Review Reports - Deciding When to Align: Computational and Neural Mechanisms of Goal-Directed Social Alignment

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors introduce a meta-learning component to dynamically regulate the propensity for alignment on the basis of existing alignment mechanisms, thereby reframing social alignment from an automatic reaction to a strategic decision process. This constitutes a significant theoretical contribution with strong empirical implications. However, the model requires further specification in the following aspects:

Figure 1 effectively describes the functions of the five brain systems, but the temporal dynamics of their interaction remain unclear. Specifically, how these systems collaborate on a time scale to process information and generate a decision is not elaborated.
While the concept of meta-learning to regulate the learning rate is proposed, the specific functional form is not explicitly defined. The authors should provide at least one plausible functional form or a more detailed discussion of its potential constitutive elements (e.g., how different inputs like confidence and expertise might be integrated). This is crucial to enhance the model's operability and to help readers understand how this core mechanism might be implemented. Specifically, given the emphasis on confidence as a key regulator of the learning rate, this claim would benefit from being supported by references to more empirical case studies.
The model presents a universal human mechanism but does not discuss how individual differences might influence the meta-learning function. A discussion on how factors such as personality traits (e.g., baseline confidence, social anxiety) or clinical conditions could modulate the parameters or efficiency of the proposed circuits would significantly strengthen the model's generalizability and explanatory power.

Author Response

Comment 1:

Figure 1 effectively describes the functions of the five brain systems, but the temporal dynamics of their interaction remain unclear. Specifically, how these systems collaborate on a time scale to process information and generate a decision is not elaborated.

Response 1:

We thank the reviewer for this helpful suggestion. In response, we have expanded Section 4.2 to provide a clearer description of the sequential temporal dynamics through which the five brain systems interact to regulate alignment. This addition clarifies how these systems operate in sequence - through bottom-up detection, contextual evaluation, and top-down control- to dynamically arbitrate alignment in a goal-directed, context-sensitive manner:

“These two systems interact dynamically with the broader alignment circuitry described earlier. Through their coordinated activity, alignment can be upregulated when others are inferred to be highly confident, accurate, or expert, and downregulated when such cues are absent or misleading. This process may unfold temporally as follows: bottom-up sensory information about self- and other-generated behavior is first processed by both the gap-detection system (dACC, AI), which monitors discrepancies between expected and perceived alignment, and the mentalizing network (dmPFC, TPJ), which infers higher-order representations of others’ mental states and traits. When a discrepancy is detected, the observation–execution system (IFG, IPL, premotor cortex) generates automatic alignment impluses by mapping observed outputs onto one’s own representations. In parallel, the executive control network (rIFG, dlPFC) integrates the higher-order representations inferred by the mentalizing network and implements top-down regulation, either permitting or inhibiting alignment impulses generated by the observation–execution system. This sequence of bottom-up detection, contextual evaluation, and top-down control enables the network to dynamically arbitrate alignment in a goal-directed, context-sensitive manner.”

In addition, we now explicitly acknowledge this issue in the Discussion’s limitation section, noting that although a plausible temporal sequence is proposed, the precise dynamics of system coordination remain to be empirically validated using high–temporal–resolution methods:

Comment 2:

While the concept of meta-learning to regulate the learning rate is proposed, the specific functional form is not explicitly defined. The authors should provide at least one plausible functional form or a more detailed discussion of its potential constitutive elements (e.g., how different inputs like confidence and expertise might be integrated). This is crucial to enhance the model's operability and to help readers understand how this core mechanism might be implemented. Specifically, given the emphasis on confidence as a key regulator of the learning rate, this claim would benefit from being supported by references to more empirical case studies.

Response 2:

We thank the reviewer for this insightful comment. To address it, we have expanded Section 4.1 to provide an explicit definition of the meta-learning function and a more detailed explanation of its constitutive elements. In the revised text, we now formalize the function using a Bayesian reliability-weighting framework, in which the learning rate is proportional to the relative confidence (precision) of the other’s information compared with one’s own. We have also strengthened the empirical grounding of this section by adding references to behavioral and neurocomputational studies demonstrating how confidence modulates social influence and advice-taking. These additions clarify the model’s functional implementation and link its computational assumptions to empirical evidence, thereby enhancing its explanatory and predictive power:

“numerous studies demonstrate that confidence critically determines the extent of social influence. When people feel highly confident in an initial judgment, they become less willing to revise their judgments or accept social advice [40]. Conversely, when individuals lack confidence in a choice, they are more likely to conform to others’ choice, and they are more inclined to request advice before finalizing their decision [41,42]. Not only is the person's decision to align influenced by their own confidence, but also by the confidence displayed by others. For example, advisors expressing their opinions with higher confidence exert greater sway over group decisions [43]. Together, these findings support the notion that relative self–other confidence governs the extent of social alignment. To formalize how this process might occur, the meta-learning function introduced in Equation 2 can be explicitly expressed as:

Equation 3:

(Equation appears in the attached PDF file)

where ? represents the context-sensitive learning rate at time ?, ? self,t and ? other,t represent the estimated confidence (or the subjectively estimated likelihood of accuracy) of self and other. The ratio between these confidence estimates defines the relative reliability of social versus personal information, such that a higher value increases the weight assigned to the other’s input. These confidence terms can be further modulated by perceived normative factors (represented in the weights ? assigned to each term). The function describes a process in which the person integrates individual and social information in a Bayesian manner, to set a learning rate that flexibly adjusts the degree of alignment according to the inferred value of copying versus maintaining independence. Neurocomputational evidence supports this formulation. Bang et al. (2022) showed that confidence signals, both self- and other-related, are represented in distinct but interacting neural populations within the dorsomedial prefrontal cortex (dmPFC), which integrates these signals to guide the degree of behavioral adjustments that a person commits to match the behavior of the other person [44]. These results converge with behavioral evidence to suggest that confidence serves as a main computational variable feeding into the meta-learning function that governs learning rates in our model (Equation 2), enabling selective, context-sensitive social alignment.”

Comment 3:

The model presents a universal human mechanism but does not discuss how individual differences might influence the meta-learning function. A discussion on how factors such as personality traits (e.g., baseline confidence, social anxiety) or clinical conditions could modulate the parameters or efficiency of the proposed circuits would significantly strengthen the model's generalizability and explanatory power.

Response 3:

We thank the reviewer for this valuable suggestion. We have addressed it by expanding Section 5 (Falsifiable Predictions and Empirical Implications) to include a discussion of how individual differences might influence the meta-learning function. Specifically, we now describe how variability in metacognitive calibration and cognitive control could modulate the efficiency of the proposed circuits.

“ The model also predicts that interindividual differences in metacognitive and cognitive control functions can affect the efficiency of the proposed circuits. For instance, individuals with overconfidence may assign excessive weight to self-confidence, leading to consistent under-alignment. Those with diminished cognitive flexibility may have a reduced capacity to switch between independence and alignment according to contextual demands. Similarly, metacognitive failures observed in certain clinical conditions, such as autism or schizophrenia, could distort confidence-based weighting, reducing the system’s ability to adaptively regulate alignment.”

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This is an interesting conceptual paper which proposes a new model to address the research problem of adaptive alignment behaviour. The aim is to better understand this problem as a context-sensitive decision process. A model is presented and described which suggests that social alignment is emergent from interactions between five brain systems, rather than being only a matter of being driven by a stimulus.

The paper is broadly well written, and my judgement is that the model presented is credible and useful as a starting point for future research. There are some places, however, in which the argument could be developed or documented more explicitly.

I therefore suggest that the authors address the following points:

In the Introduction, it would be useful to more explicitly justify (rather than merely stating) why this paper focusses on computational and neural mechanisms.
In sections 2.1 and 2.2, it would be useful to offer some more explicit critique of the papers which are reviewed. Specifically, the reader should be led to understand why a new model is required.
Section 3 would benefit from some more detailed description of a) how the model was derived, and b) what is new, by comparison with the accounts presented in sections 2.1 and 2.2.
Section 4 seems to have two subsections with the same number, and this should be rectified.
The subsections within section 4 (which will become 4.1 and 4.2) should more explicitly link the account with arguments put forward in sections 2.1 and 2.2 respectively. At the moment, it seems as though the authors are presenting a new literature review which does not link to the earlier one. It is perfectly okay to cite new papers here, of course, but the groundwork laid by the main literature review should be useful, and its use should be made explicit.
Section 6 would benefit from anticipating some objections or perceived limitations from other researchers who work on the research problem of adaptive alignment behaviour.

Author Response

Comment 1:

In the Introduction, it would be useful to more explicitly justify (rather than merely stating) why this paper focusses on computational and neural mechanisms.

Response 1:

We thank the reviewer for this insightful comment. We have revised the Introduction to more explicitly justify our focus on both computational and neural mechanisms. Specifically, we have elaborated on this point by explaining that understanding alignment decisions require linking behavior to its underlying neural computations. This approach allows us to move beyond descriptive accounts of social behavior and to specify the mechanisms. Futhermore, we explain that studying the phenomena at both the computational and neural levels of analysis provides a principled framework for integrating findings across ecological, cognitive, and brain-based studies.:

“To develop this model, we combine insights from decision neuroscience with computational models of social decision-making, extending existing accounts of automatic alignment into a goal-directed account. Focusing on both computational and neural mechanisms allows us to specify not only the cognitive computations that govern alignment decisions but also the neural systems that implement them, thereby linking algorithmic-level processes with their biological substrates.”

Comment 2:

In sections 2.1 and 2.2, it would be useful to offer some more explicit critique of the papers which are reviewed. Specifically, the reader should be led to understand why a new model is required.

Response 2:

We appreciate this constructive suggestion. To clarify the need for a new model, we have added a paragraph in Section 3 that explicitly critiques existing computational and neural accounts reviewed in Sections 2.1 and 2.2. This new paragraph summarizes their key limitations, namely, their focus on automatic, stimulus-driven alignment without mechanisms for context-sensitive regulation. And it also explains how our proposed model addresses these gaps by incorporating a meta-learning component and higher-order control systems:

“The available neural and computational accounts described above conceptualize alignment primarily as an unregulated and stimulus-driven process. The computational account captures an automatic mechanism of error minimization that drives alignment, but it does not specify how the learning rate might vary adaptively according to context, determining the degree of alignment. Similarly, the neural account describes a stimulus-driven process, whereby observing another person’s action recruits overlapping motor representations in the observer’s brain, predisposing them toward mimicry and shared outputs [29]. However, social alignment is often regulated, as observed in it’s selectivity which help maximize its benefits while preventing harmful consequences. It is therefore unlikely that alignment operates in an unregulated and purely stimulus-driven manner [30]. Addressing this conceptual limitation requires rethinking alignment as a regulated and goal-directed process. Formally, this requires extending the computational model to incorporate a hyperparameter-tuning function that allows the learning rate to vary as a function of contextual cues. It also requires extending the neural model to include monitoring and control neural mechanisms that regulate automatic alignment as a function of those informative contextual cues.”

Comment 3:

Section 3 would benefit from some more detailed description of a) how the model was derived, and b) what is new, by comparison with the accounts presented in sections 2.1 and 2.2.

Response 3:

We thank the reviewer for this helpful suggestion. To clarify, Sections 4.1 (“Extending the computational account with meta-learning”) and 4.2 (“The neural mechanisms of goal-directed alignment”) provide detailed descriptions of how the proposed model was derived and how it extends previous frameworks.

Specifically, Section 4.1 explains how our model builds upon the predictive coding and reinforcement learning accounts reviewed in Section 2.1 by introducing a meta-learning (hyperparameter-tuning) layer that allows the learning rate to vary as a function of context-dependent variables such as confidence, expertise, and normative influence. This addition transforms alignment from a fixed stimulus-driven process into an adaptive, goal-directed decision mechanism. Section 4.2 then delineates the novel neural components of our model compared to exisiting neural models reviewed in section 2.2. Namely, the interaction between the mentalizing network (dmPFC, TPJ) and the executive control network (dlPFC, rIFG) which jointly regulate automatic alignment impulses generated by the mirror system. Together, these sections outline both the derivation of the model and its theoretical innovations relative to earlier accounts focused solely on automatic predictive coding mechanisms.

In any case, we have revised Section 3 to better explain how the proposed model was derived from prior computational and neural frameworks and to explicitly highlight the novel components that distinguish it from existing accounts.

Comment 4:

Section 4 seems to have two subsections with the same number, and this should be rectified.

Response 4:

We thank the reviewer for noting this formatting issue. The subsection numbering in Section 4 has been corrected accodingly.

Comment 5:

The subsections within section 4 (which will become 4.1 and 4.2) should more explicitly link the account with arguments put forward in sections 2.1 and 2.2 respectively. At the moment, it seems as though the authors are presenting a new literature review which does not link to the earlier one. It is perfectly okay to cite new papers here, of course, but the groundwork laid by the main literature review should be useful, and its use should be made explicit.

Response 5:

We thank the reviewer for this valuable suggestion. To address it, we have strengthened the conceptual linkage between Sections 2 and 4. Specifically:

In Section 4.2, we add a linking paragraph explicitly stating that the proposed neural model builds upon the background reviewed in Section 2.2.

“The neural account reviewed earlier describes how the interaction between the observation–execution system, the error-detection and reward circuits generates the impulses to align. While this account captures the stimulus-driven aspects of alignment, it does not capture the mechanisms that determine when and with whom to align—that is, how the brain selects when alignment is adaptive and when independence should be maintained. Deciding when and whom to copy often requires reasoning about others’ mental states—including their beliefs, intentions, and attitudes—and representing their traits, such as reliability, dominance, and expertise [34]. For example, estimating relative confidence in another person’s choice depends on the ability to infer that person’s uncertainty and to represent their credibility or expertise [44]. A large body of research situates these capacities within the mentalizing network, most prominently the dorsomedial prefrontal cortex (dmPFC) and temporoparietal junction (TPJ) [45,46].“

In Section 4.2, we also add a paragraph synthesizing the interaction between the novel components in our model and the mechanisms reviewed Section 2.2:

“Together, the mentalizing network and the executive control network (ECN) provide complementary functions: while the dmPFC and TPJ represent others’ traits, accuracy, and confidence, the ECN supports the control mechanisms necessary to use these evaluations to regulate alignment (Figure 1). These two systems interact dynamically with the broader alignment circuitry described earlier. Through their coordinated activity, alignment can be upregulated when others are inferred to be highly confident, accurate, or expert, and downregulated when such cues are absent or misleading. This process may unfold temporally as follows: bottom-up sensory information about self- and other-generated behavior is first processed by both the gap-detection system (dACC, AI), which monitors discrepancies between expected and perceived alignment, and the mentalizing network (dmPFC, TPJ), which infers higher-order representations of others’ mental states and traits. When a discrepancy is detected, the observation–execution system (IFG, IPL, premotor cortex) generates automatic alignment impulses by mapping observed outputs onto one’s own representations. In parallel, the executive control network (rIFG, dlPFC) integrates the higher-order representations inferred by the mentalizing network and implements top-down regulation, either permitting or inhibiting alignment impulses generated by the observation–execution system. This sequence of bottom-up detection, contextual evaluation, and top-down control enables the network to dynamically arbitrate alignment in a goal-directed, context-sensitive manner.”

In Section 4.1, the link is made explicit with the openeing paragraph:

“The computational mechanisms described in section 2.1 conceptualize alignment as a process in which detecting a gap between self and other generates a prediction error, corrected proportionally to a learning rate. In standard formulations, this learning rate is treated as fixed, meaning that each given discrepancy between self and other is updated with equal percentage [21]. While this captures the general tendency for alignment, it does not address how the system regulates alignment, specifically, how it decides when and by how much to align and when to resist alignment.”

Together, these additions make explicit how the proposed model integrates and extends the frameworks presented earlier rather than introducing them as separate literature reviews.

Comment 6:

Section 6 would benefit from anticipating some objections or perceived limitations from other researchers who work on the research problem of adaptive alignment behaviour.

Response 6:

We thank the reviewer for this helpful suggestion. In response, we have added a paragraph to the Discussion section (Section 6) that explicitly acknowledges the model’s current limitations and outlines directions for future research.:

“While offering a cohesive framework, the model also faces limitations that warrant further empirical and theoretical work. First, although we outline a plausible temporal sequence of interactions among the five neural systems, the precise temporal dynamics of their coordination remain to be tested with high–temporal–resolution methods. Second, the proposed meta-learning function specifies how self-versus-other confidence tunes the learning rate, but its formal structure requires empirical validation and likely further refinement. Third, the model in its current formulation primarily addresses informational influences on alignment, whereas various normative factors—such as perceived closeness, social pressure, or group identity—also likely exert additional control over alignment decisions and should be incorporated in future extensions of the model. Addressing these limitations will help further specify the model’s computational and neural architecture.”