This year's El Niño may be the strongest on record, and scientists still debate what drives it. SciExam for ENSO hands AI agents raw ocean data and no answer key, then judges the models they build with hidden scientific tests.
Niño 3.4 index, monthly, 1950 to now. The map above shows the month under your pointer.
What is Niño 3.4? The average sea surface temperature anomaly inside the box on the map, 5°S to 5°N and 170°W to 120°W, the standard yardstick for ENSO. Each bar is one month. Orange bars are warmer than normal, blue bars colder, and several months beyond ±0.5 °C mark an El Niño or a La Niña. Labels mark the strongest events. Move your pointer across the map or the bars to travel from 1950 to now. Values are monthly anomalies from the 1991–2020 average in NOAA ERSSTv5, so they can differ slightly from NOAA's official index, which put Niño 3.4 at +1.8 °C in August 2026.
Chapter one · What is ENSO?
ENSO, short for El Niño–Southern Oscillation, is a natural swing of the tropical Pacific between a warm phase, El Niño, and a cold phase, La Niña, every two to seven years. It arises because the ocean and the winds above it push on each other. Scientists follow it with five numbers, marked on the diagram.
Trade winds blow from east to west and pile warm surface water in the west, so the warm layer is thick there and thin in the east, where cold water wells up from below. The five numbers measure departures from this state, so here they all sit near zero.
Westerly wind bursts, gusts lasting days to weeks over the western Pacific, turn the wind there westerly τ. They push surface water east and send warm Kelvin waves along the equator. The current u turns eastward, warm water spreads to the central TC and eastern Pacific TE, and the trade winds weaken, which slows the cold upwelling. Bursts grow stronger and more frequent when the central Pacific is warm, so each change warms the east further.
During an El Niño the equatorial Pacific slowly loses heat, and the warm layer in the west thins hW. This discharge ends the event and can tip the ocean into La Niña. The trade winds strengthen, westerly bursts τ become rare, and colder water returns to the east TE. The heat then builds up again, and the cycle repeats.
Chapter one · What is ENSO?
ENSO, short for El Niño–Southern Oscillation, is a natural swing of the tropical Pacific between a warm phase, El Niño, and a cold phase, La Niña, every two to seven years. It arises because the ocean and the winds above it push on each other. Scientists follow it with five numbers, marked on the diagram.
Normal conditions
Chapter two · The exam
Usually an AI benchmark checks answers against a key. Here the agent must build a working model of a real climate system, keep revising it against its own checks, and hand it in to be examined.
Each agent gets 35 years of gridded ocean and wind data, 1980 to 2014, and turns it into five monthly indices. A data check must pass before modelling starts.
The agent decides how a good model should be judged and writes the diagnostics. They are then frozen, so the yardstick cannot drift toward whatever the model does.
The agent writes a first model and fits it to the data. It runs its locked checks, reads the results, and revises the equations, the noise or the nonlinear terms. It repeats this loop as often as it likes within the six hours, then submits the model it chooses. It never sees a score.
After the run, the model must reproduce 70 years of ENSO statistics, forecast 2015 to 2024, years the agent never saw, and recover variables it is not shown. A published model takes the same tests.
Chapter two · The exam
Usually an AI benchmark checks answers against a key. Here the agent must build a working model of a real climate system, keep revising it against its own checks, and hand it in to be examined.
Matches 70 years of observed statistics.
Recovers the variables it is not shown.
Beats persistence 1 to 12 months ahead.
Chapter three · The scores
The reference model was published in the Journal of Climate and developed with observations from every test period. The agents had only 1980 to 2014. Hover a row for its three scores, or change the weights and watch the ranking move.
agent systems score higher than the published reference
top composite score, Claude Fable 5.1, against 0.378 for the reference
forecast lead times where the top model beats persistence for both indices
Chapter four · The discovery
El Niños grow stronger than La Niñas. Since 1950 the strongest El Niño warmed the eastern Pacific by 3.3 °C, while the strongest La Niña cooled it by only 2.0 °C, and scientists still debate why. Strip each strong model down to the terms that matter and it keeps one of two explanations, either of which makes the warm swings outgrow the cold ones. Try them yourself.
Toy model. One made-up equation for a single temperature T, run in your browser to show how each mechanism tilts the swings. It is not one of the agents' models.
Each badge is an agent's final model after removing the terms it did not need, with the composite score of that reduced model. Logos mark each model's provider.
Epilogue
@article{zhang2026sciexam,
title = {SciExam for ENSO: Can AI Agents Build Climate Models?},
author = {Zhang, Yinling and Liu, Langchen and Xiu, Dongbin and Zou, Xueyan
and Kuang, Xu and Wang, Mengdi and Liu, Shilong},
year = {2026}
}
Sea surface temperature anomalies from NOAA ERSSTv5 (1991–2020 base period), January 1950 to August 2026. The paper link will be added on release.