Conceptual Metaphor Theory as a Prompting Paradigm for Large Language Models

Oliver Kramer

REVIEW 5 major objections 5 minor 2 cited by

Conceptual Metaphor Theory as a Prompting Paradigm for Large Language Models

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Adding one Conceptual Metaphor Theory system message, which tells the model to map abstract target domains onto concrete source domains and draw stepwise inferences, improves LLM reasoning accuracy, clarity, and metaphorical coherence…

desk verdict Plausible prompt idea, but the evaluation is unblinded and metaphor-rewarding, so the reported CMT advantage is likely an artifact. read the letter →

arxiv 2502.01901 v1 pith:7F3UENIB submitted 2025-02-04 cs.CL

Oliver Kramer This is my paper · ORCID

classification cs.CL

keywords ConceptualMetaphorTheorypromptengineeringchain-of-thoughtpromptingsource-targetmappinglargelanguagemodelsautomaticevaluationreasoningbenchmarksinterpretation

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 reports that a single system-message change—instructing the model to reason through Conceptual Metaphor Theory, mapping abstract target domains onto concrete source domains and drawing step-by-step inferences—improves LLM performance on metaphor-heavy reasoning and explanation tasks. Four small models (Llama3.2, Phi3, Gemma2, Mistral) were each run with and without this CMT prompt on a 100-task benchmark covering metaphor identification, domain-specific reasoning, explanation and teaching, and metaphor reading comprehension. The central claim is that the CMT-prompted versions outperform their baselines across all four categories, as scored by a Llama3.3 70B judge on accuracy, coherence, and metaphorical alignment. If true, the result matters because it offers a cheap, task-independent way to make LLM reasoning more structured and its explanations more intuitive, with no fine-tuning and no per-task example crafting.

What carries the argument

The load-bearing object is the CMT system prompt: a fixed instruction telling the model it is a cognitive agent that interprets abstract concepts (target domains) through concrete experiences (source domains), followed by a mapping-and-inference template and three worked metaphors (time is money, heart of stone, world is a stage). The prompt is applied uniformly across tasks at temperature 0.7, so the only difference between baseline and CMT conditions is the presence of this metaphor-reasoning instruction and its examples. On the evaluation side, the machinery is an LLM judge that receives both responses plus the task description and scoring criteria, assigns 1–5 scores on three criteria per category, and selects the superior response; the CMT condition wins in the reported comparisons.

What would settle it

Run the same 100-task benchmark with the condition labels removed from the judge prompt—or with independent human raters applying the same criteria—and check whether CMT responses still win systematically; if the advantage shrinks to noise, the paper's central empirical claim collapses.

Watch

Extended reading notes

Core claim

The discovery, on the paper's own terms, is that conceptual-metaphor structure is a usable reasoning scaffold for LLMs. By encoding source-domain-to-target-domain mappings once in the system prompt—with three worked examples showing how to infer properties of an abstract concept from a concrete experience—the authors turn metaphor from a linguistic decoration into an inference procedure, in the same spirit as chain-of-thought turns arithmetic into a sequence of steps. The resulting responses are judged by Llama3.3 70B to be more accurate, clearer, and more metaphorically coherent than baseline responses, with gains largest where explanations must transfer structure from one domain to another. The paper interprets this as evidence that CMT-based prompting provides a generalizable cognitive framework rather than a set of task-specific tricks.

Load-bearing premise

The whole result rests on trusting that the Llama3.3 70B judge gives unbiased quality scores even though it can see which response came from the CMT-prompted model and is explicitly asked to reward metaphor effectiveness.

Editorial extensions

If this is right

A single CMT system message could replace per-task prompt engineering for explanation-heavy applications, because no task-specific examples are required.
Small models—including a 2B-parameter model—show gains, so the recipe may transfer to low-resource or on-device settings where fine-tuning is impractical.
If the score gap reflects genuine improvement, metaphor-guided prompting is a default option for teaching, tutoring, and explainability outputs where analogies carry the communication.
The uneven model-by-model results imply that practitioners should expect CMT prompting to help some architectures and task types more than others, so conditional use is a reasonable deployment strategy.

Reading between the lines

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

Because the judge in Figure 3 is told which response is the CMT one and is instructed to reward the 'effectiveness of analogy or metaphor,' part of the reported margin may reflect style preference rather than reasoning quality; a blind judge or human raters would test this.
CMT prompting may primarily raise rhetorical coherence rather than factual correctness; separating correctness from communicative quality would clarify what the gains are made of.
The benchmark is small—100 tasks with selected instances per category and no reported variance or significance tests—so the word 'significant' in the abstract should be read as descriptive, not statistical, until error bars are supplied.
Using an evaluator from the same model family as one of the tested models (Llama3.2 vs Llama3.3 judge) invites self-preference effects; evaluating with a different judge family would strengthen the comparison.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Colleague,

Short version: this is a modest prompt-engineering paper with a plausible idea and an evaluation that cannot support its headline claim. The CMT vs. baseline comparison is not trustworthy because the judge knows which response is which and the scoring criteria reward metaphor use—the exact behavior the CMT prompt induces. The abstract's 'significantly enhances' should be read as 'preferred by a label-aware judge using metaphor-weighted criteria.'

What is actually new: the benchmark of 100 tasks across four categories (MIM, DSR, ETT, RCM) is thoughtfully designed, and the qualitative observations in Section VI.C give a useful sense of where metaphor prompting might help, especially technical explanations and teaching-oriented output. The idea of using CMT as a system prompt is a natural extension of the author's prior cognitive prompting work. The paper is readable and honest enough to flag 'unusually high CMT-enhanced scores' as a potential evaluation artifact.

The soft spots are in the evidence, not the idea. The evaluator prompt in Figure 3 labels responses as 'Baseline' and 'CMT'; the ETT criteria explicitly include 'Effectiveness of Analogy or Metaphor'; and the CMT system prompt instructs the model to map source to target domains. A judge primed with the treatment label and metaphor-oriented criteria can prefer the CMT response for stylistic reasons, not because it is more accurate. There are no significance tests, no error bars, and no raw data. The paper mentions inter-rater agreement but never reports it. There is no control for plain chain-of-thought prompting, so we cannot attribute gains to CMT specifically rather than to longer, structured instructions. The models are small (2B–7B) and the judge is larger (70B), which is fine, but it does not fix the labeling problem.

The central argument doesn't hold as stated. The idea is worth a careful follow-up with blind evaluation, a CoT control, and some statistical reporting. For a workshop this would be a reasonable discussion piece; as a citable empirical claim it needs another pass.

For peer review: I would send it to review only with the expectation of a major revision. The benchmark and idea deserve referee time, but the current evaluation section would need to be redone. If you are considering it for a reading group, it is a good example of LLM-as-judge bias in action.

Referee Report

5 major / 5 minor

Summary. The paper proposes Conceptual Metaphor Theory (CMT) as a prompting paradigm for large language models. The intervention is a system message (Figure 1) that instructs the model to reason by mapping abstract target domains onto concrete source domains, with illustrative examples (Figure 2). The authors compare four native models (Llama3.2, Phi3, Gemma2, Mistral) against their CMT-configured versions on 100 tasks in four categories: Metaphor Identification and Mapping, Domain-Specific Reasoning, Explanation and Teaching Tasks, and Reading Comprehension of Metaphors. Responses are scored by Llama3.3 70B using an evaluation prompt (Figure 3), and the paper reports average scores in Figure 4, concluding that CMT prompting 'significantly enhances reasoning accuracy, clarity, and metaphorical coherence.'

Significance. The idea of using cognitive-linguistic metaphor theory to structure LLM prompting is interesting and connects to a live research area on metaphor understanding and reasoning. The paper gives a clear statement of the intervention and a benchmark that covers several relevant task types. If the empirical claims were supported by sound, blinded, statistically grounded evaluation, the result would be a useful practical contribution to prompt engineering. However, the current evidence does not establish the central claim: the automated evaluator is unblinded, the scoring criteria explicitly reward metaphor use, no significance tests or variance estimates are reported, and at least one displayed result contradicts the text. The paper would need substantial additional empirical work to support its abstract-level conclusion.

major comments (5)

[Section VI.A and Figure 3] The evaluation protocol is unblinded and confounded with the intervention. The evaluation prompt in Figure 3 explicitly labels the two responses as 'Baseline Model Response' and 'CMT-prompted Model Response,' and the criteria include 'effectiveness of analogy or metaphor' (Section V.C). Because the CMT system prompt instructs the model to produce source-target metaphorical mappings, the judge can prefer CMT responses simply because they match the requested rhetorical style, not because they are more accurate. The paper's central claim that CMT 'significantly enhances reasoning accuracy' therefore rests on an outcome measure that overlaps with the intervention. A blinded evaluation, or at minimum a judge that does not know which response is CMT, is necessary to support the claim.
[Section VI.C and Figure 4] No statistical evidence is provided for the word 'significantly' in the abstract. Figure 4 reports only average scores without error bars, standard deviations, per-task counts, or significance tests, and the text in Section VI.A claims inter-rater agreement was assessed but gives no values. Without variance or raw data, the reader cannot distinguish systematic improvement from noise, especially on a 100-task benchmark with small per-category samples. The manuscript should report confidence intervals or significance tests, and ideally provide the response-level data or a public artifact.
[Section VI.C, Figure 4 (right), and Section VII] There is a direct internal contradiction in the RCM results. In Figure 4 (right), the llama3.2 baseline RCM score is 3.81 while the CMT score is 1.24, a large negative difference, yet Section VI.C states that 'Llama3.2, Gemma2, and Mistral showed substantial gains' in RCM, and Section VII concludes that 'Llama3.2 benefits the most from CMT.' Either the figure labels are misaligned or the textual summary is wrong; this must be corrected and reconciled because it affects the paper's per-model conclusions.
[Section III.B and Section VI.A] The comparison lacks an appropriate control for Chain-of-Thought (CoT). The CMT prompt explicitly uses a 'CoT-like approach' with stepwise inference (Section III.B), but the baseline models receive no CoT instruction. Consequently, any observed improvement could be attributed to the presence of step-by-step reasoning rather than to conceptual metaphor mappings. A CoT-only control condition, matched for prompt length and reasoning steps, is needed to isolate the CMT contribution.
[Section V.C and Figure 3] The benchmark tasks and evaluation criteria appear to reward the very behavior the intervention is designed to elicit, which raises a circularity concern. Section V states that task instructions deliberately avoid prompting metaphorical reasoning, but the evaluation criteria in Sections V.C and V.D explicitly reward 'effectiveness of analogy or metaphor' and 'depth of interpretive insight.' The paper should either use outcome measures that are independent of metaphor use or justify why metaphor-richness is an acceptable proxy for reasoning quality on these tasks.

minor comments (5)

[Figure 2] The figure caption contains a typo: 'CMT-inspred CoT' should read 'CMT-inspired CoT.'
[Section VI.A] The claim that 'Independent evaluations by multiple annotators ensured consistency, with inter-rater agreement assessed for reliability' is not supported by any reported data, and it is unclear how these human annotations relate to the Llama3.3 evaluations that are actually presented.
[Section VI.C] The phrase 'Minor anomalies, such as unusually high CMT-enhanced scores, suggest potential refinements to evaluation methods' is vague and does not identify which models or tasks are anomalous; the RCM llama3.2 result identified above is more than a minor anomaly and needs explicit discussion.
[References] Reference [4] has a formatting error: the journal/proceedings volume information appears to be duplicated from reference [3], and the venue should be verified.
[Section III.C] Temperature is the only free parameter mentioned, and it is fixed at 0.7 without any sensitivity analysis; a brief justification or ablation would increase confidence that the results are not sensitive to this choice.

Circularity Check

1 steps flagged · score 6.0 of 10

The reported CMT advantage is largely an artifact of an unblinded, metaphor-rewarding evaluator: the outcome metric overlaps with the intervention by construction.

self definitional [Section V.C and Section VI.A, Figures 1 and 3]
"Evaluation Criteria: ... (3) Effectiveness of Analogy or Metaphor, which evaluates how well the chosen metaphor aids comprehension and aligns with the original concept. ... The evaluator received detailed instructions, including the task description, baseline and CMT responses, and scoring criteria. ... Baseline Model Response: ... CMT-prompted Model Response: ..."

The CMT system prompt (Fig. 1) instructs the model to behave in exactly the way the rubric rewards: 'As a cognitive agent utilizing Conceptual Metaphor Theory (CMT), you can interpret abstract concepts (target domains) through more concrete experiences (source domains).' The ETT rubric explicitly scores 'Effectiveness of Analogy or Metaphor', and the MIM/DSR/RCM criteria reward source-target mapping accuracy. The Llama3.3 judge is also told which response is the CMT-prompted one (Fig. 3). Thus a CMT model that follows its system prompt is rewarded for the very property the prompt induces, independent of whether reasoning accuracy improved.

full rationale

The paper's central empirical claim rests on a single unblinded auto-evaluation protocol. The CMT treatment is a system prompt that mandates source-to-target metaphorical mapping; the scoring criteria reward effective analogy/metaphor and source-target mapping; and the judge is explicitly shown the treatment label. Under these conditions, the CMT-prompted response is almost guaranteed to score higher on metaphor-related criteria simply by compliance with its system prompt, making the headline 'metaphorical coherence' advantage partly definitional. There is no fitted parameter being renamed as a prediction, no imported uniqueness theorem, and no load-bearing self-citation chain here; the one self-citation (Kramer and Baumann on cognitive prompting) is background related work, not the basis of the main claim. The circularity is therefore concentrated in the evaluation design rather than in a derivation chain. Because the reported gain on non-metaphor criteria like 'accuracy' could in principle still reflect genuine reasoning improvement, the paper is not fully circular, but the evidence as presented cannot separate prompt-following from reasoning gains. Score 6 reflects that the central claim partially reduces, by construction, to the match between the intervention and the rubric.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the validity of CMT for machine reasoning and on the reliability of the automatic evaluator. The temperature is a hand-chosen hyperparameter. No fitted constants or new entities are introduced.

free parameters (1)

temperature = 0.7
Chosen by hand for balanced creativity and coherence; not tuned to the benchmark.

assumptions (3)

domain assumption Conceptual Metaphor Theory is a valid framework for improving machine reasoning.
The paper assumes that encoding CMT instructions in a system prompt will transfer human cognitive benefits to LLMs, but the experiments do not isolate this mechanism.
domain assumption Llama3.3 70B provides fair and reliable scoring when it knows which response is the treatment.
The evaluation prompt in Figure 3 explicitly labels responses as Baseline vs CMT, and the paper reports no validation of this judge against human ratings.
ad hoc to paper The benchmark tasks are appropriate and unbiased for measuring general reasoning improvement.
The tasks are metaphor-heavy and the scoring criteria reward metaphor use, aligning with the CMT treatment.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Conceptual Metaphor Theory as a Prompting Paradigm for Large Language Models." pith.science (2026). https://pith.science/paper/7F3UENIB

@misc{pith2026250201901,
  author       = {Pith},
  title        = {Pith review of: Conceptual Metaphor Theory as a Prompting Paradigm for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7F3UENIB}},
  note         = {Machine review of arXiv:2502.01901}
}

read the original abstract

We introduce Conceptual Metaphor Theory (CMT) as a framework for enhancing large language models (LLMs) through cognitive prompting in complex reasoning tasks. CMT leverages metaphorical mappings to structure abstract reasoning, improving models' ability to process and explain intricate concepts. By incorporating CMT-based prompts, we guide LLMs toward more structured and human-like reasoning patterns. To evaluate this approach, we compare four native models (Llama3.2, Phi3, Gemma2, and Mistral) against their CMT-augmented counterparts on benchmark tasks spanning domain-specific reasoning, creative insight, and metaphor interpretation. Responses were automatically evaluated using the Llama3.3 70B model. Experimental results indicate that CMT prompting significantly enhances reasoning accuracy, clarity, and metaphorical coherence, outperforming baseline models across all evaluated tasks.

Figures

Figures reproduced from arXiv: 2502.01901 by the authors.

**Figure 2.** CMT-inspred CoT By embedding these examples in the CMT-prompts, the models acquire the ability to generalize structured metaphor- [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Prompt for evaluation with Llama3.3 B. Experimental Setup We selected representative tasks from each category. These tasks were tested across four large language models: Llama3.2 with 3B parameters, Phi3 with 3.8B parameters, Gemma2 with 2B parameters, and Mistral with 7B parameters. Each model was tested in both its baseline configuration and with CMTenhanced prompting. The responses were evaluated based on the pr… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Comparison of baseline and CMT-enhanced LLM performance across task categories. The four categories—MIM, DSR, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

Metaphor Tracer: A Theory-Informed Analysis of Hidden States
cs.AI 2026-07 conditional novelty 7.0 of 10

Hidden-state aggregator and differentiator scores, frozen on one text, track within-text organization across models and align with engineered registers and psychoanalytic marks while dissociating from information and ...
The Mind's Eye: A Multi-Faceted Reward Framework for Guiding Visual Metaphor Generation
cs.CL 2025-08 conditional novelty 6.0 of 10

A two-pipeline visual metaphor generation framework that decomposes metaphors into source, target, and meaning and refines prompts against a VLM-based reward, beating GPT-4o and Imagen on automatic scores but not on h...

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages · cited by 2 Pith papers

[1]

Metaphors We Live By

George Lakoff and Mark Johnson. Metaphors We Live By . University of Chicago Press, Chicago, IL, 1980

work page 1980
[2]

The contemporary theory of metaphor

George Lakoff. The contemporary theory of metaphor. In Metaphor and Thought, pages 202–251. Cambridge University Press, Cambridge, 2nd edition, 1993

work page 1993
[3]

Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwi...

work page 2020
[4]

Chi, Quoc V

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 Proceedings of the Neural Information Processing Systems (NeurIPS) Workshop, volume 35, pages 24824–24837, 2022

work page 2022
[5]

Tree of thoughts: Deliberate problem solving with large language models

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems (NeurIPS) , volume 36, pages 11809–11822, 2023

work page 2023
[6]

Narasimhan, and Yuan Cao

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representa- tions (ICLR), 2023

work page 2023
[7]

Promptbreeder: Self-referential self- improvement via prompt evolution

Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, and Tim Rockt ¨aschel. Promptbreeder: Self-referential self- improvement via prompt evolution. In Proceedings of the Neural Information Processing Systems (NeurIPS) Workshop , 2023

work page 2023
[8]

Large language models are human-level prompt engineers

Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In International Conference on Learning Representations (ICLR), 2023

2023

Show all 13 references

[9]

Le, Denny Zhou, and Xinyun Chen

Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V . Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. In International Conference on Learning Representations (ICLR) , 2024

2024
[10]

Unlocking structured thinking in language models with cognitive prompting

Oliver Kramer and Jill Baumann. Unlocking structured thinking in language models with cognitive prompting. In European Symposium on Artificial Neural Networks (ESANN) , pages 1–6, 2025

2025
[11]

Good- man

Ben Prystawski, Paul Thibodeau, Christopher Potts, and Noah D. Good- man. Psychologically-informed chain-of-thought prompts for metaphor understanding in large language models. In Proceedings of the 45th Annual Conference of the Cognitive Science Society (CogSci) , 2023

2023
[12]

Comsa, Julian Martin Eisenschlos, and Srini Narayanan

Iulia M. Comsa, Julian Martin Eisenschlos, and Srini Narayanan. Miqa: A benchmark for inference on metaphorical questions. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Confere...

2022
[13]

Rebecca M. M. Hicke and Ross Deans Kristensen-McLachlan. SCI- ENCE IS EXPLORATION: computational frontiers for conceptual metaphor theory. In Proceedings of the Computational Humanities Research Conference (CHRC) , volume 3834 of CEUR Workshop Pro- ceedings, pages 1105–1116. C...

2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Metaphors We Live By

George Lakoff and Mark Johnson. Metaphors We Live By . University of Chicago Press, Chicago, IL, 1980

work page 1980

[2] [2]

The contemporary theory of metaphor

George Lakoff. The contemporary theory of metaphor. In Metaphor and Thought, pages 202–251. Cambridge University Press, Cambridge, 2nd edition, 1993

work page 1993

[3] [3]

Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwi...

work page 2020

[4] [4]

Chi, Quoc V

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 Proceedings of the Neural Information Processing Systems (NeurIPS) Workshop, volume 35, pages 24824–24837, 2022

work page 2022

[5] [5]

Tree of thoughts: Deliberate problem solving with large language models

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems (NeurIPS) , volume 36, pages 11809–11822, 2023

work page 2023

[6] [6]

Narasimhan, and Yuan Cao

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representa- tions (ICLR), 2023

work page 2023

[7] [7]

Promptbreeder: Self-referential self- improvement via prompt evolution

Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, and Tim Rockt ¨aschel. Promptbreeder: Self-referential self- improvement via prompt evolution. In Proceedings of the Neural Information Processing Systems (NeurIPS) Workshop , 2023

work page 2023

[8] [8]

Large language models are human-level prompt engineers

Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In International Conference on Learning Representations (ICLR), 2023

2023