evolved-instruction-following-inductive-bias

name	evolved-instruction-following-inductive-bias
description	Evolutionary inductive bias enabling rapid instructed task learning (RITL) in humans - bridges cognitive science, neuroscience, and LLM instruction tuning
tags	["instruction-following","inductive-bias","RITL","cognitive-flexibility","zero-shot","instruction-tuning","LLM-analogy"]
activation	evolved instruction following, inductive bias, RITL, instructed task learning, cognitive flexibility, instruction tuning, zero-shot task performance, cognitive architecture
source	arxiv 2606.29792
date	2026-06-29T00:00:00.000Z

Evolved Instruction-Following Inductive Bias for Rapid Task Learning

Core Thesis

Humans possess an evolved instruction-following bias - an inductive bias shaped by evolution to interpret and execute linguistic instructions, enabling rapid instructed task learning (RITL). This bias functions analogously to how LLMs leverage instruction tuning for zero-shot performance.

Key Insights

1. RITL as Cognitive Flexibility

• Human adults perform novel tasks correctly on first attempt after verbal/written instructions
• Hallmark of cognitive flexibility, yet mechanisms underexplored
• Parallels in artificial systems understudied across disciplines

2. Evolutionary Origin

• Instruction-following bias shaped by evolution for survival advantage
• Innate cognitive architecture feature in humans (not learned)
• Enables fast generalization from language to novel behaviors

3. LLM Analogy

• Humans: innate instruction-following architecture
• LLMs: instruction tuning during training
• Both achieve zero-shot task performance via different mechanisms

4. Cross-Disciplinary Evidence

Synthesizes:

• Cognitive science (task set formation)
• Neuroscience (prefrontal control, language networks)
• Machine learning (instruction tuning, zero-shot generalization)

Testable Predictions

1. Neural signatures: Instruction-following engages specific prefrontal-language network circuits
2. Developmental trajectory: Instruction bias emerges early, independent of general intelligence
3. Comparative studies: Humans outperform primates on instruction-guided tasks even with equal associative learning
4. AI design: Explicit instruction biases improve sample efficiency in neural networks

Applications

AI/ML

• Design neural architectures with explicit instruction-following biases
• Improve few-shot learning via architectural priors
• Bridge gap between LLM instruction tuning and human cognitive flexibility

Neuroscience

• Identify neural correlates of instruction vs. associative learning
• Map prefrontal-language network interactions during RITL
• Study individual differences in instruction-following capacity

Cognitive Science

• Unify theories of task set formation and language-guided behavior
• Explain rapid cultural transmission of skills
• Model individual differences in cognitive flexibility

Methodology Recommendations

1. Behavioral experiments: Compare instruction-guided vs. learning-by-doing conditions
2. Neural imaging: fMRI/EEG during instructed vs. discovered task learning
3. Computational modeling: Architectures with explicit instruction bias modules
4. Cross-species: Human vs. primate instruction-following comparisons

Limitations

• Position paper - limited direct experimental validation
• Evolutionary claims require comparative evidence
• Instruction-following may not be a single unified mechanism
• LLM analogy may oversimplify human instruction processing

Key Questions

• Is instruction-following a domain-general capacity or domain-specific modules?
• How does instruction bias interact with working memory and executive function?
• Can we quantify "instruction-following ability" as a cognitive trait?
• What neural circuits distinguish instruction-guided from associative learning?

Connections

• Relates to task set reconfiguration (Miller & Cohen)
• Connects to language-of-thought hypotheses (Fodor)
• Parallels meta-learning in ML (learning to learn)
• Links to cognitive control and prefrontal function