Relative Neural Involvement Overview

• Relative neural involvement is a quantitative framework that estimates graded contributions of distinct subsystems to observed outcomes across diverse fields.
• It applies normalized metrics such as BERTScore recall in AI, Jacobian-derived atrophy scores in MRI, weighted arbitration in reinforcement learning, and lesion-adjusted coefficients in MS studies.
• The concept emphasizes continuous attribution over binary labels, enabling detailed analysis of system-specific contributions and improved interpretability across disciplines.

Searching arXiv for the cited papers to ground the article. arXiv search query: (Guo et al., 4 Jun 2025) Relative neural involvement denotes a class of quantitative formulations that estimate the relative contribution of distinct generators, subsystems, or anatomical regions to an observed output. In the cited literature, the term spans at least four technically distinct settings: human versus AI participation in academic text generation, region-wise attribution of longitudinal MRI changes, arbitration between hippocampal and striatal control in Markov Decision Processes, and lesion-independent versus lesion-mediated thalamic contributions to cognitive impairment in multiple sclerosis. Across these settings, the common pattern is not a single canonical metric but a family of normalized scores, weighted combinations, and covariate-adjusted effect estimates designed to move beyond binary labels and toward graded attribution (Guo et al., 4 Jun 2025, Dong et al., 2023, Fabian et al., 2018, Pooladi-Darvish et al., 26 Nov 2025).

1. Core formulations and domain-specific meanings

The phrase has domain-specific meanings. In AI-generated academic writing, human involvement, equivalently “prompt recall,” versus AI involvement is treated as a single continuous variable $y_{\rm reg}\in[0,1]$, where $0$ means no human-provided content appears in the generated text and $1$ means the generated text fully reuses the human prompt. Ground-truth labels are obtained by min-max normalizing BERTScore recall between prompt $P$ and generated text $G$ (Guo et al., 4 Jun 2025).

In longitudinal neuroimaging for Alzheimer’s disease progression, Regional Deep Atrophy defines a per-region atrophy score $A_r$ from attention-weighted Jacobian-derived shrinkage and expansion terms, then normalizes those scores to obtain a relative involvement index

$$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$

Here, relative involvement is the fractional contribution of region $r$ to overall longitudinal change (Dong et al., 2023).

In computational models of hippocampal and striatal function, relative involvement appears as a weighted combination of model-based and model-free value estimates,

$$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$

with action selection determined from the combined value. In that setting, relative neural involvement is the degree to which hippocampal planning or striatal habit learning dominates behavior under a given task regime (Fabian et al., 2018).

In multiple sclerosis, the relevant distinction is not between two controllers but between lesion-independent and lesion-mediated contributions of thalamic nuclei to information processing speed. The lesion-mediated effect percentage is defined as

$100\times \frac{\beta_1-\beta_1'}{\beta_1},$

where $0$0 is the unadjusted ROI coefficient and $0$1 is the lesion-adjusted coefficient in linear models relating ROI volume to SDMT score (Pooladi-Darvish et al., 26 Nov 2025).

Context	Quantity	Operational meaning
Academic text generation	$0$2	Human prompt reuse in generated text
Longitudinal MRI	$0$3	Fractional regional contribution to atrophy
Cognitive control	$0$4 in $0$5	Hippocampal versus striatal control
Multiple sclerosis	Lesion-mediated $0$6	Fraction of ROI effect explained by lesions

Taken together, these works treat relative neural involvement as a graded attribution problem. A plausible implication is that the concept is best understood as a methodological family rather than as a single, discipline-independent construct.

2. Continuous human–AI involvement in academic writing

The academic-writing formulation was introduced to address what the paper terms participation detection obfuscation: binary AI detectors overlook varying degrees of human involvement even though human-machine collaboration is becoming mainstream. The proposed solution uses BERTScore recall as a proxy for human involvement and trains a multi-task RoBERTa-based regressor with a token-classification auxiliary head (Guo et al., 4 Jun 2025).

Given reference tokens $0$7 and candidate tokens $0$8, BERTScore recall, precision, and F1 are defined through token-embedding similarities. In this application, $0$9 measures how much of the human prompt is “recalled” in the AI output, and the normalized recall becomes $1$0. The model architecture uses a shared RoBERTa encoder that outputs contextual token embeddings $1$1. The regression head pools the $1$2 embedding and predicts $1$3, while the token-classification head assigns, for each token, the probability of “token from prompt (1)” versus “not from prompt (0).” Training minimizes

$1$4

with class weights $1$5 and $1$6 to counter token-label imbalance.

The main evaluation resource is the Continuous Academic Set in CS (CAS-CS). It starts from 55 000 real abstracts split into $1$7 sentences; samples $1$8; forms prompts of the form “Write an abstract on the basis of [these $1$9 sentences]”; feeds prompts to ChatGPT to obtain generated text $P$0; and then computes $P$1 via normalized BERTScore recall. Token labels are derived by marking tokens in $P$2 that also appear, after lemmatization, in the prompt. Additional 1 000 texts each from Claude, Gemini, GPT-4, and Falcon were included for generalization tests. A separate Polarized Academic Set (PAS-CS) contains 2 000 ChatGPT texts with minimal prompt content and 2 000 purely human-written abstracts. Label credibility was checked using ten human judges on 55 texts, yielding Spearman correlation $P$3, or $P$4 after outlier removal.

The reported results directly target the inadequacy of binary detection. On CAS-CS, four off-the-shelf binary detectors—OpenAI, ChatGPT-D, Academic-D, and DetectGPT—plateaued around $P$5–$P$6 accuracy under varying binarization thresholds. By contrast, the dual-head model achieved MSE $P$7 and token-classification F1 $P$8 in ablation comparisons, outperforming single-head regression (MSE $P$9) and single-head token classification (F1 $G$0). On ChatGPT train/test CAS-CS splits, test MSE was $G$1 and “error within $G$2” accuracy was $G$3. Transfer performance was reported as GPT-4: MSE $G$4, ACC $G$5; Claude-3: MSE $G$6, ACC $G$7; Gemini/Bard: MSE $G$8, ACC $G$9; Falcon-7B: MSE $A_r$0, ACC $A_r$1. On polarized benchmarks at $A_r$2, the model reached $A_r$3 and $A_r$4 on PAS-CS, CHEAT, GPA, and Sa. Under prompt-template variations—“direct,” “student perspective,” “dual generation,” and “summarization”—MSE rose only modestly from $A_r$5 to $A_r$6, while $A_r$7-accuracy remained above $A_r$8 (Guo et al., 4 Jun 2025).

The significance of this framework is that it replaces a binary “human vs. AI” label with a scalar estimate plus token-wise attribution. This suggests that, in hybrid writing settings, relative neural involvement is operationalized not as authorship identity but as the extent of prompt reuse detectable in the generated surface form and semantics.

3. Regional attribution in longitudinal MRI

In Regional Deep Atrophy, relative neural involvement is localized anatomically. The model processes two 3D MR scans of the same subject, $A_r$9 and $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$0, cropped around the medial-temporal lobe and presented in arbitrary temporal order. One branch is a pre-trained VoxelMorph network $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$1 that produces a displacement field $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$2; the other is a 3D U-Net $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$3 that outputs three soft-segmentation maps for shrinkage, expansion, and background. A high-temperature SoftMax with $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$4 binarizes these masks. From the displacement field, the method computes a Jacobian map $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$5, and from the masks and Jacobian it derives pooled change signals used in self-supervised temporal inference (Dong et al., 2023).

The two learning objectives are Scan-Temporal-Order (STO) and Relative-Interscan-Interval (RISI). STO uses binary cross-entropy on the sign of pooled change $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$6 relative to temporal order. RISI samples two scan pairs from the same subject, discretizes the interval ratio into $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$7 bins $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$8, and minimizes categorical cross-entropy over a differentiable “soft-cone” mapping from $$\mathrm{RI}_r=\frac{A_r}{\sum_{i=1}^R A_i}, \qquad \sum_{r=1}^R \mathrm{RI}_r=1.$$9 to class probabilities.

For region-wise analysis, the attention mechanism can be extended to output $r$0 shrink masks $r$1 and $r$2 expand masks $r$3. The regional components are

$r$4

and the per-region atrophy score is

$r$5

Normalization then yields $r$6, the relative involvement of region $r$7. In practice, the method may threshold regions with $r$8 or $r$9 to focus on dominant contributors.

Illustrative results were reported for hippocampal subfields on a 7T template across 300 MCI$$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$0AD converters: $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$1, $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$2, $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$3, and $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$4. Correlation with manual TBM-based Jacobian integration in each subfield was $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$5 with $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$6. For cortical ROIs, ventricular expansion dominated expansion masks with $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$7, while gray-matter subregions accounted for approximately $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$8 of shrinkage. The paper reports STO/RISI accuracy of approximately $$Q_{\rm combined}(s,a)=w\,Q_{\rm MB}(s,a)+(1-w)\,Q_{\rm MF}(s,a),$$9, versus classical DBM at approximately $100\times \frac{\beta_1-\beta_1'}{\beta_1},$0 (Dong et al., 2023).

Here, relative neural involvement is not a behavioral weight or a lesion-adjusted coefficient; it is a normalized decomposition of observed longitudinal change across ROIs. A plausible implication is that the framework converts an otherwise global progression signal into an anatomically ranked explanation.

4. Hippocampal–striatal arbitration in cognitive tasks

In the computational model of cognitive tasks, the hippocampus and striatum are cast as complementary controllers. The hippocampus is the “model-based” path: a place-cell network that builds an allocentric cognitive map of states, learns transition probabilities $100\times \frac{\beta_1-\beta_1'}{\beta_1},$1 and reward contingencies $100\times \frac{\beta_1-\beta_1'}{\beta_1},$2 rapidly from single experiences, and supports forward sweeps of activity for prospective multi-step evaluation. The striatum is the “model-free” path: an egocentric stimulus-response association network that gradually learns cached action values $100\times \frac{\beta_1-\beta_1'}{\beta_1},$3 via reward prediction errors. A prefrontal cortex arbitrator receives outputs from both systems and selects actions through an effective weighted combination (Fabian et al., 2018).

The hippocampal path uses the model-based Bellman backup

$100\times \frac{\beta_1-\beta_1'}{\beta_1},$4

whereas the striatal path uses temporal-difference learning,

$100\times \frac{\beta_1-\beta_1'}{\beta_1},$5

$100\times \frac{\beta_1-\beta_1'}{\beta_1},$6

Their outputs are mixed as

$100\times \frac{\beta_1-\beta_1'}{\beta_1},$7

followed by

$100\times \frac{\beta_1-\beta_1'}{\beta_1},$8

An example reliability-based weighting,

$100\times \frac{\beta_1-\beta_1'}{\beta_1},$9

is described as not explicitly tested but consistent with theory.

The framework was applied to spatial navigation in a plus maze and to an abstract Daw et al.-style two-step decision task. In the plus maze, early in training (Day 8), $0$00–$0$01 of control rats chose the “place” solution when the start location was shifted, and model hippocampal output dominated. Late in training (Day 16), approximately $0$02 chose the “response” solution, and striatal stimulus-response associations dominated. Hippocampal blockade abolished place strategy; striatal blockade abolished response strategy; the model reproduced both effects. In the two-step task, the model-free system alone showed high “stay” probability after rewarded trials at approximately $0$03 and low after unrewarded trials at approximately $0$04, independent of common versus rare transitions. The model-based system alone showed the expected interaction: rewarded-common approximately $0$05, rewarded-rare approximately $0$06, unrewarded-common approximately $0$07, and unrewarded-rare approximately $0$08. Human data lay between these extremes (Fabian et al., 2018).

This usage makes relative neural involvement a control-allocation variable. The paper’s synthesis states that model-based hippocampal control dominates when tasks are novel, when transition structure must be explicitly used, or when practice is minimal, whereas model-free striatal control dominates after extensive overtraining in stable deterministic environments. The compact expression

$0$09

summarizes that dependence, with $0$10 increasing with task volatility $0$11 and decreasing with training duration $0$12. Relative neural involvement is therefore dynamic rather than fixed.

5. Lesion-independent thalamic involvement in multiple sclerosis

In the multiple-sclerosis study, relative neural involvement is formulated through nucleus-specific heterogeneity in the relationship between thalamic atrophy and information processing speed. The dataset comprised 100 participants with MS. Lesion load was quantified from bias-corrected, co-registered T1w and FLAIR images using the Lesion Segmentation Toolbox’s LST-AI module, described as a 3D U-Net ensemble that outputs lesion probability maps thresholded to binary lesion masks. Thalamic nuclei were delineated using HIPS-THOMAS from MPRAGE T1w and synthetic white-matter-nulled images, yielding 13 left and 13 right ROIs, including global thalamus, 9 principal nuclei, 2 perithalamic nuclei, and the MD-Pf complex. ROI volumes were normalized to intracranial volume and z-scored across subjects (Pooladi-Darvish et al., 26 Nov 2025).

Associations with SDMT were estimated using two linear models. The unadjusted model,

$0$13

captures the total association between ROI volume and IPS. The lesion-adjusted model,

$0$14

isolates the volume-IPS link independent of lesion load. The relative reduction

$0$15

is interpreted as the lesion-mediated percentage.

The reported pattern was explicitly heterogeneous. Twenty-one of 26 ROIs were significant before lesion adjustment, and 12 of 26 remained significant after adjustment. Retained lesion-independent associations included the global thalamus, sensory relay nuclei, and associative hubs. On the left, significant adjusted effects were reported for Global Thal ($0$16, $0$17), VPL ($0$18, $0$19), Pulvinar ($0$20, $0$21), MGN ($0$22, $0$23), LGN ($0$24, $0$25), and MD-Pf ($0$26, $0$27). On the right, significant adjusted effects were reported for Global Thal ($0$28, $0$29), VA ($0$30, $0$31), VLa ($0$32, $0$33), Pulvinar ($0$34, $0$35), LGN ($0$36, $0$37), and MD-Pf ($0$38, $0$39). By contrast, several nuclei lost significance after lesion adjustment, including left CM, AV, VA, VLa, and VLP, and right VLP, VPL, and AV.

The paper interprets these differences through lesion-mediated percentages. Sensory relay nuclei—VPL, LGN, MGN—and associative hubs—pulvinar and MD-Pf—showed strong lesion-independent associations with IPS, with mean lesion-mediated effect approximately $0$40. Nuclei losing significance after adjustment had mean lesion-mediated effect approximately $0$41, with $0$42. The global thalamus was characterized as partially lesion-mediated at approximately $0$43 mediated, while retaining a lesion-independent core association (Pooladi-Darvish et al., 26 Nov 2025).

Within this framework, relative neural involvement refers to the extent to which a nucleus contributes directly to cognitive slowing after lesion burden is accounted for. This suggests a distinction between intrinsic vulnerability and secondary degeneration rather than a single uniform thalamic mechanism.

6. Methodological convergences, interpretive cautions, and limits

Despite their heterogeneity, the cited formulations share several methodological principles. First, each rejects a simple binary partition. The academic-writing study argues that binary AI detectors are insufficient under participation detection obfuscation and replaces them with a continuous score plus token-level attribution. RDA converts overall longitudinal change into normalized regional fractions. The hippocampal–striatal model replaces exclusive-system accounts with weighted arbitration. The MS study separates lesion-independent from lesion-mediated components rather than treating thalamic atrophy as unitary (Guo et al., 4 Jun 2025, Dong et al., 2023, Fabian et al., 2018, Pooladi-Darvish et al., 26 Nov 2025).

Second, each framework links attribution to a specific operational substrate. In text generation, the substrate is prompt recall measured by BERTScore recall and approximated from generated text alone through a dual-head RoBERTa model. In MRI, it is the Jacobian of a deformable registration field filtered by learned shrinkage and expansion masks. In reinforcement learning, it is the relative contribution of model-based and model-free value functions to action choice. In MS, it is the residual ROI-volume association with SDMT after explicit lesion-load adjustment. A plausible implication is that relative neural involvement is only interpretable relative to the measurement operator that defines it.

Several cautions follow directly from the source papers. The academic-writing framework is limited to computer-science abstracts, and RoBERTa’s input-length cap precludes full-paper analysis; its reliance on BERTScore suggests future work could incorporate human surveys or alternative semantic-similarity measures. RDA relies on the quality of pre-trained VoxelMorph, severe misregistration can pollute $0$44, attention masks may bleed across subfield boundaries when tissue contrast is low, and the method is typically trained on MTL only rather than the whole brain. In the hippocampal–striatal model, the reliability-based weighting rule is presented as an example not explicitly tested in the paper. In the MS study, the lesion-adjustment analysis establishes heterogeneity in associations, but the distinction between intrinsic vulnerability and secondary degeneration is still an interpretive framework tied to covariate modeling rather than a direct cellular assay (Guo et al., 4 Jun 2025, Dong et al., 2023, Fabian et al., 2018, Pooladi-Darvish et al., 26 Nov 2025).

A common misconception is that “relative neural involvement” always refers to biological neural tissue. The academic-writing study uses “AI (neural) involvement” to denote contribution from LLMs relative to human prompt reuse. Conversely, the neuroimaging and cognitive-neuroscience papers concern biological systems and anatomical ROIs. Another misconception is that relative involvement is inherently a causal decomposition. In these papers it is operational: normalized prompt recall, normalized atrophy, weighted value integration, or coefficient change after covariate adjustment. The concept is therefore rigorous within each formalization, but its meaning is inseparable from the underlying task, model class, and measurement design.