Canonical decision computations underlie behavioral and neural signatures of cooperation in primates

This study demonstrates that the primate dorsomedial prefrontal cortex (dmPFC) drives cooperative behavior in marmosets by implementing a gaze-dependent drift-diffusion process that accumulates social evidence, with specific neural signatures at both single-cell and population levels directly mapping to decision variables and cooperative success.

The Gist

Imagine you and a friend are trying to dance a perfect duet. You aren't holding hands, and there's no music playing. You just have to move at the exact same time to get a reward. If you move too early or too late, you both miss out.

This is exactly what scientists asked two tiny monkeys (marmosets) to do in a lab. They had to pull levers at the same time to get a tasty treat. But here's the twist: the monkeys were free to move around, look around, and interact naturally. They weren't glued to a chair or forced into a rigid routine.

The researchers wanted to solve a mystery: How does the brain decide when to move when working with a partner?

Here is the story of what they found, broken down into simple parts.

1. The "Social Radar" (Gaze)

The monkeys didn't just stare blankly. They were constantly checking in on each other. The scientists realized that the monkey doing the pulling (the "actor") was using its eyes like a social radar.

When the actor looked at its partner, it wasn't just seeing a face; it was gathering "social evidence." It was asking, "Is my friend ready? Are they moving smoothly, or are they jittery and unsure?"

2. The "Mental Bucket" (Evidence Accumulation)

Think of the monkey's brain as having a mental bucket.

• Every time the monkey looks at its partner, it drops a "pebble" of information into the bucket.
• If the partner looks calm and steady, the pebbles are heavy and smooth. The bucket fills up quickly.
• If the partner looks jittery or confused, the pebbles are light and bumpy. The bucket fills up slowly.

The monkey waits until the bucket is full enough to decide, "Okay, I'm confident enough. I'll pull the lever now!" This process is called evidence accumulation. It's the same way your brain decides if a sound is a bird or a car, but in this case, the "sound" is your friend's body language.

3. The Brain's "Firing Squad" (The dmPFC)

The scientists looked inside the monkeys' brains, specifically at a region called the dmPFC (dorsomedial prefrontal cortex). You can think of this area as the Command Center for social decisions.

They found two amazing things happening in the neurons (brain cells) of this Command Center:

• The Slope (The Speed): Some neurons started firing faster and faster as the monkey got closer to pulling the lever. The steeper the "ramp" of activity, the faster the monkey decided to act. It was like a rocket engine revving up.
• The Baseline (The Bias): The starting level of activity depended on what happened last time. If the monkey failed to get a treat in the previous round, its brain started with a "higher baseline" of excitement, almost like saying, "Okay, let's try harder this time!"

4. The "Traffic Jam" vs. The "Highway" (Success vs. Failure)

Here is the coolest part. The scientists looked at the shape of the activity across the whole group of neurons, not just one.

• On Successful Trips: When the monkeys cooperated perfectly, the neural activity moved in a smooth, straight, predictable path. It was like driving on a highway. The brain knew exactly where it was going.
• On Failed Trips: When they missed the timing, the neural activity was all over the place—wiggly, twisting, and chaotic. It was like a traffic jam where cars are swerving and stopping randomly.

This proved that the brain's ability to stay on a smooth "track" is what leads to successful teamwork.

The Big Takeaway

This study is a big deal because it shows that cooperation isn't magic; it's math.

Even in a natural, messy environment where animals are running around, their brains are secretly doing complex calculations. They are:

1. Watching their partner.
2. Gathering data on how reliable that partner is.
3. Filling up a mental bucket until they feel safe to act.
4. Executing the move when the math says "Go."

It turns out that the same kind of brain math we use to decide if a ball is coming toward us is also used to decide if our friend is ready to high-five us. The brain is a master of turning "looking" into "doing."

Technical Summary

1. Problem Statement

Cooperation requires individuals to dynamically integrate social cues (particularly gaze) to infer a partner's intentions and time their own actions. While evidence accumulation (specifically the drift-diffusion process) is a well-established framework for perceptual and value-based decision-making in constrained, head-restrained settings, two critical gaps remain:

1. It is unknown if evidence accumulation is a conserved mechanism for complex social cognition.
2. It is unclear if these processes operate during naturalistic, unconstrained social interactions where behavior is not rigidly structured.

The authors aim to identify the neural mechanisms by which primates transform active social sensing (gaze) into decision variables to guide cooperative behavior in a freely moving setting.

2. Methodology

Experimental Design

• Subjects: Two common marmosets (Callithrix jacchus) acting as "actors," interacting with various partner marmosets in dyads.
• Task: A Mutual Cooperation (MC) task. Pairs must pull levers within a 1-second window to receive a juice reward. The task is self-paced with no external temporal cues, requiring the animals to spontaneously organize interactions into discrete "trials."
• Behavioral Tracking: Markerless facial tracking (DeepLabCut + Anipose) captured 3D coordinates of facial landmarks at 30 Hz. Eight continuous behavioral variables were extracted:
  • Spatial: Distances to partner, lever, and reward tube.
  • Gaze: Angular orientation relative to targets.
  • Dynamics: Linear speed (LS) and angular speed of the face.
• Neural Recording: Single-unit activity was recorded from the dorsomedial prefrontal cortex (dmPFC) using a custom-built, wireless, modular head implant system (allowing freely moving recording).
  • Data: 490 neurons from Animal K and 464 from Animal D.
  • Processing: Spike sorting (Kilosort4) and manual curation (phy).

Computational & Analytical Framework

1. Trial Definition: The start of a trial was objectively defined by the ramping onset of the actor's face linear speed (LS), determined via a Generalized Linear Model (GLM) as the most predictive variable for pull initiation.
2. Social Evidence Representation: Principal Component Analysis (PCA) was applied to the partner's 8 behavioral variables to create a low-dimensional "holistic" representation (PC1) of the partner's actions.
3. Behavioral Modeling (DDM): A Drift-Diffusion Model (DDM) was fitted to response times and choices. Four competing models were tested to determine how gaze mediates evidence integration:
   • M1 (Gaze-gated): Evidence accumulates only during social gaze.
   • M2: No gaze modulation.
   • M3/M4: Additive or multiplicative gaze effects.
   • Social Evidence Definition: Defined as the inverse of the variability (standard deviation) of the partner's PC1 during gaze periods. Lower variability = higher predictability = stronger evidence.
4. Neural Modeling:
   • GLM: Used to classify neurons as "predictive" (pre-action) or "reactive" (post-action) and to identify "triple-encoding" neurons (encoding pull, gaze, and partner action simultaneously).
   • Neural DDM: Linked neural dynamics to DDM parameters:
     • Firing Rate Slope $\rightarrow$ Drift Rate ($v$).
     • Baseline Firing Rate $\rightarrow$ Starting Point ($z$).
5. Population Dynamics: Single-trial neural trajectories were analyzed in a 3D state space (first 3 PCs). Path variability (tortuosity) was calculated to measure the geometric complexity of neural trajectories.

3. Key Results

Behavioral Findings

• Gaze-Dependent Accumulation: The Gaze-gated DDM (M1) provided the best fit. Social evidence (partner action predictability) is integrated only when the actor is looking at the partner.
• Nature of Evidence: The critical variable is the variability of the partner's actions. High variability (noise) slows the decision (lower drift rate), while low variability (regularity) facilitates coordination. This rules out simple threshold or trend-detection strategies.
• Holistic Processing: The model relying on the holistic PC1 representation of the partner outperformed models using single variables (e.g., partner's linear speed), indicating the brain integrates multidimensional social cues.

Single-Neuron Findings

• Mixed Selectivity: A core subpopulation of dmPFC neurons exhibited "triple-encoding," simultaneously representing the actor's pull, gaze, and the partner's actions.
• Predictive Coding: Triple-encoding neurons were significantly enriched for predictive activity (spiking before the action), suggesting an anticipatory role in decision-making rather than just reactive monitoring.
• Ramping Activity:
  • Firing rate slopes were inversely correlated with Response Time (RT) (steeper slope = faster response).
  • Slopes scaled with Social Evidence strength (steeper slope for higher evidence/low variability).
  • This encoding was robust on successful trials but significantly weaker on failed trials.

Neural-Computational Mapping

• Direct Parameter Mapping:
  • Trial-by-trial firing rate slopes predicted the DDM drift rate.
  • Baseline firing rates (during non-decision time) predicted the DDM starting point.
• History Effects: A failed previous trial increased baseline firing, which shifted the starting point bias in the behavioral model (M1+), suggesting a post-error adjustment in motivation or decision bias.

Population Dynamics

• Trajectory Geometry: The tortuosity (path variability) of neural population trajectories was negatively correlated with social evidence strength.
• Outcome Dependence: This relationship (low tortuosity = high evidence) was present only on successful trials. On failed trials, the neural geometry did not track social evidence, indicating a breakdown in the accumulation process despite similar motor actions.

4. Key Contributions

1. Naturalistic Social Neuroscience: Demonstrates that canonical evidence accumulation mechanisms operate in freely moving, unconstrained social interactions, bridging the gap between classic decision neuroscience and real-world social behavior.
2. Mechanistic Link: Establishes a direct, multi-level link between active sensing (gaze), computational parameters (drift rate), and neural dynamics (ramping slopes) in the dmPFC.
3. Social Evidence Definition: Identifies that in social contexts, "evidence" is not a static sensory feature but the reliability/variability of a partner's behavior sampled over time.
4. Predictive vs. Reactive: Distinguishes between neurons that merely react to social events and those that predictively integrate social information to guide upcoming actions.
5. Technical Innovation: Utilizes a novel wireless, modular implant system and advanced markerless pose estimation to record from the dmPFC in freely moving primates.

5. Significance

• Theoretical: Extends the evidence accumulation framework from perceptual domains to social cognition, suggesting that the brain uses a unified computational algorithm for both physical and social decision-making, modulated by active sampling (gaze).
• Clinical: Provides a potential neural biomarker for social deficits (e.g., in autism spectrum disorder) by identifying specific dmPFC dynamics (ramping slopes, trajectory geometry) required for successful cooperation.
• AI/Robotics: Offers a biological blueprint for designing cooperative multi-agent systems that rely on active sensing and dynamic evidence integration rather than static rule-based interactions.
• Future Directions: Suggests that the dmPFC acts as a gating mechanism where social attention permits evidence accumulation, opening avenues for causal manipulation (e.g., microstimulation) to test the necessity of this process.