Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in language models and post-stroke aphasia

Abstract: Mechanistic interpretability of LLMs lacks spatially resolved, falsifiable tools for testing whether internal components are specialized for distinct cognitive operations. We adapt subtraction analysis, the standard framework of human neuroimaging, from biological brains to perturbed transformers, and apply the same logic to both substrates in parallel. Building on the Brain-LLM Unified Model (BLUM), which showed that layer-perturbed LLaVA-1.6-Vicuna-13B error profiles match the lesion patterns of aphasic patients, we develop PRISM (Perturbation-based Regional Interpretability through Subtraction Mapping). PRISM maps the seven clinical Philadelphia Naming Test categories, subtracts error classes pairwise, and treats each perturbation seed as a subject in a group analysis with threshold-free cluster enhancement along the layer axis. We run a structurally matched analysis on 213 chronic post-stroke aphasia patients using correlation-difference lesion-symptom mapping, and replicate both sides on held-out splits. The designs match in subject dimension (seeds, patients), spatial dimension (layers, atlas-parcellated cortex) and thresholding, but the contrast operator differs: a within-subject error-proportion difference for the LLM, a between-subject correlation difference for the cortex. Both substrates recover a robust phonemic-favoring dissociation, a deep layer cluster and a frontal-perisylvian cortical cluster, both replicating; the semantic-favoring direction is a consistently signed but non-significant trend on both. PRISM thus gives a falsifiable, spatially resolved test of functional-specialization claims in transformer LLMs. A confirmatory ROI-level intervention (PRISM Stage 3) licensing the strongest causal-mechanism claim is left to subsequent work.

• The paper introduces PRISM, a phenotype-first method that adapts lesion-symptom mapping and subtraction analysis to transformer perturbations, using PNT naming-error categories and discovery-validation splits across 80 seeds.
• PRISM identified a reproducible phonemic-favoring cluster around layers 22–31 in LLaVA-1.6-Vicuna-13B, with layer-order permutation tests reaching empirical p<0.001, while semantic-favoring effects remained subthreshold.
• The paper found a parallel phonemic-favoring dissociation in 213 people with post-stroke aphasia, including postcentral, precentral, and superior longitudinal fasciculus regions, while moderate perturbations produced stronger selectivity than extreme damage.

Perturbation-Based Regional Interpretability and Naming-Error Dissociations

Research Problem and Conceptual Framework

“Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in LLMs and post-stroke aphasia” (2608.12717) proposes a behaviorally grounded framework for mechanistic interpretability that adapts subtraction analysis and lesion-symptom mapping from cognitive neuroscience to transformer models. The central methodological problem is that many existing interpretability techniques—probing, activation patching, causal tracing, circuit discovery, and sparse-autoencoder decomposition—identify representational or computational structure without necessarily establishing that a component is causally necessary for a behaviorally defined operation. PRISM addresses this limitation by starting from a clinically validated behavioral taxonomy and asking which spatially organized model components are preferentially implicated in distinct failure modes.

The paper treats perturbations to transformer layers as an analogue of focal neural lesions. It then compares error categories produced by a perturbed model with error categories observed in patients with chronic post-stroke aphasia performing the same picture-naming task. The relevant behavioral instrument is the 175-item Philadelphia Naming Test (PNT), whose taxonomy distinguishes semantic substitutions, phonemic errors, mixed errors, neologisms, unrelated responses, and no responses. Rather than contrasting each error category against correct performance, PRISM performs pairwise subtractions among error classes. This design is intended to isolate category-selective effects while removing variance associated with general impairment severity.

The paper’s broader theoretical claim is deliberately limited but important: transformer layers can be evaluated using the same inferential logic used to identify functionally dissociable cortical territories, provided that the artificial and biological analyses are matched at the level of behavioral phenotype, spatial organization, replication, and multiple-comparison control. The analogy is not exact. A transformer layer is not a cortical region, and a weight perturbation is not a biological lesion. Nevertheless, both systems generate spatially indexed perturbation-response maps that can support falsifiable claims about differential functional necessity.

PRISM Methodology

The LLM analysis uses LLaVA-1.6-Vicuna-13B, a 40-layer, 13-billion-parameter vision-language transformer. The model was evaluated on the PNT, with analysis restricted to the 158 items that the unperturbed model named correctly. This restriction ensures that errors observed after perturbation are attributable to the intervention rather than to baseline failures. For each administration, a single layer was selected, and a fraction of its weights was modified by multiplicative Gaussian noise:

$$w \rightarrow w(1+\varepsilon), \qquad \varepsilon \sim \mathcal{N}(0,\sigma^2).$$

The perturbation grid varied layer, noise magnitude $\sigma$, perturbation density $\rho$, and random seed. The analysis used 80 independent seeds, divided equally into discovery and validation sets. Each seed therefore functions as an analogue of a subject, while the 40 transformer layers constitute the ordered spatial axis.

For each error-category pair, PRISM computes the difference between the corresponding error proportions within each seed and layer. The resulting layer profile is submitted to a group-level one-sample analysis across seeds. Threshold-free cluster enhancement (TFCE) is applied only along the layer axis, not across $\sigma$ and $\rho$, because the latter are interpreted as perturbation-dose parameters rather than spatial coordinates. Positive and negative contrast directions are analyzed separately. The method also includes a layer-order permutation test: if meaningful layer clusters depend on the intrinsic ordering of transformer layers, randomly permuting layer labels should substantially reduce cluster mass.

The human analysis uses 213 patients with chronic post-stroke aphasia who had complete PNT and lesion-load data. Lesions were summarized over 64 left-hemisphere regions from a Johns Hopkins University atlas. For each error category and region, the authors computed a Spearman correlation between lesion load and error proportion. Pairwise subtraction was implemented as a difference between these correlations. The patient sample was split into discovery and validation halves, and replication required both sign preservation and bootstrap confidence intervals excluding zero in both halves.

This correspondence is structural rather than statistical. On the LLM side, the contrast is a within-seed difference in error proportions followed by inference across perturbation seeds. On the cortical side, it is a between-patient difference between lesion-error correlations. The paper is explicit that the contrast operators differ, while the subject dimension, spatial dimension, subtraction logic, replication strategy, and inferential emphasis are deliberately aligned.

Category-Specific Signal Before Subtraction

A necessary preliminary result is that each error category exhibits nontrivial structure before pairwise subtraction. On the LLM side, the categories show distinct layerwise profiles. Phonemic errors peak around layers 25–30, semantic errors increase toward the deepest layers, neologisms peak around layers 15–20, unrelated errors peak earlier and decline, and no-response errors are especially prominent under early-layer perturbation. On the cortical side, the error categories show distinct lesion-load associations across left-hemisphere regions.

This preliminary analysis is important because subtraction can eliminate shared signal. Semantic and phonemic errors are not represented in wholly nonoverlapping regions or layers. Both have broad late-network or perisylvian profiles. Consequently, a weak subtractive contrast does not imply that one category lacks a neural or computational correlate; it may instead indicate substantial shared loading. PRISM is designed to identify the differential component of these overlapping maps.

The paper reports that the phonemic category is associated with left perisylvian and dorsal language regions, including the superior longitudinal fasciculus, supramarginal gyrus, and postcentral gyrus. Semantic errors show stronger associations in posterior temporo-occipital territories, including middle and inferior occipital regions. These marginal patterns provide the substrate for the primary semantic-versus-phonemic dissociation.

Principal LLM Findings

The primary LLM contrast yields a robust phonemic-favoring cluster spanning layers 22–31 in the discovery split and layers 23–33 in the validation split. The mean contrast at the discovery cluster is approximately $0.05$ in magnitude, and the layer-order permutation test gives an empirical $p < 0.001$. The cluster preserves its sign and location on held-out seeds, with only a modest shift in extent.

The converse semantic-favoring direction does not produce a TFCE-surviving cluster in either split. The same asymmetry appears in the two preregistered cross-checks. The phonemic-versus-neologism comparison recovers a phonemic-favoring cluster at layers 24–31 in discovery and 24–33 in validation. The semantic-versus-neologism comparison identifies neologism-favoring clusters around layers 16–17 and 19–20, again without a corresponding semantic-favoring cluster.

Thus, the paper does not claim a symmetric semantic-versus-phonemic dissociation. Its more specific claim is that the phonemic-favoring direction is robust, spatially localized, and reproducible, whereas the semantic-favoring direction is consistently signed but statistically subthreshold. This asymmetry is central to the interpretation. The authors argue against treating the absent semantic cluster simply as a low-power result because the semantic-favoring estimates remain close to zero across discovery and validation, do not become stronger in the larger patient sample, and do not show the predicted perturbation-dose scaling.

The reported permutation results are unusually strong in the narrow sense relevant to the method: all observed layer clusters across the pairwise contrasts exceed the maximum cluster-mass distribution obtained after within-seed layer-order randomization at empirical $p < 0.001$. This supports the claim that the effects depend on the ordered organization of the network rather than merely on the marginal distribution of layerwise error contrasts. It does not, by itself, establish that a cluster corresponds to a discrete mechanism, since residual connections and distributed computation remain important constraints.

Convergence with Aphasia Lesion-Symptom Mapping

The patient analysis yields a closely corresponding asymmetry. The phonemic-greater-than-semantic direction is associated with a replicating left frontal-perisylvian territory involving the postcentral gyrus, precentral gyrus, and superior longitudinal fasciculus. In the primary contrast, the discovery and validation correlation differences for the postcentral gyrus are approximately $-0.275$ and $-0.301$, respectively, with bootstrap confidence intervals excluding zero in both splits. The precentral gyrus and superior longitudinal fasciculus show similar directional effects, although not every region satisfies the strictest replication criterion in both halves.

The semantic-greater-than-phonemic direction produces weaker and more diffuse posterior associations. Its estimates are often directionally consistent in posterior temporal and occipital regions, but no region meets the paper’s strict replication criterion in the primary contrast. The semantic-versus-neologism comparison shows the same pattern: the neologism-favoring direction recovers the frontal-perisylvian territory, while the semantic-favoring direction remains subthreshold.

The patient results therefore reproduce the qualitative structure observed in the LLM: robust phonemic-favoring dissociation, weaker semantic-favoring trend, and replication of the asymmetry across independent splits. The authors interpret this convergence as evidence that the LLM’s perturbation-response organization captures a behaviorally meaningful aspect of naming-error structure rather than merely producing arbitrary layerwise degradation.

The convergence should nevertheless be interpreted at the appropriate level. The paper does not establish anatomical equivalence between layers 22–31 and postcentral or precentral cortex. Nor does it demonstrate that the same internal computation generates the phonemic errors in both systems. It establishes parallel dissociations under matched behavioral readouts and related spatial inference procedures. That is a methodological correspondence, not a claim of one-to-one neural homology.

Dose-Response Analysis

PRISM further analyzes the effects of perturbation magnitude and density within the recovered LLM clusters. The standard deviation $\sigma$0 is treated as an analogue of lesion severity, while $\sigma$1 is treated as an analogue of lesion extent. For the primary semantic-versus-phonemic cluster, the discovery slopes are $\sigma$2 and $\sigma$3, with narrow cluster-bootstrap intervals. The validation estimates are highly similar: $\sigma$4 and $\sigma$5. The direction of scaling agrees with the cluster contrast, indicating that deeper and more extensive perturbations strengthen the phonemic-favoring dissociation over part of the perturbation range.

The dose-response surfaces are not globally monotonic. For the primary cluster, the largest absolute contrast is approximately $\sigma$6 at $\sigma$7 and $\sigma$8, with comparable maxima at alternative depth-density combinations. Both low-dose and extreme-dose conditions produce weaker category differentiation. The authors interpret this as a saturating regime: moderate perturbations selectively amplify differences among error categories, whereas severe perturbations produce broad degradation and cause error categories to rise together.

This result is theoretically relevant because it qualifies the lesion analogy. A perturbation that simply destroys performance would be expected to increase general error rates, but not necessarily to produce a selective dissociation with an intermediate optimum. At the same time, dose sensitivity is not uniquely diagnostic of lesion-like computation. Many nonlinear systems exhibit saturation under increasing noise. The dose-response analysis therefore strengthens the perturbation interpretation without resolving the underlying mechanism.

The modest explanatory power of the bilinear dose model, with $\sigma$9 values of approximately $\rho$0–$\rho$1, is appropriately acknowledged. The linear slopes summarize directional trends; they do not adequately describe the ridge-shaped two-dimensional surfaces. In particular, the inverse trade-off between perturbation depth and density suggests that total effective perturbation, rather than either parameter independently, may govern category differentiation.

Interpretability Claims and Causal Scope

PRISM’s principal contribution is inferential rather than architectural. It provides a route from an externally defined phenotype to a spatially resolved perturbational map. A significant cluster indicates that perturbing that layer range changes the relative frequency of specified error classes. This is stronger than a purely correlational representation claim, but weaker than identifying an isolated computational circuit.

The distinction between necessity and sufficiency is especially important. Because transformer layers are connected through residual pathways, perturbing one layer can affect later computation, and an implicated layer range may represent a necessary route within the intact forward pass rather than a self-contained module. The authors therefore reserve the strongest causal-mechanism claim for PRISM Stage 3, which would directly perturb the identified region of interest and compare it with matched nonsignificant control clusters. That confirmatory intervention is not performed in the present work.

This limitation also defines how PRISM should interact with finer-grained mechanistic methods. Sparse autoencoders, activation patching, causal mediation, and circuit discovery can be applied within PRISM-identified layer ranges. Conversely, component-level claims about semantic or phonological computation can be tested by asking whether perturbation of the proposed components selectively changes the relevant PNT error categories. PRISM thus functions as a phenotype-first localization method rather than a replacement for feature- or circuit-level analysis.

Methodological Strengths and Limitations

The study has several methodological strengths. It uses a clinically established behavioral taxonomy rather than an ad hoc model-generated label set; separates discovery from validation; treats perturbation seeds as independent analysis units; tests the spatial-order assumption underlying TFCE; and performs an analogous human lesion-symptom analysis on the same task. The restriction to baseline-correct LLM items is also defensible because it avoids conflating perturbation effects with baseline model errors.

The main limitations constrain the generality of the findings. The LLM analysis is confined to one vision-language architecture, one parameter scale, and one naming task. The model’s baseline performance is imperfect, and the analysis set excludes 17 of 175 PNT items. The layer axis is an ordered index but not necessarily a spatial manifold in the neuroanatomical sense. Residual connections complicate focal attribution, and random multiplicative weight noise may not correspond to any biologically realistic lesion process.

The cortical analysis is also intentionally coarse. It uses 64 atlas regions and univariate Spearman correlation differences without covariate adjustment for age, sex, lesion volume, or other potential confounds. Although this choice supports cross-substrate comparability, it may reduce clinical specificity. Only 213 of the 276 patients with PNT data entered the lesion analysis because complete lesion-load profiles were required. Moreover, the patient-side and LLM-side contrast operators are not statistically identical, despite their shared conceptual structure.

A further issue concerns the interpretation of repeated permutation results. The layer-order test establishes that the observed TFCE mass depends on layer ordering, but the null does not distinguish a specialized computation from any structured late-layer organization. Similarly, replication across seeds tests stability under the specified perturbation distribution, not invariance across perturbation mechanisms. Replication with targeted ablations, activation patching, alternative perturbation families, and other architectures would materially strengthen the causal interpretation.

Finally, the paper reports relevant financial and intellectual-property relationships involving BLUM-related patents and affiliated entities. These disclosures do not invalidate the methods or results, but they are pertinent when evaluating the broader claims about clinical digital twins and commercial deployment.

Implications for AI and Clinical Modeling

For AI interpretability, PRISM introduces a useful inferential complement to internal representation analysis. It supplies a behavioral criterion for evaluating regional specialization: a proposed layer range should produce a selective and replicable perturbational change in a clinically defined error profile. This could help distinguish components that merely encode information from components whose disruption changes the model’s capacity to execute a particular operation.

The method also suggests a hierarchy of future interpretability analyses. PRISM can first identify behaviorally relevant layer ranges; sparse-autoencoder and circuit methods can then search within those ranges; targeted interventions can subsequently test whether individual features, heads, or pathways account for the regional effect. Such a hierarchy may reduce the search space for mechanistic analyses while preserving behavioral falsifiability.

The proposed application to patient-specific digital twins is more speculative. In principle, perturbations that reproduce an individual patient’s distribution of naming errors could provide computational models of aphasic impairment. These models might support simulation of rehabilitation protocols, trial design, and treatment-response hypotheses. However, a matched behavioral error profile is not sufficient to establish that the model has the patient’s underlying lesion mechanism. Clinical use would require validation across tasks, aphasia subtypes, longitudinal trajectories, lesion topology, and treatment outcomes.

The next empirical steps are therefore clear: complete Stage 3 ROI interventions; replicate the dissociation in text-only, encoder-decoder, and larger or smaller models; evaluate additional language tasks such as sentence completion and discourse comprehension; subdivide phonemic and semantic error categories; and use multivariate lesion-symptom and causal representation analyses to resolve weak semantic effects. Increasing the number of perturbation seeds is also necessary for a more decisive assessment of the semantic-favoring direction.

Conclusion

PRISM establishes a phenotype-first framework for regional interpretability in transformer models by adapting subtraction analysis, lesion-symptom mapping, TFCE, and out-of-sample replication to perturbation-induced language errors. In LLaVA-1.6-Vicuna-13B, the method identifies a reproducible phonemic-favoring cluster in approximately layers 22–31, with corresponding phonemic-favoring frontal-perisylvian regions in 213 patients with post-stroke aphasia. The semantic-favoring direction remains consistently signed but non-significant on both substrates. Dose-response analyses further show that moderate perturbations maximize category differentiation, whereas extreme perturbations produce less selective impairment.

The paper’s strongest contribution is not an anatomical equivalence claim between transformers and cortex, but a transferable inferential protocol for testing whether model components are causally implicated in externally defined behavioral operations. Its conclusions remain coarse-grained and require Stage 3 intervention and cross-architecture replication, but PRISM provides a technically specified bridge between mechanistic interpretability and lesion-based cognitive neuroscience (2608.12717).