I Have No Mouth, and I Must Rhyme: Uncovering Internal Phonetic Representations in LLaMA 3.2

TL;DR

This work tackles whether Llama-3.2-1B-Instruct encodes internal phonetic representations capable of performing phonetic tasks like rhyming without audio grounding. It employs linear probes, causal embedding interventions of the form $E = E + c(\mu - \xi)$, and activation patching to uncover a dedicated phoneme mover head that promotes phonetic information into final predictions. The results show recoverable phoneme directions in token embeddings, cross-lingual phoneme promotion, and a PCA-driven vowel geometry that forms an emergent vowel chart, with partial alignment to human IPA but model-specific deviations. These findings illuminate a tangible internal phonetic model in Llama and have implications for interpretability and cross-lingual phonetic reasoning in large language models.

Abstract

Large language models demonstrate proficiency on phonetic tasks, such as rhyming, without explicit phonetic or auditory grounding. In this work, we investigate how \verb|Llama-3.2-1B-Instruct| represents token-level phonetic information. Our results suggest that Llama uses a rich internal model of phonemes to complete phonetic tasks. We provide evidence for high-level organization of phoneme representations in its latent space. In doing so, we also identify a ``phoneme mover head" which promotes phonetic information during rhyming tasks. We visualize the output space of this head and find that, while notable differences exist, Llama learns a model of vowels similar to the standard IPA vowel chart for humans, despite receiving no direct supervision to do so.

Figures (9)

• Figure 1: Patterns emerge from vowel representations under our methodology, revealing a world model partially inconsistent with human anatomy.
• Figure 2: Example intervention on rhymes with leet. Blue indicates that the $\xi$ phoneme /i/ is present, red indicates that the $\mu$ phoneme /E/ is present.
• Figure 3: Across three examples: Attention patterns of H13L12 at the final token shows the head attends to the rhyme target token. Running logit lens on the corresponding result vector (shown after $\rightarrow$) produces phonetically similar tokens.
• Figure 4: Phoneme vectors reduced to 2 dimensions using PCA trained on H13L12 result vectors. Phonetic patterns emerge among consonant voicedness (top) and vowel backness (bottom).
• Figure 5: Standard IPA representation of vowels in humans (left) and a proposed IPA-style representation of Llama's internal vowel model (right). The positions of some phonemes, such as /a/ and /I/, notably differ.