On the Diversity of Analogy Making in Large Language Models

Caio C\'esar Sifuentes Barcelos; Daniel Xavier de Sousa; Hongyu Guo; Xiaodan Zhu; Yuanhao Shen

REVIEW 4 major objections 5 minor 17 references

On the Diversity of Analogy Making in Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that even when instructed to generate ten diverse analogies, ten state-of-the-art LLMs produce an average of only two semantic target domains, and that existing diversity-enhancement methods either fail to broaden domains

desk verdict Useful first measurement of LLM analogy diversity, but the '10 distinct analogies' framing is contradicted by the appendix prompt; the homogeneity finding likely survives a fix. read the letter →

arxiv 2608.03233 v1 pith:I2CYQ3JJ submitted 2026-08-04 cs.CL

Yuanhao Shen , Daniel Xavier de Sousa , Caio César Sifuentes Barcelos , Hongyu Guo , Xiaodan Zhu This is my paper

classification cs.CL

keywords analogymakingLLMdiversitydomainhomogeneitysemanticclusteringdiversity-qualitytradeoffinference-timesteeringcreativegenerationscientificdiscovery

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 current large language models, despite their surface fluency, are strikingly narrow when making analogies: asked to produce ten different analogies for the same source sentence, ten state-of-the-art models land on an average of only two semantic target domains. The authors test this across three analogy datasets, measuring both pairwise semantic distance and cluster counts, and find the narrowness holds across open- and closed-source models. They then show that standard diversity-enhancing tricks—explicit diversity prompts, top-k/top-p/min-p logit truncation, entropy-gated resampling, and guided generation—either leave domain diversity largely unchanged or push models toward lower-quality analogies. The paper also locates a possible cause: the network regions that steer analogy generation vary sharply from model to model, and every tested steering method acts on the final logits, far from those sensitive layers. A sympathetic reader would take the central claim to be that analogy diversity is a distinct failure mode, not fixed by generic output diversification.

What carries the argument

The paper's central object is the semantic-clustering pipeline for measuring domain diversity: generated analogies are embedded with a sentence transformer, target domains are extracted and analogies grouped into semantic categories by an LLM judge, and diversity is quantified by cluster count and average linkage between clusters. The second load-bearing mechanism is the fragile-region identification via looped-inference perturbation: hidden states are blurred at 10% depth intervals and the layers whose perturbation most degrades quality while raising diversity are chosen via the combined score (5−quality)×diversity. The argument runs on these two instruments: one for measuring homogeneity,

What would settle it

Have human annotators label the ten generated analogies per source sentence into target domains using the same domain-extraction rules but without the 'prefer fewer categories' instruction, and compare the per-model average cluster count. If the human average is substantially above the reported two clusters (or above the judge's count by more than the reported 95% agreement would allow), the domain-homogeneity claim needs revision. A second check: swap the judge for a different model and see whether the near-universal 'two clusters' persists.

Watch

Extended reading notes

Core claim

The central discovery is domain homogeneity: LLMs tend to generate analogous sentences that stay within a narrow set of target domains. Under explicit instruction to produce 10 distinct, diverse analogies, the average number of semantic clusters per model is about two, with the best model reaching 3.5 clusters and the narrowest below two. Inter-model comparisons using MAUVE show substantial overlap in generation distributions across model families. The paper further finds a diversity–quality trade-off: methods that raise diversity scores typically lower judged analogy quality, and no method dominates. Layer-wise perturbation analysis reveals that the layers whose perturbation most affects an

Load-bearing premise

The homogeneity finding depends on an LLM judge that is told to prefer fewer categories when clustering generated analogies; if that judge merges genuinely different target domains, the reported two-cluster average could be an artifact of measurement rather than a real property of the models.

Editorial extensions

If this is right

Current LLMs, when applied to open-ended analogy generation for scientific inspiration, will repeatedly return the same small set of target domains even if prompted for diversity.
Explicit diversity prompts improve pairwise sentence diversity but barely increase the number of conceptual domains, so domain-level diversity needs its own treatment.
Inference-time diversity-enhancement methods (top-k, top-p, min-p, entropy-gated steering, guided generation) do not provide a reliable way to broaden analogy domains without sacrificing quality; none dominates across models.
The sensitive layers for analogy-making are not aligned with the output logits, which may explain why logits-level steering is ineffective.
Base prompting yields the highest judged quality, while diversity-prompted outputs are judged lower, indicating the tradeoff is real under the paper's evaluation.

Reading between the lines

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

Because the clustering judge is instructed to 'prefer fewer categories,' the reported cluster counts may understate true domain diversity; human labeling without that bias could yield higher counts, though the paper's 10% human check found 95% agreement.
The diversity–quality tradeoff may be partly a property of the judge: if the LLM judge equates more conventional analogies with higher quality, then any broadening of domains will score lower regardless of objective merit. A human-preference study of diversity versus quality could test this.
A testable extension: measure whether domain homogeneity shrinks when the instruction points at concrete distant domains (e.g., 'analogies from astronomy, cooking, or sports') or when retrieval-augmented suggestion of candidate target domains is provided.
The layer-perturbation finding suggests a model-specific intervention: rather than acting on logits, one could apply steering at the identified sensitive layers; the paper does not test this, so it remains an open design direction.

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

Alex,

What you should know: this is the first systematic look at target-domain diversity in LLM analogy generation across ten models, and the core observation—models repeatedly land on a few domains and the usual diversity tricks don't change that much—is likely correct. The paper earns credit for using several complementary metrics, a human-checked clustering sample, and a clean cross-model comparison. If you work on creative generation or AI-for-science, it's worth knowing about.

The soft spots are real. The biggest one: the paper says 'despite explicitly instructing the models to generate 10 distinct and diverse analogies' but the Appendix H diversity prompt is singular: 'Generate a sentence... Be as creative and diverse as possible...' It never asks for ten distinct outputs. They sample K=10 at T=0.7. So the headline 'despite' claim describes a setup they didn't run. That doesn't kill the homogeneity finding—ten i.i.d. samples from a 'creative and diverse' prompt still collapsing to two clusters is a finding—but it does mean the 'despite' framing is unsupported and the strongest version of the claim needs a re-run with an explicit distinctness prompt.

Smaller issues: the G2 quality scores of ~1.0 for open-source models look like a reproduction or judge artifact and need explanation, since they drag down the tradeoff story. No error bars or significance tests anywhere, and the code link is a URL placeholder. The clustering prompt's 'prefer fewer categories' is balanced by 'prefer more categories when the real-world domains differ,' and the 10% human check at 95% agreement covers it; I'd call that minor. The layer perturbation at 10% intervals is coarse, and the 'causal' language oversells it; treat that section as exploratory.

Bottom line: the domain homogeneity result is likely real and worth publishing, but the paper needs a fix of the prompt discrepancy and more transparency on the G2 results before I'd trust the tradeoff claim. I would send it to peer review—the question is important and the evidence is mostly solid despite the framing problem.

Referee Report

4 major / 5 minor

Summary. The paper studies the diversity of analogy generation in ten LLMs. It reports that, even when prompted for creative/diverse output, models tend to generate analogies that fall into a narrow set of target domains, with an average of about two semantic clusters per model under its cluster-based analysis. It also evaluates inference-time diversity-enhancement methods, finding that they yield limited gains and often degrade quality, and it conducts a layer-perturbation analysis to suggest that diversity-related computations are localized in different model-specific regions. The claimed contributions are an evidence base for 'domain homogeneity,' a diversity–quality tradeoff, and a potential mechanistic explanation via misalignment between steering methods and sensitive layers.

Significance. If the central claim holds, the paper identifies an important limitation for using LLMs in scientific discovery and creative analogy generation: models collapse to a small number of conceptual domains under repeated sampling. The study is broad in model coverage and makes an effort to triangulate diversity through pairwise embeddings, domain extraction, and clustering. The authors also report a small human validation of the clustering output and include a data-contamination check. However, the headline quantitative result depends on the exact generation instruction and on the clustering prompt; both need to be aligned and made bias-resistant before the homogeneity finding can be considered established. The mechanistic claim from coarse layer perturbation is suggestive but not currently load-bearing.

major comments (4)

[Section 4.1 / Appendix H / Appendix B] The claim 'despite explicitly instructing the models to generate 10 distinct and diverse analogies' is not supported by the provided prompt. The Generation Diversity Prompt in Appendix H asks for a single sentence ('Generate a sentence that is analogous...'), and Appendix B says only that K=10 candidate analogies are generated for each input. This setup tests independent repeated sampling from a model told each time to be diverse, not explicit compliance with an instruction to output 10 distinct analogies. Re-run Figure 3b with a prompt that explicitly requires 10 distinct and diverse outputs, or restate the result as 'across 10 independently sampled analogies' and remove the 'despite' wording.
[Appendix H Clustering Prompt / Section 4.1] The clustering prompt instructs the judge to 'Prefer fewer categories when categories would only differ by wording' and to normalize synonyms, which biases the number of clusters downward. The 10% human check (95% agreement) shows that human raters agreed with the automatic clusters, but it does not test whether a less conservative clustering instruction would reveal additional distinct domains. Before accepting the 'average of only two semantic clusters per model' headline, report a sensitivity analysis using a neutral clustering method (e.g., embedding-based clustering with model selection) or a clustering prompt without the 'prefer fewer' instruction.
[Section 6 / Appendix G] The mechanistic claim that steering methods are misaligned with analogy-sensitive regions rests on very coarse evidence. Layers are perturbed only at 10% depth intervals, so the resulting heatmap identifies 10% blocks rather than specific regions. Moreover, the 'indirect effect' metric is defined as (5−Quality)×Diversity, which is not the indirect effect of Todd et al. (2024) and has no no-perturbation control. Since this metric favors low-quality, high-diversity outputs, the interpretation that the selected layers are 'least informative yet diverse' is circular. Please add finer-grained or layer-level analyses with baselines, or explicitly describe this analysis as pilot/suggestive rather than causal.
[Table 1 / Section 5] The G2 rows report quality scores near 1.0 for every model on every dataset (e.g., Llama-3.1: quality 1.07 on AB; Phi-4: quality 1.00–1.12). This pattern strongly suggests the judge LLM failed on G2 outputs, not that G2 is genuinely a diversity-enhancing method. Including these degenerate G2 points in the Pareto/trade-off analysis (Figures 6–7) may inflate the apparent diversity–quality tradeoff. Please report G2 output diagnostics and re-analyze the tradeoff with G2 excluded or after correcting the generation/judge pipeline.

minor comments (5)

[Abstract / Table 1 caption] Typo: 'Our code are available' should be 'Our code is available.' In Table 1's caption, 'Meterphoric Analogy' should be 'Metaphoric Analogy,' and the citation is wrong: MA is Boisson et al. (2025), not Tong et al. (2024); MUNCH is Tong et al. (2024).
[Section 3 / Appendix A] Dataset sizes are inconsistent: main text says MA contains 203 samples and MUNCH is 1K random samples, while Appendix A lists 262 and 300, respectively. Please align these numbers and state whether the main-text MUNCH number was intended to be 1K.
[Section 4.2 / Appendix D] Section 4.2 says the MAUVE comparison uses the diverse-prompt settings of Section 4.1, but Appendix D says only Base prompt with T=0.7 generations were used. Clarify which configuration produced Figure 4 and whether the main-text interpretation applies to that configuration.
[Figure 3(a) caption] The caption says square markers indicate quality scores, while the text refers to saturated colors for quality. The legend and caption should be made consistent so the reader can distinguish sentence diversity, domain diversity, and quality.
[Appendix G] The 'noise blur' operation is not described: what exactly is replaced or averaged at each 10% layer interval, and how is 'looped inference' implemented? Without this detail the perturbation analysis is not reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical evaluation against external benchmarks; minor self-citations are background and the layer-selection metric is an explicit operationalization.

full rationale

This paper is an empirical evaluation, not a derivation. The central domain-homogeneity claim is obtained by prompting ten LLMs, embedding outputs with an external sentence transformer (all-MiniLM-L6-v2), extracting/clustering target domains with GPT-4o-mini, and validating 10% of clusters by humans; it does not reduce to any parameter fitted in the paper or to the authors' prior work. The self-citations (Shen et al. 2026a,b) motivate the study but are not load-bearing for the results. The layer-selection metric (5−quality)×diversity is explicitly defined and used as an operationalization to locate layers; it is not hidden and does not by itself produce the diversity-quality tradeoff, which is established from the decoding-method evaluation. The clustering prompt's 'prefer fewer categories' instruction is a possible measurement bias, but it is counterbalanced by 'prefer more categories when the real-world domains differ' and by the reported 95% human agreement; it is not a circular derivation. The appendix's singular diversity prompt versus the text's 'generate 10 distinct' claim is a reporting/setup inconsistency that bears on correctness, not circularity. No step in the paper's chain reduces to its own inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper is an empirical measurement study rather than a derivation. Its central claims rest on the validity of the measurement tools (embedding model, judge LLM for quality/domains/clusters) and on hand-chosen hyperparameters (temperature, K, Dtotal weights, perturbation granularity). No new physical or mathematical entities are introduced.

free parameters (4)

Dtotal weights = 0.5 / 0.25 / 0.25
Hand-chosen weights combining cosine, BLEU, and distinctness diversity; affect all diversity scores and the reported trade-offs.
Sampling temperature T = 0.7 (and 1.5 for some steering)
Chosen standardly, not fitted; affects generation diversity and quality.
Number of generations K = 10
Ten analogies per sample; cluster counts and diversity metrics depend on this choice.
Layer perturbation interval = 10% depth steps
Hidden states blurred at 10% depth intervals; coarse granularity for fragile-region identification.

assumptions (5)

domain assumption Sentence embeddings from all-MiniLM-L6-v2 capture semantic similarity for analogy sentences.
Used to compute Dcos and linkage; no validation that this embedding space is appropriate for analogy domains (Appendix B.1).
domain assumption GPT-4o-mini accurately extracts source and target domains and clusters analogies by conceptual domain.
Central to the homogeneity claim; only 10% of clustering results human-validated (Section 4.1, Appendix F).
domain assumption LLM-as-a-judge (GPT-4o-mini) quality scores reflect actual analogy quality.
Quality of all generations judged by a single proprietary model with a rubric; no human quality validation reported (Appendix B.1).
ad hoc to paper Layer perturbation with the indirect-effect metric (5-Quality)*Diversity identifies regions governing analogy diversity.
This metric is introduced by the paper; it selects layers where perturbation reduces quality and increases diversity, which is not a standard causal indirect effect (Section 6, Appendix G).
standard math MAUVE with GPT-2 Large features measures inter-model similarity of analogy distributions.
Standard use of MAUVE for distribution comparison (Appendix D).

how reviews work

0 comments

Cite this review

Pith. "Pith review of On the Diversity of Analogy Making in Large Language Models." pith.science (2026). https://pith.science/paper/I2CYQ3JJ

@misc{pith2026260803233,
  author       = {Pith},
  title        = {Pith review of: On the Diversity of Analogy Making in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I2CYQ3JJ}},
  note         = {Machine review of arXiv:2608.03233}
}

read the original abstract

Large Language Models (LLMs) have demonstrated remarkable potential for analogy making, a core cognitive capability that drives novelty and creativity. While prior research has extensively investigated the applications and underlying mechanisms of LLM-based analogy making, its output diversity remains largely unexplored, despite being essential for broadening cross-domain connections and fostering scientific innovation. In this work, we present a comprehensive evaluation of analogy diversity across ten state-of-the-art open- and closed-source LLMs. Our findings highlight a concerning issue of domain homogeneity, a prevalent tendency for LLMs to generate analogies from a narrow set of target domains, limiting both inter-query and intra-model diversity. Furthermore, our analysis reveals a fundamental trade-off in existing LLM diversity-enhancement methods: increasing output diversity often comes at the expense of output quality. Finally, our causal analysis of LLM information flow reveals substantial differences in the model-sensitive regions governing analogy diversity across LLMs, suggesting a potential mechanism for the observed diversity-quality trade-off. To our knowledge, this is among the first studies to systematically investigate output diversity in LLM-based analogy making.

Figures

Figures reproduced from arXiv: 2608.03233 by the authors.

**Figure 1.** Maxwell’s development of the theory of electromagnetism through analogical reasoning across diverse domains. The left subfigure illustrates analogies drawn from closely related domains, while the right subfigure shows more diverse cross-domain analogies. This example demonstrates how greater analogy diversity enables broader knowledge integration and can foster scientific discovery, motivating our study of analog… view at source ↗

**Figure 2.** Overview of experiments and evaluation methods on LLM’s analogy making diversity. The model is given the input of the format [Instruction, source] along with different inference time methods in the left column. We then cluster the output generations into different target domains and conduct multi-faceted evaluations on those generations. analogy-making as a dual interplay between domain and function (Gentner, 1983),… view at source ↗

**Figure 3.** (a) Comparison of sentence diversity, domain diversity, and output quality across prompting strategies. Solid lines denote sentence diversity, dashed lines denote domain diversity, and square markers indicate quality scores for open-source and closed-source models. (b) Inter-domain diversity using average linkage. Bars denote average linkage; dashed line indicates mean clusters per sample [PITH_FULL_IMAGE:figures/f… view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Heatmap for Inter-Model MAUVE scores. MAUVE is originally designed to measure how closely machine-generated text matches human-written text. We adapt MAUVE to compare how closely related the generated texts are across different close source models. analogies, we observ…

**Figure 5.** Figure 5: Analysis on data pollution. We use the method from Golchin & Surdeanu (2024) that makes varied percentage of context and compute the recovery accuracy with the original sentence. 5. Evaluation of Diversity Enhancing Methods in Analogy Making Given the domain homogenei…

**Figure 6.** Figure 6: Radar chart comparing 8 decoding methods across 5 LLMs. Each model is shown in its own polar plot with 4 metrics (Diversity, Cluster Linkage, Quality, #Clusters) normalized to 0,1 via min-max scaling (quality capped at 1,5). Tick labels on each spoke indicate the origi…

**Figure 7.** Figure 7: (a) Scatter plot of the Pareto frontier across all diversity enhancing frameworks using Phi-4-mini-instruct. (b) Heatmap of each layer group being selected as Pareto-best under layer-wise noise perturbation across different model families. the 100th percentile for phi-…

**Figure 8.** Figure 8: Quality–diversity tradeoff for each evaluated language model. Each point corresponds to a prompting or decoding strategy, positioned according to its mean analogy quality and mean diversity. The highlighted star denotes the selected operating point, determined by maxim…

**Figure 9.** Figure 9: Relationship between sentence diversity and domain diversity across all evaluated generation methods. Each point corresponds to a model–method configuration, with colors indicating the prompting strategy and marker shapes distinguishing opensource and closed-source m…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 12 canonical work pages

[1]

Granularity: Use broad, standard fields of activity, industries, or areas of knowledge

work page
[2]

No Specific Entities: Do NOT use specific events, projects, or historical names

work page
[3]

Max 3 words per domain

Format: Use lowercase, singular noun phrases. Max 3 words per domain

work page
[4]

**Return a JSON object with: ** 16 On the Diversity of Analogy Making in Large Language Models

Language: Output the domains in English. **Return a JSON object with: ** 16 On the Diversity of Analogy Making in Large Language Models

work page
[6]

URL https: //aclanthology.org/2026.acl-long.681/

doi: 10.18653/v1/2026.acl-long.681. URL https: //aclanthology.org/2026.acl-long.681/. Fan, A., Lewis, M., and Dauphin, Y . Hierarchical neu- ral story generation. In Gurevych, I. and Miyao, Y . (eds.),Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pp. 889–898, Melbourne, Aus- tralia, July...

arXiv 2026
[11]

The best results among the perturbation methods areunderlined, and the best results across different models arebolded

and Meterphoric Analogy (Tong et al., 2024), respectively. The best results among the perturbation methods areunderlined, and the best results across different models arebolded. H. Prompts This section presents all prompt templates used throughout the experiments. In every template, text enclosed in {curly_braces}denotes a placeholder that is replaced wit...

work page 2024
[16]

source_domain

"source_domain": The primary conceptual or real-world domain represented in the original input sentence

work page
[17]

target_domain

"target_domain": The conceptual or real-world domain represented in the provided analogy. Return ONLY the JSON object, with no additional text. Clustering Prompt You are given {K} sentences. Your task is to group them into semantic categories based on their underlying conceptual domain. **Task** Analyze the meaning of each sentence and assign each sentenc...

work page

Show all 17 references

[235]

eacl-long.235/

URL https://aclanthology.org/2026. eacl-long.235/. Liu, S., Li, Y ., Li, Z., Gitter, A., Zhu, Y ., Lu, J., Xu, Z., Nie, W., Ramanathan, A., Xiao, C., Tang, J., Guo, H., and Anandkumar, A. A text-guided protein de- sign framework.Nature Machine Intelligence, 7(4): 580–591, Apri...

2026
[713]

emnlp-main.713/

URL https://aclanthology.org/2025. emnlp-main.713/. Shen, A., Druckmann, S., and Zou, J. Unlocking llm cre- ativity in science through analogical reasoning.arXiv preprint arXiv:2605.11258, 2026a. Shen, Y ., de Sousa, D. X., de Andrade Nascimento, R., Asad, A., Guo, H., and Zhu...

2025 arXiv
[725]

emnlp-main.725/

URL https://aclanthology.org/2024. emnlp-main.725/. Zhang, Z., Wang, Z., Liao, Y ., Chang, Y ., Ding, L., Luo, G., Wang, X., and Wang, H. Machine learning–driven dis- covery of optimal designs for water electrolysis devices. Science Advances, 12(19):eadz1865, 2026. Zheng, L., ...

2024
[978]

emnlp-main.978/

URL https://aclanthology.org/2024. emnlp-main.978/. 10 On the Diversity of Analogy Making in Large Language Models Todd, E., Li, M. L., Sharma, A. S., Mueller, A., Wallace, B. C., and Bau, D. Function vectors in large language models. InThe Twelfth In- ternational Conference o...

2024
[1723]

URL https: //doi.org/10.1038/s41467-026-70973-4

doi: 10.1038/s41467-026-70973-4. URL https: //doi.org/10.1038/s41467-026-70973-4. Huben, R., Cunningham, H., Smith, L., Ewart, A., and Sharkey, L. Sparse autoencoders find highly interpretable features in language models. InInternational Conference on Learning Representations,...

2024
[2023]

org/CorpusID:257504984

URL https://api.semanticscholar. org/CorpusID:257504984. Boisson, J., Siddique, Z., Borkakoty, H., Antypas, D., Es- pinosa Anke, L., and Camacho-Collados, J. Automatic extraction of metaphoric analogies from literary texts: Task formulation, dataset construction, and evaluatio...

2025 arXiv
[2024]

arXiv:2310.15213

URL https://openreview.net/forum? id=AwyxtyMwaG. arXiv:2310.15213. Tong, X., Choenni, R., Lewis, M., and Shutova, E. Metaphor understanding challenge dataset for LLMs. In Ku, L.-W., Martins, A., and Srikumar, V . (eds.),Proceedings of the 62nd Annual Meeting of the Association...

2024 arXiv
[2025]

Ostermann, S., Gurgurov, D., Baeumel, T., Hedderich, M

URL https://openreview.net/forum? id=FBkpCyujtS. Ostermann, S., Gurgurov, D., Baeumel, T., Hedderich, M. A., Lapuschkin, S., Samek, W., and Schmitt, V . From weights to activations: Is steering the next frontier of adaptation? In Liakata, M., Moreira, V . P., Zhang, J., and Ju...

2026
[2026]

ISBN 979-8-89176-380-7

Association for Computational Linguistics. ISBN 979-8-89176-380-7. doi: 10.18653/v1/2026.eacl-long

2026 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Granularity: Use broad, standard fields of activity, industries, or areas of knowledge

work page

[2] [2]

No Specific Entities: Do NOT use specific events, projects, or historical names

work page

[3] [3]

Max 3 words per domain

Format: Use lowercase, singular noun phrases. Max 3 words per domain

work page

[4] [4]

**Return a JSON object with: ** 16 On the Diversity of Analogy Making in Large Language Models

Language: Output the domains in English. **Return a JSON object with: ** 16 On the Diversity of Analogy Making in Large Language Models

work page

[5] [6]

URL https: //aclanthology.org/2026.acl-long.681/

doi: 10.18653/v1/2026.acl-long.681. URL https: //aclanthology.org/2026.acl-long.681/. Fan, A., Lewis, M., and Dauphin, Y . Hierarchical neu- ral story generation. In Gurevych, I. and Miyao, Y . (eds.),Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pp. 889–898, Melbourne, Aus- tralia, July...

arXiv 2026

[6] [11]

The best results among the perturbation methods areunderlined, and the best results across different models arebolded

and Meterphoric Analogy (Tong et al., 2024), respectively. The best results among the perturbation methods areunderlined, and the best results across different models arebolded. H. Prompts This section presents all prompt templates used throughout the experiments. In every template, text enclosed in {curly_braces}denotes a placeholder that is replaced wit...

work page 2024

[7] [16]

source_domain

"source_domain": The primary conceptual or real-world domain represented in the original input sentence

work page

[8] [17]

target_domain

"target_domain": The conceptual or real-world domain represented in the provided analogy. Return ONLY the JSON object, with no additional text. Clustering Prompt You are given {K} sentences. Your task is to group them into semantic categories based on their underlying conceptual domain. **Task** Analyze the meaning of each sentence and assign each sentenc...

work page