AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

BIT-LLM: Brain Instruction Tuned LLM with persistent Cross-Attention for fMRI-to-Text Decoding

Sunghwan Lee, Jihun Kim, Chae Lynn Kim, Ji Yun Park, Jong-Hwan Lee

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Calibration and knowledge in LLM tuning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Stage 1 multimodal contrastive pretraining (ablation)BIT-LLM · NSD S8 (ablation)CIDEr: 6.42 -> 26.64Improved
add Stage 2->3 GRPO reward fine-tuning (ablation)BIT-LLM · NSD S8 (ablation)CIDEr: 26.64 -> 34.90Improved
add Stage 3 RFT (ablation)BIT-LLM · NSD S8 (ablation)BLEU-4: 9.57 -> 15.10Improved
add Stage 3 RFT (ablation)BIT-LLM · NSD S8 (ablation)METEOR: 15.76 -> 15.69No clear effect
use persistent cross-attention instead of prefix conditioning (ablation)BIT-LLM · NSD S8 (ablation)CIDEr: 24.80 -> 26.64Improved
use cross-attention instead of parameter-matched prefix MLP (ablation)BIT-LLM · NSD S8 (ablation)CIDEr: 23.98 -> 26.64Improved
use cross-attention instead of prefix under zeroed voxels (ablation)BIT-LLM · NSD S8 (ablation) zeroed fMRICIDEr: 0.02 -> 5.98Improved
shuffle voxel values at inference (ablation)BIT-LLM (Stage 1 encoder) · NSD S8 (ablation)CIDEr: 26.64 -> 4.46Hurt
use Stage 1 encoder instead of official MindLLM encoder (ablation)BIT-LLM · NSD S8 (ablation)CIDEr: 9.01 -> 26.64Improved
insert 7 uniform adapters instead of 4 (ablation)cross-attention adapters · NSD S8 (ablation)CIDEr: 26.45 -> 25.86No clear effect
insert 14 uniform adapters instead of 4 (ablation)cross-attention adapters · NSD S8 (ablation)CIDEr: 26.45 -> 25.59Hurt
early-biased instead of late-biased placement of 4 adapters (ablation)cross-attention adapters · NSD S8 (ablation)CIDEr: 26.47 -> 26.64No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy