Effort cost of harvest affects decisions and movement vigor of marmosets during foraging

Reviewer #1 (Public Review):

The manuscript by Hage et al. presents interesting results from a foraging behavior in Marmosets that explores the interactions of saccade and lick vigor with pupil dilation and performance as well as a marginal value theory and foraging theory-inspired value-based decision-making model thereof. The results are generally robust and carefully presented and analyses, particularly of vigor, are carefully executed.

The authors constructed a model that makes two predictions: "In summary, this simple theory made two sets of predictions: in response to an increased cost of harvest, one should work longer, but move with reduced vigor. In response to an increased reward value, as in hunger, one should also work longer, but now move with increased vigor." Their behavioral data meets these predictions. It is not clear if the model was designed and tweaked in order to make those predictions and match the data, or derived from principles. Furthermore, it is not clear what other models would make similar predictions. It would help to assess what is predicted by other simple models, as well as different functional forms for the effort costs in their model.

Line 37 page 6; the link of pupil to NE/LC is tenuous. Other modulators systems and circuits may be equally important and should be mentioned (e.g. Reimer, Jacob, Matthew J. McGinley, Yang Liu, Charles Rodenkirch, Qi Wang, David A. McCormick, and Andreas S. Tolias. "Pupil fluctuations track rapid changes in adrenergic and cholinergic activity in cortex." Nature communications 7, no. 1 (2016): 13289.)
Line 35 page 6-page 7 line 10 emphasizes a cognitive interpretation of the pupil dilations they is emphasized, in relation to effort costs. But there are also concomitant more vigorous movements. Could all of their pupil results be explained by motor correlates? This should be tested and ruled out before making cognitive interpretations.
Page 7, line 37-42: How would the model need to be modified in order to account for this discrepancy with the data? Ideally, this would be tested.
Page 9, line 2-11: In this section, it would help to also consider 'baseline' pupil size (in between trials). This would give a signal that is not 'contaminated' by movements, and may reflect control state. Relatedly, changes in control state may impact and confound the movement-related dilation magnitudes due to e.g. floor and ceiling effects on pupil size, which has a strong tendency for reversion to the mean.
Page 10 line 21-32 presents a dated view of pupil that has/had little data supporting it. They should mention other neuromodulators (Reimer et al., 2016) and related interpretations.
The hunger-related and reward-size related analyses are both heavily confounded since they were not manipulated directly and could co-vary with many latent factors. For example, why might a given Marmoset be lower weight on a given day? Could it affect sleep, stress, activity, or other factors during the preceding 24 hours? If so, could these other variables be driving the results that are interpreted as 'hunger?' Relatedly, since the reward size is determined by the animals behavior on each trial (how much they worked), factors (internal brain state, external noises, etc.) that alter how much they worked will influence the subsequent reward size. Therefore interpretations about reward expectancy are confounded. Both of these issues should be discussed and manipulations of them (different feeding schedules and reward size-work functions proposed, respectively.
A major issue is a lack of alternative models. The authors seem to have constructed a particular model designed to capture the behavioral patterns they observed in the data. The model fails in some instances, as they point out. Even more importantly, there are no results or discussion about how other plausible models could or couldn't fit the data. The lack of model comparisons makes it difficult to interpret the conclusions or put the results in a broader context.

Reviewer #2 (Public Review):

Hage et al examine how the foraging behavior of marmoset monkeys in a laboratory setting systematically takes into account the reward value and anticipated effort cost associated with the acquisition and consumption of food. In an interesting comprehensive framework, the authors study how experimental and natural variation of these factors affect both the decisions and actions necessary to gather and accumulate food, as well as the actions necessary to consume the food.

The manuscript proposes a computational model of how the monkeys may guide all these aspects of behavior, by maximizing a food capture rate that trades off the food that can be gathered with the effort and duration of the underlying actions. They use this model to derive qualitative predictions for how monkeys should react to an increase in the effort associated with food consumption: Monkeys should work longer before deciding to consume the accumulated food, but should move more slowly. The model also predicts that monkeys should show a different reaction to an increase in reward value of the food, also working longer but moving faster. The authors test these predictions in an interesting experimental setup that requires monkeys to collect small increments of food rewards for successful eye movements to targets. The monkeys can decide freely when to interrupt work and consume the accumulated food, and the authors measure the speed of the eye movements involved in the food acquisition as well as the tongue movements involved in the food consumption.

By and large, the behavioral findings fall in line with the qualitative model predictions: When the effort involved in food consumption increases, monkeys collect more food before deciding to consume it, and they move slower both during food acquisition and food consumption. In a second test, the authors approximate the effects of reward value of the food at stake, by comparing monkey behavior during different days with natural variations in body weight. These quasi-experimental increases in the reward value of food also lead to longer work times before consumption, but to faster movements during food consumption. Finally, the authors show that these effects correlate with pupil size, with pupils dilating more for low-effort foraging actions with increased saccade speed and decreased work duration. The authors conclude that the effort associated with anticipated actions can lead to changes in global brain state that simultaneously affect decisions and action vigor.

The paper proposes an interesting model for how one unified action policy may simultaneously affect multiple types of decisions and movements involved in foraging. The methods employed to measure behavior and test these predictions are generally sound, and the paper is well written. While the model and paper in their present form can clearly inspire researchers to consider this integrated perspective, and trigger further research employing such a framework, there are some conceptual and methodical shortcomings that reduce the conclusiveness of the results and the usefulness of the proposed model.

(1) The model proposed in the paper takes a very specific functional form that is neither motivated by the previous literature nor particularly useful for indexing the behavioral tendencies of individual monkeys (or of the same monkey in different contexts). For example, while it is clear that the saccade effort cost will need to outgrow the increase in the utility of the accumulated food for the monkey to start feeding, it is unclear why this needs to be modeled with a fixed quadratic exponent on the number of saccades? Similarly, why do licks deplete the food stash with the specific rate hard-coded in the model? Finally, the proportion of successful saccades and lick events is assumed to be fixed, even though it very likely to be directly influenced by movement speed (speed-accuracy trade-off), which is also contained in the model. It would strongly increase the plausibility and potential impact of the model if the authors could clearly state where these hard-coded model terms come from. Ideally, they would formulate the model in more general terms and also consider other functional forms, as briefly suggested in the discussion. This latter point would be particularly important since not all model predictions were actually borne out in the data.

(2) The authors derive qualitative predictions, by simulating their model with apparently arbitrary parameters. They then test these qualitative predictions with conventional statistics (e.g., t-tests of whether monkeys lick more for high vs low effort trials). The reader wonders why the authors chose this route, instead of formulating their model with flexible parameters and then fitting these to data. This would allow them (and future researchers) to test their model not just qualitatively but also quantitatively, and to compare the plausibility of different functional forms. The authors certainly have enough data and power to do this, given the vast number of sessions the monkey completed.

(3) The effort manipulation chosen by the authors (distance of food tube) goes hand in hand with a greater need for precision since the monkey's tongue needs to hit an opening of similar size, but now located at a greater distance. This raises the question of whether the monkeys moved slower to enhance its chance of collecting the food (in line with a speed-accuracy trade off). The manuscript would benefit from an explicit test of this possibility, for example by reporting whether for each of the two conditions, the speed of tongue movements on a trial-by-trial basis predicts the probability of food collection? At the very least, the manuscript should explicitly discuss this issue and how it affects the certainty with which effects of tube distance can be linked to anticipated effort cost alone.

(4) The authors report most of the effects on the different measures (work duration, movement vigor, lick vigor, etc) in separate analyses. However, their model predicts that all of these measures result from the same action policy (maximization of the capture rate) and should therefore be related on a trial-by-trial basis. This is so far hardly tested in the presented analyses (with the exception of the pupil correlations in Figure 5). The model's assumed action policy would appear more plausible if the authors could demonstrate these trial-by-trial interrelations with some tests of association (e.g., correlations/regressions as already done for pupil measures in Figure 5) or possibly with dimensionality reduction of the multivariate data.

(5) The manuscript measures pupil dilation in a time period ranging from -250ms before to 250 ms after saccade onset. However, the pupil changes strongly during saccade execution relative to the preceding baseline, leaving doubts as to whether the aggregated measure blurs several interesting and potentially different effects. It would be more conclusive if the manuscript could report the analyses of pupil size separately for a period prior to saccade onset and during/after the saccade.

Reviewer #1 (Public Review):

The manuscript by Hage et al. presents interesting results from a foraging behavior in Marmosets that explores the interactions of saccade and lick vigor with pupil dilation and performance as well as a marginal value theory and foraging theory-inspired value-based decision-making model thereof. The results are generally robust and carefully presented and analyses, particularly of vigor, are carefully executed.

The authors constructed a model that makes two predictions: "In summary, this simple theory made two sets of predictions: in response to an increased cost of harvest, one should work longer, but move with reduced vigor. In response to an increased reward value, as in hunger, one should also work longer, but now move with increased vigor." Their behavioral data meets these predictions. It is not clear if the model was designed and tweaked in order to make those predictions and match the data, or derived from principles. Furthermore, it is not clear what other models would make similar predictions. It would help to assess what is predicted by other simple models, as well as different functional forms for the effort costs in their model.

Line 37 page 6; the link of pupil to NE/LC is tenuous. Other modulators systems and circuits may be equally important and should be mentioned (e.g. Reimer, Jacob, Matthew J. McGinley, Yang Liu, Charles Rodenkirch, Qi Wang, David A. McCormick, and Andreas S. Tolias. "Pupil fluctuations track rapid changes in adrenergic and cholinergic activity in cortex." Nature communications 7, no. 1 (2016): 13289.)
Line 35 page 6-page 7 line 10 emphasizes a cognitive interpretation of the pupil dilations they is emphasized, in relation to effort costs. But there are also concomitant more vigorous movements. Could all of their pupil results be explained by motor correlates? This should be tested and ruled out before making cognitive interpretations.
Page 7, line 37-42: How would the model need to be modified in order to account for this discrepancy with the data? Ideally, this would be tested.
Page 9, line 2-11: In this section, it would help to also consider 'baseline' pupil size (in between trials). This would give a signal that is not 'contaminated' by movements, and may reflect control state. Relatedly, changes in control state may impact and confound the movement-related dilation magnitudes due to e.g. floor and ceiling effects on pupil size, which has a strong tendency for reversion to the mean.
Page 10 line 21-32 presents a dated view of pupil that has/had little data supporting it. They should mention other neuromodulators (Reimer et al., 2016) and related interpretations.
The hunger-related and reward-size related analyses are both heavily confounded since they were not manipulated directly and could co-vary with many latent factors. For example, why might a given Marmoset be lower weight on a given day? Could it affect sleep, stress, activity, or other factors during the preceding 24 hours? If so, could these other variables be driving the results that are interpreted as 'hunger?' Relatedly, since the reward size is determined by the animals behavior on each trial (how much they worked), factors (internal brain state, external noises, etc.) that alter how much they worked will influence the subsequent reward size. Therefore interpretations about reward expectancy are confounded. Both of these issues should be discussed and manipulations of them (different feeding schedules and reward size-work functions proposed, respectively.
A major issue is a lack of alternative models. The authors seem to have constructed a particular model designed to capture the behavioral patterns they observed in the data. The model fails in some instances, as they point out. Even more importantly, there are no results or discussion about how other plausible models could or couldn't fit the data. The lack of model comparisons makes it difficult to interpret the conclusions or put the results in a broader context.

Reviewer #2 (Public Review):

Hage et al examine how the foraging behavior of marmoset monkeys in a laboratory setting systematically takes into account the reward value and anticipated effort cost associated with the acquisition and consumption of food. In an interesting comprehensive framework, the authors study how experimental and natural variation of these factors affect both the decisions and actions necessary to gather and accumulate food, as well as the actions necessary to consume the food.

The manuscript proposes a computational model of how the monkeys may guide all these aspects of behavior, by maximizing a food capture rate that trades off the food that can be gathered with the effort and duration of the underlying actions. They use this model to derive qualitative predictions for how monkeys should react to an increase in the effort associated with food consumption: Monkeys should work longer before deciding to consume the accumulated food, but should move more slowly. The model also predicts that monkeys should show a different reaction to an increase in reward value of the food, also working longer but moving faster. The authors test these predictions in an interesting experimental setup that requires monkeys to collect small increments of food rewards for successful eye movements to targets. The monkeys can decide freely when to interrupt work and consume the accumulated food, and the authors measure the speed of the eye movements involved in the food acquisition as well as the tongue movements involved in the food consumption.

By and large, the behavioral findings fall in line with the qualitative model predictions: When the effort involved in food consumption increases, monkeys collect more food before deciding to consume it, and they move slower both during food acquisition and food consumption. In a second test, the authors approximate the effects of reward value of the food at stake, by comparing monkey behavior during different days with natural variations in body weight. These quasi-experimental increases in the reward value of food also lead to longer work times before consumption, but to faster movements during food consumption. Finally, the authors show that these effects correlate with pupil size, with pupils dilating more for low-effort foraging actions with increased saccade speed and decreased work duration. The authors conclude that the effort associated with anticipated actions can lead to changes in global brain state that simultaneously affect decisions and action vigor.

The paper proposes an interesting model for how one unified action policy may simultaneously affect multiple types of decisions and movements involved in foraging. The methods employed to measure behavior and test these predictions are generally sound, and the paper is well written. While the model and paper in their present form can clearly inspire researchers to consider this integrated perspective, and trigger further research employing such a framework, there are some conceptual and methodical shortcomings that reduce the conclusiveness of the results and the usefulness of the proposed model.

(1) The model proposed in the paper takes a very specific functional form that is neither motivated by the previous literature nor particularly useful for indexing the behavioral tendencies of individual monkeys (or of the same monkey in different contexts). For example, while it is clear that the saccade effort cost will need to outgrow the increase in the utility of the accumulated food for the monkey to start feeding, it is unclear why this needs to be modeled with a fixed quadratic exponent on the number of saccades? Similarly, why do licks deplete the food stash with the specific rate hard-coded in the model? Finally, the proportion of successful saccades and lick events is assumed to be fixed, even though it very likely to be directly influenced by movement speed (speed-accuracy trade-off), which is also contained in the model. It would strongly increase the plausibility and potential impact of the model if the authors could clearly state where these hard-coded model terms come from. Ideally, they would formulate the model in more general terms and also consider other functional forms, as briefly suggested in the discussion. This latter point would be particularly important since not all model predictions were actually borne out in the data.

(2) The authors derive qualitative predictions, by simulating their model with apparently arbitrary parameters. They then test these qualitative predictions with conventional statistics (e.g., t-tests of whether monkeys lick more for high vs low effort trials). The reader wonders why the authors chose this route, instead of formulating their model with flexible parameters and then fitting these to data. This would allow them (and future researchers) to test their model not just qualitatively but also quantitatively, and to compare the plausibility of different functional forms. The authors certainly have enough data and power to do this, given the vast number of sessions the monkey completed.

(3) The effort manipulation chosen by the authors (distance of food tube) goes hand in hand with a greater need for precision since the monkey's tongue needs to hit an opening of similar size, but now located at a greater distance. This raises the question of whether the monkeys moved slower to enhance its chance of collecting the food (in line with a speed-accuracy trade off). The manuscript would benefit from an explicit test of this possibility, for example by reporting whether for each of the two conditions, the speed of tongue movements on a trial-by-trial basis predicts the probability of food collection? At the very least, the manuscript should explicitly discuss this issue and how it affects the certainty with which effects of tube distance can be linked to anticipated effort cost alone.

(4) The authors report most of the effects on the different measures (work duration, movement vigor, lick vigor, etc) in separate analyses. However, their model predicts that all of these measures result from the same action policy (maximization of the capture rate) and should therefore be related on a trial-by-trial basis. This is so far hardly tested in the presented analyses (with the exception of the pupil correlations in Figure 5). The model's assumed action policy would appear more plausible if the authors could demonstrate these trial-by-trial interrelations with some tests of association (e.g., correlations/regressions as already done for pupil measures in Figure 5) or possibly with dimensionality reduction of the multivariate data.

(5) The manuscript measures pupil dilation in a time period ranging from -250ms before to 250 ms after saccade onset. However, the pupil changes strongly during saccade execution relative to the preceding baseline, leaving doubts as to whether the aggregated measure blurs several interesting and potentially different effects. It would be more conclusive if the manuscript could report the analyses of pupil size separately for a period prior to saccade onset and during/after the saccade.