IBM-1

The Implicit Brain Model: a multimodal foundation model of whole-brain state and dynamics, lazily materialized into explicit maps — one 3.84M kernel driving sixteen models, image-to-EEG retrieval at 127× chance, and cortical dynamics holding a musculoskeletal body upright.

2026-09-04

github ↗site ↗

The Implicit Brain Model (IBM-1) is a multimodal foundation model of whole-brain state and dynamics which can be lazily materialized into explicit maps for training and applications. By learning a perspective-agnostic representation of the brain rather than one tied to a particular measurement modality, IBM-1 can train end-to-end across heterogeneous data sources that would otherwise be difficult or impossible to combine.

try it out ↗

The substrate and its materializations: pick one on the ring to see its inputs and outputs.

The body, moving

the real anatomy crawling16.0 s · 0.920 m · worst joint excursion 11.5°3,995 of 4,000 anatomical entities bound to the driven skeleton at 99.88%, posed from a controller searched on the 22-segment scaffold. Peak muscle fibre velocity 9.69 of 10 — zero time outside the muscle model's domain.

1,047 bone and muscle surfacesa full measured gait cycleReal Rajagopal joint coordinates from a walking subject, imposed on the anatomy. The anatomy travels up to 2,234 mm; the body's own canonical run moves 6.35 mm across 30 s.

117 tissue force elementsthree arms of the same drop, side by side105 ligaments and 12 joint capsules as Blankevoort elements, with attachments derived from the per-vertex vote the segment binding is built from.

cortex in the loop0.17 mm COM drift under a 5 N pushThe trained 1024-site cortical kernel holding the body upright: the cortex is the body's motor owner.

video predictiontruth left, prediction right, at t+8Next-frame prediction from the cortical state, eight frames ahead.

cortex in the loop, beside cortex severedupright at 1.019 m · fallen at 0.590 mWith the cortex in the loop the body stays upright; severed from it, the body is on the floor by 2.90 s. The cortical dynamics carry the balance.

one step, with skinfoot airborne at 0.14% body weightClears 1.8 cm, advances 17.1 cm and lands on measured contact — a genuine step by the harness's own strict criterion.

the 22-segment scaffoldwhat the anatomy is posed fromThe body the engine integrates: 22 rigid bodies, 80 muscles and 33 coordinates, the frame the full anatomy is posed from.

**cortex in the loop, beside cortex severed**1.019 m upright · 0.590 m fallenRendered from the OpenSim model with segment colour set by force-weighted mean muscle excitation. 22 bodies, 20 segments, 98 muscles; no limb length is assumed and the geometry is read from model.osim.

**the acquired atlas**2,229 surfaces, ~350 MB, ~28 s to loadBodyParts3D 4.0 surfaces as measured, before the display promotion registered another 1,768 structures from other datasets onto them.

**multisystem physiology**the body state the brain readsDigestive, circulatory, respiratory and renal state integrated by the native engine — the substrate the 15 visceral afferent channels are transduced from.

**cardiopulmonary**the loop that does not stopHeart and lung mechanics running continuously beneath everything else. The vagal afferents that report it are ~80% afferent by declaration, and their unmyelinated C arm carries two thirds of the interoceptive signal.

**skin bioelectricity**1,326 patches, 100.00% of the exteriorEvery patch innervated: 1,199 carry a spinal root, 127 are trigeminal territory with no dermatome because the face is cranial. 29 of 30 spinal levels — C1 is absent because it has no cutaneous territory in life.

**the materialization ring**every path drawn at its measured strength151 edges from the live brain graph, each arrow's width set by its measured magnitude.

**what a materialization contributes**the signals feeding each systemClicking an annotation shows the signals feeding the highlighted system and what each contributes to it.

**asking the brain what it is seeing**63.5% top-1, 127× chanceNearest bank entries ranked by the group-mean measured evoked EEG of the image, on the THINGS-EEG2 designated test set.

§1 design

fields — F = { x(q) : q ∈ Ω } — state variables on a support. a component is a physical quantity, never an encoding. supports are not shared: the cortical sheet, the vascular tree and a sensor array are different domains.
anatomy — a(q) ∈ [0,1]ᵏ — soft membership of a position in named partitions. an atlas is evidence about where things are, not a coordinate system.
topologies — T(i, j) — which state variables may interact, through what. all spatial structure lives here; there is no universal spatial operator.
processes — P = (I, O, T, f, θ) — dynamics over a topology. (f, θ) is implementation; implementations compete for a process rather than redefining it.

M = materialize(R, r, B, F, A, T, P)

an explicit model names regions R, a resolution r(q) and a bandwidth B(q); dependency tracing instantiates only what is reachable from the target. the same field can sit at 1 mm across the brain and 50 µm around an electrode at once. the forty entries in the library differ only in R, r and B, and each carries provenance: which implementation ran, which parameters evidence moved, and where a process ran outside its regime.

§2 research

finding	measured	outcome
the shared cortex is load-bearing	bypass +324% loss, frozen +180%, association zeroed +180%. on visual retrieval, bypass costs 37 of 63 points; on balance, severing drops the body at 2.90 s	held
band-limited fusion is exact	real EEG moved the posterior 2.03 prior sd in band; out of band, exactly 0. delays as phase ramps agree to 4.4e-16	held
the thalamocortical resonance	a spindle at 13.45 Hz; held-out +5503 nats/night, p = 7.2e-11	fitted
the slow oscillation sits at the band edge	1.000 ± 0.296 Hz across 8 scored N3 nights; τ_adapt = 0.12 s reproduces it	fitted

§2b status

Where the programme stands, updated 9 September 2026. Every figure below is measured against an explicit baseline and names the set it was measured on.

what runs	measured	against
one implicit kernel, published	fused from 34 checkpoints, Procrustes-aligned, weighted by measured transfer. recovers 88.7% of a task-specific kernel on the task that kernel was trained for	0% for random
16 materializations, one 3.84M kernel	trained simultaneously to step 32,000, consolidated every 500 steps, replica drift cosine 0.998–1.000, none collapsed. visual_eeg 30.4% top-1, optic_nerve 22.8%, ten per-subject terms mean 15.4%	0.5% chance
image → cortex → EEG retrieval	63.5% top-1 on the designated test set, 200 images, median rank 1	127× chance
a declared optic nerve into occipital cortex	three retinal populations at 2.5 / 4.2 / 8.3 ms, resolved. displacing the port to a random region of equal size collapses retrieval to 0.44%	8.69% in place
the substrate resolution has an interior optimum	controlled sweep, resolution the only variable: 128 → 19.4%, 512 → 23.0%, 2048 → 23.0%, 8192 → 21.7%	both tails worse
the sheet delivers between disjoint regions	image → occipital → sheet → a precentral-only readout (ports verified disjoint) reaches 48.4× chance on real THINGS-EEG2, against a severed sheet at 1.00×	severed = chance
cortical dynamics hold a body upright	250 exchanges, five seconds, 5 N pelvis push: 0.17 mm COM displacement	held

§4 releases

forward models — av_meg_predictor, av_predictor, meg_encoder, visual_evoked, eeg_forward, meg_forward, bold_forward, fnirs_forward, ecog_forward (all WIP)
decoding — meg_to_text, speech_envelope, eeg_to_image, inner_speech (all WIP)
stimulation — tms_response, tes_response, tfus_response, dbs_response, tms_eeg_tep (all WIP)
state and disorder — sleep_dynamics, anesthesia, seizure_propagation, pharmaco (all WIP)
the embodied loop — invasive_bci, handwriting_bci, motor_mapping (all WIP)
surrogates — macro_surrogate, connectome_gnn, encoding_model, resting_state_fc (all WIP)

The body, moving

§1 design

§2 research

§2b status

§4 releases

Related