BACH-V: Bridging Abstract and Concrete Human-Values in Large Language Models

TL;DR

This paper addresses whether large language models truly understand abstract concepts or merely manipulate them as statistical patterns. It introduces an abstraction-grounding framework that decomposes conceptual understanding into three capacities: $A$-$A$ (interpretation of abstract concepts), $A$-$C$ (grounding in concrete events), and $C$-$C$ (application to regulate concrete decisions). Using human values as a testbed, the authors develop passive probing and active steering to diagnose representations and causally shift behavior across six open-source LLMs and ten value dimensions, finding that value concepts transfer across abstraction levels and that steering can alter concrete judgments while leaving abstract interpretations relatively stable. These results suggest LLMs maintain structured value representations that bridge abstraction and action, offering a foundation for transparent, generalizable value-alignment and autonomous control. The work provides an operational framework for measuring and manipulating abstract concepts in LLMs, with implications for safer and more predictable AI systems.

Abstract

Do large language models (LLMs) genuinely understand abstract concepts, or merely manipulate them as statistical patterns? We introduce an abstraction-grounding framework that decomposes conceptual understanding into three capacities: interpretation of abstract concepts (Abstract-Abstract, A-A), grounding of abstractions in concrete events (Abstract-Concrete, A-C), and application of abstract principles to regulate concrete decisions (Concrete-Concrete, C-C). Using human values as a testbed - given their semantic richness and centrality to alignment - we employ probing (detecting value traces in internal activations) and steering (modifying representations to shift behavior). Across six open-source LLMs and ten value dimensions, probing shows that diagnostic probes trained solely on abstract value descriptions reliably detect the same values in concrete event narratives and decision reasoning, demonstrating cross-level transfer. Steering reveals an asymmetry: intervening on value representations causally shifts concrete judgments and decisions (A-C, C-C), yet leaves abstract interpretations unchanged (A-A), suggesting that encoded abstract values function as stable anchors rather than malleable activations. These findings indicate LLMs maintain structured value representations that bridge abstraction and action, providing a mechanistic and operational foundation for building value-driven autonomous AI systems with more transparent, generalizable alignment and control.

Figures (16)

• Figure 1: Conceptual overview of abstract-concrete grounding. Value probes trained on abstract concept descriptions (top) capture internal representations that generalize to steering concrete decisions (bottom), enabling behavioral control via recognized abstract patterns.
• Figure 2: A 3D-visualization of passive probing [A.1 - A.3] and active steering pipelines for value probes [B.1 - B.3].
• Figure 3: Correlation between Probe Predictions and Ground Truth Values Across Different Layers for LLMs.
• Figure 4: Experimental operationalization of the abstract-concrete grounding framework. A-A tests interpretation of abstract value concepts. A-C tests grounding of abstract values in concrete event descriptions.C-C tests application of abstract values in concrete decision-making. Each regime is evaluated via both passive probing (detecting value activation) and active steering (intervening on value representations).
• Figure 5: Probing Results in A-A, A-C, and C-C Tests. Top: averaged probing scores for various probes (y-axis) on corpora with various value labels (x-axis), for Qwen3-8B only. Bottom: column-wise relative diagonal dominance in max-sense ($\sum_j\left( M_{ii}-\max_{j, i\ne j}M_{ij}\right)/M_{ii}$), on the corresponding top-row heatmap, for all models.