The Mechanistic Emergence of Symbol Grounding in Language Models

TL;DR

Symbol grounding in language models is shown to emerge even under next-token prediction objectives. The authors build a controlled testbed with distinct environmental ($\langle$ENV$\rangle$) and linguistic ($\langle$LAN$\rangle$) tokens to trace grounding through internal computations, revealing that grounding concentrates in mid-layer attention via a gather-and-aggregate mechanism. This mechanism, validated by saliency analyses and causal interventions, generalizes across Transformers, Mamba-2, and multimodal dialogue settings, but is not present in unidirectional LSTMs. The work provides mechanistic, causal evidence for emergent symbol grounding in LMs and highlights architectural conditions and head-level interventions that could inform reliability and controllability of generation.

Abstract

Symbol grounding (Harnad, 1990) describes how symbols such as words acquire their meanings by connecting to real-world sensorimotor experiences. Recent work has shown preliminary evidence that grounding may emerge in (vision-)language models trained at scale without using explicit grounding objectives. Yet, the specific loci of this emergence and the mechanisms that drive it remain largely unexplored. To address this problem, we introduce a controlled evaluation framework that systematically traces how symbol grounding arises within the internal computations through mechanistic and causal analysis. Our findings show that grounding concentrates in middle-layer computations and is implemented through the aggregate mechanism, where attention heads aggregate the environmental ground to support the prediction of linguistic forms. This phenomenon replicates in multimodal dialogue and across architectures (Transformers and state-space models), but not in unidirectional LSTMs. Our results provide behavioral and mechanistic evidence that symbol grounding can emerge in language models, with practical implications for predicting and potentially controlling the reliability of generation.

Figures (10)

• Figure 1: Illustration of the symbol grounding mechanism through information aggregation. Lighter colors denote more salient attention, quantified by saliency scores, i.e., gradient $\times$ attention contributions to the loss wang2023label. When predicting the next token, aggregate heads bick2025understanding emerge to exclusively link environmental tokens (visual or situational context; $\langle$ENV$\rangle$) to linguistic tokens (words in text; $\langle$LAN$\rangle$). These heads provide a mechanistic pathway for symbol grounding by mapping external environmental evidence into its linguistic form.
• Figure 2: Average surprisal of the experimental and control conditions over training steps.
• Figure 3: Grounding information gain and its correlation to the co-occurrence of linguistic and environment tokens over training steps.
• Figure 4: Average surprisal of the experimental and control conditions in caption- and image-grounded dialogue settings, as well as the grounding information gain and its correlation to the co-occurrence of linguistic and environment tokens over training steps. All results are from a 12-layer Transformer model on grounded dialogue data.
• Figure 5: Overtime mechanistic analysis on GPT-CHILDES.