March 2026
Bees exhibit behavioral markers of self-awareness — including metacognitive uncertainty monitoring, caste-appropriate role identity, tool-use anticipation, and general cognitive ability (GCA) — but no unified computational account of these phenomena exists. We propose that a predictive coding architecture in the central complex (CX), governed by a single precision parameter, accounts for all four domains simultaneously. We show analytically that metacognitive opt-out accuracy is an inverted-U function of precision, with a maximum at \(p^*= 0.563\). This overconfidence paradox — high precision drives task performance but degrades self-knowledge — generates a counterintuitive prediction: more intelligent bees (higher GCA) should perform worse on opt-out metacognition tasks. Simulations with \(N = 500\) agents per condition, with precision distributions empirically grounded in published reversal learning \(d'\) data, confirm a metacognition–GCA anti-correlation of \(r = -0.658\). A single precision parameter explains 83.5% of variance across the three non-metacognitive domains (GCA factor), validating the unified architecture. Crucially, circadian disruption reverses the metacognition–GCA anti-correlation (\(-0.658 \to +0.730\)), providing a third falsifiable prediction. This is the first formal computational theory of insect functional self-awareness, bridging neuroethology, predictive coding theory, and comparative cognition.
Keywords: predictive coding; metacognition; central complex; precision; bee cognition; general cognitive ability; functional self-awareness
Insects exhibit a remarkable repertoire of behaviors that superficially resemble markers of self-awareness in vertebrates: metacognitive uncertainty monitoring , caste-appropriate role identity , tool-use anticipation requiring forward models , and individual variation in general cognitive ability (GCA) . Yet these phenomena have been studied in isolation, with no unifying computational account of how a miniature brain could implement the self-referential computations they apparently require.
The central complex (CX), a conserved mid-brain structure across insects, is an increasingly compelling candidate for a shared substrate. The CX encodes body-centric spatial representations , regulates circadian timing, and gates action selection. Critically, it maintains a real-time model of the organism’s current state—a property essential to self-referential behavior. We propose that a predictive coding (PC) architecture instantiated in the CX, governed by a single precision parameter, can account for all four self-awareness markers simultaneously.
A crucial distinction must be established at the outset. We use the term “functional self-awareness” to denote a computational property: the capacity to maintain a precision-weighted generative model of one’s own state, where self-model prediction errors drive behavior. This is distinct from phenomenal consciousness—the presence of subjective experience or “what it is like to be” . We make no claim that bees are conscious; our model is agnostic to that question. The behaviors we model (opt-out on hard trials, caste-appropriate task selection, anticipatory tool orientation) can be produced by a purely functional self-model without any accompanying inner experience .
This operationalization follows the Free Energy Principle (FEP) : under FEP, agents minimize expected free energy by maintaining generative models of their environment and themselves. The precision parameter in FEP quantifies the confidence placed on prediction errors, determining whether sensory surprises update internal models or are suppressed. Our model inherits this formal apparatus while focusing on the behavioral signatures precision variation predicts across domains.
A central and counterintuitive prediction emerges directly from our model: more intelligent bees should be worse at metacognitive opt-out tasks. This “overconfidence paradox” arises because high precision, while improving task performance, simultaneously reduces the uncertainty signal that drives opt-out behavior. An agent with near-perfect internal predictions rarely experiences the uncertainty needed to trigger opt-out on hard trials—and therefore fails metacognitive calibration despite excelling at the task itself.
We derive this result analytically (Section 3) and confirm it computationally (metacognition–GCA correlation \(r = -0.658\) in simulations of 500 agents). The prediction is directly falsifiable by correlating GCA battery scores with opt-out accuracy in the same individuals.
We first formalize the CX predictive coding model and its four behavioral domain functions (Section 2). We then derive the precision–metacognition trade-off analytically and identify the optimal precision level \(p^*= 0.563\) (Section 3). Full-scale simulations (\(N = 500\) agents per condition) validate the model’s quantitative predictions across four experimental conditions (Section 4). We discuss three directly testable empirical predictions and situate the model within the broader self-awareness and FEP literature (Section 5).
We model each bee agent as maintaining a four-dimensional self-state vector: \[\mathbf{s} = (p,\; c,\; \phi,\; e) \in [0,1]^4\] where \(p\) is the precision (inverse variance of internal predictions), \(c\) is the caste identity self-representation (\(0 =\) nurse, \(1 =\) forager), \(\phi\) is the circadian phase accuracy, and \(e\) is the energy state. The derived uncertainty is \(u= 1 - p\).
Precision is the central parameter of the model. In the FEP framework , precision weights prediction errors: high-precision agents trust their predictions and update beliefs slowly; low-precision agents are driven by every sensory surprise. We implement precision as a scalar for tractability, acknowledging that the biological CX likely employs multiple precision channels.
We do not hand-tune precision distributions. Instead, we derive them from published bee psychophysics data. Reversal learning studies report \(d'\) values spanning roughly \([0.3, 2.1]\) across individual honeybees (\(d'_{\max} \approx 2.5\) in optimal conditions). We normalize to a precision scale: \[p_i = 0.2 + \frac{d'_i}{d'_{\max}} \times 0.75\] giving a precision range of \([0.2, 0.95]\) with the empirical shape preserved. For the normal condition, this yields a Beta\((5, 2)\) distribution (mean \(\approx 0.64\), right-skewed), consistent with most bees being above-chance at standard discrimination tasks.
Each behavioral domain maps precision to performance through a distinct functional form, reflecting the different computational roles of the self-model.
The opt-out task presents hard and easy stimuli. Correct metacognitive behavior requires opting out on hard trials (where internal predictions are unreliable) and staying on easy trials. The probability of opting out on a trial with difficulty \(d\) is: \[P_{\text{opt-out}}(d \mid u) = \sigma\!\bigl(u\cdot s- \theta_{d}\bigr)\] where \(\sigma(\cdot)\) is the logistic function, \(s= 4.0\) is the uncertainty sensitivity, and \(\theta_{d}\) is a difficulty-specific threshold (\(\theta_{\text{hard}} = 1.0\), \(\theta_{\text{easy}} = 2.5\)). Metacognitive accuracy is: \[\text{acc}_{\text{meta}}= \frac{P_{\text{opt-out}}(\text{hard} \mid u) + (1 - P_{\text{opt-out}}(\text{easy} \mid u))}{2} \label{eq:meta}\]
Tool use requires a forward model: the agent must anticipate the goal state before completing the initial action step. Anticipatory orientation probability scales with precision weighted by caste identity: \[P_{\text{tool}} = p\cdot \bigl(0.5 + 0.5 \cdot c\bigr)\] Foragers (\(c \approx 1\)) have higher goal-directedness, consistent with their spatial foraging role.
Learning accuracy on a task of type \(\tau\) depends on the match between task demands and caste identity. For spatial tasks (\(\tau = \text{spatial}\)), foragers excel (\(c \approx 1\)); for social tasks (\(\tau = \text{social}\)), nurses excel (\(c \approx 0\)). Formally: \[P_{\text{learn}}(\tau) = 0.5 + 0.45 \cdot p\cdot \text{match}(c, \tau)\] where \(\text{match}(c, \text{spatial}) = c\) and \(\text{match}(c, \text{social}) = 1 - c\).
GCA reflects the shared precision substrate across multiple tasks. We model a composite GCA score as linear in precision: \[\text{GCA}_i = 0.5 + 0.4 \cdot p_i + \varepsilon_i, \quad \varepsilon_i \sim \mathcal{N}(0, 0.04)\]
We simulate four experimental conditions (Table 1):
Normal: Beta\((5,2)\) precision, mixed caste (\(c \sim \mathcal{N}(0.5, 0.2)\))
Disrupted: Beta\((2,4)\) precision (circadian disruption reduces precision)
Nurse spatial: Normal precision, nurse caste (\(c \sim \mathcal{N}(0.15, 0.1)\)), spatial task
Forager spatial: Elevated precision, forager caste (\(c \sim \mathcal{N}(0.85, 0.1)\)), spatial task
We derive analytically the precision level \(p^*\) that maximizes metacognitive accuracy (Equation [eq:meta]). Substituting \(u= 1 - p\): \[\text{acc}_{\text{meta}}(p) = \frac{1}{2}\Bigl[ \sigma\!\bigl((1-p)s- \theta_{\text{h}}\bigr) + 1 - \sigma\!\bigl((1-p)s- \theta_{\text{e}}\bigr) \Bigr] \label{eq:meta_full}\]
As \(p\to 1\): \((1-p)s\to 0\), so \(P_{\text{opt-out}}(\text{hard}) \to \sigma(-\theta_{\text{h}}) \approx 0.27\). The agent rarely opts out even on hard trials. Accuracy \(\to (0.27 + 0.924)/2 = 0.597\).
As \(p\to 0\): \((1-p)s\to s\), so \(P_{\text{opt-out}}(\text{easy}) \to \sigma(s- \theta_{\text{e}}) \to 1\). The agent opts out on every trial, including easy ones. Accuracy \(\to (1 + 0)/2 = 0.5\).
By continuity and the intermediate value theorem, since \(\text{acc}_{\text{meta}}(0) = 0.5\), \(\text{acc}_{\text{meta}}(1) = 0.597\), and \(\text{acc}_{\text{meta}}\) must exceed both boundary values at some interior point (because for intermediate \(p\), the agent correctly opts out on hard trials and stays on easy trials), a maximum \(p^*\in (0,1)\) is guaranteed to exist. Numerically: \[\boxed{p^*= 0.563, \quad \text{acc}_{\text{meta}}(p^*) = 0.679}\] Figure 1 shows the full precision–metacognition curve.
The inverted-U with an interior maximum holds for any slope \(s\in [2, 6]\) and any positive threshold separation \(\theta_{\text{e}} - \theta_{\text{h}} > 0\). The value of \(p^*\) shifts modestly with slope (range \(p^*\in [0.50, 0.63]\) across \(s\in [2.0, 6.0]\)), but the directional prediction is robust: intermediate precision is optimal for metacognitive accuracy. The shaded region in Figure 1 shows this robustness band.
Let \(\text{Perf}(p)\) denote performance on any precision-monotone task (Domains 2–4). By construction, \(d\,\text{Perf}/dp> 0\) everywhere. By contrast, \(d\text{acc}_{\text{meta}}/dp< 0\) for all \(p> p^*\). Therefore, in the high-precision regime (\(p> p^*\)), performance and metacognitive accuracy are necessarily anti-correlated across individuals. Since the empirical normal precision distribution has mean \(\approx 0.64 > p^*= 0.563\), most bee individuals in the normal condition lie in this anti-correlated regime.
Table 1 presents the simulation results (\(N=500\) per condition, seed 2026).
| Condition | Precision | Meta | Tool | Learning | \(r_{\text{meta,GCA}}\) | GCA% |
|---|---|---|---|---|---|---|
| Normal | 0.728 | 0.656 | 0.550 | 0.664 | \(-\)0.658 | 83.5% |
| Disrupted | 0.302 | 0.639 | 0.226 | 0.565 | \(+\)0.730 | 81.2% |
| Nurse (spatial) | 0.737 | 0.655 | 0.425 | 0.783 | \(-\)0.682 | 79.7% |
| Forager (spatial) | 0.793 | 0.647 | 0.741 | 0.804 | \(-\)0.599 | 81.0% |
Extracting the first principal component from the three precision-monotone domains (tool use, caste-appropriate learning, GCA score) explains 83.5% of variance in the normal condition. This is a stringent test: the GCA factor emerges entirely from the shared precision substrate, with no additional factor structure imposed. The result validates the single-parameter architecture—if multiple independent mechanisms underlay these domains, we would expect a much weaker first component. The GCA factor remains between 79.7% and 83.5% across all four conditions, demonstrating robustness to large changes in the precision distribution.
In the normal condition, individual-level correlations form a distinctive pattern (Figure 3):
Meta–GCA: \(r = -0.658\) (novel prediction: more intelligent bees are worse at opt-out)
Meta–Tool: \(r = -0.725\) (strongly negative; tool use and metacognition compete for precision)
Tool–GCA: \(r = +0.696\) (positive; tool use and intelligence share the precision substrate)
These anti-correlations emerge directly from the model’s architecture: agents with high precision (above \(p^*= 0.563\)) perform better on monotone-precision tasks but are past the metacognition optimum (Section 3). The model therefore predicts that metacognition and intelligence are functionally dissociated at the individual level, despite both arising from the same precision parameter.
Caste effects on learning are substantial. Foragers achieve a mean spatial learning accuracy of 0.804, nurses 0.783, compared to 0.664 for the mixed-caste normal condition. The advantage arises from the caste-match amplification in Domain 3, not from higher precision alone. This replicates the documented nurse-forager performance difference on spatial tasks and generates a quantitative prediction: the learning advantage of matched-caste bees should be approximately \(\Delta = 0.14\) (from 0.664 to 0.804), testable in within-colony experiments.
Circadian disruption reduces mean precision from 0.728 to 0.302 (\(-\)58%). This impairs all precision-monotone domains: tool use \(-59\%\), learning \(-15\%\). Metacognitive accuracy shows a smaller non-monotonic change (\(-\)2.6%), consistent with the disrupted precision distribution (\(\bar{p} \approx 0.30\)) falling below \(p^*= 0.56\), partially restoring calibration.
The most striking finding is the sign flip: the metacognition–GCA correlation reverses from \(r = -0.658\) (normal) to \(r = +0.730\) (disrupted) (Figure 4). Under disruption, precision is uniformly low and variable. Any residual precision benefits all domains simultaneously—including both GCA and metacognitive accuracy—so the anti-correlation disappears and is replaced by a positive relationship. This sign flip is a strong falsifiable prediction: the correlation between metacognitive accuracy and GCA should switch sign in circadian-disrupted bees relative to controls.
The model generates three empirically testable predictions, currently unverified in any bee species:
Prediction 1 — Anti-correlation: Correlating GCA battery scores with opt-out accuracy in the same individuals should yield a negative correlation (\(r < -0.3\), model prediction \(r = -0.658\)). Both paradigms exist; the missing experiment is applying them to the same marked individuals.
Prediction 2 — Inverted-U dose-response: Testing opt-out accuracy across a range of precision levels—induced by varying training intensity, reward certainty, or mild pharmacological challenge —should yield a non-monotonic relationship with a maximum at intermediate precision. A concrete experimental design: train three groups of honeybees to different discrimination difficulty levels (easy, medium, hard visual discriminations), then transfer all groups to the opt-out paradigm. The medium-difficulty-trained group, with precision closest to \(p^*= 0.563\), should show the highest opt-out accuracy. Severity of neonicotinoid challenge could serve as an alternative precision manipulation with dose-dependent precision reduction.
Prediction 3 — Disruption sign flip: Under circadian disruption, the metacognition–GCA anti-correlation should reverse sign. This requires running the GCA battery and the opt-out task on the same disrupted and control animals and comparing the correlation structure.
The overconfidence paradox bears surface similarity to the Dunning-Kruger effect , but the mechanism is qualitatively different. The Dunning-Kruger effect attributes impaired metacognition in incompetent individuals to lack of metacognitive skills. Our model predicts that competent agents have impaired metacognition because high precision suppresses the uncertainty signal. This is an inverted Dunning-Kruger: skilled individuals have systematically degraded self-knowledge. The prediction can be distinguished empirically: Dunning-Kruger predicts that the worst performers are least calibrated, while our model predicts that the best performers (by GCA) are least calibrated on opt-out tasks.
Several limitations should be noted. First, precision is modeled as a scalar; the biological CX almost certainly employs multiple, partially independent precision channels (spatial vs. temporal vs. olfactory). A multi-channel extension would require different predictions for domain-specific disruptions. Second, we provide no neural implementation; connecting to actual CX topology (ring attractor, Kenyon cells) is left for future work . Third, the model has no learning dynamics: precision is fixed at initialization and does not change within trials. Fourth, while precision distributions are grounded in published \(d'\) data , empirical validation requires measuring the same individuals across all four behavioral domains, which has not yet been done. Fifth, we focus on four domains where same-individual testing is feasible; four additional behavior domains proposed in the original research agenda (play behavior, rhythm calibration, deception-adjacent behavior, and helping) are consistent with the model but require extension of the current framework and are left for future work.
Our model is a behavioral-level abstraction of the FEP . Under FEP, precision plays the role of attention: agents allocate precision to minimize expected free energy. Our scalar precision captures this role without implementing the full variational inference that FEP specifies. A natural extension would map our scalar precision to the FEP’s sensory precision hyperparameter, deriving predictions about neuromodulatory control of the CX (e.g., octopamine vs. dopamine as precision-adjusting neuromodulators in bees ).
If the predictions are confirmed, this model has implications beyond bees. The precision–metacognition trade-off may be a general property of any system using a shared precision signal across performance and self-monitoring domains. In artificial agents, an analogous tension between predictive accuracy and uncertainty quantification has been identified in Bayesian deep learning . In humans, some accounts of autism spectrum disorder describe high local precision but impaired global uncertainty monitoring—a pattern qualitatively consistent with the high-precision extreme of our model.
We present a unified predictive coding model of functional self-awareness in bees, parameterized by a single precision variable grounded in published bee psychophysics. The model yields three falsifiable predictions:
More intelligent bees (higher GCA) should perform worse on opt-out metacognition tasks (predicted \(r < -0.3\), simulated \(r = -0.658\)).
Metacognitive accuracy is maximized at intermediate precision (\(p^*= 0.563\), analytically derived), not at maximum precision.
Circadian disruption should reverse the metacognition–GCA anti-correlation from negative to positive (simulated: \(-0.658 \to +0.730\)).
These predictions can be tested with existing behavioral paradigms applied to the same marked individuals. The model provides the first formal computational account of insect functional self-awareness and bridges neuroethology, predictive coding theory, and comparative cognition.
All simulations were implemented in Python 3.x using NumPy. Random seed was fixed at 2026 for reproducibility. \(N = 500\) agents were simulated per condition. Source code is available at https://github.com/CunyiKang/ARIS-GCA-Bees .
Precision distributions were derived from published bee reversal learning \(d'\) data . The reported \(d'\) range is approximately \([0.3, 2.1]\) for individual honeybees, normalized to \([0.2, 0.95]\) via \(p= 0.2 + (d'/2.5) \times 0.75\). The resulting empirical distribution was approximated by:
Normal condition: Beta\((5, 2) \times 0.75 + 0.20\) (mean \(\approx 0.64\))
Disrupted condition: Beta\((2, 4) \times 0.60 + 0.10\) (mean \(\approx 0.30\))
The precision optimum was identified numerically on a grid of 999 points, \(p\in [0.01, 0.99]\). The existence of an interior maximum was verified by checking \(d\text{acc}_{\text{meta}}/dp> 0\) at \(p= 0.1\) and \(d\text{acc}_{\text{meta}}/dp< 0\) at \(p= 0.9\).
The GCA factor was defined as the fraction of variance explained by
the leading eigenvalue of the \(3 \times
3\) covariance matrix of [tool use, caste learning, GCA score].
Eigendecomposition was performed via
numpy.linalg.eigvalsh.
All correlations are Pearson’s \(r\) computed over \(N = 500\) individual agents per condition. No multiple comparisons correction was applied, as all reported correlations are directional predictions from the model rather than exploratory comparisons.