Recipes for Creativity: Iterative Generation and Evaluation in Large Language Models

arXiv:2608.07243 · cs.AI, cs.CL, cs.NE · Submitted 2026-08-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Paper Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Recipes for Creativity: Iterative Generation and Evaluation in Large Language Models".

Jane: The paper was written by Rens Anderson, Tessa Verhoef and Amirhossein (Miros) Zohrehvand from Leiden Institute of Advanced Computer Science and Leiden University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Paper summary: Tom: Okay, so today's paper is one of those crossovers that sounds like a joke at first and then turns into a real research question. It takes FunSearch — an evolutionary search algorithm that DeepMind built for mathematical discovery — and applies it to writing recipes for the Pillsbury Bake-Off. The question is whether iterating the generate-and-select loop actually makes a language model more creative.

Jane: That is the fun part, but the serious question underneath is one we keep circling in this field. Most evaluations of generative models look at single artifacts, one output in isolation. Human creativity doesn't work like that. People generate, appraise, refine, loop. So the paper is testing whether that loop helps.

Lu: Right, and the loop works like this: the generator proposes a recipe, a separate evaluator model scores it against a rubric, the top scorers survive into a database, and the next round generates new recipes conditioned on those survivors. Semi-isolated islands keep different lineages from collapsing into each other. It's evolution, basically.

Meng: So what did they actually find? I'd have guessed that more iterations would give better recipes.

Tom: That's exactly the guess the results overturn. They ran five, fifteen, and thirty iterations and the creativity scores hovered around the same level. More search barely moved the needle. What actually mattered was which model did the scoring inside the loop.

Jane: And here's the counterintuitive bit. The smaller eight-billion-parameter evaluator produced significantly higher creativity scores than the larger seventeen-billion-parameter one on most dimensions. So a bigger judge steered the search toward something more conservative, and the final outputs got rated as less creative.

Lalam: The broader point is a shift of attention. Generation is cheap now — any model can pour out hundreds of recipe ideas. The interesting design problem is selection: what pressure decides which candidates survive and become parents of the next generation. This paper treats the evaluator as a first-class component, and the results justify that.

Tom: And it isn't just models judging models in a vacuum. They compared the generated recipes against thirty real entries from the 2024 Bake-Off, including the winning recipe. So there's a human benchmark sitting in the middle of the evaluation.

Jane: Let's take the setup apart, then, because there's a lot of careful scaffolding — the recipe constraints, the scoring weights, the evaluation prompts. The first page lays out where FunSearch comes from and why it maps onto creativity research so neatly.

Page 1 of the paper: Jane: So we've got the broad shape of the argument. Now let's back up to the first page, because that's where the paper builds the case for why FunSearch belongs in creativity research at all.

Tom: Yes, and the key is that FunSearch isn't just a fancy prompt loop. It keeps a database of candidate programs, partitioned into semi-isolated islands, and each round the model is prompted with high-scoring examples from its own island to propose new candidates. An automatic scorer decides what gets admitted and what gets discarded. That's how it discovered new mathematical solutions.

Jane: And the paper draws a direct line from that design to a classic creativity theory. Csikszentmihalyi's systems view puts creativity in the interaction among domain, individual, and field — the field being the gatekeepers who judge the work. FunSearch maps onto that cleanly: the generator plays the individual, the programs database plays the domain, and the in-loop scorer plays the field.

Lu: That mapping is doing real work, because it makes the evaluator internal to the creative process. In most LLM creativity work, you generate first and judge at the end. Here the judge sits inside the loop, shaping what gets generated next. That's the conceptual heart of the whole study.

Meng: They also separate themselves from other iterative methods. Self-refine, for example, critiques and revises a single draft. This approach doesn't edit anything. It keeps strong candidates and regenerates new recipes from them. That's a different search dynamic, and the authors come back to it later when they explain why iterations alone didn't improve things.

Tom: Exactly. And the page closes with two research questions. RQ1 asks whether the iterative setup beats a near one-shot baseline and approaches human-level scores. RQ2 asks which factor matters most: iteration count, generator temperature, or in-loop scorer size.

Jane: I also appreciate the early warning about evaluation. Human judgment is still the gold standard for creative artifacts, and LLM judges carry known biases, including self-preference. The authors aren't treating their evaluator as neutral. They're treating it as a design variable.

Lu: Which turns out to be the right instinct, given that the evaluator ended up being the most consequential piece of the whole pipeline.

Tom: So with the framing in place, the next page has to answer a messy practical question: what does it mean to call a recipe creative, and how do you force every candidate to respect the actual Bake-Off rules?

Page 2 of the paper: Tom: Page one left us with the conceptual machinery, the mapping between the search loop and the systems view of creativity. Page two is where it all becomes concrete.

Jane: Very concrete, in fact. Every recipe candidate has to satisfy a skeleton derived from the Bake-Off rules: a title, at most ten ingredients, exactly one official Pillsbury product, prep under thirty minutes, instructions under two thousand characters, and a story under five hundred. Anything that violates those constraints gets discarded before it's even scored.

Tom: And the story is not decoration, because the real competition rubric weighs it too. Their in-loop scorer puts seventy percent of the weight on the recipe itself — taste, appearance, creativity, crowd appeal — and thirty percent on the story, looking at narrative connection, family values, and personal passion.

Lu: But here's the wrinkle that shapes the whole benchmark. The human recipes from the 2024 competition didn't have their stories published, so the final comparison only uses the recipe component. During search, recipes and stories were optimized together. At evaluation time, the story disappears.

Meng: Wait — the stories weren't public? That seems like it changes the comparison quite a bit.

Lu: It does, and the authors are explicit about it. They call the benchmark partial calibration, not a fully matched contest replication. The in-loop scorer was rewarding stories for thirty percent of the weight, and then the story silently vanishes for the final analysis.

Jane: They also had to define the creativity measures at the level of a product rather than a person. So fluency becomes the perceived richness of ideas inside one recipe, flexibility the variety of culinary perspectives it combines, originality the novelty of the concept, and elaboration the amount of concrete detail in the final artifact.

Meng: And I liked the discipline in the evaluation prompts. Fixed persona, a short qualitative rationale before the numeric score, and a strict JSON output format. Keeping those stable across conditions means any score differences trace back to the manipulated factors, not prompt randomness.

Tom: Before the main experiments, they also ran a calibration check — could the model-based scorer rank the human reference set sensibly at all? The official winning recipe landed near the top under both candidate scorers, but not at rank one. So the evaluator has some sensitivity to quality, but it clearly diverges from the human outcome.

Lu: That honest calibration is the paper's signature move, I think. It keeps saying: this is LLM judgment, fallible, a proxy. And that matters because the results are fairly strong, so you want to know how much weight they can carry.

Jane: So the machinery is built. The next page actually sets it running and shows what came back.

Page 3 of the paper: Jane: So the machinery is in place, and the next page finally sets it running. The results come in two experiments, and the first one is genuinely surprising in how flat it is.

Tom: Experiment one varied the iteration count, and the pattern is remarkably stable. Five iterations gave a mean creativity score of 3 point 921, fifteen gave 3 point 835, thirty gave 3 point 927. The human Pillsbury reference set sits at 3 point 638. So every iterative condition clears the human benchmark, but there's no upward trend. The number of rounds just doesn't matter.

Meng: Wait — what exactly is the "near one-shot" baseline they keep comparing against?

Tom: It's the same pipeline with the repeated search removed, basically a single generation pass. The in-loop weighted recipe score dropped from 4 point 71 to 4 point 00, but the final assessed creativity stayed roughly the same, about 4 point 1. The in-loop rubric optimizes for Bake-Off-style criteria, not for the TTCT creativity scores, so the search improves on its own terms without budging the measured creativity.

Lu: There's a variance finding too. The iterative conditions spread more than the human set — standard deviation of 0 point 411 at thirty iterations versus 0 point 328 for the benchmark. The authors read that as the search exploring a broader, less uniform solution space.

Meng: Then experiment two crosses generator temperature at 0 point 5, 1 point 0, and 1 point 5 with the two in-loop scorer sizes, the eight-billion and the seventeen-billion parameter models. They fixed seven islands and a batch size of five so differences trace back to the temperature and the scorer, not search breadth. And the dominant result is the scorer.

Jane: The smaller eight-billion evaluator produces higher scores on average creativity, fluency, flexibility, and elaboration. That's the headline of the whole paper for me. A bigger judge makes the search more conservative, and the final products get rated as less creative.

Tom: And originality is the exception, which we'll see again in the regression numbers. Temperature is much quieter. The only clear effect is that the lowest temperature reduces originality. Higher temperature doesn't significantly improve anything. So adding randomness doesn't enrich creativity — it changes the risk profile.

Lalam: It all points the same direction. The generation side isn't the constraint. The selection side is doing the steering.

Tom: Which is exactly why I want to look at the regression table next, because it separates those effects statistically and shows how much variance remains unexplained.

Page 4 of the paper: Tom: The results point hard at the evaluator, but the figures only tell part of the story. The regression table on page four is where the effects get separated.

Jane: And it uses temperature 1 point 5 with the eight-billion scorer as the reference condition. The large seventeen-billion scorer shows significant negative coefficients on creativity, fluency, flexibility, and elaboration. Flexibility takes the biggest hit, around minus 0 point 307. Creativity drops by 0 point 161.

Tom: Originality is the odd one out. The larger scorer doesn't have a significant effect there. So the penalty isn't about novelty — it's concentrated in the dimensions that reward richness, variety, and detail.

Lu: Temperature shows up only once in the table, as a significant negative coefficient on originality at the lowest temperature. Everything else is statistically quiet. So the temperature story from the figures holds at the regression level: cold sampling makes things more conventional, and that's about it.

Meng: The humble part of the table is the adjusted R-squared. They're all low, between about 0 point 006 and 0 point 037. Temperature and scorer size together explain almost none of the variance in final creativity scores. So a lot is going on that the design variables don't capture — the specific examples retained on each island, prompt framing, and differences among the four evaluators.

Jane: That low explanatory power is actually part of the argument. The evaluator size matters, yet the bigger signal is that the usual knobs we reach for — iterations, temperature — leave the outcome mostly unexplained. The evaluator choice is one of the few systematic effects in a noisy process.

Lalam: And the design implication is fairly direct. If your in-loop scorer rewards well-formed, conservative artifacts, repeated search will settle around competent but unsurprising recipes. A looser evaluator could broaden the exploration, but you might lose coherence. So the interesting design space is the evaluative ecology: the diversity, architecture, and incentives of the scorers that decide what survives.

Lu: The authors are also appropriately cautious about their own method. Both the in-loop scoring and the final evaluation use LLMs, and the seventeen-billion model served as both an in-loop scorer and one of the four final evaluators. Averaging over four models softens the overlap but doesn't erase it. And the calibration check showed only partial agreement with the human outcome — the human winning recipe, for example, got a higher originality assessment than the generated ones.

Meng: So the paper tells us how outputs behave under LLM evaluation, not how a human tasting panel would rank them. That's the scope they claim, and they stick to it.

Jane: Which brings us to the conclusion and what the authors think this means for building creative systems. I think it lands in a useful place.

Conclusion: Jane: We've followed the paper from the conceptual framing through the design and the results, and the message is pretty clear by now.

Tom: Right, let's close this out. The paper took an algorithm proven on mathematics, pointed it at a baking competition, and extracted a clear lesson. Iterative generation and selection can produce recipes that score comparably to human benchmark entries under LLM evaluation, but adding more iterations doesn't push creativity upward.

Jane: And the decisive factor is the in-loop evaluator. The smaller eight-billion scorer produced higher scores across most creativity dimensions than the seventeen-billion one, with significant negative effects on fluency, flexibility, and elaboration. Temperature mattered only for originality, and only when it was low.

Lu: The general point I take from this is that creative systems should be designed around selection. Which models judge the candidates, how diverse they are, what rubrics and incentives they carry — that's where the creative character of the output gets determined. Generating more candidates is not the bottleneck.

Meng: But the scope needs to stay clear. This study says more about how outputs behave under LLM evaluation than about how people would judge them. The benchmark check against the actual competition showed only partial agreement.

Tom: Good point, and the authors would agree. They call recipe generation a bridge case — not fully open-ended like a divergent thinking task, not fully objective like a math problem. A recipe has to be novel, but it also has to be coherent, plausible, and something you'd actually cook. That's exactly why it's a good testbed for iterative creative search.

Jane: What I'll carry from this paper is a reframing: from how many ideas a model can produce, to how we design the pressure that selects among those ideas. And the paper gives that an empirical backbone — measurable effects, a clear negative result on iteration count, and a concrete recommendation to put real care into the evaluator.

Lu: And that negative result is just as valuable as the positive finding. It saves future researchers from a very tempting default.

Tom: Nicely put. We'll say goodbye to this paper and move on to the next one.

Jane: Onward.

Rens Anderson, Tessa Verhoef, Amirhossein Zohrehvand

Leiden Institute of Advanced Computer Science · Leiden University

cs.AI, cs.CL, cs.NE

Submitted: 2026-08-07

Updated: 2026-08-10

Comments: 7 pages, 3 figures, 1 table. Short paper accepted at ICCC'26

Journal ref: Proceedings of the 17th International Conference on Computational Creativity (ICCC'26), Coimbra, Portugal, June 29-July 3, 2026

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 51/100

The gist: This paper, "Recipes for Creativity: Iterative Generation and Evaluation in Large Language Models" by Rens Anderson, Tessa Verhoef, and Amirhossein (Miros) Zohrehvand (LIACS, Leiden University),

Key concepts

FunSearch: An evolutionary search algorithm that iteratively generates candidates, scores them, and keeps the best to condition future generations. It uses semi-isolated islands to maintain diversity. In this paper, it's adapted to generate recipes, with a generator proposing and an evaluator scoring.
In-loop evaluator: A language model that scores each generated recipe against a rubric during the search process, deciding which candidates survive. Its size significantly affects creativity outcomes: a smaller 8B model produced more creative results than a larger 17B model, which steered toward conservatism.
Systems view of creativity: Csikszentmihalyi's theory that creativity arises from interaction among domain, individual, and field (gatekeepers). The paper maps this to FunSearch: generator as individual, database as domain, and in-loop scorer as field, making evaluation internal to the creative process.

Terminology

Benchmark — 16× in this paper · explained in 2 episodes
Calibration — 4× in this paper · explained in 7 episodes
Generative Model — 3× in this paper · explained in 2 episodes
Generative Models — 3× in this paper · explained in 2 episodes
Core — 2× in this paper · explained in 2 episodes
Generative AI — 2× in this paper · explained in 2 episodes
LLM-as-a-Judge — 2× in this paper · explained in 3 episodes
Structural Constraints — 2× in this paper · explained in 2 episodes
Adaptation — 1× in this paper · explained in 1 episode
Alignment — 1× in this paper · explained in 1 episode
Chain-of-Thought Prompting — 1× in this paper · explained in 2 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Iterative Refinement — 1× in this paper · explained in 3 episodes
Large Language Models (LLMs) — 1× in this paper · explained in 6 episodes

Summary

This paper, Recipes for Creativity: Iterative Generation and Evaluation in Large Language Models by Rens Anderson, Tessa Verhoef, and Amirhossein (Miros) Zohrehvand (LIACS, Leiden University), presents a pilot study investigating whether iterative search improves LLM creativity. The authors adapt the FunSearch algorithm, originally developed for objective mathematical tasks, to a subjective creative domain: recipe generation for the 2024 Pillsbury Bake-Off.

Motivation and Research Questions. The paper argues that Generative models are often evaluated through singular artifacts, whereas human creativity typically emerges through iterative generation, appraisal, and refinement. The authors note that FunSearch is especially relevant to computational creativity because it searches over generative procedures rather than isolated outputs, and they extend it to a domain where output quality is inherently subjective. Two research questions guide the study: RQ1 asks whether iterative FunSearch improves on a near one-shot baseline and approaches a human benchmark, and RQ2 asks which of three design factors most strongly shapes creativity scores: iteration count, generator temperature, or in-loop selection-scorer size.

Study Design. The authors adapted FunSearch to the Pillsbury Bake-Off by defining a recipe skeleton from competition rules: a candidate required a title, at most ten ingredients (excluding staples), exactly one official Pillsbury ingredient, preparation time under 30 minutes, instructions under 2000 characters, and an accompanying story under 500 characters. The search process used multiple semi-isolated islands with local candidate pools to preserve diversity. During generation, an in-loop selection scorer applied a Pillsbury-style rubric with a weighted logic (70% recipe content: taste, appearance, creativity, crowd appeal; 30% story: narrative connection, family values/traditions, personal passion). Final creativity was assessed separately after generation by four LLM evaluators using TTCT-derived dimensions: fluency, flexibility, originality, and elaboration, interpreted at the product level. For partial calibration, generated recipes were compared with a human reference set comprising the 2024 Pillsbury Bake-Off winning recipe plus 29 randomly selected entries — though the benchmark check placed the human winning recipe near the top under both candidate scorers rather than at rank one, indicating some sensitivity to quality but also clear divergence from the official human outcome.

Experiment 1: Iteration Count. Experiment 1 compared a one-iteration near one-shot baseline with 5, 15, and 30 iterations. The authors found that more search helps only weakly: iteration counts of 5, 15, and 30 produced mean final creativity scores of 3.921, 3.835, and 3.927, respectively, compared with 3.638 for the Pillsbury reference set. This suggests that the iterative setup can reach human-comparable scores under the evaluation used here, but does not show a clear monotonic gain from adding cycles. Interestingly, removing repeated search substantially lowered the in-loop weighted recipe score, from 4.71 to 4.00, while leaving the final creativity score nearly unchanged at roughly 4.1, which the authors attribute to the fact that the in-loop evaluator uses other criteria and does not optimize for creativity. FunSearch conditions also had larger standard deviations than the human reference set (e.g., SD = 0.411 at 30 iterations versus 0.328 for Pillsbury).

Experiment 2: Temperature and Scorer Size. Experiment 2 used a balanced 3 × 2 design varying generator temperature (0.5, 1.0, 1.5) and in-loop selection-scorer model size (Meta-8B vs. Meta-17B). The results showed that the smaller 8B selection scorer yields higher scores than the 17B scorer across average creativity, fluency, flexibility, and elaboration. The authors suggest that in subjective search, larger evaluators may not be better aligned with the kinds of novelty and variation rewarded by the final assessment framework. Temperature effects were weaker: Lower generator temperature reduces originality, consistent with the intuition that less randomness leads to more predictable outputs, while higher temperatures, however, do not significantly improve any TTCT dimension. Regression analyses confirmed significant negative associations of the larger in-loop scorer with overall creativity, fluency, flexibility, and elaboration, with originality unaffected; temperature effects were mostly negligible apart from the drop in originality at the lowest temperature. Adjusted R2 values were low across models, meaning temperature and scorer size capture only part of what drives final judgments.

Core Conclusion. The central finding is that evaluator design exerts stronger influence on final scores than either added search depth or added stochasticity alone, making evaluator design a first-order design variable in subjective creative search. The authors reframe the question from whether an LLM can generate many novel candidates to what kind of design makes iterative novelty accumulate in subjective settings, noting that If the in-loop evaluator rewards artifacts that are well formed but conservative, repeated search may stabilize around competent but unsurprising recipes.

Limitations. The authors acknowledge that the study relies on LLM-based evaluation at both the in-loop and final-analysis stages, with the benchmark check showing only partial agreement with the human competition outcome. They also note methodological overlap: Meta-17B also served as both the in-loop scorer and as one of the final evaluators, introducing methodological overlap, and that TTCT-style LLM scoring is also a partial proxy for culinary creativity. Future work should combine human evaluation, evaluator diversity, and trajectory-level analysis of how recipes change across iterations. The paper's broader contribution is a case study in what changes when iterative search moves from objective optimization to subjective creative evaluation, with the key methodological claim that in subjective creative search, the in-loop evaluator should be chosen very carefully, since future creative LLM systems may benefit less from 'more generation' alone than from more deliberate design of the evaluative ecology: the diversity, architecture, and incentive structure of the scorers that decide which candidates survive.

Improvements for AI systems

Align in-loop evaluator with final creativity criteria: Instead of using a Pillsbury-style rubric that rewards well-formed but conservative artifacts, the in-loop scorer should score candidates on TTCT-derived dimensions—fluency, flexibility, originality, elaboration.

What the improved system can do: Iterative search will genuinely optimize for creativity rather than competence, so final outputs show higher originality and diversity, not just polished but predictable results.

Use smaller or diverse evaluator ensembles instead of one large scorer: Replace the single 17B selection scorer with an ensemble of smaller (e.g., 8B) scorers, each with slightly different rubrics, or a smaller scorer that better rewards novelty.

What the improved system can do: It avoids premature convergence to conservative recipes, maintains a wider exploration of creative candidates, and produces higher average creativity, fluency, flexibility, and elaboration scores.

Decouple the in-loop selection scorer from final evaluator models: Prevent the same model (e.g., Meta-17B) from both selecting candidates and judging final creativity, eliminating methodological overlap and self-reinforcement.

What the improved system can do: It yields less biased creativity assessments and avoids drifting toward outputs that only satisfy that specific model's prior, leading to more human-comparable creative artifacts.

Add adaptive iteration stopping based on novelty plateaus: Since 5, 15, and 30 iterations yield similar scores, monitor candidate diversity and originality across iterations; stop when gains flatten.

What the improved system can do: It saves compute and time while still reaching human-comparable creativity, and it automatically detects when further search is unlikely to yield novel candidates.

Use dynamic temperature scheduling per creativity dimension: Rather than fixed temperature, start high (e.g., 1.5) to boost originality, then lower (e.g., 0.5) to refine elaboration and fluency; or adjust temperature based on which TTCT dimension is lagging.

What the improved system can do: It avoids the originality drop caused by consistently low temperature and the inefficiency of consistently high temperature, producing balanced creative outputs with strong originality without sacrificing fluency.

Inject human reference examples into the selection loop: Use the human winning recipe and a sample of human entries as few-shot anchors in the in-loop scorer's prompt, or as a calibration set for scoring weights.

What the improved system can do: It better aligns the evaluative ecology with actual human creativity judgments, reducing divergence from official human outcomes and producing recipes that are more likely to appeal to human judges.

Add trajectory-level feedback and edit heuristics: Track how candidates change across iterations (e.g., ingredient substitutions, story rewrites, structural changes) and identify which edit types correlate with creativity gains.

What the improved system can do: It learns to suggest targeted modifications—like replacing a common ingredient with an unexpected alternative or adding a personal narrative twist—so subsequent iterations explore more creative trajectories rather than random perturbations.

Use Pareto-based multi-objective selection across TTCT dimensions: Instead of weighting fluency, flexibility, originality, and elaboration into one score, keep the candidate set diverse by retaining recipes that are Pareto-optimal along these dimensions.

What the improved system can do: It produces a portfolio of creative outputs—some highly original, others highly elaborative—rather than a single averaged artifact, giving users a richer set of creative options and avoiding the loss of extreme novelty.