Is In-Context Learning a Type of Error-Driven Learning? Evidence from the Inverse Frequency Effect in Structural Priming

TL;DR

The paper investigates whether in-context learning (ICL) operates as an error-driven processing mechanism by exploiting the inverse frequency effect (IFE) in structural priming with dative alternations. It compares explicit gradient-based fine-tuning to concatenation-based ICL across a range of Transformer-based models using a corpus of 92,400 prime-target pairs, including a pronoun variant to modulate verb bias exposure. The results show that IFE emerges in several models, especially larger ones and under pronoun-rich conditions, suggesting an implicit error signal during the forward pass and supporting a gradient-descent-like character of ICL in at least some regimes. These findings highlight a link between human processing and machine ICL, propose the IFE as a diagnostic for error-driven learning in ICL, and outline directions to generalize the approach to other tasks and to pursue deeper mechanistic interpretations.

Abstract

Large language models (LLMs) have shown the emergent capability of in-context learning (ICL). One line of research has claimed that ICL is functionally equivalent to gradient descent, a type of error-driven learning mechanism. In this paper, we introduce a new way of diagnosing whether ICL is functionally performing error-driven learning. Our approach is based on the inverse frequency effect (IFE) -- a phenomenon in which an agent's behavior is influenced to a greater degree when presented with improbable examples as compared to more likely ones. The IFE has previously been identified in psycholinguistics where humans exhibit the IFE in the context of structural priming (the tendency for people to produce sentence structures they have encountered recently). In that context, the IFE has been used as evidence that human structural priming must involve error-driven learning mechanisms. In our experiments, we simulated structural priming with ICL and found that LLMs indeed display the IFE, with the effect being stronger in larger models. We conclude that at least in the case we studied, ICL is indeed a type of error-driven learning, supporting the hypothesis that an error signal is implicitly computed in the forward pass during ICL. Our results suggest that both humans and LLMs make use of error-driven processing mechanisms in on-line processing.

Figures (5)

• Figure 1: An overview of our experimental design.
• Figure 2: A demonstration of the IFE: a stronger priming effect of a DO prime is predicted as PD-bias increases. The numerical values of the primed log probabilities are hypothetical, for illustration purposes only.
• Figure 3: GPT2-small shows a robust IFE under the Fine-Tuning mode, shown by the two negative slope.
• Figure 4: The IFE across models of different sizes in the WithPronoun condition under the Concatenation mode.
• Figure 5: Comparison of PD biases with (left) and without (right) pronoun for GPT3. As is shown in Equation. \ref{['eq:vb']}, a high PD-bias means a larger proportion of probability assigned to the PD structure against the DO structure in LLMs.