IC-1631Binding id mechanism fidelity increases with model size in both Llama and Pythia families

Jiahai Feng, Jacob Steinhardt

SourceHow do Language Models Bind Entities in Context?

Figure 6 shows that the effectiveness of binding id interventions (measured by median-calibrated accuracy under control, entity, and attribute conditions) increases monotonically with model parameter count across both the Pythia and Llama families. Smaller models show weaker adherence to the binding id structure, while the largest models in each family show near-perfect fidelity. The paper concludes that only sufficiently large models exhibit the binding id mechanism, suggesting a convergence in representations with scale.

Evidence
correlational
Caveat
Llama-65B was not included in the figure for computational reasons. The exact threshold at which the mechanism emerges is not precisely quantified in the text.
Model
LLaMA, Pythia
Concepts
Scale-dependent behaviour
Datasets
Bias in Bios [eval]
Methods
Causal mediation analysis / Vig et al. 2020 (causal mediation analysis) [primary]
Related findings
IC-1630, IC-1632
Extraction
automatic-extraction