Enterprise AI Analysis
Sensorimotor Regularities as Alignment between Humans and Large Language Models
Jingyi Li, Jinghui Hu, and Per Ola Kristensson (2026)
This research explores how Large Language Models (LLMs) construct conceptual representations compared to human cognition, leveraging sensorimotor regularities (image schemas). It introduces a novel framework to assess human-LLM alignment and demonstrates how targeted sensorimotor priors can significantly enhance LLM output quality, making them more conceptually clear, contextually contingent, and human-like.
Executive Impact & Key Findings
Our analysis reveals critical insights into LLM capabilities and their alignment with human conceptual understanding, highlighting pathways for more intuitive and effective human-AI interaction.
0
Image Schema Instances (GPT-4)
0
Avg. Human-LLM Alignment (IS Distribution)
0
User Preference for Clarity (SRs-Guided)
0
Non-Human Co-occurrence (SRs-Guided)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Image Schema Distribution: The Building Blocks
This metric quantifies the presence and frequency of image schemas, identifying the fundamental conceptual building blocks available to LLMs compared to humans. Findings show considerable alignment in overall schema usage.
0.9294
Llama-2-70b's highest human-like IS distribution (Cosine Similarity)
| Key Finding | LLM Performance & Deviations |
|---|---|
| LLMs largely mirror human image schema distributions, demonstrating human-like conceptual building blocks. |
|
Conceptual Associations: Linking Schemas to Concepts
This metric examines how available image schemas are re-activated to structure abstract concepts, revealing how LLMs connect fundamental sensorimotor patterns to higher-level thought.
0.7285
Llama-2-70b's highest human-like conceptual associations (Cosine Similarity)
| Key Finding | LLM Performance & Deviations |
|---|---|
| LLMs show decreased humanlikeness in conceptual associations compared to schema distribution, often inventing non-human-like links. |
|
Image Schema Co-Occurrences: Relational Dynamics
This metric captures the complex interactions and co-occurrence patterns between image schemas, reflecting deeply ingrained human cognition and potential for cross-modal reasoning.
0.5334
Llama-2-70b's highest human-like co-occurrence patterns (Cosine Similarity)
| Key Finding | LLM Performance & Deviations |
|---|---|
| Significant performance decline in mirroring human co-occurrence patterns, with substantial invention of non-human-like co-occurrences. |
|
Downstream Impact: Augmenting LLMs with Sensorimotor Priors
We investigated whether closer sensorimotor alignment yields practical benefits. Through prompt-based intervention, gpt-4-1106 was augmented with human-aligned sensorimotor knowledge derived from identified misalignments.
Enterprise Process Flow: SRs-Guided Prompt Intervention
Schema-guided Instructions
→
Image Schema Knowledge Base
→
Few-shot Examples
→
LLM Output
Enhanced Sentence Continuations: A Proof-of-Concept
In a creative writing task using fictional terms, human evaluators rated sentence continuations generated by the SRs-guided gpt-4-1106 as significantly superior to baseline outputs:
- Conceptual Clarity: SRs-guided outputs received a 67.9% preference rating, compared to 32.1% for baseline.
- Contextual Contingency: SRs-guided outputs were preferred by 68.0%, vs. 32.0% for baseline.
- Humanlikeness: Users rated SRs-guided outputs as more human-like with 62.0% preference, vs. 38.0% for baseline.
- Imaginativeness: Baseline outputs were rated as significantly more imaginative (68.2%) than SRs-guided (31.8%), indicating a potential tradeoff.
This demonstrates that targeted sensorimotor alignment at the prompt level can lead to measurable improvements in output quality, making LLM responses more interpretable and aligned with human cognitive patterns.
Calculate Your Potential AI ROI
Estimate the impact of aligning your AI models with human-centric conceptualizations. More aligned AI can lead to greater efficiency and reduced errors.
Estimated Annual Savings
0
Productive Hours Reclaimed
0
Your Path to Human-Aligned AI
Implementing sensorimotor alignment into your enterprise AI is a strategic journey. Here's a typical roadmap:
Discovery & Baseline Assessment
Understand current AI conceptualization, identify key human-AI interaction points, and establish an alignment baseline using frameworks like ours. This phase leverages detailed linguistic analysis to pinpoint existing misalignments.
Targeted Intervention Design
Develop custom sensorimotor priors and prompt-based interventions based on identified misalignments. This includes curating domain-specific image schema knowledge bases and few-shot examples.
Pilot & Iteration
Implement interventions in a controlled pilot environment, evaluate performance against human benchmarks, and iterate on prompt designs and model tuning strategies for optimal alignment.
Integration & Scaling
Integrate aligned AI models into broader enterprise applications, develop monitoring tools for continued alignment, and scale successful strategies across the organization for widespread impact.
Ready to Unlock More Human-Like AI?
Schedule a free consultation with our AI alignment specialists. We'll show you how to leverage sensorimotor regularities to make your LLMs more intuitive, reliable, and powerful for your business needs.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat