Part 2 of 5 in the LibriBrain series. The project page summarises the whole project. Previous: What timing reveals about MEG decoding benchmarks, the technical account.
LibriBrain began as one dataset among many in a broader investigation of public neuroscience data. In late September I gave it a dedicated session, and on October 8 I entered both tracks of the PNPL competition: decode which of fifty words a listener heard from 306 channels of magnetoencephalography. My first question was whether a route to the top of the leaderboard existed at all. This post is about how we found one, and how far it went.
Changing the question
My opening idea came from earlier work on handwriting decoding, where guessing a word from several uncertain letters had proved far easier than guessing each letter alone. The same principle should hold for speech. Decode each word roughly, then let the sentence correct its weakest members. That intuition would return in a more disciplined form later in the month.
The first model, a classifier choosing among fifty labels, stalled near chance. An independent audit of the pipeline made the cause clear, and I redirected the work around a different question taken from published research. Instead of asking which of fifty words the brain heard, ask where that word sits in a language model’s space of meaning, in the context of its sentence. I set one condition for the rebuild: keep going until the local score passed 0.4. Two days later the leaderboard score had doubled, from 0.23 to 0.43. The architecture had barely changed. The target had.
One brain teaches the rest
I worked in two registers throughout the month. Ideas from AI assistants, the literature and my own notes went into an open-ended brainstorming session. Anything that survived scrutiny went to an execution session with a cheap test attached. Most borrowed ideas did not survive their first measurement. The value of the process was its speed: an idea could be proposed, tested and closed within an hour.
The decisive result came from a run aimed at the other track. A model trained on the one listener with eighty hours of recordings, and no one else, transferred to the thirty-nine other listeners better than any model trained on all of them. Their short recordings added more variation than information. The Broad score rose from 0.43 to 0.66 in a single evening.
That result made me interested in comparing the transfer of sensory and motor representations. The contrast with user-specific handwriting models suggests a useful question, although these two projects alone cannot separate task differences from differences in data volume or measurement.
Rules first
Near the top of the leaderboard, I paused to verify one thing before any further work: that nothing in our pipeline used the shortcut the organisers had prohibited. Every listener heard the same sentences, and the test files leaked enough timing to align rows across people. We had used none of it.
I held a second line of my own. Early on, a model given the spacing between words jumped from 0.30 to 0.51 locally. It was reading word durations from the audio, not the brain. The rules allowed it; I removed it. After the competition I measured how far that shortcut alone reaches. A model with no brain data at all scores 0.63 on our test blocks, which makes timing-only prediction an important comparison alongside chance on this evaluation split.
Pockets of predictability
The most original stretch borrowed from finance. Norman Packard, before co-founding a trading firm, designed an algorithm that searches for the conditions under which a chaotic system briefly becomes predictable. Applying it to brain recordings seemed like a reach, but conceptually close enough to test. The search found one condition with a hit rate of 0.92. The first analysis called it a fluke. I asked why we were ruling it out so early, and a second look, controlling for which words occur where, showed the effect was real: words at the start of a sentence decode far better than the same words mid-sentence, most likely because the brain responds more strongly after silence.
The home stretch
The final day and a half was a sustained effort to stack every gain that had shown signal. A self-learned grammar built from each listener’s own decoded sentences added nearly two points; that was the opening intuition about sentence correction, in its disciplined form. Correcting for recording-day drift added four points on the Deep track. A parallel session generated and screened ideas in volume while the main session handled the finals. The more ambitious ideas, ten moonshots and a turn toward formal linguistics, did not move the score. At four in the afternoon I called the lock-in and we consolidated what had worked.
The entry scored 0.722 on Broad, eighth on the public leaderboard, and 0.729 on Deep, against published references of 0.430 and 0.732. The final private leaderboard placed it 17th of 53 on Broad (0.678) and 26th of 91 on Deep (0.697).
What I take from it
Two practices carried this work. Keeping wide-open exploration separate from disciplined execution let ideas move quickly without contaminating the results. Refusing to accept a stopping point until it had been measured was right more often than it was wrong; several limits we had called hard turned out to be soft. Explicit evaluation rules helped keep that process consistent under deadline pressure.
The code, the pre-registered experiment log and the full development history are at github.com/rohit-ravi2/libribrain-2026, and the technical summary is on the project page.
Next in the series: Eighth in public, seventeenth in private, on what the private leaderboard changed.