Technical Summary: Are Humans Evolved Instruction Followers?

Problem Statement

The paper addresses a critical gap in cross-disciplinary understanding regarding Rapid Instructed Task Learning (RITL). While human adults can execute novel tasks correctly on the first attempt based solely on verbal or written instructions, this capability remains under-explored as a unified mechanism across cognitive science, neuroscience, and artificial intelligence.

Current research is siloed:

• Cognitive Science studies RITL as a product of "natural pedagogy" and social learning.
• Neuroscience investigates the frontoparietal networks enabling flexible task sets but often treats instruction-following tangentially.
• AI Research views instruction-following primarily as a model training step (Instruction Tuning) rather than a fundamental cognitive architecture.

The central problem is the lack of a unifying theoretical framework to explain whether an underlying principle of instruction-based generalization exists in both human brains and artificial models, and whether human instruction-following is an evolved inductive bias analogous to machine instruction tuning.

Methodology

This work is a position paper that synthesizes existing evidence from three distinct fields to propose a unifying hypothesis. The methodology involves:

1. Literature Synthesis: Reviewing cognitive science (developmental psychology, natural pedagogy), neuroscience (neuroimaging of task-set reconfiguration, prefrontal cortex anatomy), and machine learning (instruction tuning, zero-shot generalization).
2. Comparative Analysis: Contrasting human performance with non-human primates (chimpanzees, macaques) and domesticated animals (dogs) to isolate the specific nature of human instruction-following.
3. Analogy Construction: Drawing a structural parallel between the "inductive bias" of the human brain and the "policy" learned by Large Language Models (LLMs) through instruction tuning.

The paper does not present new empirical data but rather constructs a theoretical argument based on established findings from prior works (e.g., Cole et al., Ouyang et al., Csibra and Gergely).

Key Contributions

The paper's primary contribution is the proposal of the Evolved Instruction-Following Bias hypothesis. The authors argue that:

• Humans possess an innate inductive bias: Unlike other species, humans are neurologically "pre-wired" to interpret linguistic inputs as actionable instructions. This bias is an evolutionary adaptation shaped by the advantages of cultural transmission and social learning.
• Analogy to AI: This human bias functions analogously to Instruction Tuning in LLMs. Just as instruction tuning endows base models with the ability to generalize to zero-shot tasks by treating inputs as directives, human evolution has endowed the brain with a similar capacity to rapidly reconfigure cognitive states based on language.
• Reframing RITL: The paper reframes RITL not merely as a learning outcome but as a foundational aspect of human intelligence driven by a specific cognitive architecture (likely centered in the anterior prefrontal cortex) dedicated to parsing and implementing instructions.

Results and Evidence Synthesis

The paper synthesizes the following evidence to support the hypothesis:

• Behavioral Evidence: Humans achieve high accuracy on novel tasks on the first trial when given instructions, whereas non-human primates typically require extensive trial-and-error reinforcement. Even bonobos (e.g., Kanzi), who show some instruction-following ability, perform at levels comparable to human toddlers and lack the broad, compositional, language-mediated task recomposition seen in adults.
• Neural Evidence: Neuroimaging studies indicate that the lateral prefrontal cortex (LPFC) and frontoparietal networks act as "flexible hubs." These regions represent intended task sets before practice occurs and rapidly reconfigure connectivity to map stimuli to responses based on rules. The disproportionate enlargement of the anterior prefrontal cortex (Brodmann area 10) in humans compared to other primates suggests an evolutionary expansion of circuitry supporting abstract, instruction-derived rules.
• AI Parallel: Instruction-tuned LLMs demonstrate that explicit training on (instruction, response) pairs allows models to generalize to unseen tasks without further fine-tuning. The paper posits that for humans, this "training" was not a specific protocol but a result of evolutionary and cultural selection pressures.

Significance and Claims

The paper claims that recognizing the Evolved Instruction-Following Bias provides a conceptual bridge between cognitive neuroscience and AI. Its significance lies in:

1. Unifying Theory: It proposes a mechanism that explains rapid generalization in both natural and artificial neural networks, suggesting that "instruction-following" is a core computational strategy rather than a peripheral skill.
2. Guidance for Neuroscience: It highlights instruction-following as a distinct cognitive function worthy of focused investigation, suggesting specific neural signatures (e.g., stable LPFC patterns during instruction encoding) and evolutionary histories.
3. Guidance for AI: It suggests that understanding the human inductive bias could inform the design of AI systems that learn more efficiently, while conversely, insights from instruction-tuned models can shed light on the neural mechanisms of RITL.

The authors conclude that interdisciplinary research is necessary to investigate instruction-following as a unifying mechanism, potentially leading to a more comprehensive theory of rapid generalization in both biological and artificial systems.