What a Meta Paper Taught Me About Giving Story Notes

This one is me learning in public.

I read a machine learning paper this week that I was not supposed to understand, and halfway through I realised it was about something I have done my whole working life. It was about giving notes.

The paper is called Rethinking Rubric Generation for Improving LLM Judge and Reward Modeling for Open-ended Tasks. It came out of Meta in February 2026, and it proposes a method with an ugly name, Recursive Rubric Decomposition, or RRD. I am not an engineer. I am a storyteller who works inside a venture fund and spends a lot of time around people who train models. So I am going to explain this paper the only way I can, which is by comparing it to the thing I actually know: how a story gets better in a room.

Call it RLNR, reinforcement learning from narrative rewards: the idea that the signal we train storytelling models on should come from the structure of story itself rather than from a reader's mood. This essay is me working out why that idea needs a paper about rubrics to become real.

Here is why I think a storyteller should care. For the last two years, the frontier of AI has been about teaching models to do things well where nobody can write an answer key. Maths has an answer key. Code has an answer key. A story does not. So the people training these models have had to invent a substitute for the answer key, and the substitute they landed on is the rubric. A rubric is a checklist. A checklist is a set of notes. And the person who knows the most about which notes make a story better is not an engineer.

A reward is just a note that the writer is paid to follow.

Let me get one word out of the way, because the whole paper hangs on it. When engineers say reward, they mean the score a model gets for a draft, and the model is trained to chase that score. If the score goes up when the draft gets more dramatic, the model learns to be dramatic. If the score goes up when the draft gets longer, the model learns to pad. The reward is the teacher, and the model becomes whatever the teacher rewards.

Now picture how this works in practice. Model A writes a story. Model B reads it with a checklist in hand and gives it a score. That score becomes the reward. So Model B is a script reader writing coverage, and Model A is a writer whose whole career is being decided by that coverage.

Every writer has lived under a bad reader. You learn to write for the note instead of for the story. The paper's authors noticed that the models were doing exactly this, and they traced the problem back to the checklist itself. The notes were bad. So the writer got worse while the score got better, which is the oldest story in any writers' room.

Storytelling solved the checklist problem a long time ago, we just never called it that.

Here is what struck me. The paper is trying to figure out how to turn a vague judgment like "this is good" into small, specific, checkable criteria. Storytelling has been doing this for about two and a half thousand years.

Aristotle said a plot should move by what is probable or necessary, so that each event happens because of the last one, not merely after it. That is a checkable note. You can go through a story and ask, for each scene, does this follow because of what came before?

Robert McKee gave us the beat, the smallest unit of story, which he defines as a moment where a value changes charge. Love turns to hate, safety turns to danger, ignorance turns to knowledge. If a scene ends on the same value it began on, nothing happened. That is a checkable note too. You can mark the opening value and the closing value and see whether they differ.

Kendall Haven, who studied how the brain actually processes stories, found that what a listener keeps is not the words but a small structure: a character, a goal, an obstacle, a struggle, and the details that make it real. If those survive in a listener's head, the story worked. That is a checkable note about retelling.

And long before any of them, Bharata's Natya Shastra laid out rasa, a typed vocabulary of eight stable emotional states and the transient states that combine into them. It is, quite literally, a schema for what a scene is supposed to make you feel.

So when I read the paper's complaint that machine-generated checklists are vague and tangled, I thought: we have the untangled version. It is sitting on every screenwriter's shelf. The two fields just have not met.

The paper found four kinds of bad note, and every writer has received all four.

The authors looked at the checklists that models generate for themselves and found they fail in four ways. I am going to give each one its writers' room name.

1. Poor coverage. The checklist only talks about some of what matters. This is the reader who gives you three pages of notes on dialogue when the problem is that your second act does not exist.
2. Mixed dimensions. One criterion bundles several different things. This is "it needs more heart." Heart is character, stakes, pacing and tone all at once, and you cannot act on it. You can only nod.
3. Wrong direction. The checklist quietly prefers the weaker draft. This is the reader who likes the safe version, and if you write for them long enough, everything you make turns beige. The paper found that naive machine-written checklists actually made the judge less accurate than no checklist at all.
4. Redundancy. Three criteria that are really the same criterion wearing different clothes. So one weakness in the draft gets counted three times and drowns out everything else.

I read this list and laughed, because it is also a complete taxonomy of bad feedback from producers, from brand managers, and from me on a bad day. The machines did not invent a new way to give bad notes. They inherited ours.

RRD fixes the notes with three moves, and a good story editor already makes all three.

The method is a loop. Here it is in plain words.

First, read the drafts before you write the checklist. The paper's first move is to have the model propose criteria after looking at several sampled responses, not just the prompt. This sounds obvious and it is the thing most people skip. An editor who writes the coverage template in the abstract produces generic notes. An editor who reads five drafts first and asks "what actually separates the best one from the rest?" produces notes that bite. The paper found that grounding the checklist in real drafts was what fixed coverage.

Second, split vague notes until they bite, and fire notes that point the wrong way. If every draft passes a criterion, that criterion is not telling you anything, so the method splits it into finer sub-criteria. The paper's own example is that "good explanation" gets broken into "defines key terms," "uses examples," and "explains steps clearly." That is exactly the motion of turning "it needs more heart" into "the protagonist never wants anything specific in act one." At the same time, any criterion that keeps favouring the weaker draft gets removed, and so does any criterion that duplicates another. The loop runs until splitting stops producing useful new notes.

Third, count a note once even when it shows up in three places. The last move is a weighting scheme that notices when criteria rise and fall together and treats them as one. If "clear structure" and "easy to follow" always agree, they are one opinion, not two. This is the discipline of a good room, where someone eventually says "we have given that note already, move on."

The results were not small. The refined checklist took GPT-4o's accuracy as a judge on one benchmark from 55.6 percent to 73.3 percent. When the refined checklist was used as the reward for training, the reward rose by up to 160 percent on one small model, where earlier rubric methods managed 10 to 20 percent. And the models trained this way got better on tests the checklist was never written for, including a hard medical dialogue benchmark.

Better notes made a better critic and a better teacher at the same time. I have believed that about human rooms for years. It is strange and a little moving to see it measured.

Here is how I think RRD could make narrative rewards better, and where it stops.

Now the part I actually care about. The papers that train models to write stories mostly reward a reader's opinion dressed up as a number. A judge model reads the story and says how literary it felt. RRD gives me a way to think about doing better than that, in four steps.

1. Start the checklist from the canon, not from scratch. RRD discovers its criteria by brute search over model outputs. We do not have to. McKee's value shift, Aristotle's causal chain, Haven's five structural elements and the rasa vocabulary are already decomposed, already non-redundant, and already point in the right direction, because centuries of readers filtered them. A narrative reward should begin with those and let RRD refine them against real drafts, instead of asking a model to rediscover story theory from nothing.
2. Let decomposition push notes toward things a machine can actually check. Watch what the splitting step does to a taste word. "Good explanation" became three concrete checks. Run the same move on "the scene works" and you get "the scene opens on one value and closes on its opposite," "the protagonist wants something specific," "the outcome follows from the previous scene." Each of those is closer to a test than to an opinion. Some of them can be checked without a judge at all, especially if the story is written in a structured form. That is the direction I want narrative rewards to go: part verified, part judged, with the verified share growing.
3. Use the wrong-direction filter to fight beige. The single most damaging failure in story generation is that models converge on the safe, smooth, average draft. RRD has a step that removes any criterion that keeps rewarding the weaker response. Pointed at stories, that is a machine for noticing when a note is quietly rewarding sameness, and throwing the note out. I think this is the most underrated idea in the paper for anyone who cares about taste.
4. Count structure once and texture once. Story checklists are full of near-duplicates: coherent, consistent, logical, well-structured. The correlation step would collapse these into one weight and leave room for the criteria that actually separate a good story from a competent one, like surprise, specificity, and whether anything was risked.

And here is where it stops. RRD makes the checklist better, but the checklist is still read by a model, and a model can only see what it can see. The paper's results say the refined judge agreed with humans 73 percent of the time. That leaves a quarter of judgments where the reader was wrong, and in storytelling that quarter is where all the interesting work lives. The thing no decomposition can produce is the person who knows which note is true. That person is a storyteller, and the paper, without meaning to, is an argument for putting one in the loop.

I am going to test this, and I will share what breaks.

I am not writing this from the outside. I build narratives for a living and I am betting my own time on the idea that storytelling can be engineered and measured. So here is what I am going to try, and I will report back whether it works.

First, I am going to take the one claim I believe most, that a story which works survives retelling, and run RRD-style decomposition on it. If the machine rediscovers beats and commitments on its own, that is a result. If it finds something I did not expect, that is a better result.

Second, I am going to try writing short stories in a structured story language, so that some of the checks above can be run by a compiler instead of a judge, and see how far a reward can be verified before it has to be judged.

If you are an engineer who has run one of these training loops, or a storyteller who has sat through enough bad notes to know a good checklist when you see one, I would like to hear from you. The two fields have been working on the same problem from opposite ends, and the paper that made me see it was not about stories at all.

A reward is a note the writer is paid to follow. The least we can do is make sure the notes are ours.

The paper: Shen, Qiu, Whitehouse et al., Rethinking Rubric Generation for Improving LLM Judge and Reward Modeling for Open-ended Tasks, arXiv 2602.05125, February 2026.