Towards a Model of the Interplay of Mentalizing and Mirroring in Embodied Communication Sebastian Kahl (skahl@uni-bielefeld.de) Social Cognitive Systems Group, Faculty of Technology Bielefeld University, Inspiration 1, 33619 Bielefeld, Germany Stefan Kopp (skopp@uni-bielefeld.de) Social Cognitive Systems Group, Faculty of Technology Bielefeld University, Inspiration 1, 33619 Bielefeld, Germany Abstract are perception-action couplings modulated by social interac- tion and mentalizing? How do these processes enable coor- The social brain contains distinct networks for mentalizing as dination devices in communication like feedback, joint atten- well as for mirroring-based action observation. We present work towards a model of how these two systems may interact tion, or grounding? during embodied communication. The model connects a men- While a number of mechanisms have been hypothesized, talizing system for attributing and inferring different orders of these questions are far from being answered. What is belief about own and other’s mental states, with a hierarchical predictive model for online action perception and production. undisputed is that interacting with other agents assumed to We discuss interactions between both systems and describe be “intentional” is fundamentally different from interact- simulation experiments in which two agents equipped with this ing with non-intentional things or objects (Gangopadhyay & model engage in embodied communication. Results demon- strate how mentalizing affords higher robustness of communi- Schilbach, 2012). That is, intentionality-attribution and un- cation by enabling interactive grounding processes. derlying mentalizing influence sensory processing to become Keywords: Cognitive Modeling; Social Cognition; Mirroring; “social perception” (Wykowska, Wiese, Prosser, & Müller, Mentalizing; Coordination; Gesture; Embodied Virtual Agents 2014; Teufel, Fletcher, & Davis, 2010). A key component to trigger this intentional stance is social attention, most promi- Introduction nently signaled through gaze (Teufel et al., 2010). For ex- ample, Ciaramidaro, Becchio, Colle, Bara, and Walter (2014) A growing body of research has started to study the cogni- recently found that social gaze leads to the attribution of com- tive mechanisms of social interaction and communication in municative intent, which in turn differentially recruits the humans. Two partially overlapping networks have been iden- mirror neuron system and mentalizing system networks in tified in the “social brain” (Van Overwalle, 2009): an ac- processing the behavior of the (now considered) interlocutor. tion observation system for perceiving and recognizing oth- Clearly, these processes play an important role not only in the ers’ behavioral cues, and a mentalizing system for under- solitary observation events in which they have been studied standing others in terms of attributed mental states or theory mostly so far, but even more so in continuous online interac- of mind (ToM). Both systems have been studied a lot sep- tion (Myllyneva & Hietanen, 2015; Schilbach et al., 2013). arately. Action observation is nowadays widely assumed to In this paper, we present work towards a model of how rest upon principles of prediction-based processing (A. Clark, a mentalizing system interacts with a mirroring-based action 2013). This means that predictions about expected sensory observation system in continuous online interaction. In the stimuli (either caused by an observed stimuli or through remainder of this paper, we first review related modeling at- self-generated action) are continuously formed and evaluated tempts and then present an integrated model that formalizes against incoming sensory input to determine a prediction er- the two systems in terms of computational processes, as well ror that informs further processing. A core mechanism to de- as their roles and dynamic interplay in inter-agent communi- rive such predictions are covert simulations of the observed cation. We report results from simulations of embodied com- behavior, based on own sensorimotor action representations munication with hand gesture, gaze, and head nods/shakes, that are assumed to be shared between perception processes between two virtual agents each of which equipped with the and action production processes. This is what is also re- integrated model. We analyze how different abilities for ferred to as mirroring in behavior perception. The principle mentalizing enable increasingly complex social coordination, of prediction-based processing has also been argued to ac- from mere mimicry to eventually shared understanding. count for language production and comprehension (Pickering & Garrod, 2013) or the social brain more generally (Frith & Related work Frith, 2010). So far there have only been few attempts to combine mental- Yet, what is less clear is the full picture of how the mental- izing, perception and action control in dynamic social inter- izing system and the mirroring system work together: How action. In their MOSAIC model, Wolpert, Doya, and Kawato does behavior perception change when a behavior is assumed (2003) underline that a true communicative model needs to to be “for me”, i.e., intended to be socially meaningful? How close the communicative loop and must be perceptive to the 300 observer’s responses and ultimately her understanding. In sensory processing is associated to area STS. The mentaliz- their model, a hierarchy of paired forward and inverse models ing system is assumed to influence sensory processing in a is hypothesized as a basic mechanism for processing move- top-down fashion, but also to affect sensory gain control in ment as well as beliefs or intentions. Sadeghipour and Kopp attention mechanisms. They report the sensory gain manip- (2011) present the Empirical Bayesian Belief Update model ulation for attentional reorienting mechanisms to be stronger (EBBU), a probabilistic model that implements a mirroring- in the intentional stance than in the design stance. A key as- based account of the perception and production of iconic ges- pect in triggering this intentional stance seems to be social tures. In this model, a hierarchically organized representation gaze, which has been found to lead to the attribution of com- of motor knowledge is used during action perception by for- municative intent (Ciaramidaro et al., 2014), which in turn ward models that formulate probabilistic expectations about differentially recruits the mirroring and mentalizing system possible continuations of the observed gesture. The same rep- networks in processing the behavior of the interlocutor. resentation is used for action generation, with probabilistic interactions between both processes to model e.g. priming Towards an integrated model and resonance effects, and it is expanded by way of inverse In this paper we present a novel model of how a predictive models when an unknown action is encountered. sensorimotor subsystem for action observation and produc- A Bayesian approach to modeling intention inference is tion is coupled with a mentalizing subsystem for attributing presented by Diaconescu et al. (2014). They apply a hier- mental states, to enable situated communication between em- archical Gaussian filter approach to learn the intentionality of bodied agents. To that end, and as described below, we embed others through updating the beliefs about others’ intentions the models in virtual agents and let them interact nonverbally during an interaction game. This framework captures indi- to test how communicative coordination emerges from the in- vidual social learning differences in order to predict the par- terplay between the two subsystems (see Fig. 1). ticipants’ perspective-taking abilities and trustworthiness. A We base our modeling approach on a number of assump- similarly effective approach to modeling a mentalizing mod- tions: First, we define successful communication to be a pro- ule was implemented by Devaine, Hollard, and Daunizeau cess leading into shared communicative intentionality and es- (2014), who applied it to a gambling game scenario. Partici- tablished perceptual or conceptual common ground between pants had to play against a Bayesian mentalizing model that the participants (Tomasello, 2008). This state is achieved in a could employ several levels of ToM belief attributions, from dynamic grounding process (H. H. Clark & Brennan, 1991), zero up to a third-order ToM. Results show that only partici- in which communicating agents mutually present and coordi- pants who were framed to apply ToM could beat the Bayesian nate their subjectively held beliefs about each other as well mentalizing model. as the state of the interaction. Second, mentalizing plays a Few attempts have been made to clarify the interaction be- pivotal role in this through facilitating information integra- tween mentalizing and mirroring. A meta-analysis of studies tion and self-other distinction for coordinated action. It re- on mentalizing found that mirror areas are not recruited un- ceives information from, and affects the mirroring subsys- less the task involves inferencing intentionality from action tem, which itself processes perceived action in an immedi- stimuli (Van Overwalle, 2009). Teufel et al. (2010) present ate fashion and feeds into higher-level mentalizing. Third, the “Perceptual Mentalizing Model” which focuses on the in- we assume that coordinated action in communication highly fluence of the mentalizing system on the mirror system via depends on the order of ToM realized by the mentalizing sub- perceptual processing in the STS area. They differentiate be- system. 2nd order reasoning, i.e. beliefs about other-beliefs, tween explicit and implicit theory of mind (what we call here is minimally necessary for any cooperative behavior that goes mentalizing and mirroring, respectively) and associate the ar- beyond accidental coordination. Finally, gaze plays a special eas mPFC and TPJ to the former and the mirror neuron sys- tem to the later kind of ToM processing. Importantly, both kinds of ToM are assumed to be influenced by social sen- Agent A Agent B sory processing. Explicit ToM processes are associated with Social gaze processing of the intentionality of a movement and have a (comm. intent) Mentalizing Head nod/shake Mentalizing strong influence on area STS, which acts as a gating mech- (meta comm.) anism to increase or decrease perception-action coupling in implicit ToM processing. Mirroring Mirroring Wykowska et al. (2014) present a model of social atten- Gestures tion, the “Intentional Stance Model”, according to which the (referential comm.) mentalizing system either exhibits a design stance or an inten- tional stance. The latter would be exhibited towards agents to whom mental states are attributed, while the former is applied to objects without intentionality attribution. Again, mentaliz- Figure 1: Outline of the overall structure of the simulated ing is also associated to areas mPFC and TPJ, while social communication between two cognitive embodied agents. 301 role in signaling and regulating social attention and commu- observed gesture performance. In each time step, the model’s nicative intent. Mirror areas were found to be recruited es- predictions are compared to the actual perception in order to pecially when intentional action is expected (Van Overwalle, evaluate the corresponding motor commands, programs, or 2009) and one indicator for communicative intention is social schemas. When no corresponding representation of the ob- gaze (Ciaramidaro et al., 2014). We hence assume that gaze served behavior can be found in the knowledge structure, in- triggers mentalizing and thus also mirroring activity. Staying verse models can execute a motor learning mechanism that within the confines of the nonverbal domain, we also include imparts the novel behavior into the motor command graph head-nods and head-shakes as feedback signals for agreement and corresponding motor program and schema, when no sim- and disagreement. ilar schema exists. Our model of a sensorimotor subsystem employs this Mentalizing subsystem EBBU model and is equipped with knowledge about The mentalizing subsystem is a model of an agent’s subjec- schemas, programs and commands of different gesture tra- tive ToM, which processes definite information about itself jectories for ‘circle’, ‘square’, ‘surface’ and ‘waving’. Those and needs to infer others’ mental states from perceptual input. were learned from real human motion data. Single motor pro- A detailed depiction of the mentalizing subsystem is given in grams for a schema can take up to five seconds to produce, Figure 2, which we will refer to and describe in more detail with motor commands being activated every tenth of a sec- in the Simulation results section. In its current version, the ond. For every new observation of a hand trajectory entering subsystem utilizes a rather simple set of inference heuristics the subsystem, posterior probability distributions are updated to model how mental state attributions arise and change in using the EBBU rule (Sadeghipour & Kopp, 2011). The top- social interaction. In detail, this mentalizing model consists most level distributions about motor schemas are taken as a of three sets of mental state attributes for different orders of proxy for a gesture’s meaning, and are linked to 1st order ToM reasoning: Beliefs held about mental states of myself mental state attributes in the mentalizig subsystem. (me) or the interlocutor (you) constitute what we call 1st or- der ToM. In pursuit of a minimal cognitive model of mental- Integration and interplay izing, we assume that only one order of ToM higher is needed Our goal is to integrate mentalizing with mirroring-based ac- for what we want to model. In contrast to the classical recur- tion perception to account for how behavior and mental states sively nested beliefs (beliefs about beliefs about...), however, arise and interact dynamically in a communicative interac- we stipulate these beliefs to be held about mental states that tion. both interlocutors have in common (we). This is what we call In other models of mentalizing (Teufel et al., 2010; 2nd-order ToM. Generally, mental states consist of beliefs, Wykowska et al., 2014), the detection of social gaze plays desires, and intentions. The functional role that we ascribe to a crucial role. In our model, it triggers an inference for com- 2nd order ToM is to keep track of common ground, the desire municative intent to the gazing agent, which in turn affects to agree, and the collaborative state of communication more further processing of its behavior. In particular, as long as generally. the intention to communicate can be inferred, for any ob- served gesture processed by the mirroring subsystem, the Mirroring subsystem most likely motor schema hypothesis is immediately pro- We adopt the Empirical Bayesian Belief Update model jected into the mentalizing subsystem where it forms a men- (EBBU) (Sadeghipour & Kopp, 2011) for action observa- tal state attributed as a you-belief. This resembles the gating tion and production. It implements a probabilistic hierar- mechanism of area STS as suggested by Teufel et al. (2010). chical representation of sensorimotor knowledge about iconic Correspondingly, a me-belief would cause the mirroring sub- gestures, along with basic prediction, evaluation and activa- system to recruit the intended motor schema for production. tion processes that are used both in perception and genera- The current version of the mirroring subsystem is only capa- tion of gestures. On the lowest level motor commands are ble of processing hand gestures; gaze and head movements stored that represent single movement segments in time and are thus directly asserted to the mentalizing subsystem. space. Hand trajectories are given as directed graphs with Depending on the degree of ToM available in the agent’s their edges representing the motor commands. On the inter- mentalizing subsystem, communicative intent can trigger an mediate level, motor programs represent paths in the motor inferred desire to reach mutual agreement about the under- command graph. In that way each motor program stands for a standing of the produced gesture. This is assessed by apply- meaningful movement. The highest level of abstraction stores ing a threshold for good-enough understanding to the like- motor schemas that cluster similar motor programs to rep- lihood of beliefs about mental states of me and you (1st or- resent functional equivalence classes (e.g. depicting similar der ToM). Note, however, when this threshold is exceeded shapes). When observing a gesture trajectory, probabilistic the producer agent still cannot be certain about the correct forward models predict expectations about possible continua- understanding in the recipient unless sufficient feedback is tions of the observed gesture by running internal simulations provided. Here, we require at least one correct reproduction on all levels of the hierarchy in parallel. While simulating, of the gesture. Further, head-shake and head-nod signals are the hierarchical motor knowledge structure “resonates” to the employed for meta-communication and can either increase or 302 Communicator Gaze Recipient Probabilistic variables I M C1: DC ) ICM R1: ICY ) DC Y (Un)specific intention to communicate Y Y M P C2: IC ) DC R2: DC ) ICM Perceived or produced belief C3: D Y ^ ¬P Y ^ P M ) I M (P M ) M Y D C C Presentation phase R3: DC ^ DC ) ICW Desire about communication (me & M C4: DC Y ^ DC ) ICW R4: ICW ^ P 0Y ^ P M ) (P M = P Y ) ^ ICM (P M ) you) or agreement (we) M/Y/W W C5: IC ^ PM ^ PY ) PW Acceptance phase R5: ICW ^ P M ^ P Y ) P W Mental state attribution to me/you/we C W C6: IC ^ ¬P W ) ICM (Hshake ) ^ ICM (P M ) R6: ICW ^ ICY (Hnod ) ) (P Y ) P W ) Communication W A C7: IC ^ ICY (Hnod ) ^ ¬P Y ) ICM (Hshake ) R7: ICW ^ ICY (Hshake ) ^ ICY (P Y ) ) P M = P Y = ; Agreement H W C8: IC ^ P W ) ICM (Hnod ) R8: ICW ^ ICY (Hshake ) ) ICM (P M ) Head signal W C9: IC ^ P W ) DA W R9: ICW ^ P W ) ICM (Hnod ) 1st order Theory of Mind Agreement W W W R10: IC ^P ) DA 2nd order Theory of Mind Figure 2: Attributes and inference heuristics in the mentalizing subsystem applied during different phases of the interaction. The basis for complex inference is “Communicative Intention”, inferred from social gaze. The “Communicator” agent enters the “Presentation Phase”, followed by an “Acceptance Phase” of interactive grounding, where higher order mental attributions are needed for both agents to reach “Agreement”. decrease confidence in the respective you-belief. grounding process with presentation and acceptance phases (H. H. Clark & Brennan, 1991), where the Communicator al- Simulation results ways starts with producing a ‘circle’ gesture. In order to test the model in online interactions we im- To examplify the effect of the mental attributions and infer- plemented the model and ran simulations with two virtual ences possible in 1st and 2nd order ToM, Figure 3 illustrates agents, each of which equipped with its own integrated two typical interaction patterns from our simulation study. model. At the start of the simulation, both agents only have The overt behavior of two agents, a Communicator and a Re- a predefined set of mental states about themselves. They cipient, are shown along with the inferences drawn after per- can communicate using four gestures (‘circle’, ‘square’, ‘sur- ceiving or producing a certain behavior, with indices referring face’, and ‘waving’) that are perceived and generated as 3D to the inference rules as shown in Figure 2. hand trajectories, as well as head nods/shakes that are trans- ferred as simple timed key-value pairs. Gestures are pro- The interaction at the bottom shows a sequence of behav- duced with a configured amount of white noise, normalized ior and inferences typical for 1st order ToM mentalizing. The to the maximum movement vectors in the motor schema, so configured desire to communicate triggers rule C1, hence that 10% noise reflect only a small amount of deviation dur- gaze behavior is perceived by the Recipient (rule R1). Since ing gesturing. The amount of noise, the ability for 2nd or- the Recipient is equally configured, its reciprocal gaze behav- der ToM, and the good-enough threshold for minimal confi- ior (rule R2) triggers an inference about the Recipient’s desire dence in observing a gesture are the independent variables to to communicate in the Communicator (rule C2), and conse- parametrize the simulation. quently a gesture is produced (rule C3). We ran six simulation setups: 10%/20%/30% noise with The interaction at the top shows behavior and inferences enabled or disabled 2nd order ToM capacity, and a static con- enabled through 2nd order ToM. While in the beginning there fidence threshold of 0.8. Each of the setups was run 100 is a similarity to the 1st order ToM interaction, additionally times, always with identically configured agents. Simula- rules R3 and C4 are triggered and establish the agents’ com- tions ended either when both agents believed to have reached mon communicative intent and thus the foundation for mean- agreement, or without 2nd order ToM, as soon as the Com- ingful coordination behavior. After the initial gesture produc- municator finished its gesture production. As dependent tion the Recipient’s mirroring subsystem provides the men- variables we collected the probability distribution of the at- talizing subsystem with the most likely interpretation for the tributed you-belief about a gesture’s meaning after every pro- Communicator’s behavior. That novel behavior triggers rule cessing of a hand trajectory. We were particularly interested R4, by which the Recipient would ideally produce the under- in the effects that different degrees of mentalizing have on stood gesture back to the Communicator, but in this interac- the inter-agent coordination dynamics. The complexity of tion the gesture was understood with a likelihood above the the communication depends on inferred communicative in- good-enough threshold. This triggers rule R5 and R9 as well, tent, signaled via social gaze. As soon as mutual commu- leading to a head-nod. Since the Communicator has no idea nicative intent is established, the simulation follows a typical what the Recipient has understood the head-nod behavior is 303 C2 & C5 & With 2nd order ToM Gesture Communicator C1 Gaze C3 & C7 Head shake C8 & Head nod production C4 C9 R1 & R4 & R6 & Gaze Gesture Head nod Recipient R2 & R5 & Head nod R8 R9 & production R3 R9 R10 t Only 1st order ToM C2 & Gesture Communicator C1 Gaze C3 production R1 & Recipient Gaze R2 Figure 3: Example interactions from our simulation when both agents have 2nd order ToM (top) or 1st order ToM (bottom). Overt behavior is shown along with the triggered mentalizing inferences (gray circles; indices referring to Figure 2). answered by a head-shake (rule C7), which triggers the Re- 1.4 a) Noise comparison b) Mentalizing comparison cipient to produce its understood gesture back to the Commu- * p<0.05 ** p<0.01 1.2 nicator (rule R8). The Communicator understands the ges- 1.0 Belief likelihood ture, which triggers rules equivalent to those in the Recipient (rule C5, C8, and C9), leading to a head-nod, which is equiv- 0.8 alently answered by the Recipient (rule R6, and R9) and fi- 0.6 nalizes the interaction through mutually believed agreement 0.4 (rule R10). 0.2 0.0 Mean belief distribution differences over time 10% 30% 1st order 2nd order 2.0 Kulbach-Leibler Divergence Figure 5: Analyses show a) mean differences between noise 1.5 conditions after 5 sec., and b) mean differences between final 1.0 likelihood about another’s belief in ToM conditions. Head-shake signal 0.5 5 sec interactions 10 sec interactions the effect of noise we compared the success of both agents 20 sec interactions 0.0 reaching the target belief within last three hand position sam- 0 50 100 150 200 Time sec/10 ples before the 5 second mark, averaged over ToM conditions (Figure 5(a)). The comparison shows a significant difference Figure 4: Simulations show KL-divergence between agents’ (t(1198) = 2.4, p < 0.05, d = 0.14) between 10% (M = 0.6, beliefs during interactions of different extent, averaged over SD = 0.4) and 30% (M = 0.7, SD = 0.5) noise conditions noise and ToM conditions. on gesture understanding. Subsequently, we tested the in- fluence of 2nd order ToM, also by analyzing the success of To test the agents’ ability to coordinate with and without reaching the target belief (Figure 5(b)). A comparison of the 2nd order ToM enabled, we analyzed the Kulbach-Leibler final beliefs averaged over all noise conditions with 2nd order Divergence between the probability distributions of the Re- ToM (M = 0.9, SD = 0.27) and without (M = 0.7, SD = 0.45) cipient’s you-belief and the Communicator’s me-belief, i.e. showed that 2nd order ToM leads to significantly more likely the “target belief”. Figure 4 shows the divergence over in- success in coordination (t(598) = 6.8, p < 0.01, d = 0.56). teraction time. Without 2nd order ToM only one gesture was produced within 5 seconds. With 2nd order ToM the duration Conclusions was strongly dependend on the correct understanding of ob- In this paper we have investigated the questions how mere ac- served gestures. The more mistakes, due to noise, the more tion observation needs to be complemented by higher order correction effort emerged and hence longer interactions. Ana- mentalizing, and how those systems need to interact in order lyzed were interactions with length of at least 10 seconds and to account for the dynamic inter-agent coordination mecha- 20 seconds, respectively. These plots show the average suc- nisms that are required for successful communication. Our cess of coordination, especially in longer interactions. To test view is based on the notion that there is a strong difference 304 in the interaction with intentional, versus non-intentional en- ference When We Learn? PLoS computational biology, tities. To that end we augmented a predictive action observa- 10(12), e1003992. doi: 10.1371/journal.pcbi.1003992 tion system with a ‘minimal’ mentalizing model that enables Diaconescu, A. O., Mathys, C., Weber, L. a. E., Daunizeau, distinct mental perspectives corresponding to beliefs about J., Kasper, L., Lomakina, E. I., . . . Stephan, K. E. (2014, me, you, and we. September). Inferring on the intentions of others by hier- Our approach is to explicate possible interactions between archical bayesian learning. PLoS computational biology, mentalizing and mirroring in terms of computational process 10(9), e1003810. doi: 10.1371/journal.pcbi.1003810 models that can be implemented and tested in actual simu- Frith, U., & Frith, C. (2010, January). The social brain: lations of dynamic unfolding interactions. Here, we inves- allowing humans to boldly go where no other species has tigated whether 1st order mental state attributions are suffi- been. Philosophical transactions of the Royal Society of cient to infer the information necessary to successfully act to- London. Series B, Biological sciences, 365(1537), 165– wards a communicative goal, or whether higher order theory 176. doi: 10.1098/rstb.2009.0160 of mind can give a distinct advantage. Our results demon- Gangopadhyay, N., & Schilbach, L. (2012, July). See- strate that mentalizing affords interactive grounding and thus ing minds: A neurophilosophical investigation of the makes communication significantly more robust and efficient. role of perception-action coupling in social percep- However, even with higher order mentalizing capacities, a too tion. Social Neuroscience, 7(4), 410–423. doi: large perturbation of the communicative signals led to long 10.1080/17470919.2011.633754 interaction times due to the inefficient error correcting mech- Myllyneva, A., & Hietanen, J. K. (2015, January). There is anism emerging from both agents’ goal for successful com- more to eye contact than meets the eye. Cognition, 134, munication. Still, we established in a first prototypical mod- 100–109. doi: 10.1016/j.cognition.2014.09.011 eling attempt that mentalizing is crucial for meaningful co- Pickering, M. J., & Garrod, S. (2013, August). An inte- ordination behavior, and success in communication could not grated theory of language production and comprehension. be established without 2nd order mental state attributes. We The Behavioral and brain sciences, 36(4), 329–47. doi: are thus confident that the present framework presents a good 10.1017/S0140525X12001495 basis for further investigation of social cognitive processes, Sadeghipour, A., & Kopp, S. (2011, September). Embodied which Neuroscience is currently not yet able to elucidate. For Gesture Processing: Motor-Based Integration of Percep- one, we aim to provide an account for how agents dynami- tion and Action in Social Artificial Agents. Cognitive com- cally compensate for noise by strategically altering their sig- putation, 3(3), 419–435. doi: 10.1007/s12559-010-9082-z naling behavior. Another question to pursue is how self-other Schilbach, L., Timmermans, B., Reddy, V., Costall, A., distinctions manifest themselves in the sensorimotor system, Bente, G., Schlicht, T., & Vogeley, K. (2013, Au- e.g. when observing other agents while performing an action gust). Toward a second-person neuroscience. The oneself. Behavioral and brain sciences, 36(4), 393–414. doi: 10.1017/S0140525X12000660 Acknowledgements Teufel, C., Fletcher, P. C., & Davis, G. (2010). Seeing other minds: Attributed mental states influence percep- This research/work was supported by the Cluster of Excel- tion. Trends in Cognitive Sciences, 14(8), 376–382. doi: lence Cognitive Interaction Technology ’CITEC’ (EXC 277) 10.1016/j.tics.2010.05.005 at Bielefeld University, which is funded by the German Re- Tomasello, M. (2008). Origins of Human Communica- search Foundation (DFG). tion. Cambridge, Massachusetts: The MIT Press. doi: 10.1353/lan.0.0163 References Van Overwalle, F. (2009, March). Social cognition and the Ciaramidaro, A., Becchio, C., Colle, L., Bara, B. G., & Wal- brain: a meta-analysis. Human brain mapping, 30(3), 829– ter, H. (2014, July). Do you mean me? Communicative in- 58. doi: 10.1002/hbm.20547 tentions recruit the mirror and the mentalizing system. So- Wolpert, D. M., Doya, K., & Kawato, M. (2003, March). cial Cognitive and Affective Neuroscience, 9(7), 909–916. A unifying computational framework for motor control doi: 10.1093/scan/nst062 and social interaction. Philosophical transactions of the Clark, A. (2013, June). Whatever next? Predictive brains, Royal Society of London. Series B, Biological sciences, situated agents, and the future of cognitive science. Be- 358(1431), 593–602. doi: 10.1098/rstb.2002.1238 havioral and Brain Sciences, 36(03), 181–204. doi: Wykowska, A., Wiese, E., Prosser, A., & Müller, H. J. (2014). 10.1017/S0140525X12000477 Beliefs about the minds of others influence how we process Clark, H. H., & Brennan, S. E. (1991). Grounding in com- sensory information. PLoS ONE, 9(4). doi: 10.1371/jour- munication. In Perspectives on socially shared cognition nal.pone.0094339 (Vol. 13, pp. 127–149). Washington, DC, US: American Psychological Association. doi: 10.1037/10096-006 Devaine, M., Hollard, G., & Daunizeau, J. (2014, December). The Social Bayesian Brain: Does Mentalizing Make a Dif- 305