Brain Passage Retrieval (BPR)

• Brain Passage Retrieval (BPR) is a neural IR approach that converts EEG signals into direct document retrieval actions without explicit query formulation.
• It employs dual-encoder architectures with text and EEG encoders using strategies like [CLS] pooling and contrastive loss to map both modalities to a shared semantic space.
• BPR supports accessible, hands-free search interfaces for users with impairments and shows marked improvements in cross-sensory and multimodal retrieval performance.

Brain Passage Retrieval (BPR) is a neural information retrieval paradigm addressing the challenge of transforming internal cognitive information needs directly into document retrieval actions via electroencephalogram (EEG) signals. BPR eliminates the intermediate textual formulation of queries, mapping EEG representations—arising from reading or listening—to a shared semantic space with text passages, enabling direct brain-to-text IR. BPR holds particular significance for accessible information systems, offering search modalities for users with communication or physical impairments and supporting interfaces beyond visual text input (McGuire et al., 20 Jan 2026, McGuire et al., 2024).

1. Task Definition and Core Formalism

BPR formalizes a retrieval scenario in which a passage $D = \{w_1, w_2, \dots, w_L\}$ has a simultaneously recorded EEG segment $X \in \mathbb{R}^{L \times D}$, with $C$ channels and $T$ time samples per word, $D = C \cdot T$. For a corpus $\mathcal{C} = \{p_1, \dots, p_N\}$ and an EEG query $q \subset X$, the retrieval objective is to rank corpus passages so that the “positive” (corresponding) passage achieves maximal rank: $\text{score}(q, p)$ The top-$k$ subset

$$\{p \in \mathcal{C} \mid \text{rank}(\text{score}(q, p)) \leq k\}$$

is evaluated by standard Information Retrieval (IR) metrics, including Mean Reciprocal Rank (MRR) and Hit@$X \in \mathbb{R}^{L \times D}$0 (McGuire et al., 20 Jan 2026, McGuire et al., 2024).

2. Model Architectures and Representation Learning

BPR predominantly employs a dual-encoder (bi-encoder) paradigm. The text encoder $X \in \mathbb{R}^{L \times D}$1 maps tokenized passages to $X \in \mathbb{R}^{L \times D}$2-dimensional embeddings (e.g., via BERT-base, typically frozen during training). The EEG encoder $X \in \mathbb{R}^{L \times D}$3 projects word-wise EEG sequences to the same space:

• EEG input preprocessing: word-aligned EEG segments $X \in \mathbb{R}^{L \times D}$4 are flattened.
• Linear projection: $X \in \mathbb{R}^{L \times D}$5, $X \in \mathbb{R}^{L \times D}$6, $X \in \mathbb{R}^{L \times D}$7.
• Transformer layers: $X \in \mathbb{R}^{L \times D}$8 stacked transformers (2–6 layers) contextualize $X \in \mathbb{R}^{L \times D}$9 to $C$0, capturing temporal and spectral structure (McGuire et al., 2024).
• Pooling: Embedding aggregation leverages [CLS] token, mean, max, or multi-vector pooling, with [CLS]-based pooling showing consistently superior results for EEG inputs.

Similarity is evaluated as cosine distance, with InfoNCE contrastive loss enforcing high similarity between EEG-query–passage pairs. Training uses negative sampling strategies (in-batch, subject-aware, BM25 hard negatives), with hard negative mining shown to yield significant gains (e.g., +51.6% P@1) (McGuire et al., 2024).

3. Modalities, Datasets, and Cross-Sensory Training

Early BPR studies focused exclusively on visual (reading-induced) EEG. Auditory BPR, using EEG collected during listening, expands the paradigm to non-visual input. Key datasets:

• Alice (auditory modality): Participants listen to "Alice’s Adventures in Wonderland" (32,469 words, 1,200 queries, unique tokens ≈ 543).
• Nieuwland (visual modality): Subjects read narrative text word-by-word (28,989 words, 1,200 queries, unique tokens ≈ 674).

Data are split (80/10/10) for training/development/testing. The Inverse Cloze Task (ICT) with 30% passage masking and a 90% masking probability is used to create retrieval pairs, enforcing semantic (rather than surface lexical) matching. Cross-sensory training merges the Alice and Nieuwland datasets, despite low lexical overlap (Jaccard ≈ 0.18) (McGuire et al., 20 Jan 2026).

4. Empirical Results and Comparative Analysis

Core Test Metrics (CLS pooling, mean ± std):

Setting	MRR	Hit@1	Hit@10
BM25 (audio)	0.428	0.342	0.542
Neural BPR (audio)	0.362	0.220	0.668
Neural BPR (visual)	0.139	0.074	0.262
Neural BPR (combined)	0.474	0.314	0.858
Neural BPR (combined, visual eval)	0.256	0.141	0.515

Auditory EEG consistently outperforms visual EEG in stand-alone settings (MRR: $C$1 vs $C$2; Hit@1: $C$3 vs $C$4), with a 2.6× advantage in MRR and 3× in Hit@1. Cross-sensory (combined) training yields marked improvements: auditory MRR $C$5 ($C$6), Hit@1 $C$7, Hit@10 $C$8; similar relative gains are observed for visual evaluation. Critically, the combined auditory model surpasses the BM25 baseline on the same test set (MRR: $C$9 vs $T$0) (McGuire et al., 20 Jan 2026).

EEG-based retrieval exhibits higher robustness under increased text masking than text-only baselines, indicating learned deep semantic alignment rather than shallow lexical matching. In the reading-only ZuCo dataset, end-to-end BPR (DEEPER) outperforms EEG-to-text translation systems by several hundred percent on precision metrics and produces tight EEG–text embedding clusters (avg. cosine similarity matched pairs: $T$1) (McGuire et al., 2024).

5. Pooling Strategies and Modeling Insights

Four aggregation strategies are explored:

• [CLS] pooling: Learnable, yields bidirectional improvements in cross-modal and cross-sensory settings; preferred for robust sequence-level EEG summarization.
• Mean pooling / Max pooling: Provide weaker or asymmetric gains, especially in combined-modality settings.
• Multi-vector (ColBERT style): Retains all $T$2 vectors; enables late interaction and may capture fine-grained alignment but has not yet yielded superior results for BPR.
• This suggests [CLS] pooling is a crucial design axis for robust cross-sensory BPR.

Contrastive loss with hard negative mining enforces granularity in neural passage discrimination, which ablation studies show as critical for closing the EEG–text modality gap.

6. Practical Implications and Accessibility

BPR demonstrates that continuous EEG signals—both auditory and visual—enable effective passage retrieval, obviating the need for intermediate decoding into explicit queries. The technology supports:

• Hands-free, vision-free search interfaces for visually impaired (WHO: ~285M) or physically impaired users.
• Integration with conversational interfaces and applications such as podcast or audiobook search, where no text input is available.
• Data scarcity mitigation via cross-sensory training, leveraging heterogeneous EEG corpora for improved generalization.

A plausible implication is that BPR establishes neural queries as competitive with classical IR even under severe data scarcity and low lexical overlap (McGuire et al., 20 Jan 2026).

7. Open Challenges and Future Directions

Limitations include:

• Under-utilization of rich EEG sequence dynamics in single-vector pooling; multi-vector late interaction models are proposed as future solutions.
• Reliance on ICT proxy tasks rather than human-generated relevance signals.
• Dataset sizes lagging behind text-IR benchmarks, constraining generalizability.

Proposed future work encompasses:

• Development and evaluation of late-interaction models (e.g., ColBERT style) for enhanced brain–text alignment.
• Pre-training of EEG encoders on larger, unlabeled neural datasets.
• Creation of dedicated neural IR resources with explicit relevance judgements and diverse, enlarged subject pools.
• Extension to other neural modalities (imagined speech, cross-lingual).
• Deployment in accessible information systems and hands-free, intent-driven search interfaces (McGuire et al., 2024).

BPR thus establishes a robust neural IR interface, providing a foundation for semantically aware search capabilities rooted in electrophysiological brain activity rather than explicit linguistic formulations.