Research map / Large language models / Alignment and preferences

The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans

NeurIPS 2026 · Large language models › Alignment and preferences · approach: LLM-as-judge and feedback data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Replicates two cognitive-science property-generation experiments and a rating experiment on 21 LLMs to measure how their grounding of abstract concepts differs from humans, and probes Gemma 3 with sparse autoencoders for grounding-related features.

Problem

Whether LLMs ground abstract concepts (e.g., justice, art) in human-like experiential dimensions such as sensorimotor, emotional, and social content when generating properties freely, versus only recovering these dimensions when explicitly queried.

Approach

Applies human property-generation paradigms to 21 LLMs, codes generated properties with LLM coders into the original taxonomies, and measures per-category Pearson correlation with human distributions; contrasts this with a rating experiment and SAE feature analysis to separate failure to recognize dimensions from failure to recruit them.

Key results

• Best model reaches mean Pearson r about 0.37 on Experiment 1 versus an estimated human ceiling of about 0.97.
• On the explicit rating task, the strongest model (Claude Sonnet 4.6) reaches mean r 0.760 versus a human ceiling of about 0.78, and alignment improves with scale on sensory dimensions.
• SAE features in Gemma 3 4B/12B correlate with grounding dimensions at r of about 0.6 to 0.8, and clamping a social feature raised social property generation by 14.7%.

Surprising findings

• Smaller models sometimes outperform much larger ones on property generation alignment (e.g., Llama 3.1 70B and Qwen3 8B rank above GPT-5.4).
• Grounding gap does not shrink with scale on property generation, while rating alignment does improve with scale on sensory dimensions.

Limitations

• Limited to two property-generation datasets with public stimuli and taxonomies, so generality is constrained.
• LLM coders could introduce biases, though human re-coding changed metrics only marginally.
• SAE analysis limited to Gemma models and SAE features are approximations, so mechanistic claims are exploratory.

Experimental setup

Datasets: Harpaintner et al. property-generation norms (293 abstract nouns) Kelly et al. emotion property-generation data (236 abstract words) 751-noun rating dataset (14 ACG dimensions) psycholinguistic norm databases (426 nouns)

Benchmarks: Harpaintner property-generation norms (Experiment 1) Kelly et al. property-generation (Experiment 2) 751-noun grounding rating dataset

Metrics: Pearson r Mean r across categories Cohen's kappa percentage agreement steering category shift

Baselines: human ceiling Gemini 2.5 Flash-Lite GPT-5.4 Claude Opus 4.6

Builds on: property-listing paradigm (Harpaintner et al.) Kelly et al. emotion property generation Gemma Scope 2

Closest papers by meaning

• Divergences between Language Models and Human Brains (NeurIPS 2024 · similarity 0.87)
• Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.87)
• Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (NeurIPS 2025 · similarity 0.87)
• Understanding Annotator Perception: Modeling Psychological Inference from First- and Third-Person Annotations (Student Abstract) (AAAI 2025 · similarity 0.87)
• Uncovering Gaps in How Humans and LLMs Interpret Subjective Language (ICLR 2025 · similarity 0.87)
• Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.86)