Neurocomputational mechanisms of confidence in self and others

The study addresses how humans form confidence estimates about others’ decisions, beyond one’s own metacognitive confidence. Prior work suggests individual confidence reflects the Bayesian probability a choice is correct and engages prefrontal cortex. Social contexts require estimating another’s likelihood of being correct, which should integrate both current decision difficulty and others’ ability. Existing research shows people track others’ ability, and the theory-of-mind (ToM) network (TPJ, dmPFC) supports representing others’ attributes. The authors hypothesize that social confidence combines a sensory-based estimate of decision difficulty (e.g., motion coherence, supported by MT+/LIP) with a separate representation of others’ ability (supported by TPJ/dmPFC), enabling predictions such as high-ability agents failing on difficult trials and low-ability agents succeeding on easy trials.

• Individual confidence: Bayesian probability of correctness; neural substrates in prefrontal cortex support computing and reporting confidence, and integrating trial-wise confidence into performance estimates.
• Social inference: People track others’ performance/ability; ToM network (TPJ, dmPFC) represents others’ attributes distinct from self and maintains running estimates of others’ task performance.
• Sensory decision-making: Motion coherence manipulates difficulty; MT+ and LIP implicated in sensory processing and evidence accumulation for perceptual decisions.
• Gaps: Unclear whether and how humans integrate decision difficulty with others’ ability to compute confidence about others’ choices, beyond self-projection or simple performance tracking.

Participants: 22 healthy adults; 1 excluded for poor pre-scan performance; final N=21 (12 female; mean age 22.6 ± 4.4 years). Informed consent; NYU ethics approval; fixed fees plus bonus.

Task: Random dot motion discrimination with post-decision wagering (PDW) during fMRI. Trials signaled as self-trials (participant decides left/right) or other-trials (observe stimulus while another player decides; participant does not see their choice). After choice period, participants choose between a safe option (certain 5–20 points) and a risky option (±25–50 points contingent on correctness). Feedback reveals PDW outcome and allows inference of choice accuracy. Participants completed 3 runs × 40 trials (20 self, 20 other) = 120 trials.

Other players: Three players (low ≈0.55 accuracy, medium ≈0.75, high ≈0.95) constructed from pilot datasets; pairing was blockwise (one player per run). Motion coherence and reward differences were uncorrelated by design.

Stimuli (RDK): 0.4 s displays in a 7° aperture; coherence range ~0.005–0.5 during main task; left/right motion; dot parameters standard. Pre-scan familiarization with varied coherences.

Behavioral analysis: Multiple logistic regression on trial-by-trial PDWs with predictors: trial type (self vs other), motion coherence, reward difference, others’ ability (low/med/high), and interactions of trial type with coherence, reward, and ability. History effects tested via prior-trial accuracy.

Computational modeling: PDWs modeled via expected value difference AEV = EV(risky) − EV(safe); P(gamble) via softmax with possible separate bias/temperature for self vs other. EV(safe)=safe value; EV(risky)=P(correct)×gain + (1−P(correct))×loss. Confidence on self-trials derived from Bayesian decision theory: compute belief over stimulus states given sensory sample and noise, generate choice and P(correct). Three model classes for other-trials:

• S-model (self-projection): Use one’s own confidence as proxy; sensitive to difficulty but not others’ ability.
• Q-model (performance tracking): Rescorla-Wagner running estimate of each player’s accuracy Q(j); sensitive to ability but not current difficulty.
• ToM-model: Integrate belief over stimulus space with a representation of each player’s expected accuracy (psychometric function) parameterized by their sensory noise; update the player’s sensory noise via a Rescorla-Wagner-like rule using social prediction error (outcome − predicted success) scaled by derivative of prediction w.r.t. noise. Variants included Weber-Fechner-like rescaling of stimulus space and separate softmax parameters for self/other. Models also refit with a utility function to capture risk/loss aversion. Model fitting: Hierarchical variational Bayes in Stan; out-of-sample prediction via PSIS-LOO and WAIC; model identifiability tested via simulations.

fMRI acquisition: 3T Siemens Allegra; BOLD EPI (3 mm isotropic; TR=2.24 s; TE=30 ms; 42 slices); T1 MPRAGE structural; field maps for unwarping. Preprocessing in SPM12: slice timing, realign/unwarp, normalize to MNI (2 mm), smooth 8 mm FWHM; motion regressors + derivatives.

fMRI analyses:

• Whole-brain GLM1: regressors for decision phase (RDK onset to 3 s after decision onset) and gamble phase (gamble onset to feedback offset), each split by self/other; contrasts self>other and other>self.
• ROIs: MT+ (localizer dynamic>static), LIP (SPLD+SPLE from Mars atlas), TPJ (TPJp from Mars atlas), dmPFC (area 9 from Neubert atlas); bilateral.
• Single-trial ROI time courses: motion regressors removed; high-pass filtered; oversampled; beta-series c-HRFs for trial-wise estimates; sliding-window regressions.
• Sensory encoding: regress trial type, coherence, and interaction in MT+ and LIP; examine coherence effects separately by trial type.
• Social confidence encoding: regress trial type, model-based confidence (orthogonalized to coherence), and interaction in TPJ and dmPFC; examine confidence effects by trial type.
• PPI: generalized PPI with LIP or MT+ as seeds, psychological factor self vs other; test LIP–TPJ/dmPFC connectivity.
• Learning signals: regress social prediction error (from ToM-model) in TPJ and dmPFC; orthogonalization checks performed.

Behavioral:

• Gambling increased with motion coherence (t(20)=7.78, p<0.001) and reward difference (t(20)=5.02, p<0.001); no interaction with trial type (coherence×type: t(20)=0.18, p=0.858; reward×type: t(20)=−0.44, p=0.664).
• Gambling increased with others’ ability (t(20)=2.85, p=0.010), with a trial-type interaction (type×ability: t(20)=4.60, p<0.001). Ability had no effect on self-trials (t(20)=1.46, p=0.161) but did on other-trials (t(20)=4.52, p<0.001).
• Participants selected risky option more on self- than other-trials (t(20)=4.47, p<0.001); mean P(risky)=72% (self) vs 61% (other).
• History effect: more gambling after another player’s previous correct vs incorrect choice (t(20)=2.97, p=0.008); no such effect for self-trials (t(21)=−0.35, p=0.733). Others’ ability did not affect self-task performance (accuracy: coherence t(20)=7.90, p<0.001; ability t(20)=−0.78, p=0.444; RT coherence t(20)=−11.98, p<0.001; ability t(20)=0.43, p=0.674).

Modeling:

• ToM-model best explained PDWs (including Weber-Fechner scaling and separate softmax parameters for self/other). It captured qualitative effects and trial-by-trial evolution per player. Model identifiability confirmed; ToM-model remained best when including a utility function for risk/loss aversion; inferred confidence robust (r>0.9 with/without utility function).

fMRI whole-brain:

• Decision phase: self>other in extrastriate, posterior parietal, cingulate, and motor regions; other>self in ToM network (middle temporal gyrus, TPJ, dmPFC). Gamble phase: other>self again identified ToM network.

ROI analyses:

• Sensory ROIs: MT+ and LIP tracked motion coherence on both trial types. MT+: higher coherence → higher activity equally for self and other (c-HRF: social t(20)=−7.72, p<0.001; coherence t(20)=5.42, p<0.001; interaction t(20)=−0.85, p=0.408; coherence self t(20)=3.65, p=0.002; other t(20)=3.94, p<0.001). LIP: flip by trial type (interaction t(20)=3.07, p=0.006): coherence increased activity on other-trials (t(20)=2.94, p=0.008) but decreased activity on self-trials (t(20)=−2.14, p=0.045).
• Social ROIs: Both TPJ and dmPFC showed interaction between trial type and model-based confidence (TPJ: interaction t(20)=−2.10, p=0.049; dmPFC: interaction t(20)=−3.44, p=0.003). TPJ encoded confidence only on other-trials (lower activity with higher confidence: t(20)=−2.17, p=0.043; self-trials t(20)=−0.60, p=0.555). dmPFC encoded confidence on both: decreased with confidence on other-trials (t(20)=−2.67, p=0.015) and tended to increase with confidence on self-trials (t(20)=1.79, p=0.089).
• PPI: Connectivity between LIP and TPJ/dmPFC was higher on other- than self-trials (TPJ t(20)=2.50, p=0.021; dmPFC t(20)=2.56, p=0.019); no such effect for MT+.
• Learning signals: TPJ and dmPFC encoded social prediction errors (TPJ t(20)=−3.18, p=0.005; dmPFC t(20)=−3.12, p=0.005).

Findings show that social confidence computations integrate a sensory-based estimate of current decision difficulty with a separate representation of others’ ability. Behaviorally, participants’ PDWs reflected both coherence and player-specific ability, inconsistent with pure self-projection or simple performance tracking alone. Modeling favored a ToM-based Bayesian integration of belief over stimuli with a player’s psychometric function, learning player-specific sensory noise from feedback via social prediction errors. Neurally, MT+ and LIP encoded motion coherence across self/other contexts, providing a sensory representation of difficulty. TPJ and dmPFC—core ToM network nodes—encoded model-based confidence about others, beyond coherence, and carried social prediction error signals, consistent with maintaining and updating others’ ability. Increased LIP–TPJ/dmPFC coupling on other-trials supports an interaction whereby social regions augment sensory representations with social ability information. TPJ responses were selective to other-evaluation, while dmPFC differentiated self vs other more strongly when confidence was low, consistent with a role in separating agent-specific information. Overall, the results address the central question by demonstrating a neurocomputational interplay between decision-making and social cognition systems during social confidence computation, supporting distinct but interacting self- and other-related processes.

The study demonstrates that people compute confidence in others’ decisions by combining estimates of decision difficulty with player-specific ability and that this integration is supported by interactions between sensory decision-making regions (MT+, LIP) and ToM regions (TPJ, dmPFC). A Bayesian ToM-model best accounted for behavior and provided latent variables that mapped onto neural signals of confidence and social learning. Future work should disambiguate algorithmic implementations (e.g., heuristic/model-free mixtures vs full Bayesian), test contexts where self-evaluation requires learning or counterfactual reasoning, and investigate how predictions about others’ success interact with inferences about others’ own felt confidence.

• The paradigm addresses only one facet of social confidence—predicting others’ success—not inferring others’ subjective confidence.
• Although the ToM-model outperformed alternatives, the study did not fully arbitrate among related heuristic/model-free accounts; extensive psychophysical mapping may be needed to distinguish them.
• Social uniqueness was not tested; similar mechanisms might arise with non-human agents, though ToM network involvement suggests social inference.
• TPJ selectivity observed here may not generalize to tasks where self-ability must be learned or where counterfactuals are required.
• Modest sample size (N=21) typical of fMRI studies may limit generalizability.