Confidence acts as a reward that steers attention
The paper asks whether an internal feeling—confidence in one's own judgement—can do the job that money or food does in reward experiments: make a stimulus capture more visual processing. The short answer is yes. When observers repeatedly experienced lower error or higher reliability for one coloured stimulus, they later allocated more neural resources to that stimulus, and the bias appeared even when the reward was purely internal rather than a bonus payment.
The key qualification is that this prioritization was not optimal. Because observers could not control which stimulus would be probed on a given trial, concentrating resources on the item they felt good about increased overall estimation error relative to spreading resources evenly. The authors therefore describe the effect as simultaneously adaptive and counterproductive: locally sensible, globally costly.
Deceptive feedback and real difficulty both shifted confidence
The study ran a series of visual estimation experiments with 215 naive observers in total. On each trial, two random-dot kinematograms—patches of moving dots, one green and one blue—were shown, and observers later reported the motion direction of a cued stimulus. Reproduction error indexed how precisely that stimulus had been encoded.
Confidence was shifted in two ways. In Experiment 2, trial-by-trial feedback was made deceptive: feedback error was magnified for one stimulus colour and minified for the other, so observers came to believe they were worse at one colour and better at the other. In Experiment 3, no feedback was given; instead, one colour was presented at 85% coherence and the other at 45% on most trials, creating a genuine difference in difficulty. Crucially, precision was then tested on interleaved trials where both stimuli had equal 65% coherence, so any remaining colour advantage had to come from learned prioritization rather than from the stimulus itself.
Resources shifted roughly two-to-one toward the easier item
A population-coding model of neural resource allocation estimated how much of the limited resource each stimulus received. In Experiment 2a, observers allocated resources at 1.18:1 in favour of the error-minified item. In Experiment 3a, the ratio was 1.76:1 toward the high-coherence colour, and in the simultaneous laboratory condition of Experiment 3b it was 1.61:1.
These are allocation ratios, not accuracy scores, and the distinction matters. The same model also simulated which allocation would minimize overall response variance in each task. For most observers that optimum was close to equal allocation, because pouring resources into the low-coherence item cannot rescue it from perceptual noise without substantially damaging the high-coherence item. The observed weights were reliably larger than the optimal weights, and the resulting overall error was higher than the optimal strategy predicted (mean absolute deviation 0.627 versus 0.569 in Experiment 3a).
The bias arises when stimuli compete at encoding
The advantage for the preferred colour depended on the two stimuli being encoded together. When Experiment 3b presented them sequentially, removing direct competition during encoding, the estimated allocation was 1.2:1 in favour of the high-coherence colour but was not reliably different from equal allocation. Observers could still say which item was harder, and precision still favoured the high-coherence stimulus when coherence actually differed, yet the equal-coherence advantage disappeared.
This pattern localizes the effect to the encoding stage: the bias is not a decision-stage preference for reporting one colour, but an unequal distribution of attentional resources while the memory array is being formed. The laboratory version, which held gaze at fixation, also showed the effect, so it does not require overt eye movements toward the preferred stimulus.
Accumulated reward values predict the allocation
To explain the bias, the authors extended a normalization model of neural resource allocation with a simple reinforcement-learning rule. Rewards obtained from reproducing a stimulus—external points, feedback error, or internal confidence—become associated with that stimulus's identifying colour, raising its subjective value and biasing how resources are split on later encounters.
The model's strongest test is a comparison of two independent estimates. The allocation parameter fitted freely to behavioural errors alone, with no reward information, correlated strongly with the allocation predicted from the history of accumulated rewards (r = 0.976). A median-split analysis added trial-level support: in all four experiments, trials where the model predicted below-average resources for the probed item showed larger response error, with Bayes factors ranging from 3.74 in Experiment 3a to 5.55 × 10⁴ in Experiment 1.
The account is computational, not a direct neural measurement
The reinforcement-learning account is fitted to behaviour; the paper does not record neural activity, so the claim that intrinsic and external rewards cumulatively shape stimulus value is a model-based explanation that quantitatively matches the data rather than a measured mechanism. The resource allocation itself is likewise inferred from error distributions through the population-coding model, not observed directly.
Scope is also bounded by the task. Observers could not choose which stimulus would be probed, which is why the observed bias raised overall error; in settings where people can steer behaviour toward the more valuable item, the same value-driven allocation could pay off. The sequential-presentation result further indicates that the bias needs simultaneous competition at encoding to appear.
Watch the gap between felt competence and actual benefit
The paper's most useful distinction is between where confidence comes from and what it does. Confidence can be manufactured by misleading feedback or arise from genuine differences in stimulus reliability, and both routes produced the same directional bias: resources flowed toward the item already judged easier. The decisive comparison is not rewarded versus unrewarded trials but observed allocation versus the error-minimizing allocation, and here the two diverged in every experiment that tested it.
That gap is the paper's practical priority. A learner or operator whose sense of competence is calibrated to an easy subset of stimuli will keep feeding resources to that subset, widening the precision gap between easy and hard items. The model's reward-history prediction, which tracked freely estimated allocation at r = 0.976, suggests this drift can be forecast from the accumulated record of felt success rather than from the stimuli themselves.
Sources used in this answer
Intrinsic rewards guide visual resource allocation via reinforcement learning
Intrinsic rewards such as perceived competence bias visual working-memory resources toward stimuli judged easier, and a reinforcement-learning extension of a normalization model reproduces the observed allocation and error patterns.