Brain Carves Motivation into Three Orthogonal Axes

> Nature 2026 | Nanci Winke lab > > dmPFC is not a murky mix of "excitation" or "inhibition." It lays out valence, salience, and value on three mutually perpendicular coordinate planes.

The Problem: Three Intertangled Variables

Animals constantly make one basic judgment: should I approach this thing or avoid it? Behind this seemingly simple decision lie at least three dimensions:

• Valence: good or bad? Water reward or electric shock punishment?
• Salience: how "eye-catching" or noteworthy is the event?
• Value: how exactly good/painful? Quantified degree.

The trouble is that these three are almost always glued together in nature. A piece of meat for a starved rat—positive valence, high salience, high value. A lightning bolt—negative valence, very high salience, value (pain) depending on voltage. Researchers wanted to tease the dimensions apart, but reality rarely cooperates.

Winke's team's solution was clever: keep the physical action identical, vary only the outcome.

The 7-Second Shuttle: Action Locked, Outcomes Diverge

Mice were placed in a two-chamber box. When a light came on, they had to shuttle from chamber A to chamber B within 7 seconds.

• Sometimes shuttling → water reward (CSw)
• Sometimes shuttling → shock avoidance (CSs)
• Sometimes a third cue → larger reward or weaker punishment (CSr)

Crucially, the action was identical regardless of outcome—a shuttle within 7 seconds. With behavior locked down, neural activity could reveal what was actually being encoded.

Even better, CSs (shock escape) and CSw (reward seeking) could be tuned to equal motivational strength—the mice shuttled with the same probability. Under this condition, any neural difference between the two could not be explained by salience (both equally salient) or value (both equally drive behavior). The difference had to come from valence—positive versus negative.

Orthogonal Axes: A Cartesian Coordinate System in the Brain

After calcium imaging plus machine learning decoding, the team found three things:

1. Population level: value is the dominant code

At the population level, dmPFC activity most accurately decoded value—the specific magnitude of reward or punishment. Not a vague "good/bad," but a quantified "how good / how bad."

2. Subpopulations: valence and salience go separate ways

Looking at individual neurons:

• One set of populations encoded valence—positive vs. negative, appetitive vs. aversive
• Another set encoded salience—how noteworthy
• These coding spaces were orthogonal—approximately perpendicular in a mathematical sense

Geometrically: dmPFC holds a 3D map. Value is one principal axis; valence and salience expand along two other mutually perpendicular directions. The three dimensions don't interfere.

3. Orthogonality enables recomposition

Orthogonal coding has a key benefit: you can adjust one dimension without touching the others.

• Reward size changes (value axis stretches) → valence and salience judgments unaffected
• Environment suddenly becomes dangerous (salience axis rises) → doesn't automatically mean it's judged as bad (valence needn't flip negative)

This decoupling lets the brain flexibly reorganize motivational judgments. A mouse can learn "this tone used to mean water, but now means shock"—valence flips while salience stays. dmPFC's geometry permits this kind of precise surgery.

Technical Notes: Why This Experiment Is Hard

Freely moving + calcium imaging

Most neuroscience experiments head-fix animals and implant electrodes. Winke's team let mice freely shuttle while recording dmPFC calcium signals through head-mounted miniature microscopes. Consequences:

• Actions were real and spontaneous
• But optical recording is slower than electrophysiology (calcium indicator kinetics)
• The team ran electrophysiology follow-ups confirming the sustained representations seen in calcium signals were genuine neural activity, not artifacts

Linear classifier decoding

The team used combinations of linear classifiers to ask the same question:

• If dmPFC encodes salience, equal reward and equal punishment should activate similar population patterns (both "salient")
• If it encodes valence, reward and punishment should activate orthogonal patterns (opposite directions)
• If it encodes value, large reward and small punishment could look similar (intensity-matched)

Results ruled out pure salience and pure valence models, supporting a hybrid scheme: value-dominated coding with valence/salience subspaces orthogonally separated.

An AI Metaphor: Orthogonal Representations and Disentanglement

The finding has an interesting projection onto machine learning.

A core challenge for current LLMs and RL agents is entangled representations—reward, risk, and curiosity signals crammed into the same vector space, interfering with each other. dmPFC's solution suggests an architectural idea:

• Orthogonal subspace constraints: make different motivational dimensions approximately perpendicular in vector space
• Decoupled value estimation: compute "how important is this" separately from "is this good or bad"
• Flexible recomposition: an orthogonal basis allows interpolation and rotation—one representation can "rotate" to change valence judgment without affecting salience

Work already exists in this direction (e.g., β-VAE, FactorVAE in disentangled representation learning), but dmPFC provides biological validation: the brain genuinely uses orthogonal geometry for motivation, and evolution kept this solution.

Core Conclusions

1. dmPFC populations encode value (quantified appetitive/aversive magnitude), not fuzzy "emotion." 2. Subpopulations orthogonally separate valence from salience—the brain lays "good/bad" and "how noteworthy" on independent coordinate planes. 3. Orthogonal structure allows flexible recomposition—when circumstances change, the brain can flip the valence card without moving the salience card. 4. The 7-second shuttle paradigm is a methodological breakthrough—lock the behavior, decompose the motivation, with calcium imaging + machine learning decoding as a triple play.

References

• Winke N, et al. (2026). *Prefrontal neural geometry of learned cues guides motivated behaviours*. Nature, 651(8104):164-173. DOI: 10.1038/s41586-025-09902-2
• Winke N, et al. (2023). Preprint on bioRxiv: 10.1101/2023.03.03.530871
• Jercog D, et al. (2021). Sustained threat representations in dmPFC. *(Related prior work)*
• Laubach M. (2018). dmPFC anatomy definitions in mice.
• Tye KM. (2018). Neural circuit motifs in valence processing. *Nature Neuroscience*.
• Dalley JW, et al. (2004). Prefrontal cortical control of executive functions.