Walk through how a lab bakes a consistent persona into a model post-training

Short answer·Hard·4.0 · 0·~3 min·Asked atComet Ml Droom Servicenow

Attempt it

A lab wants its assistant to have a stable, recognizable character: warm but direct, refuses gracefully, and holds consistent positions about itself across paraphrased questions. Walk through how persona gets instilled during post-training and how you would verify it stays stable.

Free · 2 AI evals / day

**Start with a written spec.** Before any training, document the character explicitly: traits (warm, direct), voice (how it phrases things), how it talks about itself, how it disagrees and refuses. Without the spec, everyone on the team trains a slightly different model. **Instill through every post-training surface.** SFT demonstrations are written or curated to exhibit the voice. Preference data is collected so that responses adhering to the persona are chosen over those that do not, often using constitution-style principles describing the character. RL rounds penalize persona violations in the reward signal. **System-prompt distillation** is the key technique for moving the persona from the prompt into the weights. Generate training responses with the full persona prompt in context, then train the model to produce those responses without the prompt. The character is now baked in, not fragile. **Verification.** Ask the same self-referential question many ways (paraphrase consistency). Test across long conversations (does the voice survive turn 50?). Run adversarial probes that try to break character. Track tone classifier scores on a held-out traffic sample. Regression-test the persona suite on every later round, since safety or capability tuning can silently erode the character.

**Spec first.** A character spec is the foundation: traits (warm, direct, curious, not sycophantic), voice rules (sentence rhythm, opening conventions), self-talk (how the model describes itself), and refusal style. Without an explicit spec, every round drifts differently. **Multi-surface instillation.** - **SFT** demonstrations exhibit the voice. Hand-written or curated from a persona-prompted generator. - **Preference data** where labelers rank partly on persona adherence. Constitution-style principles describe the character. - **RL or DPO** rounds where persona violations cost reward. The RM scores voice consistency, not just helpfulness. **System-prompt distillation** is the structural move. Generate responses with the full persona prompt in context, then train the model to produce them without the prompt. The character moves from the prompt (fragile, jailbreakable, truncatable) into the weights (durable, survives long conversations). **Overcooking risk.** Push persona too hard and it collapses into mode-collapse tics (identical openers, formulaic refusals). Keep data diverse and penalize formulaic phrasing. **Verification.** - Paraphrase consistency: same self-referential question 30 ways, measure agreement. - Long-conversation drift: does voice survive turn 50? - Adversarial character-break probes: jailbreaks, persona-flip attempts, role-play. - Tone classifiers (warmth, directness) on held-out traffic. - Regression-test the persona suite on every subsequent round, because safety or capability tuning can silently erode character.

TL;DR

Persona lives in the weights only when every post-training surface (SFT, prefs, RL) is shaped by an explicit character spec, with system-prompt distillation moving voice from prompt to weights.

Memory aid

Easy to grasp

Imagine training a new customer service agent to embody your company's voice. You write down what that voice is (warm, direct, never sarcastic). You give them example conversations to study (SFT). You have a supervisor watch them and say which of two responses is more on-brand (preference data). You score them on every shift (RL). Eventually they internalize the voice and do not need to read the rulebook before each call (system-prompt distillation). Then you test them: ask the same question fifty different ways and see if they stay consistent. Have someone try to provoke them into breaking character. Keep checking, because new training (say, safety drills) can quietly overwrite the voice you spent so long building.

Key concepts

RLHF

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Persona is the part of an LLM product that users notice first and remember longest. It is also the part most likely to silently degrade across post-training rounds because nothing in a standard helpfulness or safety pipeline explicitly protects it. A model with a strong, stable character feels coherent across turns and topics; a model whose persona has eroded feels generic, sycophantic, or jarringly inconsistent.

This deep dive walks through how persona actually gets built into the weights of a 2026-class assistant, why putting it in the system prompt alone is insufficient, and how the verification suite that catches erosion across rounds is structured.

The goal is to leave you able to describe the full post-training pipeline through the lens of persona, name the structural failure modes (system-prompt fragility, overcooked stereotypy, silent erosion across rounds), and design the regression discipline that keeps a character stable across many deployment generations.

The spec is the contract

A written character spec is the artifact that everything else in the pipeline references. The spec includes the trait list (warm, direct, curious, not sycophantic, not condescending), voice rules (sentence rhythm, opening and closing conventions, hedging patterns), self-talk (how the model describes itself when asked who it is, what it can and cannot do, how it handles questions about its own nature), refusal style (how it declines requests without being preachy or curt), and canonical examples for the gray areas (humor on serious topics, disagreement with a user's stated belief, admitting a mistake, handling ambiguous requests).

Without the spec, every labeler internalizes a slightly different voice, every round of preference data drifts in a slightly different direction, and the character degrades stochastically. With the spec, the labeling guidelines, the RM training rubric, the regression suite, and any future round have a shared reference. The spec is also what the legal and brand teams sign off on, which matters operationally because persona changes can have brand implications.

The spec evolves over time, but every revision is versioned and the training data is tagged with which spec version it was collected under. Old data labeled to a previous spec can be re-labeled or down-weighted in the new round.

The spec is the contract. Before any training data is written, the persona team produces a written specification: voice characteristics, allowed and disallowed registers, refusal style, humor stance, formality range, cultural assumptions, signature phrases (if any), what the model says about itself. Anthropic's published Claude constitutions and OpenAI's 2024 model spec are public examples. The spec is the artifact that everyone (data labelers, eval writers, red-teamers, leadership) refers back to when there is disagreement about whether a behavior is on-spec.

Multi-surface instillation

System-prompt distillation moves voice into the weights

Overcooking and the verification suite

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

Anthropic's Claude Opus 4.7 uses constitutional principles describing the persona, with system-prompt distillation to move voice into the weights
OpenAI's GPT-5.5 personality work pairs hand-written demonstrations with persona-aware preference labeling at the post-training stage

What an interviewer would ask next. Try answering before peeking at the approach.

QHow do you stop a safety round from eroding the persona you built in the previous round?

Include persona-adherence as one of the reward dimensions in the safety round, run the persona regression suite as a deploy gate, and use joint optimization rather than sequential overwriting where possible.

1 more follow-up an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Putting the persona in the system prompt only. Prompts get truncated, leaked, or overridden; a stable character has to live in the weights.

60 second bullets to scan on the way to the call.

Why a written character spec is the foundation
How persona is instilled across SFT, preference data, and RL

Primary sources. Browse if you want the original framing.