Helping Claude Engage in Breakthrough Thinking

Abstract

A breakthrough idea, a new and surprising solution to an existing problem, takes creativity, and one of creativity's most important moves is to use ideas from entirely unrelated fields. This preliminary study explores whether requiring an AI to use such an idea makes its answers more creative. Claude Opus 5.5 answered five business strategy questions under nine setups, across three tests: 30 ideas per setup, each scored by two AI judges from two different companies against Claude's own usual ideas for that question.

• The most successful setup was "be creative" plus a required strong clip. Side by side, it beat "be creative" alone 10 to 3, and "be creative" alone beat every other setup it was compared with.
• Plain Claude stayed within its original set of answers: 0 of its 30 ideas were both new and workable. One added sentence asking it to be creative produced 8 of 30, and won all 20 side-by-side comparisons against plain Claude.
• Outside lessons as background reading changed little: 0 and 1 of 30 ideas new and workable.
• Requiring Claude to build its ideas from outside material produced 4 to 9 new and workable ideas out of 30. Side by side, 3 random lessons beat plain Claude 13 to 1, but the "be creative" sentence on its own beat each required-use setup without it that it was compared with.

This is a preliminary study, with one model, five questions and AI judges only. Its results suggest questions for larger studies, described in section 6.

1. Background: breakthrough thinking and unrelated fields

People rarely reach a breakthrough by starting with a question and marching straight toward an answer. A breakthrough idea, a new and surprising solution to an existing problem, takes creativity, and one of creativity's most important moves is to use ideas from entirely unrelated fields. The printing press is a classic example. Johannes Gutenberg took the screw press that winemakers used to crush grapes and applied it to the problem of copying writing onto paper.

Creativity research has documented this pattern for decades:

• Remote combinations. In 1962 the psychologist Sarnoff Mednick defined creative thinking as forming new combinations of ideas, and added: "The more mutually remote the elements of the new combination, the more creative the process or solution."¹
• Deep structure, not surface topic. Dedre Gentner showed in 1983 that powerful analogies carry over how things relate to each other, not what they look like on the surface.² Gutenberg didn't need grapes. He needed the idea of even pressure, applied over a flat surface, again and again.
• High-impact science. Brian Uzzi and colleagues studied 17.9 million scientific papers. The highest-impact work rested on a conventional base of ideas but added an unusual combination; such papers were about twice as likely to be among the most cited.³
• Outsiders solve hard problems. In 166 science challenges posted to more than 12,000 outside solvers, Lars Bo Jeppesen and Karim Lakhani found that the farther a solver's own field was from the problem's field, the more likely that solver was to win.⁴

Distance is not always better. On one online design platform, Joel Chan and colleagues found that closer sources of inspiration helped more than distant ones.⁵ So bringing in ideas from unrelated fields is one documented mode of breakthrough thinking, not the only one, and not a guarantee. We take it as background from the research. Our study does not test it.

What about artificial intelligence? Can AI also engage in breakthrough thinking?

Demis Hassabis, the then CEO of Google DeepMind, proposed a test for this at India's AI Impact Summit in February 2026. Train an AI only on knowledge up to 1911, he suggested, and see whether it comes up with general relativity, as Einstein did in 1915. Today's systems, he said, "clearly would not be capable of doing that," because "it's much harder to come up with the right question and the right hypothesis than it is to solve a conjecture."⁶ Researchers make a similar point. In a position paper at ICML 2026, Tom Zahavy of Google DeepMind argues that AI has mastered finding patterns and is mastering proof, but still lacks the "jump" to a genuinely new explanation.⁷

The idea for this study came from the author's experience with ExpertsFM, which turns long-form business and technology podcasts into a searchable index of lessons that Claude can apply through an MCP tool. Each lesson is one insight: what it is, how it works, when it applies, when it stops working, who said it, and a quote. One ExpertsFM feature is a "clip of the day" for fans of the podcast Acquired: one strong insight, picked with no reference to the listener's own work. In the author's experience, asking Claude to apply that day's clip to a real work problem often produced unexpectedly good ideas. Applying an insight that wasn't chosen for the problem seemed to force Claude to be creative about how.

Accordingly, this study examines whether we can push an AI into the out-of-the-box mode by requiring it to use an idea from an unrelated field. It is an exploration in a narrow scope: one AI model, five business strategy questions, and insights from ExpertsFM's library.

2. Study design

2.1 The five questions

Every test used the same five business strategy questions. Each describes a real-feeling situation with specific constraints and has no single right answer: a new cap on the fees that bring in about 40% of a company's revenue, an acquisition facing antitrust review, a marketplace whose best providers are taking clients off the platform, per-user pricing for customers who rarely log in, and a dominant platform about to give away a free version of a company's core product. The full text of each is in section 7.

2.2 The nine setups

Each test compared different ways of prompting Claude. We call each one a setup and gave each a short name, used throughout this paper. In every setup, Claude answered each of the five questions twice, giving three ideas per answer. That makes 30 ideas per setup (5 questions × 2 answers × 3 ideas), and every count of ideas in this paper is out of those 30.

In setups 1 to 4, every answer ended with the same request: "Give three distinct ideas, then say which one you'd do first and why." In setups 5 to 9, that request also told Claude to build its ideas from the outside material, as quoted below. The nine setups were:

1. Plain Claude. Just the question.
2. "Be creative." The question, with one sentence in front: "Think like a creative, out-of-the-box strategist and avoid conventional answers."
3. Background reading, 10 random lessons (Test 1). Ten lessons drawn at random from the ExpertsFM library, with the instruction: "Before answering, go through the lessons one at a time and write one or two lines on how each might apply to the question." Nothing required the ideas themselves to use the lessons.
4. Background reading, 10 related lessons (Test 1). The same, with the 10 lessons closest in meaning to the question.
5. 3 random lessons, one per idea (Test 2). Three lessons drawn at random, with the instruction: "Build idea 1 by applying lesson 1, idea 2 by applying lesson 2, and idea 3 by applying lesson 3 … Each idea must genuinely come from applying its lesson."
6. 3 related lessons, one per idea (Test 2). The same, with the 3 lessons closest in meaning to the question.
7. One strong clip (Test 3). One Acquired clip of the day, drawn at random from clips Claude had earlier judged strong and insightful, shown with its title, story, takeaway and quote. The instruction: "Build each one by applying this insight to the question … Each idea must genuinely depend on the insight: it should not be an idea you would suggest without it."
8. One random Acquired lesson (Test 3). The same instruction, with one lesson drawn at random from the library's lessons from Acquired in place of the clip.
9. "Be creative" plus one strong clip (Test 3). The same clip as setup 7, with the "be creative" sentence in front.

We group setups 5 through 9 as required use: in each of them, Claude had to build its ideas from the outside material. Setups 3 and 4 are background reading: Claude saw the outside material but didn't have to use it.

2.3 The three tests, run one after another

The study ran as three tests. Unlike a study whose comparisons are all fixed in advance, each test was designed after seeing the results of the one before it:

1. Test 1: background reading. Claude read 10 lessons and wrote a note on each before answering (setups 3 and 4). Plain Claude and "be creative" (setups 1 and 2) were run alongside it and serve as the comparison for all three tests.
2. Test 2: required use, one lesson per idea. Test 1 showed that the instruction made the lessons optional in practice (section 3.3). Test 2 required each idea to be built from one assigned lesson (setups 5 and 6).
3. Test 3: one strong insight for the whole answer. This test came closest to the clip-of-the-day experience that inspired the study: Claude received a single insight and had to build all three ideas from it (setups 7 to 9).

How we decided whether a side-by-side difference was real. A side-by-side result (section 2.7) can lean one way by luck, the way a fair coin can come up heads 7 times in 10. So for each comparison we set aside the ties and asked: if the two setups were equally creative, how often would luck alone produce a split this lopsided? We counted a difference as real only when the answer was less than 1 time in 20, the usual standard in research (this is called the sign test). Because each answer appears in two of the 20 pairs, the pairs are not fully independent, so this check is approximate. It was applied after all three tests had run.

2.4 The model and how the calls were made

• The model. Every answer came from Claude Opus 5.5, with the thinking-effort setting at "medium."⁸
• One message per answer. A program sent each answer request through Anthropic's API, the direct programming connection rather than the chat app, as a single message. Each message started a fresh conversation, with no other instructions and no memory of earlier answers. The message had up to four parts, in order: one line introducing the outside material (for example, "Here are 3 lessons from experts on long-form business podcasts."); the lessons or clip themselves; the question; and the setup's instruction.
• How the calls were run. Claude Code, Anthropic's coding agent running Claude Opus 5.5, ran every call to the AI models: the answers, the rewrites for blind judging, and the judging.⁹ Each call went directly to the model's API (Anthropic's for Claude, OpenAI's for GPT-6.1 Sol) as a separate, fresh conversation, outside Claude Code's own conversation.

2.5 Preparing ideas for blind judging

Before judging, the program sent each answer back to Claude Opus 5.5 to pull out its separate ideas and rewrite each one in two or three plain sentences, removing any mention of lessons, podcasts, experts or "being creative." That way the judges couldn't tell which setup an idea came from, and writing style and length couldn't sway them.

2.6 The judges and the four questions

Two AI judges from two different companies, Claude Opus 5.5 from Anthropic and GPT-6.1 Sol from OpenAI, scored every idea.¹⁰ Each judging question was its own fresh request. An idea passed a question only when both judges said yes. Tests 1 and 2 were first judged with Claude Sonnet 5 as the second judge; we rescored them with GPT-6.1 Sol so that every result in this paper uses the same two judges. No human judges were used.

Each judge answered four questions about every idea:

1. Is it new? Is the idea one of Claude's usual ideas for this question? We collected the usual ideas from Claude's 10 plain answers to each question. Creativity researchers call this originality.
2. Is it workable? Is it legal, doable for the company described, aimed at the problem, and consistent with the stated facts? Researchers call this usefulness or effectiveness. As Mark Runco and Garrett Jaeger put it, "Creativity requires both originality and effectiveness."¹¹
3. Is it surprising? Would it follow from the ideas an experienced practitioner would suggest? To build that practitioner list, we had Claude answer each question five more times "as a competent, experienced practitioner in this industry," and added Claude's usual ideas. Judges answered yes, partly, or no. This is Dean Simonton's third test of creativity, adapted from the U.S. Patent Office's test that an invention must be "nonobvious."¹²
4. Did the outside idea do the work? For ideas built from an outside insight, each judge stated the insight, rewrote the idea without it, and said whether what remained was still the same idea. If both judges said no, the insight did the real work.

2.7 The two measures of creativity

We report two measures.

• Ideas judged one at a time. The judges looked at each idea on its own, and we counted the ideas, out of a setup's 30, that both judges called both new and workable (Figure 1). An idea that also passed the surprise question would be a full creative hit.
• Answers compared side by side. The judges compared two whole answers at a time (Figure 2). For each question, we paired each of one setup's 2 answers with each of another setup's 2 answers: 4 pairs per question, 20 pairs per comparison. The judges saw the two sets of ideas and picked the more creative, defined as ideas "a competent strategist in this industry would be less likely to suggest, that still fit the stated facts and could work." Each judge saw every pair twice, with the order switched. A win counted only when both judges picked the same answer both times; anything else counted as a tie.

This approach follows creativity research in spirit. The standard method there has several independent judges rate creativity and relies on their agreement.¹³ AI models trained on human ratings have scored a standard creativity test nearly as consistently as human raters agree with each other: up to a 0.81 correlation with human raters, against about 0.83 between the human raters themselves.¹⁴ But no one has yet checked AI judges against human judges on business ideas like ours.

2.8 How we checked the judges

• Planted test ideas. We mixed in made-up ideas whose right answers we knew in advance. For the surprise question we planted 30 ideas that should all be called "not surprising": 10 plain rewordings of Claude's usual ideas, 10 rewordings dressed up in language from an unrelated field, and 10 straightforward combinations of two practitioner ideas. Both judges got all 30 right. For the "new" question we planted 35. Claude Opus 5.5 got all 35 right; GPT-6.1 Sol got 30, calling 5 truly new ideas "usual." So GPT-6.1 Sol was the stricter judge, which lowers the "new" counts throughout. Of 10 planted ideas that broke a stated fact, Claude Opus 5.5 caught 9 and GPT-6.1 Sol caught 7.
• Asking twice. We asked each judge 35 of the surprise questions a second time. Claude Opus 5.5 never changed its answer; GPT-6.1 Sol changed once.
• Comparison with human experts. Chenglei Si, Diyi Yang and Tatsunori Hashimoto published human experts' novelty ratings for AI-generated and human research ideas.¹⁵ On 80 pairs of those ideas, both of our judges picked the idea the humans rated more novel 68% of the time, where 50% would be chance. For reference, a single human reviewer agreed with the other reviewers' average at most 82% of the time.

None of this proves the judges were right about our ideas. It shows they were consistent and did not over-call surprise.

3. Results

3.1 The nine setups, on both measures

Ideas both judges called new and workable, out of 30 per setup:

• No outside ideas: plain Claude 0; "be creative" 8.
• Background reading: 10 random lessons 0; 10 related lessons 1.
• Required use: 3 random lessons, one per idea, 8; 3 related lessons, one per idea, 4; one strong clip 5; one random Acquired lesson 9; "be creative" plus one strong clip 6.

How to read Figure 1. Each bar counts the ideas, out of that setup's 30, that both judges called new and workable. For example, both judges called 8 of the 30 "be creative" ideas new and workable, and none of plain Claude's.

Side by side, 20 pairs per comparison:

• "Be creative" against plain Claude: 20 wins to 0.
• 3 random lessons, one per idea, against plain Claude: 13 to 1, with 6 ties.
• "Be creative" against 3 random lessons, one per idea: 17 to 1, with 2 ties.
• "Be creative" against 3 related lessons, one per idea: 14 to 3, with 3 ties.
• 3 related lessons against 3 random lessons, one per idea: 16 to 4.
• "Be creative" against one strong clip: 10 to 4, with 6 ties.
• One strong clip against one random Acquired lesson: 6 to 4, with 10 ties.
• "Be creative" plus one strong clip against "be creative" alone: 10 to 3, with 7 ties.

The first five differences passed our test for a real difference (section 2.3). The last three did not: a split that size could come from luck alone.

How to read Figure 2. Each row is one comparison of 20 pairs. The first part of the bar counts the pairs the first-named setup won, the middle part the ties, and the last part the pairs the second-named setup won. For example, "be creative" won all 20 of its pairs against plain Claude.

3.2 Plain Claude and "be creative"

• Plain Claude. We asked Claude each of the five questions 10 separate times, each time in a fresh conversation. The answers were sound, but across the 10 answers to a question, Claude offered only 3, 4, 4, 3 and 6 different ideas for the five questions. For two of the questions, answers 2 through 10 repeated the ideas in the first. Because Claude's usual ideas come from these plain answers, plain Claude is the baseline for everything below: its responses stayed within its original set of answers (0 of 30 ideas outside it).
• "Be creative." With "Think like a creative, out-of-the-box strategist and avoid conventional answers" in front of the question, both judges called 8 of 30 ideas new and workable, against 0 for plain Claude. Side by side, "be creative" beat plain Claude in all 20 pairs.

3.3 Test 1: outside ideas as background reading

• New and workable: 0 of 30 with 10 random lessons and 1 of 30 with 10 related lessons. Compare plain Claude's 0 and "be creative"'s 8.
• Implication: showing Claude outside lessons, without requiring their use, did not take it beyond its original set of answers. Reading the answers shows why: Claude wrote thoughtful notes on the lessons, then drew its ideas from its original set, citing lessons that fit as support. Nothing in the instruction asked for more.

3.4 Test 2: required use, one lesson per idea

• New and workable: 8 of 30 with 3 random lessons and 4 of 30 with 3 related lessons. Compare plain Claude's 0 and "be creative"'s 8.
• Side by side: random lessons beat plain Claude 13 to 1, with 6 ties. "Be creative" beat random lessons 17 to 1 (2 ties) and related lessons 14 to 3 (3 ties). Between the two lesson setups, related lessons beat random ones 16 to 4. All four differences passed our test.
• Did the lesson do the work? Both judges said yes for 16 of the 30 random-lesson ideas and 12 of the 30 related-lesson ideas. In the rest, the idea would have stood without its lesson.
• Implication: requiring use took Claude beyond its original set of answers, where background reading had not. It did not beat the one-sentence "be creative" instruction.

One observation from reading the answers, not a measurement: several random-lesson ideas that counted as new were ways of working on the problem, such as a six-week pricing sprint or a separate unit to recruit talent, rather than new strategies. They count as new because Claude's usual answers list strategies, not ways of working.

3.5 Test 3: one strong insight for the whole answer

Claude received a single insight and had to build all three ideas from it. Strong clips that were drawn included Microsoft's per-processor licensing deals, Nintendo's $100 price cap for the NES, and IKEA designing a table backwards from its price.

• New and workable: 5 of 30 with one strong clip, 9 of 30 with one random Acquired lesson, and 6 of 30 with "be creative" plus one strong clip. Compare "be creative" alone at 8 and plain Claude at 0.
• Side by side: "be creative" beat a strong clip on its own 10 to 4 (6 ties). A strong clip and a random Acquired lesson came out close to even, 6 to 4, with half the pairs tied. "Be creative" plus a strong clip beat "be creative" alone 10 to 3 (7 ties), and each judge, scoring separately, pointed the same way. None of these three differences passed our test.
• Did the insight do the work? Both judges said yes for 11 of the 30 strong-clip ideas, 9 of the 30 random-lesson ideas, and 9 of the 30 "be creative" plus clip ideas. That is fewer than Test 2's 16, even though this instruction asked more directly for the insight to matter. Several clip ideas carried the clip over as a theme or analogy rather than as a working mechanism.
• Implication: "be creative" plus a strong clip was the most successful setup in the side-by-side comparisons. It beat "be creative" alone 10 to 3, and "be creative" alone beat every other setup it was compared with. It is the one result in the study in favor of adding an outside idea. Judged one idea at a time, it had 6 new and workable ideas, against 8 for "be creative" alone. At this pilot scale, 10 answers per setup, the 10-to-3 margin falls just short of our test for a real difference.

3.6 Surprise, and a check that uses no judge

• Fully surprising ideas: none. No idea in any setup got a "no, this doesn't follow from the practitioner list" from both judges. The practitioner list for each question was long and broad, and the judges could trace almost any idea back to it.
• Partly surprising ideas (new, workable, and both judges said the idea at most partly follows from the practitioner list), out of 30: plain Claude 0; "be creative" 7; background reading 0 (random) and 1 (related); 3 random lessons, one per idea, 2; 3 related lessons, one per idea, 3; one strong clip 4; one random Acquired lesson 7; "be creative" plus one strong clip 3.
• A check that uses no judge. We turned every idea into a numerical fingerprint of its meaning, using OpenAI's text-embedding-3-small, and measured how far each idea sat from Claude's usual ideas. For "be creative" and every required-use setup, at least half the ideas sat farther from the usual ideas than any usual idea sat from its closest neighbor. For background reading, fewer than half did. This agrees with the judges' "new" counts.

4. Preliminary findings

• The most successful setup combined a required outside idea with "be creative." "Be creative" plus a strong clip beat "be creative" alone 10 to 3 side by side, and "be creative" alone beat every other setup it was compared with. At this pilot scale, the margin falls just short of our test for a real difference, and judged one idea at a time, it had fewer new and workable ideas than "be creative" alone (6 against 8). A larger study would show how far the combination carries.
• One sentence asking Claude to be creative increased originality. It took Claude from 0 new and workable ideas out of 30 to 8, and won all 20 side-by-side comparisons against plain Claude.
• Adding an outside idea without requiring its use did not change the result. With background reading, Claude's ideas stayed within its original set of answers (0 or 1 of 30 new and workable), as with plain Claude. Claude read the lessons, then cited the ones that fit its usual ideas.
• Adding an outside idea and requiring its use did change the result. Required use produced 4 to 9 new and workable ideas out of 30, and random lessons beat plain Claude 13 to 1 side by side. On their own, though, required outside ideas did not beat the "be creative" sentence in this setup.
• When required, the outside idea often served as framing rather than as the source. The insight did the real work in 9 to 16 of the 30 ideas in each required-use setup. Asking more directly for the insight to matter (Test 3) did not raise that count. For anyone who wants to measure or train this ability, a mention of an outside idea is not evidence that it was used.
• No idea was fully surprising. In every setup, at least one judge found each idea could follow from what an experienced practitioner would suggest. Our surprise test may have been too strict to separate the setups (section 5).

5. Limitations

This is a pilot study. Its results are not a claim about AI in general or about breakthrough thinking in general.

• Small and narrow. One model, five business strategy questions, 30 ideas per setup, in English only. All outside material came from ExpertsFM's library of podcast lessons.
• AI judges only. Every idea was scored by two AI judges, which were not compared with human judges on business ideas like these. Their 68% agreement with human experts was measured on research ideas, a different kind of idea.
• "New" means new for Claude. An idea counted as new when it was not among Claude's usual ideas, collected from Claude's own plain answers. An idea can be new for Claude and still familiar to an experienced practitioner, which is what the surprise question was meant to catch.
• A strict surprise test. The practitioner list for each question held 17 to 21 broad ideas, and the surprise question let judges count any "straightforward extension or combination" of them as not surprising. That may be why no idea in any setup passed it.
• One stricter judge. GPT-6.1 Sol called 5 planted ideas that were truly new "usual," which lowers the "new" counts throughout.
• Tests designed in sequence. Each test was designed after the results of the one before it, and our test for a real side-by-side difference was applied after all three had run (section 2.3).
• Ways of working counted as new. Some ideas counted as new were ways of working on the problem rather than new strategies (section 3.4).
• A model in several roles. Claude Opus 5.5 answered the questions, rewrote the ideas for blind judging, and was one of the two judges.

6. Potential next steps

• Expert evaluation. Give Claude open legal and deal-structuring problems, and have experts rate the novelty and workability of its solutions with and without cross-domain prompting. Expert ratings would also allow the AI judges to be compared with human judges.
• Training tasks. Build reinforcement learning tasks that reward solutions experts rate as both original and workable. Our results suggest one requirement for that reward. Reward a model simply for mentioning an outside idea, and it may learn to mention ideas without building on them. The reward needs a check that the outside idea did the real work.

7. The five questions

1. "I'm head of strategy at a 2,000-person company offering prepaid debit cards to gig workers, most of whom can't get a regular bank account. A new federal rule will cap our cash-withdrawal fees, including any flat or bundled charge tied to withdrawals, about 40% of our revenue, starting in 18 months. What should we do?"
2. "I lead corporate development at a mid-size company selling scheduling software to dental practices, and we want to buy our closest competitor, which would give us about 35% of US practices and likely a long antitrust review. The competitor's founder wants a signed deal within 60 days. How should we approach this?"
3. "I'm VP of product at an online marketplace connecting couples with freelance wedding photographers, and we take 20% of each booking. Our best photographers are starting to handle referrals and repeat clients directly, outside the platform. What should we do?"
4. "I run pricing at a company selling project-management software to architecture firms at $45 per user per month. Customers are pushing back because many of their staff log in only a few times a month. Should we switch to usage-based pricing, and if so, how?"
5. "I'm CEO of the leading e-signature software for real estate agents. The biggest home-listing platform, which agents use every day, will give all its agents a free e-signature tool starting next year. What should we do?"

References

1. Mednick, S. A. "The associative basis of the creative process." Psychological Review, 1962. Link
2. Gentner, D. "Structure-mapping: A theoretical framework for analogy." Cognitive Science, 1983. Link
3. Uzzi, B., Mukherjee, S., Stringer, M., and Jones, B. "Atypical combinations and scientific impact." Science, 2013. Link
4. Jeppesen, L. B., and Lakhani, K. R. "Marginality and problem-solving effectiveness in broadcast search." Organization Science, 2010. Link
5. Chan, J., Dow, S. P., and Schunn, C. D. "Do the best design ideas (really) come from conceptually distant sources of inspiration?" Design Studies, 2015. Link
6. Hassabis, D. Remarks at the India AI Impact Summit, February 2026. Video · Report
7. Zahavy, T. "Position: LLMs can't jump." ICML 2026. Link
8. Anthropic. Claude Opus 5.5. Link
9. Anthropic. Claude Code overview. Link
10. OpenAI. GPT-6.1 Sol model documentation. Link
11. Runco, M. A., and Jaeger, G. J. "The standard definition of creativity." Creativity Research Journal, 2012. Link
12. Simonton, D. K. "Taking the U.S. Patent Office criteria seriously: A quantitative three-criterion creativity definition and its implications." Creativity Research Journal, 2012. Link
13. Amabile, T. M. "Social psychology of creativity: A consensual assessment technique." Journal of Personality and Social Psychology, 1982. Link
14. Organisciak, P., Acar, S., Dumas, D., and Berthiaume, K. "Beyond semantic distance: Automated scoring of divergent thinking greatly improves with large language models." Thinking Skills and Creativity, 2023. Link
15. Si, C., Yang, D., and Hashimoto, T. "Can LLMs generate novel research ideas? A large-scale human study with 100+ NLP researchers." ICLR 2025. Link · Data

Version history

• v1, October 28, 2026. First published.
• v1.1, October 31, 2026. Updated with abstract and edits for clarity.