Distinct Neural Representations of Decision Uncertainty in Metacognition and Mentalizing

Metacognition and mentalizing are both associated with meta-level mental state representations. Specifically, metacognition refers to monitoring one’s own cognitive processes, while mentalizing refers to monitoring others’ cognitive processes. However, this self-other dichotomy is insufficient to delineate the two high-level mental processes. We here used functional magnetic resonance imaging (fMRI) to systematically investigate the neural representations of different levels of decision uncertainty in monitoring different targets (the current self, the past self, and others) performing a perceptual decision-making task. Our results reveal diverse formats of intrinsic mental state representations of decision uncertainty in mentalizing, separate from the associations with external information. External information was commonly represented in the right inferior parietal lobe (IPL) across the mentalizing tasks. However, the meta-level mental states of decision uncertainty attributed to others were uniquely represented in the dorsomedial prefrontal cortex (dmPFC), rather than the temporoparietal junction (TPJ) that also equivalently represented the object-level mental states of decision inaccuracy attributed to others. Further, the object-level and meta-level mental states of decision uncertainty, when attributed to the past self, were represented in the precuneus and the lateral frontopolar cortex (lFPC), respectively. In contrast, the dorsal anterior cingulate cortex (dACC) consistently represented both decision uncertainty in metacognition and estimate uncertainty during monitoring the different mentalizing processes, but not the inferred decision uncertainty in mentalizing. Hence, our findings identify neural signatures to clearly delineate metacognition and mentalizing and further imply distinct neural computations on the mental states of decision uncertainty during metacognition and mentalizing.


41
Humans are social beings. We interact with others not only in the physical world 42 but also in the mental world. Differing from objects, humans are free and 43 intentional agents who can hold their mental states that are not direct reflections 44 of reality. The human brain thus needs to concurrently represent mental states 45 reflecting the physical world and the mental worlds of both the self and others [1- representations.

51
A principal criterion to distinguish non-social activities from social activities is 52 whether the activity is conducted towards the self or others [6][7][8][9]. A 53 corresponding distinction is also drawn for mental state attributing processes: 54 monitoring one's own cognitive processes is referred to as metacognition [10], 55 but when the target subject is an intentional agent other than the self it is referred 56 to as mentalizing [11][12][13]. Although both metacognition and mentalizing involve 57 meta-representations of the mental world [1][2][3], the representational formats and 58 available information is only from external cues. That is, the momentary mental 70 states of the past self cannot be inspected as in metacognition but can be only 71 inferred. It thus becomes ambiguous whether the mental state representations 72 are similar to metacognition or mentalizing. Like the two-order hierarchy of type 1 73 (object-level) and type 2 (meta-level) mental states and processes in 74 metacognition [10], the mental states attributed to others in mentalizing could 75 also be hierarchically categorized (Fig 1). Thereby, the representations of the 76 two-level mental states even for the same target in mentalizing might be separate. 77 Further, during monitoring others' object-level performance, the evoked meta-78 level mental states are actually tied to the observer, rather than the target 79 subjects. In this sense, the mental state representations are more similar to those 80 in metacognition than mentalizing. Hence, in many situations the distinctions 81 between metacognition and mentalizing along the self-other dichotomy are 82 ambiguous. Therefore, we looked to neural signatures to clearly delineate 83 metacognition and mentalizing.
84 Surprisingly, although a number of disparate studies on the neural 85 mechanisms of metacognition and mentalizing have been conducted in cognitive 86 neuroscience [18,19] and social neuroscience [20,21], respectively, a direct 87 comparison of the two neural processes is so far lacking. This situation might be 88 primarily due to the lack of an appropriate experimental paradigm applicable for 89 both processes. The mental state that is mainly addressed in studies of 90 metacognition is decision uncertainty (the meta-level mental state), that is, the 91 extent to which one subjectively believes that one's own decision is incorrect 92 (decision inaccuracy, the object-level mental state). Decision uncertainty is an 93 endogenous control signal for improving one's decision even with no external 94 feedback [18,22]. Importantly, it also serves as a critical social signal for 95 efficacious decision improvement in joint decision-making [23,24]. Hence, it is of 96 great importance to understand how to attribute the mental states of decision 97 uncertainty to the target subjects other than the current self in mentalizing. 98 In the current study, we aimed to delineate the neural representations of 99 different levels of decision uncertainty attributed to different target subjects: the 100 current self, the past self and others. To reveal the relationship between 101 metacognition and mentalizing, we adapted a task paradigm often used in 102 metacognition to apply to mentalizing. We used functional magnetic resonance 103 imaging (fMRI) to characterize the neural representations of object-level (type-1) 104 and meta-level (type-2) intrinsic mental states of decision uncertainty attributed to 105 others and the past self in mentalizing, separate from the associations with

123
Task paradigm 124 We carried out three fMRI experiments investigating the mental state 125 representations of decision uncertainty in metacognition and mentalizing (Fig 2a).

126
Twenty-eight healthy participants took part in all of the experiments (see 127 random moving dots and rated her subjective uncertainty about the decision (Fig   129   2b). There were four different task difficulty levels randomly mixed in the task (S1  To avoid evoking the participant's own decision uncertainty, the stimuli presented 141 to the participant were noiseless: Only coherently-moving dots were moving, 142 whereas randomly-moving dots remained stationary. By virtue of this altered 143 stimulus presentation, the participant could perceive the task difficulty without 144 evoking her own decision uncertainty (Fig 2c). This was a necessary condition to 145 dissociate the neural representations of decision uncertainty in mentalizing from 146 those in metacognition. Further, the AO's response time (RT) was reported to the 147 participant by a progress color bar, whereas neither the AO's choice nor the 148 reported decision uncertainty was presented to the participant. Hence, the 149 participant could only use the external information of the task difficulty and the 150 RT to estimate the AO's decision inaccuracy. In a parallel task, the participant 151 alternatively observed performance on the RDM task previously done by the self 152 and judged the past-self decision inaccuracy (PS-DI). Otherwise, the 153 experimental procedure was identical to the AO-DI task. Notably, as the decision 154 variables of the past self were also inaccessible, the underlying process in both 155 tasks should be mentalizing. We refer to these tasks as the type-1 mentalizing 156 tasks.
In experiment 3 (Fig 2d), the experimental procedure was identical to the 158 type-1 mentalizing tasks (experiment 2), but the participant estimated the 159 AO/PS's meta-level mental states of decision uncertainty in each trial (AO-DU 160 and PS-DU). That is, "I believe that the target is this uncertain in her decision." 161 The two tasks thus also entailed mentalizing to attribute mental states of decision 162 uncertainty to the AO/PS. We refer to these two tasks as the type-2 mentalizing 163 tasks.

164
The task sequences of all the mentalizing tasks were identical, only the 165 instructions differed. Thereby, any behavioral and neural differences between 166 them should be caused by different mentalizing processes. The task sequences 167 of the metacognition task and the mentalizing tasks were also quite similar.

168
However, the differences between the two types of tasks existed in both the 169 perception (object-level) phase and judgment (meta-level) phase. We mainly According to the decision-making theories [25], decision inaccuracy is crucially 182 dependent on both task difficulty and RT (Fig 2e). The higher the task difficulty 183 and the longer the RT, the higher the decision inaccuracy. For the sake of 184 simplicity, decision inaccuracy is assumed to be a sigmoid function of task 185 difficulty and RT (Fig 2f). Hence, it is plausible to estimate decision inaccuracy 186 and decision uncertainty in the mentalizing tasks merely from the external 187 information. However, one indispensable process to distinguish mentalizing from 188 non-social inferences or associations is the target subject's perspective taking.

189
For this purpose, in the type-1 mentalizing tasks (AO-DI and PS-DI), the 190 participant should consider that the target subject has unique internal noise ( ! ) 191 during the perceptual decision-making process as described by the drift-diffusion   Notably, the estimates in the mentalizing tasks relied more on task difficulty than 218 RTs (two-tailed paired t-test, t 27 = 6.2; P = 2.5 × 10 -8 ), while the estimates of 219 decision uncertainty in the metacognition task relied more on RT than task 220 difficulty (two-tailed paired t-test, t 27 = 4.1; P = 5.9 × 10 -5 ; Fig 2e), probably 221 because the task difficulty (by the nature of the experimental design) were clearly 222 discerned in the mentalizing tasks, but not in the metacognition task. The  However, all these correlations disappeared after the associations with the 227 external information of task difficulty and RT were regressed out (two-tailed t-test, 228 Ps > 0.30; Fig 2f). That is, the residuals in each of the mentalizing tasks did not  However, after the associations with the external information were regressed out, 236 the residual AUROCs (measured by the estimate residuals) were no longer 237 significantly different from the chance level in each of the mentalizing tasks (two-238 tailed t-test, Ps > 0.20), but in the metacognition task it remained significant and 239 as large as 0.62 (two-tailed t-test, t 27 = 11.8; P = 3.6 × 10 -12 ; Fig 2g). Thus, 240 reliable estimates of decision inaccuracy/uncertainty in the mentalizing tasks 241 were crucially dependent on the external information provided by the task 242 difficulty and RT, which had stable associations with the inaccessible mental 243 states. In striking contrast, the reported decision uncertainty in the metacognition 244 task relied heavily on internal information that the task difficulty and RT could not 245 explain. Similar to the previous argument in animals [16], it was hard to discern whether the mental state attributions in mentalizing were merely achieved 247 through association between external information and the covert mental states.

248
Nonetheless, the estimate residuals accounted for about half of the total variance 249 in each of the mentalizing tasks, though much lower than the ratio of 0.74 in the 250 metacognition task (two-tailed paired t-test, t 27 = 3.5; P = 3.3 × 10 -4 ; Fig 2h).

251
Importantly, the residual variances in the type-2 mentalizing tasks were 252 significantly larger than those in the type-1 mentalizing tasks (two-tailed t-test, 253 AO: t 27 = 2.5; P = 0.0096; PS: t 27 = 2.1; P = 0.023). These extra variances in the 254 type-2 mentalizing tasks were probably generated by the additional process of 255 the target subject's perspective-taking, as suggested by the theoretical analyses 256 described above.

257
Common neural representations of external information in mentalizing 258 As the external information of task difficulty and RT contributed equivalently to 259 the estimates across the mentalizing tasks, we first tested the hypothesis that 260 there are stable neural correlates of the two external social cues. Further, as 261 utilizations of the external information were the same across the mentalizing 262 tasks, the neural representations of each external cue might be also the same.

263
To test these hypotheses, we regressed the trial-by-trial fMRI activity during the 264 judgment phase with the order of the task difficulty and RT across the whole 265 brain in each task (see Methods). Across all mentalizing tasks, fMRI activity in 266 the primary visual cortex (V1) was negatively correlated with task difficulty 267 (conjunction analysis, z > 2.6, P < 0.05 after cluster-level family-wise error (FWE) 268 correction; Fig 3a), decreasing as the number of moving dots was reduced 269 (increasing task difficulty). On the other hand, fMRI activity in the right IPL was 270 positively correlated with task difficulty (conjunction analysis, z > 2.6, P < 0.05 271 after cluster-level FWE correction; Fig 3a). In contrast, the fMRI activity in a wide 272 range of brain regions was positively correlated with RT (conjunction analysis, z 273 > 2.6, P < 0.05 after cluster-level FWE correction; Fig 3b). Among these brain 274 regions, the right IPL region also overlapped with the regions associated with 275 task difficulty: the same right IPL region responded to both task difficulty and RT in the mentalizing tasks (Fig 3c and 3d). Thus, integration of the two pieces of 277 external information in the right IPL partially predicted decision 278 inaccuracy/uncertainty.

279
Distinct neural representations of estimate residuals in mentalizing 280 According to our theoretical analyses as described above, additional unique 281 processes should be involved in each of the mentalizing tasks besides the 282 associations with external information. To explore the underlying neural 283 correlates, we regressed the estimate residuals with trial-by-trial fMRI activities 284 during the judgment phase across the whole-brain voxels in each task. The   Table). In the type-1 mentalizing task, where the 295 participant estimated the AO's decision inaccuracy (AO-DI), the estimated 296 residuals were also correlated with the fMRI activities in the left TPJ and the left 297 IFJ (z > 3.1, P < 0.05 after cluster-level FWE correction, Fig 4b and 4e; see also 298 S1 Table), but not in the dmPFC (t 27 = 1.1, P = 0.12; Fig 4e), suggesting that 299 dmPFC was selectively involved in type-2 mentalizing. To further test the 300 reliability of the dmPFC selectivity in type-2 mentalizing, we repeated the same 301 GLM analysis on the dmPFC and TPJ regions independently defined by meta-302 analytical maps from the NeuroSynth database [27], as well as the conjunction 303 regions between the meta-analytical regions and those in the current study. In 304 both analyses, we obtained results consistently supporting that dmPFC but not 305 TPJ showed activity selective to type-2 mentalizing (S4 Fig). 306 Instead, the estimate residuals of the PS decision uncertainty in the type-2 307 mentalizing task were selectively correlated with the fMRI activities in the right 308 lFPC (z > 3.1, P < 0.05 after cluster-level FWE correction, Fig. 4c  with the metacognition-associated areas reported in prior studies [18,19].

314
Consisting with our theoretical account, selective neural representations of 315 estimate residuals in the mentalizing tasks showed that these residuals were 316 partially but reliably correlated with diverse brain activities during mentalizing in 317 different contexts. Notably, the right IPL that encoded task difficulty and RT did  Table), as repeatedly 327 observed in previous studies [18,19,28]. Although the lFPC region was shared 328 with type-2 mentalizing (PS-DU), the dACC region selectively represented the 329 estimate residuals in the metacognition task but not in the mentalizing tasks.

330
Notably, the components of decision uncertainty associated with task difficulty 331 and RT in the metacognition task were also represented in the dACC (S5b Fig). 332 Thus, the dACC uniformly represented the components of internally-generated 333 decision uncertainty in metacognition.

334
Common neural representations of estimate uncertainty in mentalizing 335 In the mentalizing tasks, the use of associations with external cues could not 336 provide sufficient information for the trial-by-trial estimation of AO/PS decision 337 inaccuracy/uncertainty. Further, the intrinsic mental state attributions also did not 338 rely on the target subject's attributes that were unknown for the participant in the 339 current study. Thereby, the estimating processes often brought out estimate 340 uncertainty [29,30]. The estimates were usually more uncertain at the middle 341 levels: longer RTs in reporting ratings at the middle levels than at the lowest and 342 highest levels (inverted U-shape, Fig 5a). We divided all the trials equally into 343 eight bins according to the quantity of external information, calculated by a 344 sigmoid function of the task difficulty and RT (equation 1 in Methods). RTs were 345 also longer at the middle bins than at the lower and higher bins (inverted U-346 shape, Fig 5b). The estimate uncertainty was the second-order variable of the 347 estimates, similar to decision uncertainty about the decisions in metacognition. 348 We then operationally defined the trial-by-trial estimate uncertainty as -  (Fig 5c). Critically, the dACC region associated with estimate 356 uncertainty across the mentalizing tasks was largely overlapping with the dACC 357 region associated with decision uncertainty in the metacognition task (Fig 5d).

358
According to estimation theory [29,30], the variance of the estimate residuals 359 should be larger in the central range of the external information (Fig 2f). Exactly 360 consistent with this prediction, the variance of the estimate residuals in each bin 361 of the external information was a negative parabolic function in each task 362 (inverted U-shape, Fig 5e). Accordingly, the mean dACC activity in each bin of 363 the external information was also a negative parabolic function in each of the 364 mentalizing tasks (inverted U-shape, Ps < 0.05 ; Fig 5f and 5g), but not in the 365 metacognition task (two-tailed t test, t 27 = -1.9; P = 0.07; Fig 5g). Altogether, 366 these results suggest that the dACC also plays an important role in monitoring 367 the mentalizing process. Notably, this is complementary with prior findings that with the former prediction across the mentalizing tasks (Fig 5l), supporting that 407 the dACC was involved in monitoring the residual variances (estimate 408 uncertainty), rather than in generating these residuals. That is, metacognition 409 monitors mentalizing (Fig 6a). Importantly, the same format of mental state 410 representations of decision uncertainty in metacognition and estimate uncertainty 411 in mentalizing were registered in the dACC. On the contrary, mentalizing was not 412 subject to metacognition and was not associated with the dACC activities.

413
Instead, different forms of mentalizing had diverse formats of mental state 414 representations of inferred decision uncertainty in the human brain (Fig 6b).  The type-1 mentalizing task in monitoring an anonymous other's decision 444 inaccuracy (AO-DI) was similar to false-belief tasks (e.g., the "Sally-Anne" task)  When the participant monitored the AO/PS's task performance in the type-1 492 mentalizing task, similar to monitoring the current-self task performance in the 493 metacognition task, the reported AO/PS's decision inaccuracy was the subjective 494 beliefs of the participant, rather than the target subject. However, as described 495 above, the neural representations of these mental states in the type-1 496 mentalizing tasks were entirely different from those in the metacognition task. To 497 this end, the critical distinction between metacognition and mentalizing should 498 depend on the accessibility of sources to be monitored, rather than the agents to 499 whom the mental states are tied or the target subjects to whom the mental states 500 are attributed. Altogether, our results illustrate that the human brain diversifies 501 separate neural systems to represent the different mental states of decision 502 uncertainty in monitoring the current self, the past self, and others in performing 503 the same perceptual decision-making task. plausibly an effective strategy to manipulate influences on others when they are 522 uncertain, rather than when they are highly confident, since the odds of success 523 in changing others' minds should be higher in the former case. Even for 524 preverbal infants, when they feel uncertain, they are willing to seek caregiver's 525 helps [43]. Therefore, across the mentalizing tasks, the target subject's decision 526 inaccuracy/uncertainty, rather than the decision accuracy/confidence, were 527 predominantly positively correlated with the brain activities, which might be used 528 to guide appropriate social control [44,45]. Hence, the dACC involved in However, estimates of others' mental states and even those of the past-self 538 mental states were not predictable for the target's intrinsic mental states and 539 performance (after the associations with external information were regressed out).

540
These findings thus implicate that volatile momentary mental states are quite 541 difficult to predict, probably due to the fact that the AO's or PS attributes in both 542 object-level and meta-level performance were unknown to the participant in the 543 current study. One potential approach to improve the predictability of mentalizing

565
The participant was required to discriminate the net motion direction and rate her uncertainty about the decision by pressing a corresponding button. The task 567 difficulty was determined by the percentage of coherently moving dots, with four 568 levels of task difficulty randomly mixed in the task. The easiest and hardest levels 569 were fixed at a coherence of 1% and 30%, respectively, while the two middle 570 levels were set for each participant to achieve accuracy of 50% and 80%, 571 determined by a staircase procedure in a practice session conducted several 572 days prior to the fMRI experiments [18,28].

573
Mentalizing tasks. A pair of participants who had similar stimulus coherences at 574 the 50% and 80% accuracy level in the practice session jointly took part in the 575 task (the stimulus coherences at the two task difficulty levels used in the 576 mentalizing tasks were the means of their stimulus coherences, respectively).

577
One (the target subject) performed the RDM task outside the scanner, and 578 another (the participant) observed the target's performance from inside the 579 scanner. The two computers separately presented the stimuli but were were almost identical. In the metacognition task, each trial started with a green 594 cross cue to indicate that the task stimulus would be presented 1 s later. The 595 stimulus was then presented for 2 s, and four options of the moving directions 596 were presented. The participant made a choice within 2 s. After a choice was 597 made, four ratings from 1 (most uncertain) to 4 (most certain) were presented, 598 and the participant reported the rating by pressing the corresponding button 599 within 2 s. The inter-trial interval (ITI) was jittered uniformly between 2 s and 6 s.
where ! represents the estimate of decision inaccuracy/uncertainty at the trial , 613 ! and ! represents the task difficulty and the response time at the trial , 614 respectively, and ! represents the estimated residual. Task difficulties and RTs 615 were separately normalized within each participant. Further, for the mentalizing 616 tasks, we also considered an autoregressive (AR) model to account for the 617 associations of the estimates between the contiguous trials. However, the 618 autocorrelation of the estimates between the contiguous trials in each 619 mentalizing task was not significant, indicating that the estimates were not 620 dependent on the history, but only on the current trial. 697 where X !"# and X !"#$%& represent the design matrix for the stimulation phase and 699 the judgment phase, respectively. ! and ! are the mean activities of the two 700 events, ! and ! are the estimate uncertainty modulation effects on the two 701 events, respectively. ℎ is the canonical hemodynamic response function with 702 two-gamma kennels, ℇ is an additional Gaussian noise. The values of ! and ! , 703 as well as ! or ! were independently and randomly drawn from a uniform 704 distribution in the range of [0.2,0.8], while the alternative ! or ! was set to zero.