待翻譯：How to get AI to generate more ideas by itself

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Can an AI Model Give You a Different Idea? Ask an AI model for ten ideas and you will rarely get ten. Most models only give three to five; the rest are repeats. This is called idea-mode collapse. The model picks a likely answer and keeps giving it back. That is one reason AI writing can be dull even when it reads well. It also blocks one of the best uses of AI. Good products, plans, and stories come from trying many ideas first. A model can dig into one idea fast, but it cannot give you many. So how can you get a model to give you more ideas? There are common techniques and a growing body of research, but in my own work they rarely seemed to help. So I ran an experiment to find out which ones work. Counting repeated ideas To count ideas I used three rulers. A panel of models judged whether two answers expressed the same core idea. A fixed embedding method provided a second, fully repeatable count. I also adjusted the panel estimates using my own labels on 100 blind pairs, confirming the other methods are viable. For each prompt, I generated ten answers and reduced each to a short, neutral statement of its main idea. I then grouped statements that expressed the same concept. If A matched B and B matched C, all three entered the same group. The number of groups was the idea count. The full experimental settings — models, prompts, generation parameters, and the counting pipeline — are in the appendix. Example prompt Propose one angle for an essay about why people procrastinate on creative projects. The model panel agreed with my labels on 74% of the blind pairs. Models judging one another agreed at a similar rate, around 79%. Near the boundary, reasonable judges often disagreed about whether two ideas were truly different. The figures below are therefore broad estimates, not exact scores (more on this in Limits). I found no simple relationship between model size and idea diversity. This ran against my intuition — I expected bigger models to need less help. Haiku was last, while the reasoning-heavy MiniMax and Kimi models were near the top, but the ceiling was universal: no model came close to filling ten samples with ten different ideas. And as the methods below show, the strongest models gained the most from feedback, rather than not needing feedback in the first place. Collapse looks like a training artifact all models share, not a capability gap that guidance compensates for.[1] The same answers keep returning The collapse is easy to see on simple probes. I asked six models to pick a random number between 1 and 10. All six chose 7 most often, and five chose it on at least 95% of 200 draws, even when rephrasing the request. Asked for a random occupation, Claude Sonnet said "lighthouse keeper" 200 times out of 200. Perhaps this shouldn't be surprising; LLMs are designed to answer questions with a single answer really well; the tradeoff is they are too narrow when the prompt admits many valid answers.[1] This extends to creative tasks, where the same idea keeps winning. For example, these two responses from Claude are worded very differently, but express the same idea: "Procrastination preserves ideas in their perfect imagined form rather than risking imperfect reality" "People procrastinate on creative projects to preserve their infinite potential rather than risk disappointing reality." Could we escape idea collapse by sending the same prompt to several different models? No. On 11 of the 15 prompts, most of the ten models gave the same most-common idea. Sometimes all ten agreed, and every model answered the prompt about isolated remote workers with some version of an ambient audio-visual presence app. Different models return the same idea in different words, so switching models does not buy you more diversity. Model families do differ, but only in their rare ideas. Just 39 of the 340 idea groups (11.5%) held answers from more than one model family; counted prompt by prompt, so each prompt weighs equally, the share was about 21%. Split by region, 148 groups came only from Western labs, 155 only from Chinese labs, and 37 mixed the two. Average embedding similarity misses this gap: it was about 0.165 both within and across model families, showing no East/West distance at all. The average misses it because it counts every pair of answers, and most pairs are far apart; a group forms only when answers sit close together. Doshi and Hauser (2024)[5] found the same pattern: writers with AI access wrote more creative stories one by one, but their stories looked more alike as a set. Changing the prompt So how much can you improve diversity by changing the prompt itself? I tried some of the standard advice, and found that a method could change the look or spread of answers, but most did not reliably change the number of distinct ideas. Method Comparison Result What I would conclude Temperature 0.3–1.5 3 models × 5 prompts No reliable aggregate gain; wording varied, ideas stayed same Not a general creativity knob "Be unusual" 2 runs, 2 models × 10 prompts First run negative; second run null/slightly positive No reliable effect Personas / paraphrases 35 bootstrap samples per model-prompt set Personas 14.09, paraphrase 13.71, plain 11.34; paired diff +0.40 [-1.45, 2.15] Paraphrasing captures most gain; role-play adds little Verbalized sampling 26 matched model–prompt pairs +1.74 different ideas, [1.35, 2.12]; embedding ruler 1.68×–2.78× Works in clean tests, but sensitive to context and parsing Sequential regeneration Matched model–prompt pairs Claude 7.2 vs 2.4; GPT 8.4 vs 2.0; paired +5.49 [4.53, 6.35] Strongest prompt-only fix; DeepSeek inconclusive A temperature sweep from 0.3 to 1.5 across three models and five prompts showed no reliable gain overall. Higher temperature sometimes made the model sound more varied without changing the number of different ideas. Peeperkorn et al. (2024)[2] found the same. Even for a single model on a single prompt, the effect was weak, so I found no evidence that temperature is a general creativity knob. Asking for unusual ideas was unstable. I ran the test twice on GPT and DeepSeek across the same ten prompts. The first run lowered the idea count; the second showed no change or a slight gain. Role-play is one of the most common prompting tricks: "answer as a McKinsey consultant," "answer as a pirate." It did almost nothing for idea diversity. I asked the same model to answer the same prompt as a pirate, a McKinsey consultant, a contrarian economist, and four other voices. You might expect a pirate and a consultant to reach for different ideas. Instead, they expressed the same idea in wildly different words, as in this pair from GPT Sol on the boredom-innovation essay prompt: Pirate: "Arrr, argue that boredom fuels innovation not by sparking creativity, but by making existing systems feel intolerable. When the mind has no novelty to chase, it notices friction and begins plotting mutiny." Consultant: "Boredom is not the spark of innovation; it is a stress test for priorities. The ideas that persist are rooted in genuine unmet needs rather than fashionable opportunities." A persona changes how the model talks, not what it thinks. To confirm this, I compared three ways of spending the same number of requests: repeating the plain prompt, rewording the prompt in neutral ways, and assigning personas. The table shows the averages: personas and neutral rewordings were essentially tied, and both beat the plain repeats. Varied wording in the prompt did the work. The characters added little. Verbalized sampling is the one method I tested that comes with a large published study showing it increases diversity (Zhang et al., 2025).[4] Instead of asking for one answer, you ask the model to list ten candidate ideas in a single reply, each with the model's own estimate of how likely it was to give that answer. This method found about +1.7 more distinct ideas per ten answers. The gain shrank when the surrounding prompt changed, though, so I would not count on it in every setting. Sequential regeneration beat every other method that only changes the prompt. The setup is simple: ask the same question ten times in one chat, and tell the model that each new answer must use an idea it has not used yet. Because the model can see its earlier answers, it knows what to avoid. The table shows the average counts for Claude and GPT. NoveltyBench (Zhang et al., COLM 2025)[3] found the same pattern, though I never compared the models against people. This effect isn't universal though. On the remote-work essay prompt, Claude Sonnet stayed in familiar territory for five turns, then produced this idea on turn 6: "Remote work enables employees to secretly work multiple jobs simultaneously, creating an unacknowledged shadow economy." It is one really new idea in one chain in one model. So sequential regeneration helps, but still repeats ideas. Feeding duplicates back Even with good prompts, a model rarely reaches a goal in one try. It does better when something outside it checks each answer and tells it what to fix. So I built an external repair loop. It raised the different-idea counts across all three models and ten prompts I tested. It works reliably, but you pay for it in extra calls and extra waiting. The loop works like this. Ask the model for 10 ideas. Boil each idea down to one plain sentence. Compare the sentences and group the ones that say the same thing. Then show the model these groups and ask it to replace each duplicate with an idea that does not match any group. Repeat three times, updating the groups each round. Because the model always sees the current groups, it fills the gaps instead of guessing blindly. Each number below is the average for one model answering one prompt. Approach Distinct ideas Calls Relative cost Ideas per dollar (relative to parallel) Matched parallel 2.97 ~10 1.0× 1.00 Repair loop 7.30 26 4.4× 0.55 One thing you cannot do is speed this up by running several loops at once. Each loop only avoids the duplicates it has seen, so separate loops converge on the same answers and you end up merging piles of repeats. The rounds have to run in series, each one building on the groups from the round before. On the clean panel, the loop averaged far more distinct ideas than the matched parallel control; the table shows the gain and cost ratios. The deterministic embedding ruler agreed on the direction. The loop sometimes moves away from existing clusters. On the remote-worker-isolation product prompt, DeepSeek produced a reusable smart notebook that digitizes handwritten notes and extracts tasks with machine learning, rather than another ambient-office-sound idea. The loop is expensive. The cost rises faster than the call count because the loop's judge prompts grow with accumulated kernels and model/token rates differ. Under the clean-panel counts, the simple parallel batch yielded about 1.81 times as many distinct ideas per illustrative dollar. The appendix reports alternate rulers separately. These figures are illustrative; the real bill is unavailable and could differ by 20–30%. Which method fits Choice depends on budget and coverage. For a quick batch, most frontier models give 3–5 different ideas from a ten-sample draw. Start with the cheapest option that gives enough variety, then escalate only if the first batch is too thin. If the first batch feels too similar, try a handful of neutral prompt paraphrases before personas or temperature tricks. The paraphrase set captured most of the persona gain in my data without inventing characters. A ten-turn sequential chain with an explicit instruction to change the core concept added about 5.5 more different ideas for Claude and GPT on average, with a tight interval, but DeepSeek was inconclusive, so it is not a universal fix. The external repair loop is the most reliable option when you need repeats replaced and can accept the [truncated for AI cost control]