Research map / Architectures and efficiency / RNNs and normalization
Abstraction Induces the Brain Alignment of Language and Speech Models
Emily Cheng, Aditya R. Vaidya, Richard Antonello
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use intrinsic dimension instead of surprisal to explain layerwise encoding | layerwise brain encoding of LLMs and speech models · fMRI, all models | Spearman rho with encoding performance: -0.53 -> 0.72 | Improved |
| use intrinsic dimension instead of surprisal to explain layerwise encoding | layerwise brain encoding · ECoG, all models | Spearman rho with encoding performance: 0.21 -> 0.43 | Improved |
| add I_d to a mixed-effects model with surprisal (ablation) | mixed-effects model of encoding performance · fMRI | effect size: beta_surprisal approx 0 vs beta_Id = 0.63 | Improved |
| add I_d to a mixed-effects model with surprisal (ablation) | mixed-effects model of encoding performance · ECoG | effect size: beta_surprisal = 0.26 vs beta_Id = 0.97 | Improved |
| fine-tune WavLM-base-plus layer 9 on fMRI responses | WavLM-base-plus · fMRI | encoding performance, semantic content and I_d | Improved |
| replace trained model layers with random Fourier features of matched I_d (ablation) | random features · fMRI | encoding performance: R approx 0.04 vs R > 0.1 | Hurt |
| train Pythia-6.9b longer (more checkpoints) | Pythia-6.9b · fMRI | correlation of I_d with encoding performance: rho = 0.96 | Improved |
| train Pythia-6.9b longer (more checkpoints) | Pythia-6.9b · ECoG | correlation of I_d with encoding performance: rho = 0.64 | Improved |
| change GRIDE scale k by factor 2 (ablation) | I_d estimator · fMRI/ECoG | I_d peak layer and correlation: Pythia-6.9b peak layer 13 unchanged | No clear effect |
| use linear dimension (PR) instead of nonlinear I_d (ablation) | dimensionality measure · fMRI, Pythia-6.9b | correlation with encoding performance: 0.86 (PCA) / -0.05 (PR) | Hurt |
Closest papers by meaning
- The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities (ICLR 2025 · similarity 0.90)
- Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement (NeurIPS 2025 · similarity 0.89)
- The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.89)
- Improving Semantic Understanding in Speech Language Models via Brain-tuning (ICLR 2025 · similarity 0.89)
- Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models (NeurIPS 2025 · similarity 0.89)
- fMRI predictors based on language models of increasing complexity recover brain left lateralization (NeurIPS 2024 · similarity 0.88)
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.
