The error that moves through time
Imagine two vending machines. Both deliver the same drink. Machine A has worked on every visit. Machine B failed yesterday and has flickered all morning. When both finally deliver, the liquid has the same objective amount. Yet the informational event is different. Delivery from A confirms a strong prediction. Delivery from B repairs a weak one. A learning system should update more after B.
Keep reward magnitude fixed. Change only expectation. If a signal represents reward amount, both machines should evoke the same response. If it represents signed reward prediction error, the unreliable machine should evoke the larger positive response. The manipulated property is expectation, not pleasure, consumption or sensory intensity.
This distinction made the classic dopamine result unusually powerful. In primate conditioning, phasic responses were stronger for unpredicted than predicted rewards, transferred towards reward-predicting cues during learning, and fell below baseline when an expected reward failed to arrive Schultz, Dayan and Montague, 1997. Later work showed that the timing of a negative response could follow the expected time of reward Hollerman and Schultz, 1998, that firing could scale quantitatively with the difference between obtained and expected value Bayer and Glimcher, 2005, and that identified dopamine neurons could share a common monotonic response function within a defined task Eshel et al., 2016.
The smallest useful formalisation is a one-step temporal-difference error:
Suppose a tone predicts one drop exactly one step later, and let γ equal 1. Before learning, both state values are zero. Delivery gives δ = 1 + 0 − 0 = +1. The tone state is then updated. Once its value reaches 1, delivery gives δ = 1 − 1 = 0. If the expected drop is omitted, δ = 0 − 1 = −1. The reward did not become less sweet. Its capacity to revise a prediction changed.
With a learning rate α = 0.2, an initial value of 0 and error +1, the cue value becomes V ← 0 + 0.2 × 1 = 0.2. The next identical reward is now only 0.8 better than predicted, so the next update is 0.16. Repetition shrinks the outcome error while building value at the predictor. The load-bearing idea is not that dopamine detects good things. It is that a transient can teach by reporting how wrong a value prediction was.
Temporal transfer is the signature that made RPE compelling
Reward amount stays fixed while informational value changes
Three signatures that should travel together
A reward prediction error claim becomes persuasive when three signatures converge. The first is sign. Better-than-expected and worse-than-expected outcomes should push the signal in opposite directions when sensory intensity is matched. The second is calibration through learning. As the prediction changes, the response should change with it, including transfer towards earlier predictors and a timed response to omission. The third is teaching consequence. Perturbing the transient should alter the next prediction or action in the direction implied by the error.
Any one signature can mislead. An unexpected stimulus may produce a burst without carrying value sign. A cue response may shift because action timing shifts. A perturbation may reduce behaviour by changing arousal rather than learning. Convergence matters because shape, computation and causal role are different claims.
Even sign needs care. Dopamine neurons often express negative errors as pauses from a limited baseline, so positive and negative responses need not be mirror images. Sensors can also saturate differently for increases and decreases. The right test is not visual symmetry. It is whether response magnitude changes monotonically with signed prediction error within the instrument’s usable range.
A cue predicts one drop with probability 0.5. Its expected immediate value is 0.5 drops. On rewarded trials, the simple outcome error is 1 − 0.5 = +0.5. On omission trials, it is 0 − 0.5 = −0.5. A second cue predicts reward with probability 0.9. Receiving the same drop produces only +0.1, while omission produces −0.9. The physical event is identical within each outcome class, yet the learning signal differs because expectation differs.
This also separates RPE from uncertainty. The 0.5 cue is more uncertain before the outcome, but the sign of the outcome response depends on what happens. A signal that rises for both reward and omission under the uncertain cue may encode unsigned surprise or uncertainty rather than signed value error.
Reward probability is not the only source of expectation. Amount, delay, effort, internal need and inferred task state can all alter value. These additions make a model more realistic but also more vulnerable to post hoc rescue. The value function should be stated before the decisive neural result is inspected, or several pre-specified value models should compete on held-out trials.
Why the cue response is more than a correlation
A cue-locked burst might merely announce predicted reward. Temporal-difference theory makes a stronger claim: the transient is an error that trains earlier predictors. Causal tests can separate these roles. In second-order conditioning, suppressing dopamine at the start of an already valued cue impairs learning about a new predictor in a pattern consistent with removing a TD error Maes et al., 2020. Experimentally imposed dopamine transients can also produce learning effects expected from positive prediction errors Steinberg et al., 2013.
The bounded conclusion is strong: particular phasic dopamine signals can carry and causally implement TD-like teaching errors. It does not follow that every dopamine neuron, projection, release event or behavioural epoch computes the same scalar.
The hidden state inside the equation
The equation looks complete, but it quietly assumes that the system knows which state it is in. Consider a cue followed by reward at an uncertain time. As seconds pass without reward, has the animal entered a “late but still possible” state, or a “this trial is probably unrewarded” state? Those representations assign different values and therefore different errors to the same physical event.
Manipulations of temporal uncertainty and medial prefrontal cortex activity support the idea that dopamine errors depend on inferred hidden state, not only current sensory input Starkweather, Gershman and Uchida, 2018. Contingency-degradation experiments further show that temporal-difference models can account for behaviour and ventral-striatal dopamine when they use an appropriate prospective state representation Qian et al., 2025.
An RPE is always conditional on a predictive state model. A poor fit may mean that dopamine is not carrying an RPE. It may instead mean that the analyst supplied the wrong states, timing basis, outcome features or reward function. A decisive design must distinguish failure of the signal hypothesis from failure of the modeller’s representation.
flowchart LR A["Opening case"] --> B["The error that moves through time"] B --> C["How the waveform is counterfeited"] C --> D["The value–salience–movement assay"] D -. "re-test" .-> A classDef first fill:#0a1424,stroke:#007D68,stroke-width:2px,color:#e6f3ff classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff class A first class B,C,D step
How the waveform is counterfeited
The classical pattern becomes overinterpreted when three steps collapse into one: a trace resembles an error; therefore the recorded signal encodes that error; therefore dopamine as a system has that function. The first step is descriptive. The second is causal. The third generalises across cells, projections and measurements. Each requires new evidence.
Counterfeit one: movement walks into the cue
During learning, animals do not remain mechanically constant. They orient, lick, accelerate, approach and initiate actions at new times. Distinct dopaminergic axons can carry rapid locomotion-related and reward-related signals in dorsal striatum Howe and Dombeck, 2016. Identified dopamine neurons can also show movement-related excitation or inhibition, and these components can depend on reward context. In one learning dataset, cue-related and movement-initiation-related components summed as action timing became more precise, producing apparent RPE correlates Coddington and Dudman, 2018.
Two animals hear the same cue and receive the same reward. Animal A must initiate an approach. Animal B is carried towards the reward on a motorised platform while its limbs remain still. Now delay the movement in A without changing reward probability or delivery time. A scalar RPE should follow changed value prediction, not kinematics alone. A movement-linked signal should follow action onset. A mixed signal should split. The single changed property is self-initiated movement.
A useful intervention changes one parent of the trace
Movement correlation does not imply that dopamine is a motor command. A transient may reinforce a state–action relation, alter future vigour, or reflect a prediction about action without directly initiating the movement being recorded. Dorsolateral striatal dopamine transients can shape the later statistical structure of spontaneous behaviour Markowitz et al., 2023. Conversely, disrupting rapid dopamine release can leave basic movement initiation intact while impairing reward-oriented conditioned behaviour Cai et al., 2024. Motor execution, action selection, action reinforcement, movement prediction error and motivational invigoration are distinct causal claims.
Counterfeit two: unsigned salience wears a signed mask
Unexpected events capture attention, trigger orienting and alter arousal. Those reactions often co-occur with positive reward errors because an unpredicted reward is also surprising. If a study contains only neutral baseline trials and surprising rewards, an unsigned surprise signal and a positive RPE make the same prediction.
After a reward-predicting tone, insert a bright but value-neutral flash on rare trials. Match its novelty and sensory intensity to the unexpected reward as closely as the task permits. An RPE account predicts no positive value-signed response to the flash. A salience account predicts a transient to both. Now omit a strongly expected reward. A signed error should dip; unsigned salience should rise or remain positive. The changed property is value sign while surprise remains high.
Local dopamine release in nucleus accumbens can track perceived salience under conditions where value and salience are experimentally separated Kutlu et al., 2021. Yet “salience” is not a single mechanism. Novelty, sensory intensity, uncertainty, arousal and behavioural relevance can diverge. Salience is a rival only when it is operationalised strongly enough to make a different prediction. Calling every unexplained transient salient simply replaces one catch-all label with another.
Counterfeit three: the population is not one wire
Midbrain dopamine neurons differ by genetic subtype, projection target, input pattern and task response. Large-scale recordings have found specialised mixtures of sensory, motor and cognitive variables within ventral tegmental dopamine populations Engelhard et al., 2019. Genetic subtypes can map onto distinct functional response profiles Azcorra et al., 2023. Population averages are therefore useful descriptions of samples, not proof of a transmitter-wide scalar.
“Dopamine” is a family name
Recent work extends this diversity beyond a simple “reward versus movement” opposition. Dopamine input to the tail of striatum can carry an action prediction error, defined by the discrepancy between an action taken and the action predicted in that state, and can causally support value-free action learning Greenstreet et al., 2025. Striatal dopamine has also been reported to signal errors in predictions of value-neutral sensory relationships as well as valued outcomes Costa et al., 2025. The sharper distinction is reward prediction error versus broader predictive teaching, not prediction error versus every non-reward signal.
Counterfeit four: the measurement layer changes the message
A spike recorded at a dopamine cell body is not the same object as axonal calcium, transmitter release, extracellular concentration, receptor occupancy or downstream plasticity. Propagation can fail. Local cholinergic and presynaptic mechanisms can gate release. Uptake reshapes concentration. Receptors integrate over different timescales. A behavioural read-out adds another circuit transformation.
This matters empirically. Dopamine-cell spiking and nucleus-accumbens dopamine release can dissociate across learning and motivational epochs Mohebi et al., 2019. A statement such as “dopamine increased” therefore omits the very boundary needed to interpret the claim.
Record the same identified cell bodies while measuring release in two projection targets. Keep task, reward and movement fixed. Suppose soma spiking follows signed RPE, release in target X follows it closely, and release in target Y rises with movement regardless of error sign. Has the neuron stopped carrying RPE? No. Has “dopamine release” as a whole carried RPE? Also no. The changed property is the sampled measurement layer and target.
“Dopamine activity” is not one observable
Suppose half the sampled axons respond +1 to positive value error and 0 to movement. The other half respond 0 to value error and +1 to acceleration. In a task where positive errors always trigger acceleration, the pooled signal is +1 on every relevant trial. Regressing it against RPE produces an excellent fit. Regressing it against movement produces the same fit. More trials reduce uncertainty around the wrong causal conclusion because the task never separates the predictors.
The value–salience–movement assay
The practical contribution of this paper is a compact causal protocol for deciding what a dopamine trace may be called. The value–salience–movement assay, or VSM assay, treats RPE attribution as an identifiability problem. It does not ask which model produces the prettiest average waveform. It asks which model keeps predicting when its usual correlates are pulled apart.
The assay has five commitments. Define the biological and measurement boundary. Manipulate signed value error independently of unsigned surprise. Manipulate movement timing independently of both. Compare explicit rival generators on held-out trials. Add a causal perturbation at the moment when the models disagree.
The VSM cube turns a naming dispute into an experiment
1. State the candidate boundary before fitting anything
A useful claim names the cell class or projection, target region, recording method, task epoch and timescale. “VTA dopamine soma spikes during cue onset in this conditioning task” is testable. “Dopamine encodes reward prediction error” is too broad to fail cleanly. The boundary also determines which transformations have already occurred. A terminal-release claim cannot be inferred directly from a cell-body recording.
2. Build trials that separate the generators
The minimum design is not a large sample. It is a set of contrasts that changes one candidate cause while holding the others as steady as possible. The table below converts the three rival families into trial requirements.
| Contrast | Hold steady | Change | RPE prediction | Rival prediction |
|---|---|---|---|---|
| Expectation contrast | Reward magnitude, movement, sensory event | Predicted probability or amount | Signed response scales with actual minus expected value | Reward detector stays similar |
| Salience-only probe | Expected reward and movement | Novelty or unsigned sensory surprise | No value-signed response | Salience or sensory-error response appears |
| Movement-time shift | Cue, outcome probability and timing | Action onset or kinematics | Trace follows value error | Movement-linked component follows action |
| Signed surprise pair | Absolute surprise and sensory intensity | Better versus worse than expected | Opposite-signed responses | Unsigned salience responds similarly |
3. Pre-register rival generators, not just extra regressors
A regression coefficient is not a mechanism. Write each candidate as a generator with an input, temporal rule and failure condition. The model should predict trials not used to choose its features. It should also specify when it is allowed to abstain because the relevant state or movement variable was not measured.
| Candidate generator | Minimal form | Strengthening observation | Weakening observation |
|---|---|---|---|
| Signed value error | βvδvalue(t) | Opposite sign for matched positive and negative errors; generalises across expectation levels | Same response to equally salient opposite-sign events |
| Unsigned salience | βs|surprise(t)| | Novel neutral and valued surprises evoke comparable transients | Reliable below-baseline response to negative value error |
| Movement-linked signal | βmM(t) | Response shifts when action timing changes but reward statistics do not | Trace stays fixed under large movement-time shifts |
| Multiplexed signal | βvδ + βsS + βmM | Components separate under orthogonal contrasts and transfer across tasks | A single simpler generator wins held-out interventions |
4. Use a perturbation where the models disagree
Observation tells us what covaries. A teaching-signal claim requires downstream consequence. Perturb the transient in a narrow window and test the update predicted by each model. Suppressing a candidate positive RPE should reduce subsequent value updating. Suppressing a movement-linked teaching signal may instead alter action repetition, stimulus–action association or future vigour. The critical outcome is more than less behaviour. It is the pattern of what fails to be learned.
5. Fit on one set of contrasts and decide on another
Train models on familiar conditions, then decide on counterfactual trials that break the usual covariance. Report held-out predictive performance, coefficient stability and uncertainty. If multiple models remain tied, the conclusion is underidentified. Do not resolve the tie by choosing the most familiar theory. Add a trial type that separates them.
Why prediction and intervention answer different questions
Imagine that the full VSM model predicts forty held-out trials better than the RPE-only model. The gain is concentrated in trials where movement begins before the cue. That result earns a descriptive claim: the measured trace contains movement-related information that improves prediction. It does not yet tell us whether that component teaches an action, energises performance, reflects an efference copy or merely covaries with a hidden state. Prediction ranks generators by compression of the observations. Intervention asks which variable the circuit actually changes.
Now suppress a cue-locked transient for 150 milliseconds only after unexpectedly good outcomes. Three patterns matter. If later choices reveal a smaller update in expected value while action execution remains intact, the result strengthens an RPE teaching interpretation. If value estimates remain stable but the animal is less likely to repeat the just-completed action, an action-learning account gains ground. If latency, speed and choice all deteriorate immediately, the manipulation may have disrupted performance rather than learning. The causal signature is the selective downstream update, not simply a behavioural decrement.
A perturbation can change at least three things: the stored estimate learned from the event, the expression of an already learned policy, or the inferred state used to compute both. Separate these with a delayed choice test, a probe that changes motor demands, and a transfer trial in which the same inferred value guides a different action. Without those readouts, “learning was reduced” may be only a new label for slower movement.
Temporally precise stimulation and inhibition studies show why intervention can strengthen a teaching-signal claim, while newer work separates reward-guided updating from movement initiation and identifies value-free action prediction errors Steinberg et al., 2013; Maes et al., 2020; Cai et al., 2024; Greenstreet et al., 2025. The practical consequence is a crossed experiment: specify which transient is changed, when it is changed, and which learned quantity should fail. A model that predicts the waveform but not the selective consequence has described a correlate, not yet the teaching mechanism.
Train mice on two auditory cues. Cue H predicts a water drop with probability 0.8; cue L predicts the same drop with probability 0.2. On a minority of trials, a novel neutral light appears. Independently, a gate either permits an approach movement immediately or delays it by 600 milliseconds while reward timing remains fixed. Record optogenetically identified VTA soma spikes and dopamine release in one specified striatal target.
A reward after cue L creates a larger positive value error than the same reward after cue H. An omission after H creates a larger negative error than omission after L. The neutral light changes salience without reward value. The gate shifts movement without changing cue contingency. A time-locked perturbation on only one contrast then tests whether the transient teaches value, action or both. The design is small, but every included trial earns its place by forcing a disagreement.
Executable lab: generate a counterfeit dopamine trace
Choose a hidden generator, vary predictor collinearity, and compare held-out model performance. All data are synthetic and generated in your browser.
| Model | Held-out R² | Key coefficients |
|---|
The interactive lab intentionally lets movement correlate with RPE. When collinearity is high, a movement-generated trace can look well explained by an RPE-only model, especially under noise. Lowering the correlation creates the counterfactual trials that reveal the hidden generator. Increasing sample size in a confounded task is therefore not the same as improving identification.
Better identification can lower the wrong models’ scores
The failure case: a perfect fit to the wrong task
Suppose every surprising reward causes approach, pupil dilation and an RPE at the same instant. A large dataset can estimate their common waveform with great precision. It still cannot identify which variable caused the recorded signal. The problem sits in the covariance structure of the task, not in the optimiser.
The adversary is the task’s covariance structure
Python reference implementation and assumptions
The browser lab uses ordinary least squares on a 70:30 chronological split. The code below reproduces the core comparison. It assumes linear mixing, observed predictors and independent Gaussian measurement noise. A positive result permits a candidate attribution within those assumptions. A negative result may indicate the wrong generator, the wrong state representation, non-linearity or insufficient separation. It cannot establish a biological teaching role without intervention.
from __future__ import annotations
from typing import Literal
import numpy as np
Mechanism = Literal["rpe", "salience", "movement", "multiplex"]
def make_trials(
n: int = 480,
confound: float = 0.70,
noise: float = 0.45,
mechanism: Mechanism = "multiplex",
seed: int = 6801,
) -> dict[str, np.ndarray]:
"""Generate synthetic trials with a known hidden dopamine generator."""
if not 0 <= confound < 1:
raise ValueError("confound must be in [0, 1)")
rng = np.random.default_rng(seed)
expected = rng.choice([0.2, 0.5, 0.8], size=n)
outcome = (rng.random(n) < expected).astype(float)
rpe = outcome - expected
novelty = (rng.random(n) < 0.22).astype(float)
salience = np.abs(rpe) + 0.8 * novelty
rpe_standard = (rpe - rpe.mean()) / max(rpe.std(), 1e-9)
independent_move = rng.normal(size=n)
movement = confound * rpe_standard + np.sqrt(1 - confound**2) * independent_move
weights = {
"rpe": (1.0, 0.0, 0.0),
"salience": (0.0, 1.0, 0.0),
"movement": (0.0, 0.0, 1.0),
"multiplex": (0.8, 0.45, 0.6),
}[mechanism]
dopamine = (
weights[0] * rpe
+ weights[1] * salience
+ weights[2] * movement
+ rng.normal(scale=noise, size=n)
)
return {"rpe": rpe, "salience": salience, "movement": movement, "dopamine": dopamine}
def fit_and_score(data: dict[str, np.ndarray], names: list[str]) -> tuple[float, np.ndarray]:
"""Fit ordinary least squares on 70% of trials and score the held-out 30%."""
n = len(data["dopamine"])
split = int(n * 0.7)
X = np.column_stack([np.ones(n)] + [data[name] for name in names])
y = data["dopamine"]
beta, *_ = np.linalg.lstsq(X[:split], y[:split], rcond=None)
pred = X[split:] @ beta
truth = y[split:]
denominator = np.sum((truth - truth.mean()) ** 2)
r2 = 1 - np.sum((truth - pred) ** 2) / denominator
return float(r2), beta
if __name__ == "__main__":
trials = make_trials()
for label, predictors in {
"RPE": ["rpe"],
"Salience": ["salience"],
"Movement": ["movement"],
"Full": ["rpe", "salience", "movement"],
}.items():
r2, beta = fit_and_score(trials, predictors)
print(f"{label:9s} held-out R²={r2: .3f} coefficients={beta}")
stateDiagram-v2 [*] --> Claim state "Three signatures that should<br/>travel together" as Claim state "Why the cue response is more<br/>than a correlation" as Test state "The hidden state inside the<br/>equation" as Revision Claim --> Test: expose mechanism Test --> Decision: evidence survives Test --> Revision: rival survives Revision --> Claim: narrow or rebuild Decision --> [*]
What survives the alternatives
The VSM assay is deliberately strict, but it can fail in two directions. It can over-credit RPE because movement and salience remain confounded. It can also under-credit RPE because the analyst uses the wrong value model, state representation or temporal basis. The remedy is not to weaken the identification standard. It is to make the alternatives predict something new.
The RPE may be richer than one scalar
Individual dopamine neurons can differ in how strongly positive and negative errors change firing. One influential interpretation is distributional reinforcement learning: heterogeneous neurons learn different expectiles of the future reward distribution, allowing a population to represent more than the mean Dabney et al., 2020. A scalar model can then fit poorly even when the population remains value-error-based.
Other extensions use belief states, outcome features, action-specific errors or several predictive channels. These are legitimate alternatives when they make new predictions. They become immunising stories when every deviation is absorbed after the result. A richer state model should be penalised for extra flexibility, fitted before the decisive contrast and tested on a different task structure. Perturbing a circuit proposed to carry the hidden-state inference should change the error in the direction that model predicts.
A projection-defined dopamine population may multiplex a signed value error with a domain-specific error, while local release mechanisms reweight the mixture according to target state. This proposal is strengthened if the components separate under VSM contrasts, generalise across tasks and produce dissociable learning effects when perturbed. It is weakened if one simpler generator predicts all held-out interventions. A rival explanation is an unmeasured state representation. The smallest useful implementation is a two-target, two-measurement VSM task.
Do not infer subjective reward from a teaching signal
RPE evidence does not show that a transient is pleasure, wanting or conscious value. It identifies a computational relation between expectation, outcome and learning. Motivation, hedonic impact, effort allocation and felt experience require their own evidence. Likewise, a movement-linked dopamine signal does not establish that dopamine issues motor commands. The allowed conclusion stays at the level tested.
This boundary matters beyond neuroscience. Reinforcement-learning systems often name any scalar feedback “reward”, then slide from an optimisation variable to language about preference or desire. The biological case demonstrates why that move is unsafe. A teaching signal may alter future policy without representing what the organism likes, what it consciously feels, or what it would endorse under reflection.
An attribution ladder, not a verdict stamp
Call a trace a reward prediction error only when it shows value-signed responses, predicts held-out counterfactual trials, resists salience and movement alternatives, and causes the learning update expected from that error. Otherwise name the observation directly: cue transient, movement-linked activity, salience-correlated release or mixed predictive signal.
sequenceDiagram participant Q as Question participant M as Mechanism participant E as Evidence participant D as Decision Q->>M: Opening case M->>E: Three signatures that should travel together E-->>M: compare rival explanations E->>D: Make rival dopamine mechanisms disagree D-->>Q: revise scope or act
Make rival dopamine mechanisms disagree
Stop asking whether dopamine “is” reward prediction error at the level of an entire transmitter system. Decide at the level where an intervention can be specified: this cell class, this projection, this measurement, this task state, this temporal window and this downstream update.
For experimental neuroscience, the changed decision is methodological. Do not spend the next cohort increasing power on a task in which reward surprise, movement and salience remain locked together. Add the counterfactual trial first. Pair observation with a perturbation whose learning consequence differs across models. Report the narrowest conclusion that survives.
For reinforcement-learning and agent designers, the changed decision is architectural. Do not build a single global variable called “dopamine” that mixes reward error, novelty, action surprise and response vigour. Keep those quantities typed and separately inspectable. Combine them only where the update rule states how and why. Biological multiplicity is not permission for an unstructured scalar.
The classical RPE account remains one of neuroscience’s most productive theory–experiment correspondences. Its strength comes from precise predictions about expectation, timing and learning. Preserving that strength now requires the same discipline that created it: make rival mechanisms disagree, then let the intervention decide.
Glossary
- Reward prediction error
- The signed difference between obtained and expected reward value, often including discounted next-state value.
- Temporal-difference learning
- A family of learning rules that updates predictions from differences between successive value estimates and observed reward.
- Phasic dopamine
- Brief changes in dopamine-neuron firing or dopamine concentration over subsecond to short timescales.
- Salience
- Behavioural importance or attention capture. It must be operationalised as novelty, surprise, arousal or another measurable construct.
- Action prediction error
- The discrepancy between an action that occurred and the action predicted in that state, independent of reward value in the strict formulation.
- Hidden state
- A task condition not directly given by current sensory input and therefore inferred from history or context.
- Candidate boundary
- The exact biological population, projection, measurement layer, task and time window to which a mechanism claim applies.
- Multiplexing
- Carrying more than one task variable in the activity of a population or pathway.
References
- Schultz, W., Dayan, P. and Montague, P. R. (1997). “A neural substrate of prediction and reward.” Science, 275, 1593–1599. DOI.
- Hollerman, J. R. and Schultz, W. (1998). “Dopamine neurons report an error in the temporal prediction of reward during learning.” Nature Neuroscience, 1, 304–309. DOI.
- Bayer, H. M. and Glimcher, P. W. (2005). “Midbrain dopamine neurons encode a quantitative reward prediction error signal.” Neuron, 47, 129–141. PubMed.
- Eshel, N., Tian, J., Bukwich, M. and Uchida, N. (2016). “Dopamine neurons share common response function for reward prediction error.” Nature Neuroscience, 19, 479–486. PubMed.
- Steinberg, E. E. et al. (2013). “A causal link between prediction errors, dopamine neurons and learning.” Nature Neuroscience, 16, 966–973. DOI.
- Starkweather, C. K., Gershman, S. J. and Uchida, N. (2018). “The medial prefrontal cortex shapes dopamine reward prediction errors under state uncertainty.” Neuron, 98, 616–629.e6. DOI.
- Maes, E. J. P. et al. (2020). “Causal evidence supporting the proposal that dopamine transients function as temporal difference prediction errors.” Nature Neuroscience, 23, 176–178. DOI.
- Howe, M. W. and Dombeck, D. A. (2016). “Rapid signalling in distinct dopaminergic axons during locomotion and reward.” Nature, 535, 505–510. PubMed.
- Coddington, L. T. and Dudman, J. T. (2018). “The timing of action determines reward prediction signals in identified midbrain dopamine neurons.” Nature Neuroscience, 21, 1563–1573. DOI.
- Engelhard, B. et al. (2019). “Specialized coding of sensory, motor and cognitive variables in VTA dopamine neurons.” Nature, 570, 509–513. DOI.
- Mohebi, A. et al. (2019). “Dissociable dopamine dynamics for learning and motivation.” Nature, 570, 65–70. DOI.
- Dabney, W. et al. (2020). “A distributional code for value in dopamine-based reinforcement learning.” Nature, 577, 671–675. DOI.
- Kutlu, M. G. et al. (2021). “Dopamine release in the nucleus accumbens core signals perceived saliency.” Current Biology, 31, 4748–4761.e8. DOI.
- Azcorra, M. et al. (2023). “Unique functional responses differentially map onto genetic subtypes of dopamine neurons.” Nature Neuroscience, 26, 1762–1774. DOI.
- Markowitz, J. E. et al. (2023). “Spontaneous behaviour is structured by reinforcement without explicit reward.” Nature, 614, 108–117. DOI.
- Cai, X. et al. (2024). “Dopamine dynamics are dispensable for movement but promote reward responses.” Nature, 635, 406–414. DOI.
- Greenstreet, F. et al. (2025). “Dopaminergic action prediction errors serve as a value-free teaching signal.” Nature, 643, 1333–1342. DOI.
- Costa, K. M. et al. (2025). “Striatal dopamine signals errors in prediction across different informational domains.” Science Advances, 11. DOI.
- Qian, L. et al. (2025). “Prospective contingency explains behavior and dopamine signals during associative learning.” Nature Neuroscience, 28, 1280–1292. DOI.