BrainCoT: A Multi-Task Zero-Shot Brain Signal Foundation Model with Neurometric-Anchored Chain-of-Thought Reasoning
Organizations: New Cornerstone Science Laboratory, Guangdong Provincial Key Laboratory of In-Memory Computing Chips, Shenzhen Graduate School, Peking University, Shenzhen, China · Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University, Shenzhen, China · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE
Abstract
Brain signals exhibit substantial multi-center heterogeneity, strong inter-subject variability, and highly diverse task settings, which makes it difficult to build models that generalize beyond a single task. While recent brain signal foundation models leverage large-scale pretraining to learn generalizable representations, they often fall short in practice because they still require target-data-specific supervised adaptation, lack multi-task zero-shot capability, and rarely provide verifiable evidence to support their predictions. In this work, we present BrainCoT, a multi-task zero-shot Brain signal foundation model with neurometric-anchored Chain-of-Thought (CoT) reasoning. BrainCoT unifies heterogeneous tasks into an instruction-driven framework so that, after pretraining, a single model can be directly applied to multiple tasks in zero-shot settings without additional adaptation. It further introduces neurometric-anchored chain-of-thought grounded in neurometric evidence for verifiable reasoning, and adopts Decision-Consistent Rationale Training to keep process learning aligned with correct predictions. Across extensive evaluations, BrainCoT achieves an average ACC of 67.41% and AUC of 74.33% in zero-shot classification without downstream training data, outperforming the strongest linear-probing baseline trained with 1% labeled downstream data by 3.85% in ACC and 5.97% in AUC.
Figures & tables
| Signal modality | Dataset | Task | Label | Channels | Duration | Subjects | Origin sample rate |
| iEEG | Mayo [ 29 ] | Interictal epileptiform activity detection (IEA detection) | 2-class | 1 | 3s | 15 | 500 Hz |
| IEDS [ 30 ] | Interictal epileptiform activity detection (IEA detection) | 2-class | 1 | 3s | 25 | 1000 Hz | |
| EEG | ISRUC [ 31 ] | Sleep stage classification | 5-class | 2 | 30s | 100 | 200 Hz |
| Dreams [ 32 ] | Sleep stage classification | 5-class | 3 | 30s | 20 | 200 Hz | |
| Siena [ 33 ] | Epileptic state recognition | 4-class | 16 | 10s | 14 | 512 Hz | |
| SanDiego [ 34 ] | Parkinson’s disease detection | 2-class | 32 | 2s | 31 | 512 Hz |
| Dataset | ISRUC | Dreams | Siena | SanDiego | ADHD | Schizo | APAVA | ADFSU | IEDS | Mayo | NMT | SEE |
| ACC | 82.77 ±1.38 | 77.66 ±2.51 | 47.31 ±16.20 | 62.64 ±9.65 | 59.13 ±3.37 | 62.23 ±6.49 | 73.20 ±12.71 | 80.05 ±8.67 | 67.80 ±6.32 | 77.26 ±6.76 | 57.47 ±0.00 | 61.45 ±0.00 |
| AUC | 95.19 ±0.62 | 92.49 ±1.07 | 64.85 ±9.68 | 63.61 ±11.13 | 59.44 ±3.83 | 73.53 ±8.35 | 82.36 ±6.60 | 71.79 ±15.45 | 61.00 ±4.98 | 88.31 ±4.24 | 75.10 ±0.00 | 64.24 ±0.00 |
| Model | BrainBERT | BIOT | LaBraM | CBraMod | NeuroLM | ELM-MIL | CSBrain | BrainCoT | |
| Metric | Training data | ICLR’ 23 | NeurIPS’ 23 | ICLR’ 24 | ICLR’ 25 | ICLR’ 25 | ICML’ 25 | NeurIPS’ 25 | ours |
| 0% (zero-shot) | - | - | - | - | - | - | - | 67.41 | |
| ACC | 1% | 62.47 | 63.56 | 59.24 | 57.58 | 57.43 | 59.56 | 55.06 | 70.74 |
| 10% | 67.69 | 67.15 | 63.15 | 64.61 | 62.52 | 63.27 | 56.00 | 73.78 | |
| 0% (zero-shot) | - | - | - | - | - | - | - | 74.33 | |
| AUC | 1% | 68.36 | 63.02 | 57.54 | 59.43 | 58.15 | 61.36 | 52.92 | 75.59 |
| Variants | Stage-I | Stage-II | Reasoning | Correct reasoning trajectory | ACC | AUC |
| Stage-I Direct | 57.59 ↓ 9.82 | 63.18 ↓ 11.15 | ||||
| Stage-II Direct | 62.91 ↓ 4.50 | 69.56 ↓ 4.77 | ||||
| Stage-II Intervention | 60.71 ↓ 6.70 | 67.96 ↓ 6.37 | ||||
| Stage-II Reasoning (full BrainCoT) | ✓ | 67.41 | 74.33 |
| Metric | BrainBERT | BIOT | LaBraM | CBraMod | NeuroLM | ELM-MIL | CSBrain | BrainCoT | |
| Full fine-tuning | ACC | 68.98 | 75.28 | 73.43 | 75.78 | 68.06 | 74.64 | 71.15 | \textbf{77.40}_{{\color[rgb]{1,0,0}\uparrow\,1.62}} |
| AUC | 76.46 | 84.12 | 79.11 | 82.82 | 74.34 | 85.95 | 74.52 | \underline{85.01}_{{\color[rgb]{0,1,0}\downarrow\,0.94}} | |
| Linear probing | ACC | 68.55 | 65.44 | 64.05 | 70.51 | 60.39 | 64.36 | 67.37 | \textbf{75.42}_{{\color[rgb]{1,0,0}\uparrow\,4.91}} |
| AUC | 76.55 | 75.33 | 65.01 | 72.82 | 71.00 | 70.62 | 57.33 | \textbf{82.62}_{{\color[rgb]{1,0,0}\uparrow\,6.07}} |
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Task | Dataset | Question templates | Canonical answers |
| Epileptic state recognition | HUP | T1 Which epilepsy state does this signal belong to? T2 Identify the epilepsy state of this recording segment. T3 Classify the epilepsy state for the given neural signal. T4 What epilepsy state best describes this signal? | Interictal; Ictal; Preictal; Postictal |
| Epileptic state recognition | UPenn | T1 Which epilepsy state does this signal belong to? T2 Identify the epilepsy state of this recording segment. T3 Classify the epilepsy state for the given neural signal. T4 What epilepsy state best describes this signal? | Interictal; Ictal |
| Epileptic state recognition | SWEC_ETHZ | T1 Which epilepsy state does this signal belong to? T2 Identify the epilepsy state of this recording segment. T3 Classify the epilepsy state for the given neural signal. T4 What epilepsy state best describes this signal? | Preictal; Ictal; Postictal |
| Interictal epileptiform discharge detection | FNUSA | T1 Does this signal contain interictal epileptiform activity? T2 Is interictal epileptiform activity present in this recording segment? T3 Determine whether this signal shows interictal epileptiform activity. T4 Does this signal show evidence of interictal epileptiform activity? | Non-epileptiform activity; Interictal epileptiform activity |
| Epileptic state recognition | SeiZeIT2 | T1 Which epilepsy state does this signal belong to? T2 Identify the epilepsy state of this recording segment. T3 Classify the epilepsy state for the given neural signal. T4 What epilepsy state best describes this signal? | Interictal; Ictal; Preictal; Postictal |
| Epileptic state recognition | CHB-MIT | T1 Which epilepsy state does this signal belong to? T2 Identify the epilepsy state of this recording segment. T3 Classify the epilepsy state for the given neural signal. T4 What epilepsy state best describes this signal? | Interictal; Ictal; Preictal; Postictal |
| Class | Canonical answer string |
|---|---|
| 0 | "Burnout" |
| 1 | "Subjective memory complaints" |
| 2 | "Normal" |
| 3 | "Dyslexia" |
| 4 | "Chronic pain" |
| 5 | "Major depressive disorder" |
| Class | Canonical answer string |
|---|---|
| 0 | "Focal nonspecific seizure" |
| 1 | "Generalized nonspecific seizure" |
| 2 | "Simple partial seizure" |
| 3 | "Complex partial seizure" |
| 4 | "Absence seizure" |
| 5 | "Tonic seizure" |
| Hyper-parameter | Value |
| Transformer layers | 12 |
| Hidden dimension | 768 |
| FFN dimension | 3072 |
| Attention heads | 12 |
| Key–value heads | 2 |
| RoPE base |
| Dataset | Data type | Subject | Samples | Duration (hours) |
| HUP [ 26 ] | iEEG | 57 | 3,322,242 | 2768.54 |
| UPenn [ 27 ] | iEEG | 12 | 1,691,209 | 469.78 |
| SWEC_ETHZ [ 28 ] | iEEG | 16 | 1,004,999 | 837.50 |
| FNUSA [ 29 ] | iEEG | 12 | 147,030 | 122.53 |
| SHHS [ 12 ] | EEG | 6,441 | 5,789,088 | 48,242.40 |
| SeiZeIT2 [ 13 ] | EEG | 125 | 4,519,290 | 12,553.58 |
| Model | BrainBERT | BIOT | LaBraM | CBraMod | NeuroLM | ELM-MIL | CSBrain | BrainCoT | Zero-shot better | ||
| Dataset | Metric | Training data | ICLR’ 23 | NeurIPS’ 23 | ICLR’ 24 | ICLR’ 25 | ICLR’ 25 | ICML’ 25 | NeurIPS’ 25 | ours | than baselines |
| 0% (zero-shot) | - | - | - | - | - | - | - | 82.77 | |||
| ACC | 1% | 65.45 | 43.27 | 37.95 | 35.09 | 38.60 | 37.01 | 25.68 | 84.07 | 7 / 7 | |
| ISRUC | 10% | 73.80 | 59.60 | 53.36 | 35.23 | 57.50 | 58.61 | 30.47 | 85.26 | 7 / 7 | |
| 0% (zero-shot) | - | - | - | - | - | - | - | 95.19 | |||
| AUC | 1% | 85.48 | 70.47 | 57.96 | 67.47 | 62.73 | 65.58 | 51.32 | 95.30 | 7 / 7 | |
| Variants | Stage-I Direct | Stage-II Direct | Stage-II Intervention | Stage-II Reasoning | ||||
| Datasets | ACC | AUC | ACC | AUC | ACC | AUC | ACC | AUC |
| ISRUC | 83.30 | 94.92 | 81.80 | 94.76 | 75.63 | 90.91 | 82.77 | 95.19 |
| Dreams | 73.27 | 90.10 | 75.77 | 91.38 | 67.83 | 85.85 | 77.66 | 92.49 |
| Siena | 46.69 | 60.13 | 47.45 | 58.99 | 47.77 | 56.58 | 47.31 | 64.85 |
| SanDiego | 50.40 | 57.64 | 56.74 | 59.87 | 58.66 | 58.67 | 62.64 | 63.61 |
| ADHD | 47.95 | 45.32 | 47.72 | 44.13 | 50.92 | 50.68 | 59.13 | 59.44 |
| Dataset | Logistic Regression | KNN | MLP | SVM | Random Forest | XGBoost |
| ISRUC | 91.74 | 85.29 | 90.46 | 92.38 | 91.90 | 92.93 |
| Dreams | 91.16 | 83.67 | 88.94 | 91.15 | 91.15 | 91.59 |
| Siena | 59.88 | 53.43 | 56.85 | 59.11 | 58.31 | 59.32 |
| SanDiego | 60.92 | 63.00 | 64.58 | 64.84 | 63.26 | 67.57 |
| ADHD | 79.11 | 71.07 | 75.73 | 77.21 | 80.31 | 79.57 |
| Schizo | 67.37 | 70.71 | 69.09 | 75.36 | 74.99 | 73.64 |
| Dataset | Metric | BrainBERT | BIOT | LaBraM | CBraMod | NeuroLM | ELM-MIL | CSBrain | BrainCoT |
| ISRUC | ACC | 79.18 | 84.10 | 79.76 | 84.39 | 75.27 | 84.68 | 77.95 | 86.66 |
| AUC | 93.60 | 95.50 | 92.91 | 95.85 | 91.23 | 95.92 | 92.36 | 97.00 | |
| Dreams | ACC | 75.55 | 72.54 | 74.56 | 75.10 | 40.73 | 30.97 | 64.67 | 80.14 |
| AUC | 91.58 | 89.99 | 90.29 | 90.45 | 58.89 | 81.45 | 51.12 | 94.09 | |
| Siena | ACC | 51.63 | 49.12 | 51.94 | 48.13 | 52.68 | 56.63 | 48.46 | 46.78 |
| AUC | 55.98 | 67.39 | 67.29 | 60.41 | 66.14 | 73.84 | 63.33 | 66.08 |
| Dataset | Metric | BrainBERT | BIOT | LaBraM | CBraMod | NeuroLM | ELM-MIL | CSBrain | BrainCoT |
| ISRUC | ACC | 62.78 | 57.14 | 33.87 | 76.46 | 35.48 | 65.04 | 66.39 | 86.08 |
| AUC | 91.80 | 88.22 | 76.09 | 75.88 | 83.79 | 86.14 | 59.83 | 96.69 | |
| Dreams | ACC | 68.56 | 71.59 | 41.75 | 70.24 | 62.82 | 40.93 | 73.11 | 80.69 |
| AUC | 88.43 | 89.39 | 49.66 | 60.78 | 82.78 | 90.40 | 50.34 | 94.08 | |
| Siena | ACC | 52.79 | 50.95 | 36.88 | 50.68 | 50.76 | 49.87 | 52.08 | 49.43 |
| AUC | 61.49 | 60.83 | 59.52 | 59.33 | 63.14 | 56.29 | 55.42 | 65.09 |
EEG Foundation Models
Electroencephalography (EEG) foundation models learn reusable representations of brain signals for transfer across people, tasks, and recording setups.
See all 120 papers