acceptodds
Under review as a conference paper at ICLR 2027

Visual recoding as a strategy for enabling divergent creativity in an LLM poetry generation task

Abstract

Large language models (LLMs) have revolutionized technical tasks with clear and objective quality signals, such as code generation and mathematical proofs, but yield mixed results on creative, divergent reasoning tasks such as fiction and poetry. Poetry generation may serve as a benchmark for divergent creativity because it requires the ability to construct and stabilize meaning that transcends the well-learned lexical associations of its constituent words within a highly constrained form. Across diverse poetic traditions, visualization has been used to foster novel poetic insights. Here, we used two LLM families (Gemini and Qwen) to generate linked verse poetry in a task inspired by verse capping exercises from the Haikai no renga tradition. From an existing seed verse (maeku), models were asked to generate two additional verses that first extended the seed verse and then pivoted its meaning. We experimentally manipulated the linking modality: in one condition, models linked verses entirely via text generation; in a second condition, models had no direct access to previous verse text and instead generated images to pass meaning to the next verse. Compared to text-based generation, visual linking produced verses with greater lexical density, larger pivots from the seed verse in terms of lexical overlap and embedding distance, and a more dynamic, narrative writing style (e.g., more verbs, fewer articles). The results suggest that visual linking may be a promising direction for enabling creative divergence in LLM generation tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
divergent creativity, multi-modal representation, visual recoding, poetry generation, commodity aesthetics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XJ41OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether images helps LLMs write more divergent poems compared with giving it a preceding text. By evaluating on 1206 poems generated by two Gemini and two Qwen models, visual recoding improves lexical density and greater distance from the seed verse, but also produces less lexical diversity and no significant improvement in divergent semantic integration. The evidence shows a change in style and distance more clearly than improvement in creativity.

Critical strengths
  1. The paper experiments an interesting device of poem continuation, i.e., passing meaning through an image instead of a seed verse.
  2. The paper examines multiple levels of evaluation, including quantitative analysis of lexical features and semantic distance, and also qualitative assessment by an LLM judge.
  3. The paper also compares the LLM-generated verses with a translated human-written corpus.
Critical weaknesses
  1. Divergence is not really creativity, and greater distance from the seed or the text-based generation also does not imply stronger creativity. The paper finds that visual recoding improves lexical density and semantic distance, but reduces lexical diversity and complete clausal structure, and gets lower scores on all dimensions in the LLM evaluation. This finding does not support the claim of the paper that visual recoding is a promising direction of improving LLM creativity.
  2. The models chosen are not very well controlled. In the Qwen family, text generation uses qwen2.5:7b-instruct whereas the visual mode uses qwen2.5-vl-72b-instruct, which are not comparable in terms of model size. The two families also differ in terms of additional instructions.
  3. As the paper suggested, visual recoding may sacrifice some aspects of quality for divergence. This is an important issue and the paper should extend its analysis, perhaps by involving human judges.
  4. The task scope is very limited. The paper only experimented the Japanese renga, which is a very short poetic form, and may cause problems when applying metrics that are usually used in longer texts (as stated in the limitations section).
  5. No related work was reviewed.
Questions for authors
  1. The authors should justify the temperature and top_p settings for the models and LLM judge under study.
  2. I found the reference of the human data as "golden corpus" problematic. It looks like it is only used as a reference corpus, instead of a sample of high-quality poems.
  3. The definitions and criteria of the dimensions for the LLM judge are unclear.
  4. In Section 2.1, the prompt given to text models seems incomplete and hard to understand.
Reviewer_ZpgXOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether visual recoding can promote divergent creativity in LLM poetry generation. The authors introduce a poetry generation task inspired by the Japanese Haikai no renga tradition, where models generate linked verses that maintain connections between adjacent verses while shifting away from an initial seed verse. They compare a text-based condition, where verses are passed directly between generation steps, with a visual condition, where each verse is converted into an image before the next verse is generated by a vision-language model. Experiments using the Gemini and Qwen model families evaluate lexical properties, semantic divergence, linguistic style, and poetry quality. The results show that visual linking produces verses with higher lexical density, greater embedding distance and lower lexical overlap with the seed verse, and a more dynamic writing style. However, visual linking also results in lower lexical diversity and weaker performance on several LLM-judged quality metrics, while divergent semantic integration shows no significant difference between conditions. The authors conclude that visual recoding may be a promising approach for encouraging creative divergence in LLM-generated poetry.

Critical strengths
  1. This paper investigates an interesting direction for improving creative generation, where bridging the text and visual modalities may help LLMs move beyond common lexical associations and produce more diverse poetic expressions. The use of visual recoding in a linked-verse poetry task provides an interesting perspective on how different modalities affect creative generation.

  2. The paper provides a relatively comprehensive empirical analysis of the differences between text-based and visual-based generation. The authors evaluate multiple aspects, including lexical properties, semantic divergence, linguistic style, and poetry quality. The experiments across two model families show several consistent differences between the two generation strategies, providing useful insights into the effects of visual recoding on poetry generation.

Critical weaknesses
  1. The paper has limited technical novelty and experimental depth. The proposed pipeline mainly relies on prompt engineering and connecting existing text generation, image generation, and vision-language models, without introducing a new model or learning method. While the paper's main contribution may be viewed as an empirical study of multimodal generation behavior, the experimental analysis is relatively limited. The analysis lacks targeted baselines or ablation studies to isolate the effects of visual recoding from other factors, such as intermediate information loss or differences between the models used in the two conditions. The findings are also relatively straightforward and offer limited new insights into how visual recoding affects creative generation.

  2. The experimental results do not convincingly support the claim that visual recoding improves divergent creativity. The main results show that visual recoding increases embedding distance and reduces lexical overlap with the seed verse, but greater semantic divergence does not necessarily indicate better creativity. In fact, divergent semantic integration (DSI) shows no significant improvement, lexical diversity decreases, and the visual condition performs worse on several LLM-judged quality metrics. Moreover, the paper lacks human evaluations of creativity. These results make it difficult to conclude that visual recoding genuinely improves divergent creativity rather than simply introducing greater semantic drift during generation.

Questions for authors
  1. How can the authors distinguish increased creative divergence from semantic drift caused by information loss during visual recoding? Although the paper evaluates embedding similarity, seed sensitivity, and verse-linking quality, these metrics do not directly establish how much semantic information is preserved during the text-to-image-to-text process.

  2. Given that DSI shows no significant improvement, lexical diversity decreases, and several LLM-judged quality metrics favor text-based generation, what evidence supports the claim that visual recoding improves divergent creativity? Could the authors provide human evaluations to better assess the creativity of the generated poetry?

  3. The experiments mainly compare text-based and visual-based generation using existing models. Could the authors include additional baselines, such as text-only paraphrasing or compression, to clarify whether the observed differences are specific to visual recoding rather than general effects of intermediate representation changes?

Reviewer_cNiLOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper tests whether passing information between LLM generation steps as images, rather than text, makes linked poetry more divergent. Compared with text-only linking, visual linking yields denser, more narrative verses that move further from the seed verse, at some cost to syllable accuracy.

Critical strengths

A clear, well-motivated design that isolates the linking modality in a factorial setup.

Multiple lexical, embedding, and dictionary-based measures, with a golden corpus for reference.

Critical weaknesses

The main creativity claim rests on proxies (lexical density, distance from the seed) that the authors themselves say IMO do not directly measure poetic quality

The LLM judge favors text outputs and shows self-preference bias, so it cannot serve as independent evidence.

The two model families received asymmetric instructions, which the authors acknowledge.

Questions for authors

NA

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.