| name | llm-semantic-convergence-human-neural-representations |
| description | LLM-Human neural semantic convergence methodology - dimension-resolved interbrain encoding modeling comparing LLM-derived and human-shared neural semantic representations across 10 semantic dimensions. Use when: LLM brain alignment, semantic representation analysis, interbrain synchronization, dimensional semantic space, neural encoding modeling, shared semantics, human-LLM comparison, MEG encoding analysis. Activation: LLM convergence, semantic alignment, neural representation, interbrain encoding, semantic dimensions, brain-LLM comparison, shared neural semantics |
| metadata | {"arxiv_id":"2606.11598","published":"2026-06-10","authors":["Chen Hong","Ximing Shao","Gangyi Feng"],"tags":["neuroscience","llm","semantic","representation","alignment","meg","interbrain","encoding","convergence"]} |
| license | Complete terms in LICENSE.txt |
LLM Semantic Convergence with Human Neural Representations
arXiv: 2606.11598 | Published: 2026-06-10 | Categories: q-bio.NC
Context
Interpersonal communication requires building shared semantics that enable listeners to understand speakers' meanings. LLMs increasingly approximate human language capability and neural responses. This research addresses: Do LLMs capture the same semantic structure shared between human brains?
Key Question: Are LLM semantic representations converging with human-shared neural semantics, and if so, is this convergence selective or complete?
Core Methodology
1. Ten-Dimensional Semantic Space Framework
Semantic dimensions rated for each content word:
| Dimension | Description | LLM Alignment Status |
|---|
| Perception | Sensory experiences | Moderate alignment |
| Motor | Action-related | Moderate alignment |
| Space | Spatial relations | High alignment |
| Time | Temporal concepts | High alignment |
| Socialness | Social interactions | Partial divergence |
| Animacy | Living vs non-living | Moderate alignment |
| Emotion | Emotional content | Partial divergence |
| Attention | Attentional focus | Moderate alignment |
| Causality | Cause-effect relations | Moderate alignment |
| Drive | Motivational states | Partial divergence |
2. Dimension-Resolved Interbrain Encoding Modeling
Procedure:
for word in narrative_content_words:
human_ratings[word] = [perception, motor, space, time,
socialness, animacy, emotion, attention,
causality, drive]
llm_ratings[word] = [LLM_rate(word, dim) for dim in semantic_dimensions]
for dimension in semantic_dimensions:
ns_model = fit_encoding(
speaker_semantics[dimension],
listener_meg_response,
baseline_features=[acoustic, phonological]
)
r2_dimension[dimension] = ns_model.score()
geometry_alignment = compare_semantic_geometry(
human_semantic_space,
llm_semantic_space,
metrics=[RSA, CKA, procrustes_distance]
)
for model_size in [small, medium, large, largest]:
alignment[model_size] = measure_convergence(
model_size,
human_neural_semantics,
dimensions=semantic_dimensions
)
3. Individual Differences Prediction
Link neural alignment to story comprehension:
comprehension_score = predict_from_alignment(
neural_synchronization_strength,
listener_behavioral_measures
)
Key Findings
Finding 1: Multidimensional Structure (Not Single Global Signal)
Shared semantics is characterized as multidimensional neural structure, not a single global semantic signal. Different dimensions contribute differentially to neural synchronization.
Finding 2: LLM Dimension-Dependent Alignment
- High alignment dimensions: Space, Time (abstract, structural)
- Moderate alignment: Perception, Motor, Animacy, Attention, Causality
- Partial divergence dimensions: Socialness, Emotion, Drive (agency, affect, social experience)
Pattern: Dimensions closely tied to agency, affect, and social experience show largest divergences.
Finding 3: Model Scaling Improves Approximation
Larger LLMs → closer alignment with human semantic structure:
- Greater overlap in semantic geometry
- Higher NS prediction accuracy
- BUT: incomplete convergence even for largest models
- Convergence remains selective, not complete
Finding 4: Comprehension Prediction
Neural alignment patterns predict individual differences in listeners' story comprehension, linking neural synchronization to cognitive performance.
Implementation Steps
Step 1: Define Semantic Dimensions
semantic_dimensions = {
'perception': 'Sensory experiences (visual, auditory, tactile)',
'motor': 'Action-related concepts',
'space': 'Spatial relations and locations',
'time': 'Temporal concepts and sequences',
'socialness': 'Social interactions and relationships',
'animacy': 'Living vs non-living entities',
'emotion': 'Emotional content and valence',
'attention': 'Attentional focus and salience',
'causality': 'Cause-effect relationships',
'drive': 'Motivational states and goals'
}
Step 2: Dimension Rating Protocol
def rate_semantic_dimension(word, dimension, rater_type='human'):
"""
Rate word on semantic dimension.
Args:
word: Content word from narrative
dimension: One of 10 semantic dimensions
rater_type: 'human' or 'llm'
Returns:
rating: Continuous score (e.g., 0-10)
"""
if rater_type == 'human':
return human_rating_process(word, dimension)
else:
prompt = f"Rate '{word}' on {dimension} dimension (0-10)"
return llm_generate_rating(prompt)
Step 3: Interbrain Encoding Analysis
def dimension_resolved_encoding(speaker_data, listener_meg):
"""
Test dimension contribution to neural synchronization.
Args:
speaker_data: Semantic dimension ratings
listener_meg: Listener MEG responses
Returns:
dimension_r2: Dict of dimension → explained variance
"""
results = {}
baseline_features = extract_baseline(speaker_audio)
for dimension in semantic_dimensions:
full_model = fit_encoding(
features=[baseline_features, speaker_data[dimension]],
response=listener_meg
)
baseline_model = fit_encoding(
features=baseline_features,
response=listener_meg
)
results[dimension] = full_model.r2 - baseline_model.r2
return results
Step 4: Representational Geometry Comparison
def compare_semantic_geometry(human_space, llm_space):
"""
Compare semantic space geometry between human and LLM.
Methods:
- RSA (Representational Similarity Analysis)
- CKA (Centered Kernel Alignment)
- Procrustes distance
"""
rsa_score = compute_rsa(human_space, llm_space)
cka_score = compute_cka(human_space, llm_space)
procrustes_dist = compute_procrustes(human_space, llm_space)
return {
'rsa': rsa_score,
'cka': cka_score,
'procrustes': procrustes_dist
}
Pitfalls
Pitfall 1: Single Global Signal Assumption
Error: Treating semantic alignment as single global measure.
Fix: Use dimension-resolved analysis. Test each dimension's contribution separately beyond baseline features.
Pitfall 2: Overlooking Dimension-Dependent Divergence
Error: Assuming complete LLM-human alignment from overall scores.
Fix: Check dimension-specific divergences, especially social/affective dimensions. These show largest gaps even for large models.
Pitfall 3: Ignoring Model Scaling Effects
Error: Testing only one model size.
Fix: Compare multiple model sizes. Larger models improve alignment but remain incomplete.
Pitfall 4: Acoustic/Phonological Confound
Error: Not controlling for acoustic and phonological features.
Fix: Include baseline model with acoustic/phonological features. Measure unique semantic dimension contribution.
Pitfall 5: Individual Differences Ignored
Error: Focusing only on group-level alignment.
Fix: Link neural alignment patterns to individual comprehension performance.
Verification
Verification 1: Dimension Independence
correlation_matrix = compute_dimension_correlations(ratings)
Verification 2: NS Prediction Significance
for dimension in dimensions:
p_value = test_significance(
full_model_score,
baseline_model_score
)
Verification 3: Geometry Alignment Scaling
alignment_scores = []
for model in models_increasing_size:
score = measure_alignment(model, human_space)
alignment_scores.append(score)
assert is_monotonic_increasing(alignment_scores)
Verification 4: Comprehension Link
correlation = correlate(neural_alignment, comprehension_scores)
References
- Paper: Hong, C., Shao, X., Feng, G. (2026). Large language models selectively converge with human-shared neural semantic representations. arXiv:2606.11598
- Method: Pseudo-hyperscanning MEG + dimension-resolved interbrain encoding
- Semantic Dimensions: 10-dimension framework (perception, motor, space, time, socialness, animacy, emotion, attention, causality, drive)
- Related Skills: See
brain-llm-alignment, semantic-representation-analysis, neural-encoding-modeling
Activation Keywords
LLM convergence
semantic alignment
neural representation
interbrain encoding
semantic dimensions
brain-LLM comparison
shared neural semantics
dimension-resolved encoding
MEG semantic analysis
LLM human brain alignment