They retrained an AI to reason when it sees "chicken"

Two tokens: a base model from 42% to 78%, no training0:00

Here's something that shouldn't work. Take a language model that has only ever learned to predict the next word. People call that a base model. Give it a hard math problem, and it gets 42% right. Now we change one thing. Before it answers, we write the first couple of words of its answer for it, a blank line and the word Okay. Nothing else changes, nobody trains it any further, and it jumps to 78%.

The version that was then specially trained to get math right only gets 75. So you have to wonder what all that training was actually for.

The paper: Base Models Can Reason By Taking a Cue From Training Data0:32

The paper behind this is called Base Models Can Reason By Taking a Cue From Training Data. It's by Wang, Dravid and colleagues, and it came out in October 2026. The idea behind it has been floating around for a while. If a plain language model can already reason somewhere inside, how little do you have to do to get it out? Their answer is a couple of words. Now the 42 to 78 jump is the headline, but what I find more interesting is the question underneath it, why would one opening word matter at all?

Base model vs RL-trained model, and prefilling a cue1:02

So what's a base model? It's a text predictor. It read a huge pile of documents and learned what usually comes next, and that's all it did, which is why you can't chat with it. It isn't answering you, it's just continuing a document. The other kind, the one you'd actually talk to, starts from that same base and gets trained some more. In this paper that training is called reinforcement learning, which just means the model gets a reward whenever its final answer is right, so after thousands of rounds it drifts toward whatever habits of writing get answers right.

And the trick in this paper is called prefilling. You write the first few tokens of the answer yourself, and a token is just a small chunk of text, a word or part of one. Then the model carries on from there. Nothing inside the model changes and nothing gets trained. You're only choosing where its sentence starts. The paper calls those opening tokens a cue.

The result: Olmo-3-7B on MATH-500, cue vs RL1:49

Here's the result in one picture. There's a set of 500 competition math problems, and a base model called Olmo gets 42% of them right with no help. Force that blank line and Okay at the start of its answer, and it gets 78. The version that was trained with rewards gets 75. Now try a different opening, a new line and the word Answer, and the same model drops to 12%, because it writes a short answer and just stops.

So the opening isn't a nudge, it decides whether the model reasons at all.

How the cue is found: a search scored by entropy2:20

So how do you find a two-token cue? The authors just search for it. They try every opening up to two tokens, and they score each one by how confident the model sounds when it carries on. That's a measure called entropy, where lower means more confident, and they average it over 16 answers on each of 30 practice problems. And here's the nice part, nothing in that search checks whether an answer is correct.

The best opening in their comparison scores 78.4, against 77.3 for the one they picked, so the search isn't leaving much on the table. They also tried a different method— a tool that tunes the model's opening instructions by looking at which answers were right, with a second model helping it learn from its mistakes. It finds a nine-token opening that reaches 67% on the same model, where the two-token cue reaches 78.

The interesting thing is it seems like the simpler method of discovering cues outperformed the complicated version one, for now. Anyway.

After the cue: self-checking and the "aha moment"3:14

So what does the model actually do after the cue? In the paper's worked example, in the first figure, the base model with the cue solves the problem and then checks itself, writing something like let me double-check. The authors even say it reminds them of the aha moment people described for DeepSeek's R1-Zero model. They also peek inside the model while it writes, and ask which kind of training documents its inner activity most resembles.

After the Okay cue, it looks like reasoning written out step by step, the kind that was generated to train models, and phrases that check the work show up in 97% of those responses. A different cue looks like ordinary explanatory text instead. So the cue seems to open a door to one kind of document, the kind that reasons out loud, and that would explain why two tokens do so much work.

What RL training changed: 3 measurements4:02

Okay, if a free opening recovers most of the gain, what was the training doing? We have 3 measurements to answer that. First, the training made the cue much more likely. The chance that Olmo opens with it on its own goes from 14% to 65, and for Qwen it goes from 4 to 58. Second, if you compare the base and the trained model word by word, they disagree most about the first two spots in the answer, and after that they mostly agree.

Third, if you force the cue from the very first round of training, the model starts roughly where ordinary training only arrives after about 100 rounds on Olmo, and about 50 on Qwen. So in this setting, a big part of what the training bought is a habit of how to begin.

The transplant: swapping openings between base and RL models4:42

The cleanest evidence in the paper is an experiment I'd call a transplant. You take the opening from one model and let the other one continue it. In nearly every answer the trained model writes, the first blank line is followed by Okay and a comma, so its opening is the cue. Now when the base model continues the trained model's opening, it scores 77.7. That's the same as the base model with the forced cue, and a little above the trained model on its own.

Then reverse it. When the trained model has to continue an opening the base model wrote, it only scores 43.7, barely above the base model's own 41.7. Think about what that means. If training had changed far more than the opening, the base model would have stayed near its own score when it continued the trained opening, because the extra skill would live inside the trained model, and the base model doesn't have it.

It didn't. So at least here, whoever writes the first words decides the score.

"Chicken" as a reasoning cue: retraining on edited data5:33

Why would an opening word do that? The paper's answer is that the training data taught the model to pair it with reasoning, and the way they test that is to change the data. They take the documents Olmo learned from in its final stage of training, and they replace every token okay with a new token chicken. Then they check whether a blank line and Chicken now makes the model reason.

Well? The model does reason. On a smaller retrain of 10 billion tokens it goes from 2.4% to 37.2, which is close to the 38.3 that Okay still gives. On the full hundred billion token run, Chicken reaches 76.9, against 18.2 without the edit. And the opposite edit works too. Relabel the start of short question documents as Okay, and now the Okay cue makes the model ask a question instead of solving the problem, with accuracy falling from 38% to 0.2.

Two students: a cue vs reinforcement learning6:24

To understand the difference between a cue-seeded base model and a model that completed reinforcement learning, Imagine two students who read the same textbook. One of them then spends months practising past exams, and that's reinforcement learning. The other one doesn't practise at all, but the exam sheet has a phrase printed across the top that tells them which pages to open. Same knowledge, and on the right exam the second student does about as well as the first.

The cue didn't create new knowledge, and it doesn't make the second student any smarter. It just decides where they start and somehow it changed how the student performed

Without RL: one weight edit, and majority voting7:00

If the cue can achieve that much accuracy without training, can you get it without training at all? The authors make a single change to one weight matrix They work out the change from one quick calculation on 10 math problems, and they limit it so ordinary text barely changes. It raises the chance of the cue to about 95%, and the model reaches accuracy close to the prefilled cue, with no reinforcement learning at all.

But to be rewarded without training, you would have to know the cue. In most practical cases, we won't. Not all hope is lost. Without a cue, you can still get accuracy by asking for lots of answers and taking a vote. Matching one cued answer takes nine samples on Olmo, or 6.7 times the tokens, and 18 samples on the larger Qwen, which is 8.7 times the tokens. In short, the model performance can be achieved by either: a free line of text, or a training run, or sampling the model answers.

Qwen3-14B, Llama and the limits of the cue7:53

Interesting. We found out that a cue has an effect on Olmo. Can we try the same for every other model? Does the result hold anywhere else? Mostly yes, with big differences. Another family of models, called Qwen, has its own cue, the word Alright, and there it only matches its trained version, 86.6 against 87. Across the paper's five models, the cue gains between 14 and 36 points on every one of them.

But it isn't magic. On Llama, another family of models, 94% of attempts produce no usable answer under the same setup, and no opening lifts it. On a Qwen model with extra math training before it ever saw these problems, the cue actually makes things worse, because it already reasons without any help. Those are two different reasons why the cue didn't work, and neither involved any reinforcement learning, since these are base models.

Does model size predict the gain?8:42

You might guess that bigger models gain less from the cue. The author tests that guess using 5 models. In the Olmo family, the cue gains 35.7 points on the smaller one, with 7B parameters, and 15.4 on the bigger one, with 32B parameters. But the smallest model in the paper, called SmolLM3, has the cue trailing by 12 points. So size doesn't predict who gains. The model and its data do

Isn't this just "Let's think step by step"? (Kojima 2022)9:08

Okay, so here's the question you're probably already asking, isn't this just let's think step by step? It's a fair objection. There's a famous result from 2022 the paper that found adding let's think step by step before each answer took a model from 17.7% to 78.7% on one benchmark. So is this just a confirmation? Partly, and the authors say so themselves. They frame their question as how little intervention you need to trigger reasoning From the prompt hacking experiment here, their cue is two tokens, with no demonstration and no instruction.

But this time, we add a measurement of what reinforcement learning changed, a way to build and remove a cue by editing data, and the one-matrix edit. They even run the old phrase through the same test. Replace step by step with duck duck goose in the training data, and the new phrase works, 6 to 17%, against 12 to 15 for the old one. And a separate study from 2025 points the same way.

Give base models enough attempts and they overtake their trained versions, which suggests the ability was in the base model all along. Whether that's new enough is your call, and I'd say it is if you care about the mechanism and not just the headline.

What RL bought: a habit of how to begin10:16

So what did spending millions of dollars on GPU training buy? In this paper's setting, a big part of it was a habit of how to begin, and a good deal of the ability was already there, waiting for the right first words. There are two more lessons we can take away. First, a cue is a learned association that the data built, and it can be edited.

Second, the same cue can help on one model and do nothing on another. It depends on the model and on what its data paired with those words during training. So before paying for more training, maybe check out what a free opening does. First impressions matter, even for a model.

Sources

1. Wang, Dravid, Shao, Farhat, Min & Efros, "Base Models Can Reason By Taking a Cue From Training Data" (2026)
2. Wang et al., the paper's repository (cue search, evaluation scripts, MIT)
3. Kojima et al., "Large Language Models are Zero-Shot Reasoners" (NeurIPS 2022)
4. Yue et al., "Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?" (2025)