Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

RABBiT: Rapidly adaptive BOLD foundation model via brain-tuning for accurate zero-shot and few-shot prediction of speech-elicited responses in the brain

NeurIPS 2026 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Zero-shot robustness and transfer · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RABBiT, a compact brain-tuned audio-to-fMRI encoder (Wav2Vec2.0 with LoRA, temporal brain transformer, and shared-idiosyncratic decomposition) for zero-shot group-level and few-shot subject-level prediction of speech-evoked fMRI.

Problem

Brain encoding models for language fMRI either need participant-specific voxel-wise fitting or fail to generalize zero-shot to unseen listeners and adapt efficiently to new participants at high cortical resolution.

Approach

A single audio-to-fMRI encoder that combines LoRA brain-tuning of a speech backbone, learned ROI-query cross-attention over speech tokens, and a low-rank decomposition of ROI responses into shared and subject-specific components, so few-shot adaptation only updates a small deviation pathway.

Key results

• Zero-shot RABBiT surpasses leave-one-out inter-subject consistency on both held-out cohorts (324 unseen participants) and substantially outperforms TRIBEv2 and linear baselines in auditory and temporal ROIs.
• Removing brain-tuning drops zero-shot performance by roughly 60%.
• Few-shot RABBiT improves over zero-shot with 5 minutes of calibration and outperforms voxel-wise ridge baselines at 5 to 40 minutes while updating ~115K parameters versus ~200M ridge parameters.

Surprising findings

• Mixing the Moth Radio Hour dataset into training slightly reduced zero-shot performance (story-mean r 0.108 vs 0.117 for Friends-only), so more data did not help.
• RABBiT's group correlation in A1 (~0.55) exceeds the LOO ISC (~0.27), which the authors attribute to noiseless predictions versus noisy cohort-mean targets.

Limitations

• Evaluated primarily on English naturalistic listening and auditory/language regions.
• Training cohort is small (6 participants) and mixing datasets did not yield gains.
• Only partial exploration of larger speech backbones.

Experimental setup

Datasets: CourtoisNeuroMod Friends Narratives Le Petit Prince Moth Radio Hour

Benchmarks: Narratives Le Petit Prince

Metrics: group-level Pearson correlation inter-subject correlation contrast t-statistic

Baselines: TRIBEv2 Linear baseline Per-Subject Full-Rank head voxel-wise ridge (pretrained features) voxel-wise ridge (brain-tuned features) No SID RABBiT

Builds on: Wav2Vec2.0 LoRA TRIBEv2 voxel-wise ridge encoding Universal Brain Encoder

Closest papers by meaning

• Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models (NeurIPS 2025 · similarity 0.87)
• Improving Semantic Understanding in Speech Language Models via Brain-tuning (ICLR 2025 · similarity 0.87)
• BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.86)
• fMRI predictors based on language models of increasing complexity recover brain left lateralization (NeurIPS 2024 · similarity 0.86)
• MIRAGE: Adaptive Multimodal Gating for Whole-Brain fMRI Encoding (NeurIPS 2026 · similarity 0.86)
• The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning (ICML 2025 · similarity 0.86)