NARC

Narrative Augmented Reasoning Challenges

What is NARC?

NARC is a kind of abstract reasoning puzzle. Each puzzle presents a sequence of colored grids that tell a visual "story." One or more grids in the sequence are hidden, and your goal is to reconstruct them pixel-perfectly.

The catch: the grids alone aren't enough. Each puzzle comes with a short narrative clue. The clue never mentions grids; it helps only when read figuratively, a character as a bar, a mood as its height. Without it, the missing grid is ambiguous — several answers seem plausible. With it, one answer is right.

This is the NARC property: neither the grids nor the narrative suffice alone, but together they determine the answer. For AI models we measure it directly: a puzzle has the property for a given model when that model solves it with grids and clue together but not with either alone.

How to play

  1. Look at the grid sequence. Each grid uses a 10-color palette. One or more grids are hidden, shown as a ? placeholder.
  2. Try to guess what the missing grid looks like from the visual pattern alone. You can submit a guess before seeing the clue.
  3. Reveal the clue. A short narrative provides the insight that disambiguates the puzzle.
  4. Draw your answer. Set the colors of each cell in the missing grid and submit.

Color-blind? The Numbers button in the top bar prints each color's digit in every cell.

Why does this matter?

NARC asks a narrow question: can a short story supply what a sequence of pictures leaves undetermined, and when it does, which part of the story is doing the work — its words, its order, or its form? For AI models we can test this with ablations: swap the clue for its keywords, shuffle the grids, hide a different grid, tell the same story a different way.

The human side is different. People solve puzzles on this site for fun, and we record those attempts anonymously, but we have not run a human study and make no claims about human difficulty or about how humans and models differ. That comparison is future work.

NARC is a sibling project to MARC2 (Metaphor Abstraction and Reasoning Corpus), which explored how figurative language helps AI solve abstract reasoning tasks from the ARC-AGI benchmark.

How much to trust what you see here

This site began as a research proof of concept and is now used by visitors as well as by our team. Not everything on it carries the same weight.

  • Documented. The Masking results: three conditions plus order and keyword ablations, 9 models, every active puzzle. A paper on this battery is under review (September 2026). Verdicts are per model and single-sample; on re-runs about one NARC verdict in six flips, so read a puzzle's profile across models, not a single cell.
  • Pilot. The Ordering, Odd-One-Out and Stances tabs, and the narrative-variant studies from spring 2026 (Story Prism, ABT, focalization): smaller designs, not part of the paper's claims. Each tab states its own scope and dates.
  • Exploratory. Human solving on this site: anonymous, self-selected, uninstructed, our own team included. We show attempt counts only, never human difficulty or solve rates.
  • Retired. Early proof-of-concept fields — predicted "human difficulty" and "AI difficulty" ratings and "spectrum" tags written by the generating model — were never measured and are no longer shown. The only difficulty signal on the site is measured: on how many tested models a puzzle has the NARC property.

The puzzle corpus

319
Active puzzles
972
Narrative variants
9
Models tested
3–8
Grids per puzzle

Another 314 puzzles are drafts under review and visible only to the team.

Puzzles come from three sources, labelled on every card: Human-written by our team or visitors, AI-generated by a language model working from a strategy prompt or as an autonomous agent, or AI + human, an AI draft revised by a person. Every active puzzle was approved by a reviewer and then run through the model battery.

Puzzles span literary classics (Hemingway, Kafka, Shelley), scientific concepts (natural selection, entropy, plate tectonics), philosophical thought experiments (the trolley problem, Plato's cave, the ship of Theseus), and more.

Research context

NARC draws on:

  • ARC-AGI — abstract visual reasoning as an intelligence benchmark
  • MARC — figurative language as a bridge between human and machine reasoning
  • Econarratology and the Story Prism — Erin James's framework for how narratives construct "storyworlds," which shaped the narrative-variant pilots
  • Focalization — how the same story told from different perspectives changes what is foregrounded; our spring 2026 pilot on this gave mixed results

Get involved

  • Browse and solve puzzles. The landing view shows our team's picks.
  • Create and submit a puzzle. A reviewer looks at it, then it runs through the model battery; accepted puzzles are published here with their answers and model results.
  • Inspect how models perform across puzzles and narrative variants.
  • Narrate is an invited study: write the story that lets a model solve a puzzle.

What this site records

Participant information for the NARC study (University of Idaho IRB 26-225). Adults, 18 or older, only; we do not ask your age.

What you do. Solving: reconstruct a hidden grid, before and after reading the story, 2 to 15 minutes per puzzle. Creating: draw a sequence of grids and write its story, 15 to 60 minutes. Narrating (by invitation): write a story for a puzzle, test it against an AI model, revise, 20 to 90 minutes.

What is recorded. When you solve a puzzle we store the grids you submit, whether they were correct, whether you had revealed the story, and how long you took, under a random session id kept in your browser. Page actions (reveals, submits, retries) are logged the same way. Votes are stored with a random id kept in your browser and your IP address, used only to limit repeat voting and deleted after 30 days. If you submit a puzzle, we store the puzzle and, only if you enter one, the name you want to be credited by. In Narrating we store every story you submit, when, how long you spent, whether you had viewed the existing story, the model's full response, a random participant id, and the shared invite code you used (which tells us which event or mailing you came from). We do not ask for your email or anything else about you.

How it is used and shared. Puzzles and stories you create may be published here with their answers and model results, in research papers, and in public datasets, word for word. A puzzle is credited by the name its contributor entered, or as by an anonymous contributor if none was entered; Narrate stories are never credited and invite codes are never published. Solving records are released only with hashed session ids and no IP addresses. Records are kept indefinitely as part of the published benchmark.

Risks and benefits. No greater than solving or writing a puzzle for a magazine. The main ones are that something you wrote is published in a way you did not expect, and, if you chose to be credited, that your name appears publicly with your puzzle; the notices on each page are meant to make both clear before you submit. No direct benefit beyond the puzzles; the research benefit is a public, human-authored benchmark.

Voluntary and confidential. Taking part is voluntary; stop at any time by leaving the site. Participation is unrelated to any course, grade, or employment. Data live in a database on a University of Idaho server behind a login; Narrate data are kept in a separate database that only the principal investigator can read. Credit names are visible to the principal investigator and team reviewers before a puzzle is published, and publicly once it is.

Why records cannot be removed. We do not record who you are, so we have no way to tell which records are yours, and we cannot remove a person's records after they are submitted. If you do not want something recorded, do not submit it. The one exception: if you published a puzzle under a name, you can ask the principal investigator to remove the name or withdraw the puzzle.

Questions. Contact Bert Baumgaertner, University of Idaho, bbaum@uidaho.edu. Questions about your rights as a participant: University of Idaho Office of Research Assurances, irb@uidaho.edu, 208-885-6340.

Paper, code, and data

A paper describing the masking battery and its results is under review (September 2026). A citation and a data release will be posted here when they exist. Until then, please cite the site by URL and date.

Code: github.com/bertybaums/narc.

Bert Baumgaertner · University of Idaho · Page updated September 22, 2026 · Model results as of September 17, 2026