How can interpretable LLM-based semantics predict human brain responses to language?

Context and interpretability were previously a trade-off

The paper starts from a concrete failure of earlier semantic models. Feature-based representations built on cognitive and neuroscience theory, such as the Binder et al. (2016) space, describe a concept along named properties and are directly interpretable, but they assign each word a fixed vector regardless of the sentence it appears in. The paper's own illustration is the word "knife": it may carry harm and negative affect in "he cut himself with a knife" but a more neutral, functional meaning in "he chopped vegetables with a knife". Fixed norms cannot express that shift, and they depend on manual annotation and fixed vocabularies, which limits them to arbitrary natural language.

Distributed representations from large language models solve the context and scale problem but create an attribution problem. Their internal vectors are contextual and predict neural responses during natural language comprehension, yet they entangle semantic, syntactic and other lexical properties, so an alignment with brain activity cannot be cleanly assigned to meaning. SCRIBE is the paper's attempt to keep the interpretable dimensions of the first approach and the contextual coverage of the second.

LLMs act as annotators, not as the representation

SCRIBE does not use an LLM's internal activations. Instead, it uses an ensemble of open-source LLMs as annotators inside the semantic feature space of Binder et al. (2016), a space of 65 properties spanning perceptual dimensions (vision, audition, somatosensation, taste, smell), motor effectors, spatial and temporal properties, and abstract social, cognitive and affective features. Each property corresponds to a described aspect of meaning, for example something you can easily see, something associated with pain, or something that makes one feel happy.

For a given target concept and its surrounding linguistic context, the models are prompted separately for each property and asked to return a single number from 0 (not at all associated or not applicable) to 6 (very strongly associated). The property-wise scores are then aggregated across models to form the final representation, which is fed to a brain-encoding model that predicts fMRI responses. Because every output dimension is a named property, a reader can inspect which property drove a prediction; because the prompt includes the sentence, the same word can receive different scores in different contexts. The paper notes that three bipolar properties in the original space (Temperature, Texture, Weight) are queried as their two poles and recombined by taking the maximum, yielding the final 65 properties.

The scores track human ratings above chance

The first validation asks whether the generated scores behave like human semantic judgments for isolated, out-of-context concepts, using the existing human norms from Binder et al. (2016). A permutation test shuffles concept–property assignments 1,000 times while preserving the marginal distributions of annotations, producing a null distribution for each metric. Across all models the real annotations separate clearly from the permuted ones, with lower mean squared error and higher Pearson correlation. For the ensemble, the real mean MSE is 1.07 against a permuted mean of 3.88, and the real mean Pearson correlation is 0.77 against a permuted 0.12. Individual models range from 0.61 (mistral-7b) to 0.74 (llama3.1-70b) in real correlation, and the ensemble is the strongest, which the paper reads as support for aggregating across models in the way one aggregates human judgments.

A new human dataset shows judgments shift with context

The more demanding test is contextual. The authors collected a new human-annotated resource of context-dependent semantic feature norms, which they describe as the first of its kind and plan to release. They selected 54 of the 57 properties used in a synthetic dataset, excluding Temperature, Weight and Texture because those are derived from pairs of opposing dimensions that would both need to be evaluated within the same sentence. For each retained property they chose six concepts stratified by original out-of-context rating strength (two low, two mid-range, two high), giving 324 concept–property pairs. Each pair appeared in three conditions: a high-target context designed to increase the association, a low-target context designed to decrease it, and an out-of-context condition, for 972 unique items.

Participants were recruited through Prolific with English fluency, a 98–100% approval rate and at least 100 previous submissions, paid USD 2 for a roughly 10-minute study. After quality control, including exclusion for failed attention checks, mean response times below 7 seconds per item and uniform responding, and removal of one pair shown with an incorrect high-target sentence, the final sample is 921 participants and 11,983 ratings, averaging 12.4 ratings per item. Human ratings were highest in the high-target condition (M = 4.23, SD = 1.07), confirming that the controlled contexts systematically moved judgments along the targeted dimensions. SCRIBE captured these shifts, reaching 75–81% of the estimated human noise ceiling, which the paper compares to the agreement between two independently collected sets of human ratings.

Context adds predictive power in language regions

The neural test uses brain activity recorded while participants listen to naturalistic stories. Here the relevant comparison is internal to SCRIBE: its contextual representation against its own out-of-context representation. Incorporating context significantly improves prediction of fMRI responses, particularly in language-related cortical regions. This is a within-model comparison, so it speaks to whether the contextual scores carry information the static scores lack, not to whether SCRIBE beats every alternative representation.

The paper also compares SCRIBE with the closest existing method, QAEmb, which likewise uses LLMs as annotators to turn stimuli into explicit, cognitively motivated, interpretable word-level features for brain encoding. The difference is what the features are designed to represent: QAEmb spans neuroscience-motivated linguistic properties generally, while SCRIBE is built specifically to represent conceptual semantics within the independently defined Binder et al. space and to estimate how those properties vary with context. In the reported comparison, contextual SCRIBE explains more unique neural variance than QAEmb, with significant differences in semantic regions such as lateral temporal cortex and angular gyrus.

Unique variance survives a syntax control

A contextual representation can predict brain activity for the wrong reason, so the paper adds a variance-partitioning analysis with independently defined syntactic features. The result is a split verdict that matters for interpretation. Contextual Llama2-7B embeddings achieved higher overall brain-encoding performance than SCRIBE (0.2387 ± 0.0215 versus 0.1837 ± 0.0144), so raw predictive accuracy favors the uninterpretable embedding. But higher accuracy does not establish which linguistic information drives the advantage.

Variance partitioning shows SCRIBE shares significantly less variance with the syntactic feature space than Llama2-7B in inferior frontal and middle frontal gyri, with the same numerical direction in late language and lateral temporal cortex. This is notable because left inferior frontal regions are associated with syntactic processing. Permutation tests confirmed that the variance uniquely explained by SCRIBE exceeds chance across all four language-related regions of interest (late language, lateral temporal cortex, IFG/MFG and angular gyrus; all p ≤ .002). The paper's reading is that SCRIBE's representations capture aspects of linguistic meaning not accounted for by the measured syntactic features, while the higher-scoring Llama2-7B embedding remains harder to attribute.

Relative ordering holds, absolute scores need calibration

The limitations are specific and worth carrying into any use of the framework. SCRIBE captures the relative ordering of human judgments, but its scores are systematically lower than human ratings on average, so applications that depend on absolute values may require calibration. Individual human judgments of context-dependent properties are noisy, and the conclusions hold at the level of averaged ratings rather than individual annotators. The contextual stimuli are controlled, LLM-generated sentences designed to shift specific properties; validating SCRIBE against human judgments of naturalistic text remains an important next step. The fixed semantic space also constrains representations to a predefined set of dimensions and may miss aspects of meaning, a trade-off the authors accept for a principled, interpretable basis comparable with prior work. Prompt ablations showed performance largely robust across tested variations, and the ensemble correlated more strongly with human ratings than any individual model.

Interpretability changes which model looks better

The paper's most decision-relevant result is not a single accuracy number but a change in how competing representations should be compared. If the only criterion is whole-brain encoding correlation, the contextual Llama2-7B embedding wins (0.2387 versus 0.1837 for SCRIBE). If the criterion is whether the brain-predictive signal can be attributed to meaning rather than syntax, SCRIBE's advantage appears: significantly less shared variance with syntactic features in IFG/MFG and significant unique variance in all four language-related regions. For a reader choosing a representation to study semantics in the brain, that distinction sets the priority: use SCRIBE when the question requires named, context-varying semantic dimensions and defensible attribution, and treat its lower absolute scores and uncalibrated scale as properties to correct for rather than as evidence against the approach.

Sources used in this answer