| name | heterogeneous-neural-predictivity-lm |
| description | Analysis framework for evaluating language model neural predictivity during naturalistic comprehension. Identifies heterogeneous brain-language alignment patterns across participants and brain regions, separating predictive usefulness from shared neural organization claims (arXiv: 2606.26880) |
| metadata | {"arxiv_id":"2606.26880","authors":["Xiao Jia"],"published":"2026-06-25","source_url":"https://arxiv.org/abs/2606.26880","tags":["neuroscience","language models","brain-language alignment","naturalistic comprehension","neural predictivity","EEG","MEG","ECoG"],"categories":["cs.CL","q-bio.NC","cs.LG"],"methodology":"Multi-dataset frozen LM analysis with controlled feature ablations and participant-level matching","datasets":["Brain Treebank","MEG-MASC","Podcast ECoG"],"models_tested":["8 frozen language models","blocked encoding models"],"key_findings":"67/432 evaluable rows met controlled predictive criterion; participant-level advantages localized not uniform"} |
Heterogeneous Neural Predictivity from Language Models
Core Contribution
Demonstrates that language model representations can annotate neural activity during natural speech/text comprehension, but with critical caveats about interpretation.
Methodology
Multi-Dataset Framework
- Analyzed locked derived data from 3 datasets: Brain Treebank, MEG-MASC, Podcast ECoG
- Used 8 frozen language models with matched controls
- Controlled for temporal, nuisance, and representation-capacity confounds
Evaluation Strategy
- Source-level summaries for held-out prediction
- Comparison against low-level baselines
- Feature ablations to test component-level sensitivity
Control Conditions
- Brain-derived timing-linked controls
- Acoustic controls
- Implanted-signal controls
- Matched-capacity controls
Key Findings
Predictive Success
- 67 of 432 evaluable rows met controlled predictive-only criterion
- Positive held-out prediction widespread in source-level summaries
- Model-side feature ablations changed prediction scores in most evaluable rows
Heterogeneity Pattern
- Participant-level matched-control advantages were localized, not uniform
- Response-profile and feature-specificity contrasts bounded interpretations
- Complete co-indexed integrated interpretation requires future jointly indexed coverage
Critical Distinction
Separates two claims:
- Predictive usefulness: LM features can predict neural activity ✓
- Shared organization: LM features reveal neural computation mechanisms ✗ (not proven)
The framework explicitly distinguishes between using LMs as neural predictors vs. claiming they share neural computational principles.
Implementation Guidance
When to Use
- Evaluating LM neural predictivity in naturalistic settings
- Designing controlled brain-language alignment studies
- Interpreting heterogeneous alignment patterns across participants
Methodology Steps
- Select multiple frozen LMs across architecture families
- Establish matched temporal and capacity controls
- Test held-out prediction against low-level baselines
- Perform feature ablations to verify component sensitivity
- Analyze participant-level heterogeneity (don't assume uniform effects)
- Separate predictive utility from mechanistic claims
Pitfalls
- Over-interpretation: Predictive success ≠ shared computational mechanism
- Uniformity assumption: Effects are localized, not uniform across participants
- Insufficient controls: Must control for temporal, acoustic, and capacity confounds
- Single-dataset claims: Multi-dataset validation required for robust conclusions
Applications
- Naturalistic neuroscience experimental design
- Brain-computer interface feature selection
- Language model evaluation for cognitive modeling
- Cross-participant neural pattern analysis
Related Concepts
- Brain-language alignment
- Neural encoding models
- Naturalistic stimulus processing
- Frozen feature analysis
- Participant heterogeneity