Acquisition of predator knowledge in sooty mangabeys

How do primates learn to communicate? An in ﬂ uential, but largely untested model proposes that primates go through a pruning process, guided by social learning, during which they increasingly restrict alarm calling from initially broad ranges of animals to a few dangerous predators. To test this model, we conducted an experiment with free-ranging sooty mangabeys, Cercocebus atys atys , in which we systematically exposed different age groups to models of dangerous vipers and nonvenomous colubrid snakes. We found that young juveniles perceived all snakes as dangerous and indiscriminately alarm called, although they had the longest response latencies. In contrast, adults showed antipredator behaviours faster to vipers than colubrids but never alarm called to the latter, unlike juveniles. Finally, all young and some older juveniles alternated their gaze between the snake models and other group members, suggesting they engaged in social referencing, that is, gazing at others to assess their reactions to external events. Our study provides a systematic, empirical demonstration that, in nonhuman primates, predator learning starts with overgeneralization, followed by subsequent re ﬁ nement via social learning during the juvenile phase. © 2023 The Author(s). Published by Elsevier Ltd on behalf of The Association for the Study of Animal Behaviour. This is an open access article under the

Most animals share their habitats with species that can be dangerous to them and require specific antipredator responses, often accompanied by acoustically distinct and predator-specific alarm calls (Collier et al., 2017;Kalb et al., 2019;Manser et al., 2002;Suzuki, 2016;Zuberbühler, 2001).Detecting and recognizing predators early is critical for effective antipredator responses, including appropriate alarm calling, but how do young individuals learn to recognize their predators?Predation triggers some of the most basic interactions between organisms and is often considered a major force of natural selection (Abrams, 2001).As predation risk varies in time and space, it is unsurprising that in many species antipredator behaviour appears to result from a combination of rigid species-specific and flexible learned responses (reviewed in Edwards et al., 2021;Griffin, 2004;Kelley & Magurran, 2003;Mery & Burns, 2010).
Research has shown some degree of hardwired recognition and responses to predators for a diverse range of taxa, including invertebrates (Dalesman et al., 2006), fish (Zanuzzo et al., 2019), amphibians (Epp & Gabor, 2008), reptiles (Baxter-Gilbert et al., 2018), birds (Dessborn et al., 2012;G€ oth, 2001) and rodents (Apfelbach et al., 2005).For primates, there is experimental evidence for seemingly hardwired responses towards leopards, Panthera pardus (Schel & Zuberbühler, 2009) and snakes (Barros et al., 2002;Cook & Mineka, 1990;Weiss et al., 2015).Nevertheless, if antipredator responses are fully hardwired, it is not clear what kind of feedback might encourage novel signalecontext associations (e.g.novel predators).A second way to acquire competence in antipredator behaviour is through learning, for example via trial-and-error or other forms of conditioning (Pika et al., 2003).This can be very rapid, as demonstrated in adult West African green monkeys, Chlorocebus sabaeus, exposed repeatedly to an abnormal threat, a remotely operated drone (Wegdell et al., 2019).Subjects rapidly learned to avoid the drone, presumably building on preexisting knowledge from natural encounters with raptors.
Generally, however, predation events are infrequent, fast and unpredictable, which poses the question of how inexperienced individuals can learn to avoid predators if there are no regular learning opportunities and if eventual mistakes are extremely dangerous.One way by which group-living animals can overcome this predicament is by rapid social learning (for a review of social learning about predators by fish, birds, eutherian mammals and marsupials see Griffin, 2004), a process by which naïve individuals learn from interacting with more experienced group members (Deshpande et al., 2022;Galef & Laland, 2005;Heyes & Galef, 1996;Heyes & Galef, 1996;Le on, Thiriau, Bodin, et al., 2022).This is especially beneficial for young individuals, which are often especially vulnerable to predators due to small body size, poor motor skills and inexperience (Curio, 1976;Genovart et al., 2010;Janson & van Schaik, 1993).Learning from more experienced individuals is adaptive because it spares inexperienced individuals from the expensive (and potentially lethal) consequences of learning by trial-and-error.
The bulk of social learning studies, and the theories that have emerged from them, have focused on mechanical problem solving, typically how chimpanzees, Pan troglodytes, learn to open complex puzzle boxes that contain food and require sequences of manipulations, which can be learned most efficiently by observing others (Horner & Whiten, 2005).A much less studied, but arguably equally important form of social learning is by observing how others react to unfamiliar objects or events.This process of actively seeking information from more experienced individuals' responses to external events is termed 'social referencing' (Klinnert et al., 1983).In humans, social referencing has been observed in children assessing the facial expressions of parents to unfamiliar situations (Feinman et al., 1992).Social referencing is especially important for infants and juveniles and has the following features: (1) referential glances between the stimulus and another individual and (2) behavioural changes in response to the information received (Russell et al., 1997).Beside humans, social referencing has been studied mainly in other primates (chimpanzees: Evans & Tomasello, 1986;Itakura, 1995;Russell et al., 1997;vervet monkeys, Chlorocebus pygerythrus: Mohr et al., 2023; capuchin monkeys, Sapajus apella: Morimoto & Fujita, 2012; Barbary macaques, Macaca sylvanus: Roberts et al., 2008;orang-utans, Pongo spp.: Schuppli et al., 2016), domesticated animals (dogs, Canis lupus familiaris: Merola et al., 2012;cats, Felis cattus: Merola et al., 2015; horses, Equus ferus caballus: Schrimpf et al., 2020) and bats, Pteropus spp.(Hall et al., 2011).
Here, we were interested in how terrestrial forest monkeys, sooty mangabeys, Cercocebus atys atys, rely on social learning to acquire knowledge about novel predators encountered naturally.Pioneering research on vervet monkeys suggested that young primates are born with some degree of predisposition towards discriminating between broad predator classes and then, as they grow up, go through a cognitive pruning process during which predator categories and corresponding alarm call behaviour become increasingly refined, guided by social learning from other group members (Seyfarth & Cheney, 1980, 1986).Vervet monkeys produce acoustically distinct alarm calls in response to leopards, eagles and snakes (Price et al., 2015).Field observations suggested that, around 3 months of age, vervet infants begin to give their first alarm calls, but often in incorrect ways until they are older than 2 years, although never in completely arbitrary ways.For example, leopard alarms are produced to a wide range of terrestrial mammals, eagle alarms mainly to airborne objects and snake alarms mainly to snake-shaped objects (e.g.long shaking vine) and also other reptiles, such as tortoises (Seyfarth & Cheney, 1980).But as the infants age, they become more selective and finally restrict their alarm calling and antipredator behaviour to a small number of dangerous local predators.This process is guided by social learning insofar as infants were more likely to respond to alarm calls appropriately if they first looked at adults (Mohr et al., 2023;Seyfarth & Cheney, 1986).Seyfarth and Cheney's (1986) maturational model has been very influential in developmental theories of animal communication and has become somewhat the default theory of primate vocal development, despite the fact that it has hardly been tested empirically in natural settings (for related research see Meno et al., 2013aMeno et al., , 2013b)).This lack of evidence is problematic because communication and cognition have likely coevolved and close to nothing is known about the corresponding developmental patterns.In particular, it is unclear whether ontogenetic changes in primate alarm call behaviour are governed by underlying changes in conceptual organization, that is, an increase in an individual's knowledge of the local fauna (but see Gursky, 2003;Le on, Thiriau, Crockford, et al., 2022;Perry et al., 2003).Also unclear is how effective learning is, particularly whether competence develops rapidly with one or few key experiences (Brodbeck et al., 1992) or gradually over multiple trials, as predicted by standard animal learning theory (Rescorla & Wagner, 1972).There is recent experimental evidence that, in the predation context, monkeys can acquire predator knowledge very rapidly, requiring only one or a small number of experiences (Deshpande et al., 2022;Le on, Thiriau, Bodin, et al., 2022;Wegdell et al., 2019).Other learning appears to be more gradual (Castro & Snowdon, 2000;Fichtel, 2008;Fischer et al., 2000;McCowan et al., 2001;Ramakrishnan & Coss, 2000), although it is often unknown what sorts of experiences individuals have had throughout their early lives.
For most primate species, the main groups of predators consist of carnivores, raptors, snakes and human hunting (Isbell, 1994;Urbani, 2017).In the evolutionary history of the primate lineage, reptiles were most likely among the oldest predators, suggesting that the primate cognitive system has been adapted for snakes before other predator categories ( € Ohman, 2009).Snakes are often cryptically hidden in vegetation, so that effective defence requires adaptations in the visual system ('snake detection theory': Isbell, 2006Isbell, , 2009)).It comes as no surprise that primates, including humans, are especially skilled at detecting snakes, even under difficult visual conditions (Kawai & He, 2016;Kawai & Koda, 2016;Ohman & Mineka, 2001;Shibasaki & Kawai, 2009), with further evidence of efficient processing, attentional prioritization and rapid fear acquisition in response to snakes ( € Ohman, 2009).Although no species of snake specializes in hunting primates, some do occasionally predate on them (Headland & Greene, 2011) and lethal accidents are known to happen (e.g.Foerster, 2008).Since most snakes hunt by stealth and ambush (Greene, 1997), their dangerousness decreases drastically once their location is known and they can be monitored.Interestingly, during interactions with detected snakes, primates generally remain at safe distances and spend considerable time monitoring and sometimes mobbing the snake.Some species have evolved specific alarm calls to snakes but calling is usually highly selective and restricted to a few dangerous species (Ouattara et al., 2009;Ramakrishnan et al., 2005).In chimpanzees and bonobos, Pan paniscus (and probably other species), alarm calls to snakes inform other group members about the snake's location and usually cause careful approach and a reduction in startle responses, indicating less surprise about the presence of the snake (Crockford et al., 2012;Girard-Buttoz et al., 2020;Schel et al., 2013).Such combined responses of prolonged monitoring, alarm calling and low predation risk provide an ideal learning environment for young primates to refine their detection, recognition and response skills (Curio et al., 1978).
Two highly venomous species are common in the Taï Forest, Gaboon, Bitis gabonica, and rhinoceros, Bitis nasicornis, vipers which are frequently encountered by mangabeys, two to three times per week (Le on, Thiriau, Bodin, et al., 2022;Range & Fischer, 2004).Importantly, although there are more than 50 different snake species in the Taï Forest (Ernst & R€ odel, 2002;R€ odel & Mahsberg, 2000), only pythons and the two vipers normally elicit antipredator-specific responses and alarm calls from adult mangabeys.All three snakes are ambush hunters, which do not pursue their prey and are no longer dangerous once discovered and monitored at a safe distance (Crockford et al., 2015).Once mangabeys detect a viper, they typically jump aside or show other types of startle responses.Then, they often stand bipedally and cautiously approach and scan the area around the snake and produce snakespecific alarm calls that attract other individuals (Penner et al., 2008;Range & Fischer, 2004; Fig. A1, Supplementary Video 2).
Here, we examined how mangabeys from different age groups reacted to models of dangerous (viper) and nonvenomous (colubrid) snakes (Fig. 1).Following Seyfarth and Cheney's (1986) refinement model, we predicted that young individuals would be less attentive and less able than adults to recognize viper models as highly dangerous and generally be less proficient in distinguishing between dangerous and nondangerous snakes.We also predicted that, relative to adults, young individuals would engage in more social referencing and be more indiscriminate in both alarm calling and antipredator behaviour.The full definition of social referencing requires evidence of social attention, that is, inexperienced individuals observing more experienced individuals' reactions to external events, followed by some form of behavioural adjustments by the former.In this study, we only report on social attention (Johnson & Karin-D'Arcy, 2006), the key component of social referencing (Klinnert et al., 1986), that is, alternations of gaze between the snake and nearby adults.Finally, we predicted gradual age-related changes, with viper models being increasingly more effective in triggering antipredator behaviour and alarm calling than colubrid models with increasing age.conducted on August 2022.During the main study period, the group was around 87 individuals, including 30 adults (23 females, seven males; >5 years old), 12 subadults (six females, six males; 4e5 years old), 17 old juveniles (nine females, eight males; 3e4 years old), 22 young juveniles (10 females, 12 males; 1e2 years old) and six infants (two females, four males; <1 year old).Snake models were presented to young and old juveniles and adults.Infants were not used for experiments due to the difficulty of presenting the model without the mothers seeing it first.

Presentation of Snake Models
We conducted six trials per age group and model type (viper or colubrid) for a total of 36 trials distributed over an equal number of individuals (12 young juveniles; 12 old juveniles; 12 adults).J.L. conducted 22 trials (four viper model trials; 18 colubrid model trials), F.Q. conducted 13 viper model trials; Auriane Le Floch conducted one viper model trial.We used eight different viper models, crafted from wood, acrylic paint and varnish, and two rubber prefabricated colubrid models (Fig. 1).Models resembled real snakes, positioned with realistic postures to match real-life encounters in the wild.For logistical reasons, two of the viper models had elongated postures, which may be rarer than other postures for a viper in the wild, but subjects did not show biased responses to these two models (Table A1).
Each trial consisted of a snake model presentation to a focal subject.For each trial, we strategically placed the model on the anticipated travel path of the focal subject, in the leaf litter behind a log or bush, prior to the subject's arrival.Two experimenters conducted each trial: the first one placed the model, while the second continuously filmed the focal animal using Panasonic SDR-26 and HC-V500 video cameras, ca.30 s before, during and ca.30 s after the detection of the model.We ended a trial by removing the model after the subject had left the area and no other group member was within visual range.Trials were conducted by locating individuals suitable as subjects, provided they were alone or in small parties away from other group members, usually in the periphery of the group, ensuring that they would be the first individual in the group to see the model.Finally, trials were conducted only if no other predator-related event had occurred during the previous hour.
To avoid habituation and minimize disruption, we conducted only one trial with a viper model and one trial with a colubrid model per day.We did not present snake models more than once every 3 days and usually only twice per week, which was within the monkeys' natural encounter frequencies with Gaboon and rhinoceros vipers (Le on, Thiriau, Bodin et al., 2022).In total, individuals saw snake models while in the audience during one to three trials.During trials, we were careful not to use snake models that the subject had previously seen while in the audience of previous trials.
For each trial, we noted the identity, age and behavioural response of the subject, the type of model and the audience size.As soon as the subject detected the model, we scored the subject's response and determined the composition of the audience.Audience size was defined as the number of individuals within a 10 m range of the subject between the time of model detection and 30 s afterwards, as this radius and time window, respectively, should capture nearby group members that are at potential risk but also ignorant of the snake.

Coding Behavioural Responses
We used a Solomon coder (https://solomoncoder.com) to analyse video recordings on a frame-by-frame basis (25 frames/s) during the first 30 s after model detection.J.L. extracted eight behavioural responses from each video (see Table A2 for detailed definitions).We measured attentiveness by the number of looks, body orientations and looking duration towards the model.Proficiency in snake recognition was measured by the latency from snake model detection to the first occurrence of the onset of any of the extracted behavioural responses.We measured cautious behaviour by the difference in the number of pauses 30 s before and after a model was detected.As indications of social referencing, social attention was measured by counting the scans and scanning duration towards other individuals after gazing at the model.A scan was defined as a change in the head position to the left or right, up or down but beyond a 45 arc of the snake model, provided this was towards another individual and immediately after looking at the snake model.Only abrupt changes in direction were noted, indicated by a prior cessation in head movement.As explained earlier, we only measured social attention (the key element of social referencing), but not any subsequent behavioural changes in response to other's behaviour, the second component of social referencing.Finally, alarm-calling behaviour was measured by assessing whether focal subjects produced snake-specific alarm calls after spotting the models.As the difference between the number of alarm calls given as response to the two model types was too great (59 snake alarm calls in eight viper trials versus four snake alarm calls in four colubrid trials), we chose not to analyse the total number of alarm calls.Behavioural variables were selected to assess (1) how much attention was given to the snake model (attentiveness), (2) how efficient individuals were in detecting and classifying the snake model (recognition proficiency), (3) how much alertness was shown (cautious behaviour), (4) how much information was sought from others (social attention as indication of social referencing), and ( 5) how much alarm-calling behaviour occurred.To estimate interobserver reliability, a second observer (Sasha C ardenas) blindcoded (18/36) 50% of the trials, resulting in excellent interrater reliability (intraclass correlation coefficient range 0.79e1.00;Cohen's kappa for alarm calling ¼ 1; see Table A3 for results for each behavioural response).

Statistical Analysis
To investigate which factors had an impact on snake recognition and antipredator behaviours towards snakes in mangabeys, we used a series of linear (LM) and generalized linear (GLM) models (Bolker et al., 2009) using R version 4.0.3(R Core Team, 2020) and the functions 'lm' (for numerical data) and 'glm' (for count data, family ¼ Poisson; for binary data, family ¼ binomial) of the packages stats and nlme.To reduce multiple testing and redundancy between the behavioural variables, we conducted a factor analysis utilizing the 'factanal' function (also in R stats package).Factor analyses do not reduce the original number of variables (unlike principal component analyses) but allow the identification of key variables (factors) that are causing variation in the observed data.We calculated the number of factors to extract by utilizing the 'fspe' function from the fspe package (Haslbeck & van Bork, 2022).We then chose the variable with the strongest loading from each factor (Factor 1: Number of scans; Factor 2: Number of pauses; Factor 3: Alarm calling; Factor 4: Latency; Table A4).In total, these four factors accounted for almost 75% of the variance in the data (Number of scans: 27.8%; Number of pauses: 20.3%; Alarm calling: 17.0%; Latency: 9.5%).We tested these four variables as response variables in four separate models ('lm': Latency; 'glm': Number of scans, Number of pauses and Alarm calling).To achieve a normal distribution in the latency model, we calculated the square-root of the latency.
Initially, we observed particularly high standard errors for the Alarm-calling model (GLM), indicating complete separation or a convergence issue (HauckeDonner phenomenon).To solve the convergence problem and fit the model, we utilized the 'brglmFit' function from the brglm2 package (Kosmidis, 2023), which fits GLMs using implicit and explicit bias reduction methods (Kosmidis, 2014).We included focal age (i.e.young juvenile, old juvenile, adult) and the type of model (i.e.viper or colubrid) as fixed factors and test predictors.To test the significance of the fixed factors and their relations, we used the 'Anova' function (car package) in each model to perform a type III or type II ANOVA Wald chi-square test, depending on whether there was a significant interaction in the model.Originally, we included the interaction involving the two fixed factors in the full models, to detect whether the behavioural response towards one particular snake model changed with age.We then removed nonsignificant interactions to simplify the models (Engqvist, 2005).Moreover, to check whether 'Audience size' influenced subjects' behavioural responses, we reran all the analyses including this variable as control predictor.Since the results remained unchanged (Table A5), we decided not to include this variable in the models to avoid overparameterization. Finally, for each selected reduced model, we checked for overdispersion and verified homogeneity of variance.
When post hoc analyses were necessary, we conducted pairwise post hoc comparisons between levels of statistically significant control predictors by computing estimated marginal means for each model, using the 'emmeans' function (emmeans package).For these comparisons, we included a Tukey honest significant difference adjustment to account for running multiple tests on the same data of the LM (Feise, 2002).Finally, to account for multiple testing of the same behavioural response, we determined critical P values following the BenjaminieHochberg procedure:

Ethical Note
This research adhered to the regulatory requirements and was approved by the Centre Suisse de Recherches Scientifiques (CSRS), the Minist ere de la Recherche Scientifique and the Minist ere de l'Agriculture et des Ressources Animales and the Office Ivoirien des Parcs et R eserves (OIPR) of Côte d'Ivoire (Permit 19_855).The monkeys were habituated to human observers and had been studied by the research team for 25 years before this study began.We found no indications that the animals were disturbed by the observers' presence.

General Patterns
We were able to observe subjects' responses to the snake models for at least 30 s in all 36 trials (see Supplementary Material).A subject was defined as the first individual to detect the snake model in a given trial.After detecting the model, subjects often grabbed leaves and other vegetal material near the snake to smell them, as if trying to gather additional information through olfactory cues.At least one test predictor (focal age and type of snake model) showed a significant influence on the response variable in three of four statistical models (Table 1).

Development of Predator Knowledge
We first tested latencies to reaction (latency model), as a basic measure of predator knowledge, and found a significant interaction between subject age and model type (latency LM: c2   Table 1, Table A6) or to vipers (post hoc tests: latencies to reaction to vipers: adults versus young juveniles: t ¼ e3.25, P ¼ 0.03; Fig. 2, Table 1, Table A6).
As a second measure of snake knowledge, we tested whether viper models elicited more cautious behaviour (measured by the number of travel pauses taken) than colubrid models across age groups (number of pauses model).Partially supporting our prediction, subjects of all age groups showed a tendency to pause more and move more cautiously after encountering viper models compared with colubrid models (number of pauses GLM: c 2 2 ¼ 3.54, P ¼ 0.059; number of pauses mean ± SE: vipers 3.05 ± 0.23 versus colubrids 2.05 ± 0.23; Fig. 3, Table 1, Table A6;  Supplementary Videos 3, 4).However, contrary to our predictions and despite the fact that old juveniles paused on average more than young juveniles in response to viper models, the differences between the number of pauses elicited by viper and colubrid models were the same in both juvenile age groups (Fig. 3, Table A6).Nevertheless, this difference in cautious behaviour between viper and colubrid models in adults was 62% larger than in juveniles (Fig. 3, Table A6).

Development of Alarm Calling
Subjects alarm called in 12 of 36 trials (33.3%; four of 12 colubrid model trials and eight of 12 viper model trials).Crucially, no adult alarm called to the harmless colubrid model, while two of six (33.3%) young and two of six (33.3%) old juveniles alarm called to this model.In contrast, five of six (83.3%) adults alarm called to the viper models, whereas again only two of six (33.3%) old juveniles and one of six (16.7%) young juveniles alarm called (see Supplementary Material).In the Alarm-calling model, we found a significant interaction between subjects' age and model type (alarm calling GLM: c 2 2 ¼ 8.16, P ¼ 0.016).Partially supporting our prediction, post hoc tests showed that young juveniles tended to produce alarm calls in fewer viper model trials than adults (number of young juveniles versus adults that gave alarm calls during viper model trials: z ¼ 1.84, P ¼ 0.06; Fig. 4, Table 1, Table A6; Supplementary Video 3).On average, callers gave 4.91 calls per encounter during viper model trials, but when counting only viper trials on adults the average number of calls per individual per trial increased to 7.66 calls (an increment of 64%).Moreover, adults produced 2.7 more alarm calls than juveniles (46 calls in five trials versus 17 calls in seven), although they restricted their alarm calls to viper models (number of adults that gave alarm calls during viper versus colubrid trials: z ¼ e2.06, P ¼ 0.03; average number of calls produced by adults per viper encounter: 7.66; range 0e13 calls; Fig. 4, Table 1, Table A6, Supplementary Videos 3, 4), while young juveniles alarm called evenly to colubrid (two calls in two trials) and viper models (two calls in one trial).Similarly, old juveniles produced alarm calls in two trials for each type of snake model, but the number of calls elicited by viper models varied considerably (range 1e10).Every time an individual (regardless of age) produced snake alarm calls, other individuals responded by approaching and looking actively for the snake model.When other group members arrived and detected the model, they cautiously approached further and scanned the surrounding area, and then passively monitored the model from about 1e2 m, sometimes bipedally and sometimes by sitting on elevated places close by.We did not observe specific snake-directed behaviour (e.g.mobbing behaviour, threat gestures) towards the models.Finally, during both natural snake encounters and trials, young individuals, especially infants, showed particular interest in the conglomeration of mangabeys around the snake location, became more aroused, usually approached the area by climbing low branches around it and stayed around the snake's whereabouts longer than older individuals.

Development of Social Attention
For social attention, that is, number of head scans towards other individuals immediately after looking at the snake model, we found a significant age effect (c 2 2 ¼ 6.23, P ¼ 0.04): once young juveniles detected the snake, they searched for and looked at other group members more than adults did (number of scans GLM: estimate ¼ 0.62, SE ¼ 0.25, z ¼ 2.4, P ¼ 0.015; number of scans mean ± SE: young juveniles 3.58 ± 0.28 versus adults 1.91 ± 0.52; Fig. 5, Table 1, Table A6; Supplementary Videos 3, 4).

Audience Effects
Finally, we also looked at audience effects since audience size may impact on social learning and social attention opportunities.We found that audience size did not vary between model conditions (viper models trials: mean ± SE: adults 1.16 ± 0.40 versus old juveniles 1.66 ± 0.56 versus young juveniles 2.16 ± 0.65, range 0e4; colubrid models trials: mean ± SE: adults 1.66 ± 0.42 versus old juveniles 1.66 ± 0.49 versus young juveniles 2 ± 0.44, range 0e3; see Supplementary Material).Moreover, when used as a control predictor, audience size had no significant influence in any of the statistical models and was therefore excluded from the analyses to avoid overparameterization.

DISCUSSION
West African forest primates are exposed to dozens of different snake species.Most of them are not dangerous but a small number can cause fatal accidents, and previous research has suggested that primates recognize the truly dangerous species.We investigated the development of this knowledge with an experimental field study on free-ranging sooty mangabeys in the Taï National Park, Côte d'Ivoire.In doing so, we tested an influential model of primate vocal development, the hypothesis that individuals undergo a pruning process of cognitive refinement with corresponding changes in antipredator and alarm-calling behaviour.We tested this learning model by using realistic snake models of highly dangerous vipers and nonvenomous colubrids to explore agerelated snake recognition and antipredator behaviour.Our results showed that adults showed signs of competence by reacting to viper models faster than juveniles and by only producing alarm calls to the dangerous vipers.However, we also found that, across age groups, viper models elicited more cautious behaviour than colubrid models.When exposed to viper models, all young juveniles and some old juveniles alternated their gaze between the viper models and other individuals and they also had longer response latencies, suggesting that they relied on more experienced individuals before deciding how to respond to snakes, an indicator of social referencing.
Mangabeys are vulnerable to snakebites by vipers which can cause lethal venom poisoning (J.Le on & F. Quintero, personal observations; Supplementary Video 1), which may have caused strong selection pressure ( € Ohman, 2009).Snake predator core knowledge, in the form of fast recognition and identification, is essential for preventing accidents, particularly for camouflaged species such as vipers.Despite being harder to detect, viper models elicited the shortest response latencies in all age groups (twice as fast as to colubrid models).Some studies have shown that naïve primates with no prior experience with snakes exhibit fear responses when presented with snake stimuli (Barros et al., 2002;Cook & Mineka, 1990;Weiss et al., 2015).Therefore, this difference in the latency of response to viper and colubrid models could be seen as part of a strong selection for the evolution of innate viper recognition in mangabeys.However, this disparity was caused by the shorter latency exhibited by adults towards viper models compared to subjects of all ages that were exposed to colubrid models (Fig. 2).Additionally, adults also reacted faster than young juveniles when spotting viper models.On the other hand, the fact that even 1e2-year-olds moved more cautiously in response to viper than colubrid models suggests that monkeys are prepared to recognize dangerous snake species and that discrimination learning starts early (Mineka, 1992;Olsson & Phelps, 2007).Altogether, these results highlight the importance of experience in the development of snake predator knowledge in mangabeys.Future studies on infant mangabeys should address the extent to which this difference is explained by a 'module of fear' (Ohman & Mineka, 2001), where a fearful predisposition towards vipers might have facilitated their rapid and efficient recognition by mangabeys.
Overall, subjects gave an average of 4.9 alarm calls per encounter during viper model trials, similar to what has been reported in a previous study (3.7 calls; Mielke et al., 2019).In line with our predictions, both young and old juveniles (occasionally) alarm called but then to both dangerous and harmless snake models, whereas adults regularly alarm called, but only to dangerous vipers (overall alarm call response rates: young juveniles: 25.0%; old juveniles: 33.3%; adults: 41.7%; N ¼ 12 each).In sum, adults not only produced more snake alarm calls but also restricted their calls to viper models.These findings are in line with the idea that subjects go through a process during which they become more selective in their alarm-calling behaviour during predator encounters (Seyfarth & Cheney, 1980, 1986).Interestingly, infant vervet monkeys begin to correctly produce alarm calls at around 4e6 months of age, but do not regularly use them in the appropriate context until they are older than 2 years.In contrast, it takes more than 4 years for mangabeys to restrict their snake alarm calls to dangerous snakes.These differences suggest that, independently from eventual innate predisposition, vocal production and usage might develop in species-specific ways, with some species alarm calling from early on and other species refraining from alarm call production until cognitively competent.
Our study is in line with previous research showing that social attention plays an important role in how inexperienced animals learn to associate their signals with the appropriate referents (Deshpande et al., 2022;Le on, Thiriau, Crockford et al., 2022;Mohr et al., 2023;Seyfarth & Cheney, 1986).In these studies, young primates were more likely to respond appropriately if they first looked at more experienced group members, a form of social referencing (Baldwin & Moses, 1996;Evans & Tomasello, 1986).Additionally, rhesus monkeys, Macaca mulatta, that were naïve and initially fearless of snakes came to avoid them after observing the aversive response of a more experienced individual towards a snake (Mineka & Cook, 1993).Altogether, these results suggest that the process through which young monkeys acquire predator recognition, avoidance and alarm call knowledge is driven by active social learning.
In the case of mangabey snake-specific antipredator behaviour, Seyfarth and Cheney's (1986) learning model appears to contain at least three distinguishable steps of refinement.First, an overgeneralized and quasi-silent acquisition phase, where young individuals are born with general knowledge of predator classes that allows them to classify, via overgeneralization, all snake-like species and objects around them as a snake type of danger (e.g.lizards, tortoises, crocodiles, amphibians, small logs), and cautiously respond to them (Fig. 3).During this phase, infants produce a few snake alarm calls to a broad range of stimuli that resemble the colour or the shape of vipers, such as a toad (Fig. 4, Supplementary Video 5).In the second step, young mangabeys learn that not all snakes are dangerous and start to discriminate between vipers and other snakes.During these first two phases, social attention (Fig. 5) seems to be important for the subsequent refinement of snake antipredator behaviour, suggesting that social learning is involved in this fine-tuning process.Therefore, it is likely that learning about predator classes begins when infants start moving independently between 6 and 12 months old.Later, by the time juveniles are 2 years old, they have already learned to ignore snakes other than vipers, probably from observing more experienced individuals.Finally, in the third step, after developing cognitive competence, mangabeys direct snake-specific antipredator behaviour, including fast snake detection and classification and active alarm-calling behaviour, almost exclusively to vipers (Figs 2 and 4).Experiments on wild capuchin monkeys revealed similar results.Even though capuchins started to exhibit predator detection and alarmcalling behaviour as young as 4 months old, snake-species discrimination did not become apparent until the juvenile stage (2e5 years old; Meno et al., 2013a).Future studies should address whether young mangabeys sharpen their snake antipredator behaviour rapidly in response to a very few key experiences or gradually over multiple experiences across the juvenile stage.
Social beings are affected by their social environments, and one manifestation of this is audience effects.Empirical studies providing evidence of changes in the signalling behaviour of individuals caused by the mere presence of others have been conducted with fish (Doutrelant et al., 2001), birds (Karakashian et al., 1988;Marler et al., 1986;Meaux et al., 2023) and mammals (Dunlop, 2016;le Roux et al., 2008), including nonhuman primates (Crockford et al., 2012;Girard-Buttoz et al., 2020;Wich & Sterck, 2003).In mangabeys, the first individual to detect a snake is very often also the first caller (Mielke et al., 2019).Snake alarm calls seem to facilitate snake detection by other individuals by increasing awareness and distance detection.Other experimental studies with viper models in mangabeys have shown that subjects are more likely to call if they had not heard other alarm calls before and if they were with small audiences, reinforcing the notion that during viper encounters alarm calling is crucial to ensure that as many individuals as possible become aware of the snake's location (Mielke et al., 2019;Quintero et al., 2022).Although mangabeys clearly adjusted their snake antipredator behaviour to their audiences, we are confident that this variable did not have any important undetected effect on our results since audience size, when used as a control predictor, showed no significant influence on any of the statistical models and audience sizes did not vary considerably between conditions (see Supplementary Material).
We have shown that mangabeys' snake-specific discrimination and competent snake antipredator behaviour, including alarm calling, towards dangerous vipers arise, and are refined, during the juvenile stage.However, we did not determine the exact age at which infants begin to recognize the different snake species they are exposed to, nor the exact age at which they start to display snake antipredator behaviour.This should be addressed in future research to test which socioecological and cognitive features may shape mangabeys' snake species discrimination from early on.Moreover, our stationary snake models could be more similar to Gaboon and rhinoceros vipers, which are highly static snakes, unlikely to move even in response to agitated monkey mobbing behaviour (Quintero et al., 2022), than to more mobile colubrid species.Studies in vervet monkeys and wild jackdaws, Corvus monedula, have shown that the animacy of the objects (e.g.potential predator models) influences subjects' responses (Deshpande et al., 2022;Greggor et al., 2022).Future experiments, taking the movement pattern of snake models into account, are needed to examine the effect of animacy in monkeys' snake species discrimination.
One shortcoming of our study was not to include python models.Although mangabeys probably only rarely come into contact with pythons, they are confirmed primate predators (Headland & Greene, 2011;Jaffe & Isbell, 2010;Khamcha & Sukumal, 2009;Struhsaker, 1967;van Schaik et al., 1983).If encounter rates and perceived threat influence recognition learning and antipredator responses, python models would have been a valuable addition.Finally, because of the recruitment nature of snake alarm calls in mangabeys and the way they acquire their snake-specific antipredator behaviours, the mangabeyeviper system may be an interesting model to investigate the conditioning and evolution of fear in primates and the role of social learning in how they become communicatively competent in the usage and comprehension of their alarm calls.
Cognition is often defined as the capacity to gather, process and retain information, which can subsequently be used to make decisions (Shettleworth, 2010).Studying cognitive evolution, however, is notoriously hard because neither brains nor behaviours fossilize.Thus, the highly selective and restricted nature of snakespecific antipredator behaviour exhibited by adults allowed us to assess the development of cognitive competence in mangabeys.Snakes were among the earliest and most tenacious threats during the evolutionary history of primates and require, as demonstrated by this study, cognitive development to allow a safe coexistence with them (Isbell, 2006;€ Ohman & Mineka, 2003).In the Taï Forest, mangabeys are exposed to a large number of snake species, but only three elicit antipredator responses, suggesting that primates have evolved a capacity to learn to avoid false positives.
In sum, we have empirically corroborated Seyfarth andCheney's (1980, 1986) learning model.Our results indicate that learning about snakes may begin with an overgeneralized response to a wide variety of species, including some nonthreatening ones, that is winnowed down, probably via social learning, into a response directed towards specific dangerous species.This finding highlights that snake species recognition and threat assessment are prerequisites for effective snake avoidance.Our results also indicate that snakes, particularly vipers, are perceived by nonhuman primates as an important threat from an early age, but also that fast identification and alarm calling towards dangerous snakes are sharpened with experience.Furthermore, our study also builds on the role of the social environment in the development of snakedirected antipredator behaviour in primates (Meno et al., 2013b).Overall, these findings are aligned with the theories that state that competence in antipredator behaviour is acquired through social learning during the juvenile stage by refining preexisting snake fear (Griffin, 2004;Mineka & Cook, 1988;Seyfarth & Cheney, 1980), and that snakes as predators have played a central role in the evolution of primate cognition (Isbell, 2006(Isbell, , 2009)).

Figure 1 .
Figure 1.Snake models.Illustrations of (a, b) two viper models and (c, d) two colubrid models.Viper models were custom-made using wood, acrylic paint and varnish, while colubrid models were prefabricated and made of rubber.Photo credits: (a) Fredy Quintero, (b) Ferdinand Bele and (c, d) Cl ementine Bodin.

Figure 2 .
Figure 2. Development of predator knowledge: latency to reaction.Violin plots illustrating the kernel density distribution of the latency to exhibit a behavioural response during the first 30 s after the detection of the snake model, depending on age.Embedded black dots and vertical lines indicate means and 95% confidence intervals from model estimation, respectively.Young and old juveniles and adults N ¼ 12 each.

Figure 3 .Figure 4 .Figure 5 .
Figure 3. Development of predator knowledge: cautious behaviour.Violin plots illustrating the kernel density distribution of the number of pauses during the first 30 s after the detection of the snake model, depending on age.Embedded black dots and vertical lines indicate means and 95% confidence intervals from model estimation, respectively.Young and old juveniles and adults N ¼ 12 each.

Figure A1 .
Figure A1.Representative samples of sooty mangabey alarm calls produce in response to snakes, depending on age: (a) young juvenile, (b) old juvenile and (c) adult.Spectrograms were made using Raven and the following settings: 512 FFT, Hamming window, 75% overlap, 22.05 kHz sampling frequency.

Table 1
Influence of predictor variables on behavioural responses to snake models Parentheses denote the variable level that reflects the estimate when tested against the alternative level: old and young juvenile versus adult, viper versus colubrid.All P values < 0.05 are in bold and are below the critical value determined by the BenjaminieHochberg correction for multiple testing and hence considered as evidence for a significant effect.*P < 0.05; **P < 0.01.a t values are shown for the Latency model (LM).b Omitted given that there is no interpretable result.

Table A1
Subjects' behavioural responses in viper trials (N ¼ 18), depending on model posture (coiled versus elongated) Number of times the subject's head turned and paused within a 30 arc of the snake model location Number of body orientations Number of times the subject turned more than halfway from the direction of travel towards the snake model location as a sign of high attentiveness Looking duration Total duration of all looks within a 30 arc of the snake model location Latency Time between snake model detection and first occurrence of onset of response Number of pauses Difference in the number of pauses during the 30 s after and before the model was detected.Pauses were defined as halts in walking or climbing caused by all four limbs stopping movement at the same time Number of scans Number of times the position of the subject's head changed while looking left or right, up or down outside a 45 arc of the snake model, gazing towards other individuals immediately after looking at the model.Only abrupt changes in direction were noted, indicated by a prior cessation in head movement Scanning duration Total duration of all scans after the detection of the snake model Alarm calling Whether or not subjects produced snake-specific alarm calls after spotting the snake model Intraclass correlation coefficient (ICC) is used for continuous variables and Cohen's kappa (K) for the binary variable.To estimate interobserver reliability, a second observer blind-coded 50% of the trials.Results indicate excellent interrater reliability.A4 Factor loadings of behavioural response variables Table A6 Response variables and descriptive statistics from 36 trials with colubrid and viper snake models SE per age group per model type.