Cognitive capacity shapes both the "whether" and "how" of social learning

Through simulation experiments, this paper demonstrates that cognitive capacity determines both the likelihood and the specific strategy of social learning, revealing an inverse-U relationship where moderate capacity maximizes social information use and a shift from success bias to conformism as capacity increases.

The Gist

The Big Question: When is it smart to copy others?

Imagine you are walking through a dense, magical forest filled with mushrooms. Some are delicious and give you energy; others are poisonous and will make you sick. You have to figure out which is which to survive.

You have two options:

1. The Solo Explorer: You taste a mushroom yourself. If it's good, great! If it's bad, you get sick. This is risky and tiring.
2. The Social Learner: You watch what others eat. If they eat it and look happy, you eat it too. This saves you from the risk of tasting poison yourself.

For a long time, scientists thought: "The less smart you are, the more you should copy others." The logic was simple: If you are bad at figuring things out alone, you need to copy. If you are a genius, you can figure it out yourself.

But this paper says: That's not quite right.

The author, Max Taylor-Davies, ran computer simulations to test this. He discovered that the relationship between "how smart you are" and "how much you should copy" isn't a straight line. It's actually a Goldilocks Zone (an "Inverse-U" shape).

The "Goldilocks" Discovery

Imagine a curve on a graph:

• Too Dumb (Low Capacity): If everyone in the forest is very confused and can't process information well, don't copy them.
  • The Analogy: Imagine a room full of people who are all guessing randomly. If you copy the person next to you, you are just copying a random guess. Their advice is likely wrong because they are too confused to figure it out. In this case, it's actually safer to try to figure it out yourself (or just avoid eating mushrooms entirely).
• Too Smart (High Capacity): If everyone is a genius, don't copy them.
  • The Analogy: Imagine a room full of brilliant scientists who can perfectly identify mushrooms. They don't need to look at each other; they already know the answer. Copying them adds no value because they are already solving the problem perfectly on their own.
• Just Right (Medium Capacity): This is the sweet spot.
  • The Analogy: Imagine a group of people who are "okay" at figuring things out. They make mistakes, but they also get lucky sometimes. If you watch them, you can spot the ones who are doing well and copy them. Their combined knowledge is better than what any single person could figure out alone. This is where social learning (culture) actually takes off.

The Takeaway: Social learning doesn't emerge when a species is stupid, nor when it is a genius. It emerges when the species is moderately capable relative to how hard the world is.

---

How We Copy Changes as We Get Smarter

The paper also looked at how we choose who to copy. There are two main ways to copy:

1. The "Copy the Winner" Strategy: You look for the one person who is eating the most mushrooms and staying healthy, and you copy them.
2. The "Follow the Crowd" Strategy: You look at what the majority of people are doing and do that, even if you don't know who is the "best."

The simulation showed a fascinating shift:

• In the "Low Capacity" (Confused) Group: The "Copy the Winner" strategy wins.
  • Why? When everyone is struggling, a few people might get incredibly lucky and find the safe mushrooms by accident. They stand out like a beacon. If you copy them, you win. The crowd is just a mess of noise, so following the majority is useless.
• In the "High Capacity" (Smart) Group: The "Follow the Crowd" strategy wins.
  • Why? When everyone is smart, the "lucky winners" aren't that much better than the average person. Everyone knows most of the answers. In this case, the "Winner" might just be having a lucky streak. But the Crowd represents the collective wisdom of the group. Following the majority protects you from individual mistakes (noise).

The Co-Evolution Dance

Finally, the paper simulated what happens if a population evolves over time.

1. Start: Everyone is a bit confused. A few "Copy the Winner" mutants appear and take over because they find the lucky geniuses.
2. Middle: As the population gets slightly smarter, the "Follow the Crowd" strategy starts to look better because the group becomes more reliable.
3. End: Eventually, the population becomes smart enough that "Following the Crowd" becomes the dominant way to learn.

Why Does This Matter?

This helps explain why humans are so good at social learning. We aren't just "smart" in a vacuum. We live in a world we made ourselves (cities, laws, complex tools) that is incredibly hard to learn.

Even though our brains are huge, the world is so complex that we are always in that "Goldilocks Zone." We are smart enough to generate good information, but the world is so complicated that we still desperately need to copy each other to survive.

In short:

• If you are too dumb, your friends' advice is useless.
• If you are too smart, you don't need your friends' advice.
• Culture thrives in the middle, where we are smart enough to give good advice, but the world is hard enough that we still need to listen to it.

Technical Summary

1. Problem Statement

The paper addresses a fundamental paradox in the evolution of social learning.

• The Cost-Avoidance Hypothesis: Social learning is traditionally viewed as a resource-rational adaptation that allows agents to avoid the metabolic costs and risks of individual trial-and-error exploration. Under this view, social learning should be most beneficial to agents with low cognitive capacity (who struggle to learn asocially).
• The Reliability Paradox: However, in peer-to-peer transmission, the quality of social information depends on the cognitive capacity of the source agents. If the population has low capacity, the social cues they generate are noisy and unreliable, potentially making social learning less effective than asocial learning.
• The Gap: Previous research has focused on environmental variability or task difficulty, but few studies have explicitly modeled how information-processing capacity limits (cognitive constraints) affect the utility of social learning when both the learner and the information source are constrained.

The central question is: How does the interplay between an agent's cognitive capacity ($\kappa$) and environmental complexity ($n$) determine the fitness advantage of social learning, and how does this shape the evolution of specific learning strategies?

2. Methodology

The authors employed a series of computational evolutionary simulations using a Contextual Multi-Armed Bandit framework.

A. The Task Environment (Mushroom Foraging)

• Model: A deterministic contextual bandit problem where agents must classify $n$-feature binary vectors (representing mushroom species) as edible ($+1$) or poisonous ($-1$).
• Complexity ($n$): The environment is defined by the number of features $n$. The rule determining edibility is a logical function over a subset of these features. Higher $n$ implies a larger search space and higher complexity.
• Fitness Dynamics: Agents gain fitness for eating safe mushrooms and lose fitness (risk of death) for eating poisonous ones. Fitness is capped at $h_{max}$ and decays at a baseline rate.

B. Agent Cognitive Capacity ($\kappa$)

• Information Bottleneck: To model limited capacity, agents do not process the full context vector $x_t$. Instead, they pass inputs through an encoder that randomly zeros out features, retaining only $\kappa$ bits of information on average.
  • $\kappa = 0$: No information processed.
  • $\kappa = n$: Full information processed (unlimited capacity).
• Learning Algorithm: Agents use Thompson Sampling (TS) to learn the rule $\theta$. They maintain a posterior belief over possible rules and sample a hypothesis to make decisions based on their lossy representation $z_t$.

C. Social Learning Mechanisms

Social learning is modeled as a pre-action belief update where an agent samples a cue from another agent.

• Unbiased Transmission: Randomly selecting a peer (linear imitation).
• Conformist Bias: Selecting a peer who holds the majority opinion in the population.
• Copy-Fittest Bias: Selecting the peer with the highest current fitness.
• Evolutionary Dynamics: The study uses a modified Moran process. Agents die based on fitness (or randomly), and new agents are born inheriting the "learning type" (strategy) of a parent, with a small mutation probability allowing strategy shifts.

D. Experimental Design

1. Experiment 1 (Fitness Advantage): Measured the lifespan difference between a single mutant social learner and an asocial learner in a population of asocial learners across varying $\kappa$ and $n$.
2. Experiment 2 (Emergence): Allowed the entire population to evolve, tracking the proportion of social learners over $10^4$ trials.
3. Experiment 3 (Strategy Competition): Introduced multiple social strategies (Unbiased, Conformist, Copy-Fittest) to see which dominates at different capacity levels.
4. Experiment 4 (Co-evolution): Simulated the simultaneous evolution of cognitive capacity ($\kappa$) and learning strategies over time.

3. Key Results

A. The "Inverse-U" Relationship

The study found that the fitness advantage of social learning is not monotonic with respect to cognitive capacity. Instead, it follows an inverse-U shape:

• Low Capacity: Social learning is maladaptive or neutral. The information generated by peers is too noisy to be useful; agents are better off relying on their own (limited) direct experience or avoiding the risk of bad cues.
• Moderate Capacity: Social learning provides the maximum fitness advantage. The environment is complex enough to make asocial learning difficult, but the population's capacity is high enough to generate reasonably reliable social cues.
• High Capacity: The advantage of social learning diminishes. Agents with high capacity can learn the environment effectively on their own (asocially), reducing the marginal benefit of potentially noisy social information.

B. Strategy Shift: From Success Bias to Conformity

As cognitive capacity increases, the optimal social learning strategy shifts:

• Low Capacity Regime: Copy-Fittest strategies dominate. In low-capacity populations, fitness variance is high due to "luck" (some agents get lucky streaks). Copying the fittest agent provides a strong signal of a "good" rule, even if the rule is imperfect.
• High Capacity Regime: Conformist strategies dominate. As capacity rises, most agents perform above chance, reducing the variance in "luck." The primary risk becomes individual noise (idiosyncratic errors). Conformity acts as a filter against individual noise, making it more robust than copying a single "fittest" individual.
• Surprising Finding: At very low capacities, Unbiased (Random) transmission can outperform Conformist learning. This contradicts standard theories suggesting conformity is always superior; here, conformity amplifies the noise of a low-capacity majority.

C. Co-evolutionary Trajectory

In simulations where capacity and strategy co-evolved:

1. Populations initially evolved toward Copy-Fittest strategies.
2. As capacity slowly increased over generations, the population transitioned to Conformist strategies.
3. This confirms a dynamic shift in the "optimal" social learning mechanism as the species' cognitive baseline improves.

4. Key Contributions

1. Resolution of the Capacity Paradox: The paper resolves the conflict between "social learning helps the weak" and "social learning requires reliable sources" by identifying a Goldilocks zone (intermediate capacity) where social learning is most adaptive.
2. Mechanism for Strategy Evolution: It provides a theoretical mechanism explaining why different species (or populations) might favor different social learning strategies (e.g., copying the successful vs. following the crowd) based on their cognitive constraints.
3. Refinement of Resource-Rational Analysis: It extends resource-rational analysis to social contexts, demonstrating that the "cost" of social learning includes the degradation of signal quality due to the cognitive limits of the source population.
4. Implications for Cumulative Culture: The findings suggest that cumulative culture may only emerge once a species breaches a specific threshold of individual problem-solving capacity, not because they need social learning more, but because they finally generate cues reliable enough to be worth copying.

5. Significance and Biological Implications

• Predicting Species Behavior: The model predicts that species heavily reliant on social learning should be found at an intermediate level of cognitive capacity relative to their ecological complexity. Species with very low capacity may fail to evolve social learning not due to a lack of need, but due to the unreliability of conspecific cues.
• Human Exceptionalism: The paper addresses why humans, despite having high cognitive capacity, rely heavily on social learning. It suggests a feedback loop of cultural niche construction: humans increase environmental complexity through culture, which keeps them within the "adaptive regime" of relative capacity where social learning remains beneficial, even as absolute intelligence rises.
• Evolutionary Dynamics: The shift from "copy-fittest" to "conformist" as capacity grows offers a new perspective on the evolution of cultural norms and conformity, suggesting they are not just social phenomena but adaptive responses to the noise levels inherent in high-capacity cognitive systems.

In summary, the paper argues that cognitive capacity is a fundamental constraint that dictates whether a species learns socially and how it does so, with the optimal strategy shifting systematically as the population's information-processing abilities evolve.