaper substitute for some parameter growth.
  • The template search's success implies best-practice prompting advice should be treated as per-model and per-task data rather than a universal default; the 32-template successive-halving procedure is small enough to be re-run at deployment time for each new task domain, possibly with more training examples to reduce selection noise.
  • A natural stress test the paper does not run: apply the same pipeline to tasks it was not designed for, such as open-ended generation or tool use. If the gains persist there, the blueprint mechanism is about task structure generally; if not, it is specific to closed-form reasoning benchmarks.
  • The paper's own note that template search sometimes picks a suboptimal template due to small evaluation budgets suggests a direct follow-up: increase the per-candidate examples during successive halving and measure how much of the 5-of-9 winning margin survives.
  • The style-preference findings for three models hint that per-model prompt style could matter for any prompt-engineering technique, not just blueprints, so a broader survey of styles across more small models would test how general the preference divergence is.
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

3 major / 4 minor

Summary. The paper proposes a training-free framework that aims to improve the reasoning of small language models (SLMs) by (1) generating reusable 'blueprints' with an LLM from a small set of examples, (2) selecting among twelve blueprint styles on training examples, (3) refining the selected blueprint with Automatic Prompt Optimization (APO), and (4) searching over 32 prompt templates with successive halving. The method is evaluated on GSM8K, MBPP, and the BBH suite for GPT4o-mini, Mistral-7B, and Phi3-mini, reporting consistent accuracy gains over CoT and APO baselines.

Significance. If the reported gains are robust, the framework is a practically useful, lightweight alternative to model fine-tuning for resource-constrained deployments, and the paper provides a broad evaluation across 28 task categories and three SLMs. The qualitative example in Section 4.3 is a genuine strength: it shows the blueprint being followed step-by-step and contrasts this with an incorrect unguided response. The paper is also honest about the instability of template search on small evaluation sets. However, the experimental design does not yet isolate the contribution of blueprints from the contribution of the prompt search itself, and the absence of uncertainty estimates and budget-matched baselines makes the headline claims weaker than the abstract states.

major comments (3)
  1. [§3.3, Appendix B, Table 1] The central claim that blueprints improve SLM reasoning is underdetermined by the experiments. The template search space in Appendix B jointly varies four parameters: number of in-context examples (0-3), task-description ordering, Include Blueprint, and Include CoT. The BP (w. APO) + TS row in Table 1 can therefore differ from the BP (w. APO) row by changing the number of shots, adding or removing the CoT phrase, or even dropping the blueprint entirely. The CoT and APO baselines receive no template search, so the comparison bundles blueprint content together with a much larger search budget and a broader search space. A necessary control is a template-search-only condition: the same successive-halving search with the same SLM-call budget and with Include Blueprint forced to False, compared against a condition with Include Blueprint forced to True. Without this, the results do not show that blueprints, rather than prompt search, drive the gains.
  2. [§4.2, Appendix E, §A.2, §B] The optimization budget is strongly asymmetric, which weakens the claim of 'outperforming state-of-the-art methods'. Per task-model, the proposed method spends about 120 SLM calls on blueprint-style selection, 125 SLM calls on APO refinement, and 310 SLM calls on template search, for a total of roughly 555 SLM calls, while the CoT baselines use no search and the APO baseline uses a smaller budget. The abstract and introduction attribute the improvement to blueprints and prompt template optimization without acknowledging that the comparison is not budget-matched. The paper should either add equal-budget baselines (e.g., the same successive-halving search over templates with no blueprint, or APO with an equivalent number of candidates) or explicitly discuss the computational cost and qualify the comparison.
  3. [§4.2, §4.4, Appendix E] The paper reports no confidence intervals, significance tests, or resampling-based stability analysis. Although temperature=0 makes inference deterministic, the selection procedure depends on sampled training examples (10 per style, 25 and 20 during APO, 5 per template during successive halving) and on stochastic LLM blueprint generation, so the reported differences of 1-3% in Table 1 could easily be within the range of selection noise. The paper's own caveat in Section 4.2 and Appendix C that template search 'does not always provide the best result' due to the small evaluation set reinforces this concern. I ask for bootstrap confidence intervals over the test examples and, for at least a subset of tasks, repeated runs with different training-example draws to show that the selected blueprint/template and the resulting accuracy are stable.
minor comments (4)
  1. [Introduction, Section 2] There is a typo in the introduction: 'problem-solving viablueprints' should read 'problem-solving via blueprints'.
  2. [Table 1] The abbreviation 'w.o.' should be written as 'w/o' for consistency with standard usage.
  3. [Abstract, Section 4.2] The abstract's phrase 'outperforming state-of-the-art methods' is too broad; the experiments compare against CoT and APO baselines only, so the claim should be limited to those baselines or supported with additional comparisons.
  4. [Appendix E] The setup states that 50 training examples are used per dataset, but the style selection, APO, and template search use 10, 25/20, and 5 examples respectively; it would help to clarify how these subsets are sampled and whether they overlap across the stages.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: blueprint and template selection use training examples, and all reported gains are on held-out test splits.

full rationale

The derivation chain is a standard train/valid/test pipeline. Blueprint style selection uses 10 training examples per style (Appendix A.1), APO refinement evaluates on 25 and 20 training examples (Appendix A.2), and template search uses 5 training examples per successive-halving round (Appendix B); all final numbers in Table 1 and Fig. 9 are reported on held-out test sets of 200-300 datapoints per task (Section 4.1 and Appendix E). No fitted parameter, optimized template, or selected blueprint is reused to compute the evaluation metric on the same examples, so there is no self-definitional or fitted-input-called-prediction step. The blueprint claim is independently supported by the BP(w.o. APO) row, which precedes APO and template search. APO (Pryzant et al., 2023) is used as an external algorithm with full prompts in the appendix; there is no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in via citation. The absence of a template-search-only control is a possible attribution confound for the BP(w. APO)+TS row because the 32-template space also varies shots, task/example order, and the CoT phrase, but this is an experimental-design limitation, not circularity: the searched configuration is still selected on training data and measured on held-out data. The paper's own caveat that template search 'does not always provide the best result' due to few evaluation examples is a stability concern, not a circular one.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the transfer of prompt choices from small training samples to larger held-out test sets, and on the stability of proprietary LLM/SLM APIs. No physical entities are invented; the 'blueprint' is a text artifact. The framework introduces several hand-chosen hyperparameters but no fitted constants in the equation sense.

free parameters (6)
  • M (number of training examples for blueprint generation) = 50
    Set to 50 for all tasks (Appendix E). Larger or smaller sample may change blueprint quality and downstream accuracy.
  • K (number of blueprint styles) = 12
    Hand-selected set (Appendix A.1). The best achievable blueprint is bounded by the style set; other styles could shift results.
  • APO rounds = 1 (with 2 gradients, 2 edits, 2 paraphrases)
    A single APO round; more rounds would increase SLM/LLM calls and could improve or overfit the blueprint.
  • Successive halving reduction factor f = 2
    Template search halves candidates each round (Appendix B). Other reduction factors would change the selected template.
  • Training examples per template candidate per search round = 5
    Used to rank 32 templates (Appendix B). This small sample makes selection noisy, a key threat to reproducibility.
  • SLM decoding temperature/top-p = 0/1
    Greedy decoding reported for SLM evaluation (Appendix E). Blueprint generation by GPT-4o does not report temperature.
assumptions (4)
  • domain assumption Held-out splits of GSM8K, MBPP, and BBH are clean and representative.
    Standard benchmark assumption; no contamination analysis is provided, despite GPT-4o being used to generate blueprints.
  • domain assumption GPT-4o generates blueprints and APO gradients of sufficient quality and stability for single-run evaluation.
    The method's core generator is a proprietary model; no seeds, decoding parameters, or repeated generations are reported.
  • domain assumption Regular-expression answer extraction gives unbiased accuracy estimates across methods.
    Used for 26 of 27 BBH categories; word sorting is excluded (Appendix E).
  • domain assumption API versions of GPT-4o, GPT-4o-mini, Mistral-7B, and Phi-3-mini are stable during the evaluation window.
    Model updates could change results; no version or date hashes are given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Reasoning Capabilities of Small Language Models with Blueprints and Prompt Template Search." pith.science (2026). https://pith.science/paper/SE43IUS5

@misc{pith2026250608669,
  author       = {Pith},
  title        = {Pith review of: Enhancing Reasoning Capabilities of Small Language Models with Blueprints and Prompt Template Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SE43IUS5}},
  note         = {Machine review of arXiv:2506.08669}
}
read the original abstract

Small language models (SLMs) offer promising and efficient alternatives to large language models (LLMs). However, SLMs' limited capacity restricts their reasoning capabilities and makes them sensitive to prompt variations. To address these challenges, we propose a novel framework that enhances SLM reasoning capabilities through LLM generated blueprints. The blueprints provide structured, high-level reasoning guides that help SLMs systematically tackle related problems. Furthermore, our framework integrates a prompt template search mechanism to mitigate the SLMs' sensitivity to prompt variations. Our framework demonstrates improved SLM performance across various tasks, including math (GSM8K), coding (MBPP), and logic reasoning (BBH). Our approach improves the reasoning capabilities of SLMs without increasing model size or requiring additional training, offering a lightweight and deployment-friendly solution for on-device or resource-constrained environments.

Figures

Figures reproduced from arXiv: 2506.08669 by the authors.

Figure 1
Figure 1. Illustration of our framework. During training, for each SLM and task category, an LLM is used to generate blueprints using various input styles, from which the best blueprint is picked, and refined via APO. Then we perform template search to find the best prompt template. During inference, the chosen blueprint and template are reused across problems within the task category to form the input prompt to the SLM. 1. U… view at source ↗
Figure 2
Figure 2. Example LLM generated blueprint. The blueprint in￾cludes high-level reasoning steps, and optionally an in-context example question and solution process as a guide for how to apply the reasoning steps. consistently improves SLM performance, outperforming state-of-the-art methods. These findings highlight the effec￾tiveness of blueprint-guided reasoning and prompt template optimization in overcoming the inherent limit… view at source ↗
Figure 3
Figure 3. Sensitivity of SLMs to prompt variations across tasks - Even simple changes in the template, like swapping the order of <task-description> and <in-context-example>, can significantly affect performance, with impacts varying per SLM and task, e.g., in Sports understanding (Snarks, Sports understanding and Temporal sequences are sub-categories from the BBH dataset), GPT4o-mini performed better with examples before the… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Example blueprint from the BBH geometric shape reasoning task, and example GPT4o-mini responses without and with blueprint, respectively. We can observe that the blueprint listed the reasoning steps, as well as showing an example application, which guides the SLM how t…
Figure 5
Figure 5. Figure 5: Accuracy of SLM task performance across blueprint styles. For each blueprint style and SLM, the performance is averaged over 280 examples (10 samples per task category for a total of 28 task categories spanning the GSM8K, MBPP and BBH datasets.) in bullet pointed forma…
Figure 6
Figure 6. Figure 6: Blueprint generation prompt, blueprint styles and their corresponding generating instructions. When generating a blueprint, take a blueprint style and the corresponding generating instruction, which replaces the <style-instruction-placeholder> to form part of the bluep…
Figure 7
Figure 7. Figure 7: Sensitivity of SLMs to prompt variations across tasks - the number of in-context examples can significantly impact the SLM performance across different tasks. For instance in the temporal sequences dataset (BBH), GPT4o-mini shows relatively stable a d slightly improved…
Figure 8
Figure 8. Figure 8: Prompts used for blueprint refinement via Automatic Prompt Optimization (APO). During each APO round, the LLM first performs error analysis and generates a textual gradient using the Gradient Prompt, then refines and updates the blueprint according to the textual gradi…
Figure 9
Figure 9. Figure 9: Per-category Performance on the BBH dataset 14 [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. AdaHome: An Adaptive Smart Home Assistant using Local Small Language Models

    cs.AI 2026-07 conditional novelty 4.0 of 10

    A local small-model smart home assistant that routes simple commands to a fast prompt and vague ones to brief draft reasoning, then personalizes actions from a feedback-driven preference memory.

Reference graph

Works this paper leans on

26 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    A., Awan, A

    Abdin, M., Jacobs, S. A., Awan, A. A., Aneja, J., Awadallah, A., Awadalla, H., Bach, N., Bahree, A., Bakhtiari, A., Behl, H., et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024

    arXiv 2024

  2. [2]

    Promptwizard: Task-aware agent-driven prompt optimization framework

    Agarwal, E., Dani, V., Ganu, T., and Nambi, A. Promptwizard: Task-aware agent-driven prompt optimization framework. arXiv preprint arXiv:2405.18369, 2024

    arXiv 2024

  3. [3]

    Program synthesis with large language models

    Austin, J., Odena, A., Nye, M., Bosma, M., Michalewski, H., Dohan, D., Jiang, E., Cai, C., Terry, M., Le, Q., et al. Program synthesis with large language models. arXiv preprint arXiv:2108.07732, 2021

    arXiv 2021

  4. [4]

    Brown, T. B. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020

    arXiv 2005

  5. [5]

    W., Sutton, C., Gehrmann, S., et al

    Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H. W., Sutton, C., Gehrmann, S., et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24 0 (240): 0 1--113, 2023

    2023

  6. [6]

    Training verifiers to solve math word problems

    Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021

    arXiv 2021

  7. [7]

    P., and Hu, Z

    Deng, M., Wang, J., Hsieh, C.-P., Wang, Y., Guo, H., Shu, T., Song, M., Xing, E. P., and Hu, Z. Rlprompt: Optimizing discrete text prompts with reinforcement learning. arXiv preprint arXiv:2205.12548, 2022

    arXiv 2022

  8. [8]

    M., Tong, S., Lepikhin, D., Xu, Y., Krikun, M., Zhou, Y., Yu, A

    Du, N., Huang, Y., Dai, A. M., Tong, S., Lepikhin, D., Xu, Y., Krikun, M., Zhou, Y., Yu, A. W., Firat, O., et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pp.\ 5547--5569. PMLR, 2022

    work page 2022

Show all 26 references
  1. [9]

    Promptbreeder: Self-referential self-improvement via prompt evolution

    Fernando, C., Banarse, D., Michalewski, H., Osindero, S., and Rockt \"a schel, T. Promptbreeder: Self-referential self-improvement via prompt evolution. arXiv preprint arXiv:2309.16797, 2023

    2023 arXiv

  2. [10]

    Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D

    Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. d. l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023

    2023 arXiv

  3. [11]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp.\ 1207--1216, Stanford, CA, 2000. Morgan Kaufmann

    2000

  4. [12]

    The power of scale for parameter-efficient prompt tuning

    Lester, B., Al-Rfou, R., and Constant, N. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021

    2021 arXiv

  5. [13]

    Large language model guided tree-of-thought

    Long, J. Large language model guided tree-of-thought. arXiv preprint arXiv:2305.08291, 2023

    2023 arXiv

  6. [14]

    Are large language models good prompt optimizers? arXiv preprint arXiv:2402.02101, 2024

    Ma, R., Wang, X., Zhou, X., Li, J., Du, N., Gui, T., Zhang, Q., and Huang, X. Are large language models good prompt optimizers? arXiv preprint arXiv:2402.02101, 2024

    2024 arXiv

  7. [15]

    C., Mallinson, J., Adamek, J., Malmi, E., and Severyn, A

    Magister, L. C., Mallinson, J., Adamek, J., Malmi, E., and Severyn, A. Teaching small language models to reason. arXiv preprint arXiv:2212.08410, 2022

    2022 arXiv

  8. [16]

    Hello gpt-4o

    OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024 a

    2024

  9. [17]

    Gpt-4o mini: Advancing cost-efficient intelligence

    OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024 b

    2024

  10. [18]

    gradient descent

    Pryzant, R., Iter, D., Li, J., Lee, Y. T., Zhu, C., and Zeng, M. Automatic prompt optimization with" gradient descent" and beam search. arXiv preprint arXiv:2305.03495, 2023

    2023 arXiv

  11. [19]

    W., Chowdhery, A., Le, Q

    Suzgun, M., Scales, N., Sch \"a rli, N., Gehrmann, S., Tay, Y., Chung, H. W., Chowdhery, A., Le, Q. V., Chi, E. H., Zhou, D., et al. Challenging big-bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261, 2022

    arXiv

  12. [20]

    Efficient large language models: A survey

    Wan, Z., Wang, X., Liu, C., Alam, S., Zheng, Y., Qu, Z., Yan, S., Zhu, Y., Zhang, Q., Chowdhury, M., et al. Efficient large language models: A survey. arXiv preprint arXiv:2312.03863, 1, 2023

    2023 arXiv

  13. [21]

    V., Zhou, D., et al

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 0 24824--24837, 2022

    2022

  14. [22]

    Larger language models do in-context learning differently

    Wei, J., Wei, J., Tay, Y., Tran, D., Webson, A., Lu, Y., Chen, X., Liu, H., Huang, D., Zhou, D., et al. Larger language models do in-context learning differently. arXiv preprint arXiv:2303.03846, 2023

    2023 arXiv

  15. [23]

    Zhongjing: Enhancing the chinese medical capabilities of large language model through expert feedback and real-world multi-turn dialogue

    Yang, S., Zhao, H., Zhu, S., Zhou, G., Xu, H., Jia, Y., and Zan, H. Zhongjing: Enhancing the chinese medical capabilities of large language model through expert feedback and real-world multi-turn dialogue. In Proceedings of the AAAI Conference on Artificial Intelligence, volum...

    2024

  16. [24]

    React: Synergizing reasoning and acting in language models

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022

    2022 arXiv

  17. [25]

    I., Han, Z., Paster, K., Pitis, S., Chan, H., and Ba, J

    Zhou, Y., Muresanu, A. I., Han, Z., Paster, K., Pitis, S., Chan, H., and Ba, J. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910, 2022

    2022 arXiv

  18. [26]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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