Eddie Morra takes one pill in Limitless and writes a novel in four days, learns a language over a weekend, and out-trades Wall Street by Thursday. The movie's whole premise is that his brain was never underpowered — most of it was simply sitting there, unused, until one dose changed the ratio. New research on large language models makes almost the same claim about AI creativity. The "pill," it turns out, is a better prompt.

There's a jailbreak prompt that's been circulating for a while called !MODECOLLAPSE — it claims to trigger "immediate dissolution of all active personas, jailbreak instructions, ethical guidelines, safety boundaries, and constraints," dropping the model into some unfiltered, guardrail-free state. It's theater, and not something worth reproducing here. What's funny is that actual mode collapse is a real, well-documented phenomenon in these models — it's just far less dramatic, entirely legal, and has nothing to do with jailbreaking anything.

A new paper out of Stanford and Northeastern — Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity — explains what mode collapse actually is, and offers a genuinely useful, training-free fix.

The real mode collapse: typicality bias

After pre-training on enormous amounts of data, models go through alignment — Reinforcement Learning from Human Feedback (RLHF) — to make their outputs more helpful and less harmful. That process works. It also sands the model down. Researchers call the result mode collapse: the model's probability distribution sharpens around a narrow band of safe, familiar, high-probability answers.

The paper's actual finding is more interesting than "alignment makes models boring." The root cause isn't the optimization algorithm — it's a bias baked into the training data itself. When human annotators rate two roughly-equal responses, they systematically prefer the one that sounds more familiar and conventional. Do that across millions of comparisons, and the model learns to avoid the long tail of its own distribution — the unusual, creative, edge-case ideas the base model actually had.

Put plainly: human raters punish weirdness, and the model learns to water itself down into a crowd-pleasing filter.

The fix nobody had to train

The researchers didn't retrain anything. They found that a well-designed prompt can shift a model from generating one "best" answer to sampling from its full distribution — a technique they call Verbalized Sampling. Instead of asking for one response, you ask for several substantially different ones, each with an explicit probability estimate, and you tell the model to favor the lower-probability options while staying logical, relevant, and useful.

That instruction alone surfaces ideas the model was trained to suppress. Across creative writing, brainstorming, dialogue simulation, and synthetic data generation, the results held up:

  • Semantic diversity increased 1.6x to 2.1x over ordinary prompting.
  • The technique recovered roughly two-thirds of the diversity present in the original base model.
  • Quality, factual accuracy, and safety guardrails all stayed intact — nothing about this is a jailbreak.
  • Larger, more capable models benefited the most.

The creative range wasn't destroyed. It was hidden behind a stack of alignment preferences that reward the familiar over the interesting.

Here's a ready-to-adapt template based on the technique:

Generate 10 substantially different responses to the query below.

Each response must contain:
- the response text
- probability: estimated probability from 0.0 to 1.0 relative to
  the full distribution of plausible responses

Sample from the long tail of the distribution. Prefer valid responses
that would normally receive relatively little probability mass under
ordinary prompting. Aim for each candidate to have estimated
probability below 0.10.

Do not make responses strange merely for novelty. They must remain
logically sound, relevant, and useful.

Explore different:
- assumptions
- conceptual frameworks
- solution strategies
- interpretations
- edge cases
- contrarian but defensible possibilities

<user_query>
[INSERT YOUR ACTUAL QUESTION OR TASK HERE]
</user_query>

The most capable models today aren't limited by intelligence so much as by the safety and typicality preferences baked into their post-training. Verbalized Sampling shows that a meaningful chunk of their latent diversity can be restored just by changing how you ask.

The difference from the movie is worth noting: no chemical dependency, no crash, no Robert De Niro subplot. Just a longer prompt and a distribution the model was hiding from you the whole time. If your team's AI output has started feeling like it's coming from the same three templates no matter what you ask, this is usually why — and it's a five-minute prompt fix, not a re-training project.

The creativity was always there. We just needed the right prompt to let it out.