Chatbots keep naming the heroine Elara because alignment narrowed their range
Reward training concentrates a model's probability on a few favored answers, and turning up temperature samples that narrowed range more loosely without restoring the variety the base model started with.
Why this matters
Every time you ask a chatbot to brainstorm names, pitch story ideas, or pick something at random, you are drawing from a model trained to give the answer people rate highest, not the widest range of answers. That is why the results feel same-y, and why the temperature setting most people reach for to loosen them up only partly works. This lesson starts from the behavior you can see, the heroine who keeps being named Elara, and works back to the training step that produces it. It defines the spread of answers a model draws from, shows where the spread gets narrowed, and explains why one slider cannot widen it again. By the end you can say why "AI writing all sounds the same" is a fact about how these models are trained.
- 01
- Two answers to one prompt diverge at a single draw: how a model turns its weights into a next word, and what the temperature knob does at the moment of the draw.
- 02
- Chatbots overuse ‘delve’: another fingerprint that training leaves on model prose, and a popular explanation for it that failed its one direct test.
The heroine is almost always named Elara
Ask a chatbot for a short story and there is a fair chance the hero is a young woman named Elara. Open a fresh chat, ask again, and she comes back. The habit is common enough that a naming expert made Elara her 2025 name of the year, reporting that AI systems reach for it "no matter the context, no matter the genre," down to math workbooks about Elara's adventures in geometry.1 It is a strong tendency rather than a fixed law. On a widely shared thread cataloguing the pattern, one model version returns Elara while others default to Aldric, Lila, or Elinor. Models lean hard on a few names, but no model always picks the same one.2
Names are the kind of thing memory embroiders, so the sturdier evidence is measured. Padmakumar and He had 38 writers compose argumentative essays three ways: alone, alongside the base model GPT-3, and alongside its aligned version, InstructGPT. The two models contributed about the same amount of text, roughly a third of the characters, yet only the InstructGPT essays grew more alike. The average similarity between different authors' essays rose from about 0.154 for writers working alone to 0.166 with InstructGPT. GPT-3 left it near 0.158, and the jump for InstructGPT was statistically significant.3 Different people, handed the aligned model, drifted toward the same essay.
The narrowing shows up in tone as plainly as in wording. Wu, Black, and Chandrasekaran had models generate book reviews and counted how many landed in the most positive sentiment band. Real human reviews land there only 2.8% of the time. The base model stays close to people. Each further step of alignment piles the reviews onto the positive end.
| Source | Training stage | Most-positive share |
|---|---|---|
| Human reviews | the source data | 2.8% |
| Llama-2 | base, pretrained only | 5.9% |
| Vicuna-13b | supervised fine-tuning | 44.7% |
| Llama-2-chat | RLHF | 82.1% |
| GPT-3.5, GPT-4 | aligned, proprietary | 100% |
Many people pin this sameness on the temperature setting and try to cure it by turning that setting up. The cure mostly fails, and seeing why means following an answer back to where it is born.
A model draws each answer from a spread of options
A language model does not keep answers on a shelf. At each step it assigns a probability to every word that could come next, and a sampler draws one word from those odds. An earlier lesson walks through that single draw and the temperature knob that tunes it. Two answers to one prompt diverge at a single draw is the place to start if the draw is new to you. What matters here sits one level up: not the odds on the next word but the spread of whole answers the model tends to give. Call that spread the output distribution. When it is wide, repeated requests wander across many answers. When it is narrow, they keep landing on the same few.
Two words describe the shape of that spread. A mode is a peak, an answer the model returns far more often than its neighbors. Elara is a mode. Entropy measures how spread out the whole distribution is: high entropy means the odds are shared across many options, and low entropy means a few options hold most of the weight.
Before any alignment, the spread is wide, because the base model is trained to imitate an enormous range of human writing, and human writing is full of different names and different sentences. Behnam Mohammadi compared a base model with its aligned version and found the base model's next-word odds spread across many words, leaving its token-level entropy higher than the aligned model's.5 The book-review test tells the same story from the other end: the base Llama-2 put 5.9% of its reviews in the top sentiment band, close to the human 2.8% and far from the 82.1% of its aligned sibling.4 Left as a plain imitator, a model roughly reproduces the variety of the text it learned from.
Reward training piles the odds onto a few answers
That variety does not survive the next stage. After pretraining, a model is aligned so it follows instructions and answers the way people prefer. The standard method, reinforcement learning from human feedback, trains a separate reward model to score answers by how much human raters like them, then tunes the model to earn higher scores. The published InstructGPT objective states exactly what that tuning maximizes.6
- r_\theta(x,y)The reward: how much a scorer trained on human preferences likes the answer. Pushing this term up is what shifts probability onto a few high-scoring answers.
- \beta \, \log \tfrac{\pi^{\mathrm{RL}}}{\pi^{\mathrm{SFT}}}The leash: a penalty for straying from the pre-reward model. It caps how far training drifts and does not, on its own, widen the range.
Read it as one instruction: produce answers the reward model scores highly, and do not drift too far from the model you started as. The reward term is the pressure that narrows. If a handful of answers reliably earn high scores, raising the score means putting more probability on those answers and less on everything else, so the distribution contracts around the reward's favorites. That is where a recurring Elara comes from. A smooth, pleasant name with no real-world person attached is a safe high-scorer, so a model trained to please gravitates to it.1
It is tempting to blame the penalty term instead. It measures how far the aligned model has moved from the pre-reward one, so it reads like a brake on variety. It is a leash, not the hand that narrows. The InstructGPT authors added it to stop the model from over-optimizing the reward and wandering into nonsense, not to shrink the range.6 Kirk and coauthors tested the leash directly and found that tightening it, by raising the penalty coefficient, lowered diversity further rather than restoring it.7 The narrowing lives in the reward chase. The penalty only bounds how far the chase can run.
The measured drops line up across studies. Kirk and coauthors compared full RLHF against the supervised model it starts from. They found that "RLHF substantially decreases the diversity of outputs sampled for a given input," and reported what they call the first rigorous demonstration of across-input mode collapse from RLHF: even across many different prompts, the answers cluster onto a few. That across-input effect is real but smaller than the per-input drop, and the per-input variety is what falls most clearly.7 Padmakumar and He's aligned essays carried the fewest distinct wordings and the fewest distinct ideas at every threshold they checked.3 Mohammadi's aligned model ran at lower token-level entropy throughout.5
A hotter setting cannot restore what was trimmed
This is where the temperature guess falls apart. Temperature flattens or sharpens the odds at the moment of the draw, as the earlier lesson lays out. Turn it up and the sampler reaches further down the list of options the model is offering. But it can only reach among options that still carry weight. If reward training already stripped the probability off the alternatives, a hotter setting spreads the draw across a distribution that is itself narrow. It samples the surviving range more loosely. It cannot rebuild a range the training removed.
Two findings mark the limit. Mohammadi found that the aligned model's favorite completions, its attractor states, persist even at a temperature of 1, a high setting, while the base model shows no such trapping even at low temperatures.5 Wu and coauthors tried four ways to widen the reviews, including higher temperature and prompts that explicitly asked for variety. The best of them moved the most-positive share from 82.1% down to 58.1%, still far above the 5.9% the base model produced on its own.4 The knob buys a little room. It does not undo the training.
What the narrowing does not touch
A loose version of this claim does not survive one careful study. Shypula and coauthors measured diversity only among the answers that were actually good, and by that measure preference-tuned models came out ahead of base and supervised models, by a wide margin.8 Alignment raises the share of usable answers, and counting only the usable ones can flip the verdict. The same paper still finds aligned models lower on plain word- and sentence-level variety.8 So alignment trims the variety of words and sentence shapes and concentrates the model on a few modes. Whether you call that "less diverse" depends on whether you weight by quality.
Two things about this are settled engineering. Post-training measurably lowers word- and sentence-level diversity and token-level entropy relative to the base model, and a hotter temperature does not restore what it removed. Two things are still open, even for the people who build these systems. Which pressure dominates the narrowing, reward over-optimization or homogeneous raters or something else again, is unresolved, though the KL penalty has been ruled out as the main cause. And how to hand back the base model's range without giving up the quality alignment bought remains unsolved. When a chatbot keeps offering you Elara, that open question is what stands between you and the variety you asked for.
The takeaway
The variety you want lived in the base model, which imitated the full range of human writing. Alignment trained the model to prefer the answers people rate highly. In doing so it moved probability off the alternatives and onto a few favorites, and separate requests keep returning the same names, the same tone, the same shapes. Turning up temperature samples that narrowed range more loosely. It cannot restore options the training removed. The fix people reach for is a setting, and the cause sits one level deeper, in what the model was rewarded to do.
- 01
- Kirk et al., "Understanding the Effects of RLHF on LLM Generalisation and Diversity": the study that first demonstrated across-input mode collapse from RLHF, and tested the KL penalty as a cause.
- 02
- Wu, Black, and Chandrasekaran, "Generative Monoculture in Large Language Models": the sentiment and factuality measurements behind the table, and the temperature and prompting tests that only partly widen the output.
Sources
- TODAY · Laura Wattenberg on Elara, AI’s favorite name and 2025 Name of the Year
- Hacker News · “When you ask ChatGPT ‘Tell me a story’ it’s always about a girl named Elara”
- arXiv · Padmakumar and He, “Does Writing with Language Models Reduce Content Diversity?”
- arXiv · Wu, Black, and Chandrasekaran, “Generative Monoculture in Large Language Models”
- arXiv · Behnam Mohammadi, “Creativity Has Left the Chat: The Price of Debiasing Language Models”
- arXiv · Ouyang et al., “Training language models to follow instructions with human feedback”
- arXiv · Kirk et al., “Understanding the Effects of RLHF on LLM Generalisation and Diversity”
- arXiv · Shypula et al., “Evaluating the Diversity and Quality of LLM Generated Content”