Cortex-wide Dynamics of Internal Decisions About Behavioral Context

Reviewer #1 (Public review):

This paper presents another excellent, sophisticated analysis from this group of brain-wide neural activity correlated with the tracking of belief about the generative state of a stochastic visual environment under volatile conditions. Whereas previous work focussed on the normative belief-updating dynamics mainly in brain areas related to motor planning, under conditions where the environmental state translates directly to a correct action, here, they abstract the belief-updating DV from a specific action by instead associating the environmental state to a stimulus-response mapping rule, to be used in a simple perceptual decision coming up after the environmental state cues. A decoding analysis shows that a remarkably large portion of the brain has activity correlated with the normatively evolving belief about environmental state and the evidence samples feeding into that belief. What the authors were trying to achieve, however, seems far more general than the above, namely, to study "the algorithmic and neural basis of higher-order internal decisions about behavioural context, formed under multiple sources of uncertainty", and I think that the loose implication of such grand notions (such phrasing brings to mind someone's choice to believe in God, to regulate their behaviour depending on whether they are on a rugby pitch or at church, etc, not how grating orientations link to left/right hand movements) muddies the value of the study. The authors thus may have overestimated the generality of the findings. I hope my impressions are a useful guide to focus the interpretations more.

Strengths:

One of the main strengths of the study is that it is a technical tour de force. As reflected in an unusually extensive methods section, the authors put an extraordinary amount of work into rigorous data collection and analysis, and all of it is described in excellent detail. The study also builds in a very valuable way on previous landmark studies on tracking of volatile environmental state linked to correct actions using MEG (Murphy et al 2021) and tracking of volatile stimulus-response mappings using fMRI (van den Brink et al 2023). Here, the environmental state is not directly linked to actions during the cues informing about the state, but instead linked to a stimulus-response mapping rule.

Weaknesses:

It is surprising, given this main innovation of abstracting the decision about visual position-distribution from particular actions, that the authors do not engage with the literature using EEG and fMRI to study such 'abstract,' 'motor-independent' or 'domain-general' (synonymous terms) decisions. The discussion, for example, mentions the curious lack of involvement of the frontal cortex, and the possibility of intermingled opposites being represented there; motor-independent EEG decision signals have been characterised by regressing against the absolute value of the differential belief-updating process for this very reason (e.g., see Pares-Pujolras et al 2025). Single-unit studies like Bennur & Gold (2011) have also found activity related to a decision about environmental state (non-volatile motion) even when that state does not yet translate directly to an action, and, like the current study, is instead specified in a later frame of the trial.

Another weakness, as mentioned above, is that of overgeneralisation. It is not clear how "higher-order, internal decisions" are generally defined, and terms more concretely grounded in the paradigm at hand (as in van den Brink et al (2023)), e.g., 'tracking of environmental state dictating a sensory-motor mapping rule,' would seem more useful. Since this task tracks a belief about a sensory feature and how it maps to motor actions, it may not be as surprising a revelation that a range of sensorimotor areas correlate with it, as compared to more general, truly internal decisions about behavioural context involving no sensory input (e.g., deciding one has become hungry). Similarly, the authors paint the belief-tracking process of Murphy et al (2021) as "lower-order" and the current one as higher-order, but both cases are the same in that a hidden binary generative state needs to be inferred on a continual basis from a series of discrete spatial positions presented visually. The only difference is that in the current case, the belief about the current binary state is not transformed directly into an immediate action choice but rather utilised to map a follow-up stimulus to its appropriate action. These decisions then happen one after the other in sequence, with a contingency, but I'm not sure this constitutes a 'high-level' and 'low-level' in the way implied by the authors.

The paper left me confused on the question of what these widespread decoding effects reflect - whether all areas directly compute and represent the normative DV in concert, or whether at least some areas reflect other processes that may correlate with the DV. Although the discussion mentions things like feedback modulation in V1, which seems to allow for the possibility that it is not directly involved in DV computation, the phrasing used ('encoding' and 'representation' and never 'secondary modulation') from Abstract to Results tends to imply direct involvement.

Related to this, it seems that the extensive model comparison was done for behaviour, but not for the activation in each area, which may have suggested some dissociations in role - for example, for areas that showed decoding of the evidence (LLR), at least some of them may more closely correspond to the related lower-level quantity of simply spatial position itself, or the higher-level quantity of the transformed belief update (the change in prior from before to after the current cue). There is a map of areas that correlate with the difference of new vs old prior (if I understand correctly - Figure 4D), but not of areas for which activity conforms better to this belief update than to the objective LLR or location. Aside from such model-defined quantities, a critical factor is spatial attention. The authors highlight that the correlated activation of visual regions may reflect feedback modulations akin to attention in nature, but it might actually reflect attention itself, since it is plausible that subjects would pay more attention to the upper field when it is more likely that the centre of the generative distribution is up there (i.e., belief leans upwards). It seems the data could provide insight into this: If the visual cortical effects reflect a spatial attention modulation towards the likely generative source (upper/lower), then the relationship with prior, coded so that upper and lower have opposite sign, should flip in ventral versus dorsal visual cortex. Figure 4A seems like it could be positioned to answer this, but I can't fully interpret it because the prior coding is not explicit in the methods - the relevant section (lines 989-1001) refers back to the normative model description (without pointing to specific equations), which does not say what states S1 and S2 mean (upper and lower? Correct and incorrect? The former is needed to test for this spatial-specificity expected of attention). Even if there are reasons not to perform extra analyses related to the above, the impressions could guide edits to clarify what the data can and cannot say about what these DV-decoding effects reflect. Finally, it could be acknowledged that because the environmental state (upper or lower field generative source) is directly linked to stimulus-response mapping, even decoding effects that are not spatially-specific could equally reflect a representation of either one of these.

The motivation for the decoding analysis running up to the response is not clear - what are the hypotheses here? Is the idea that if these areas truly represented the belief about the currently active context, then they should continue to do so during the response and beyond, since the next trial will begin in the same context as the previous ended? Or is this section tackling a different question? Is it that there is a potential confound in finding the significant decoding during the cue tokens, because it could be driven by the visual responses to the different spatial positions, and there are no such visual responses later at the response?

Reviewer #2 (Public review):

Summary:

Calder-Travis et al. investigate how people form decisions about abstract rules in environments that may change over time. They show that individuals adaptively accumulate information, adjusting how much weight they give new evidence depending on how surprising or uncertain the environment is. Using whole-brain recordings (MEG), they further report that signals reflecting beliefs about the current rule are broadly distributed, particularly in visual and parietal regions. They further argue that these belief-related signals cannot be reduced to representations of momentary sensory evidence alone.

Overall, the behavioral results convincingly demonstrate adaptive evidence accumulation consistent with the normative model. The neural data provide solid evidence for temporally structured belief-related signals that are broadly distributed across cortical regions. However, the evidence for sustained belief maintenance "across" cues and for full dissociation from gaze-related influences in visual cortex is less definitive. These issues temper, but do not undermine, the central conclusions.

Strengths:

A major strength of the study is the integration of normative modeling with temporally resolved neural data. The authors exploit the fine temporal scale of the recordings to examine belief updating across distinct task epochs, and they show that neural signals evolve in a manner consistent with the normative model that best captures behavior. This alignment between behavioral modeling and neural dynamics is carefully executed and conceptually coherent.
Another strength is the authors' cautious interpretation of their findings. They explicitly acknowledge limitations in distinguishing between direct representation of a latent variable and neural modulation driven by that variable. This restraint strengthens the credibility of the conclusions and avoids overstatement.

Weaknesses:

(1) Evidence for sustained belief representation across cues

Behaviorally, the data clearly demonstrate accumulation across sequential cues. However, the neural analyses primarily focus on responses around individual samples (from pre-cue to late post-cue windows). While these analyses demonstrate belief updating following each sample, they do not fully establish whether belief representations are maintained continuously across cues.

Specifically, it remains unclear whether the neural representation of the prior belief is sustained from the late post-cue period of cue t-1 into the pre-cue period of cue t. Without explicit evidence of such continuity, it is difficult to conclude that the neural signals reflect a maintained belief state rather than repeated sample-locked updating processes. This distinction is important for interpreting the neural mechanism of accumulation.

(2) Interpretation of belief signals in the visual cortex

The claim that belief-related signals in the visual cortex cannot be explained by gaze position requires stronger support. The distribution of gaze positions across contexts appears largely non-overlapping, raising the possibility that context-related gaze biases could contribute to the observed neural effects.

In particular, the "gaze-inconsistent" analysis based on a median split may not fully dissociate belief from gaze if the absolute gaze positions remain systematically different between contexts. As currently presented, the evidence does not fully rule out the possibility that gaze-related modulation contributes to the belief-related signal in visual areas. This affects the strength of the interpretation regarding abstract belief representation in early sensory cortex.

(3) Clarity and transparency of task and model description

Several aspects of the task and modeling framework would benefit from clearer exposition. The description of the noise distribution in the context cue would be easier to interpret if the overlapping distributions were visualized explicitly, allowing readers to assess how much accumulation is required versus reliance on strong individual cues. Similarly, the main text would benefit from a clearer explanation of how change point probability and uncertainty are computed (not just in Methods), as these quantities are central to the analyses and interpretation.

In addition, temporal epochs (e.g., pre-cue, early post-cue, late post-cue) are not clearly defined with specific time ranges in the main text, making it difficult to compare across figures.

(4) Interpretation of neural dynamics

Several neural findings are intriguing but underinterpreted. For example, the absence of clear sensory evidence representation in early post-cue epochs in any regions (Figure 4B) is surprising and not discussed. The relative stability of belief-related signals in visual cortex compared to parietal regions (Figure 4E) is also unexpected and warrants interpretation. Additionally, the temporal dynamics of change point probability and uncertainty representations appear different from each other, but such a pattern was not described in detail.

Clarifying these points would strengthen the interpretability of the results and help readers understand the mechanistic implications.

Reviewer #3 (Public review):

Summary

In this study, the authors investigated how inference about the current task context is encoded in the cortex, using MEG measurements. Using the same behavioral task that was initially developed for an fMRI study to identify the loci of task context representation, the current results complement and extend the previous study by identifying the candidate regions that are important for the inference process, not just for encoding the end product. They reported widespread modulation of cortical activity by uncertainty in evidence and volatility of task context changes. In comparison, modulation correlated with the decision variable underlying the task context inference process was more restricted to the parietal and visual cortices, particularly in alpha-band activity.

Strengths:

(1) The normative model provides a solid computational foundation for disambiguating quantities related to decision variables from those related to task factors (e.g., uncertainty and volatility).

(2) The MEG technique allows examination of cortical activity that is modulated by the temporally evolving decision variable.

(3) Rigorous modeling efforts, including comparisons of well-reasoned alternative/reduced models and examinations of diagnostic features using participant-matched simulations.

Weaknesses:

(1) There are two major surprises in the results that raise concerns about how to interpret these data. The first is the absence of modulation of prefrontal cortical activities by prior or posterior. As the authors acknowledged, there are extensive single-neuron recording data (e.g., from the Miller group) demonstrating the presence of task rule modulation in the monkey PFC and prior representation in the PFC in the mouse study that they cited. The second surprise is that the strongest modulation of prior/posterior/evidence was almost always observed in the visual cortex, in contrast to the common embodied cognition assumption. A more elaborated discussion about these discrepancies would help contextualize the current results.

(2) It is not clear why the effects in Figures 2D and E dipped before responses, which is not expected from any of the models. This could potentially affect the interpretation of the MEG signals in late-post-cue or pre-response periods.

(3) The definitions of the different periods (e.g., early/late post-cue) are vague, making it hard to assess the functional relevance of the signals. For example, is the difference between the early pre-response map in Figure 5B and the late evidence map in Figure 4B due to completely non-overlapping time periods? A diagram of the timing definitions for different task periods would be helpful.

(4) Perhaps related to #2, it is puzzling that evidence encoding is absent in the visual cortex during the early post-cue period.

(5) The presentation and discussion of results related to correlated variability assume that the readers have already read their previous paper. A little more elaboration of the significance of this measurement would be helpful.