To Copy or Not to Copy: Copying Is Easier to Induce Than Recall

TL;DR

This paper investigates how language models choose between recalling stored knowledge and copying from context, introducing an arbitration vector to steer this behavior and analyzing the mechanisms behind copying and recall processes.

Contribution

The study presents a novel arbitration vector for controlling model behavior between copying and recalling, with detailed mechanistic insights into these processes.

Findings

Copying is easier to induce than recall in language models.

Inducing copying can be triggered at various input locations.

Restoring recall requires more fragile suppression techniques.

Abstract

Language models used in retrieval-augmented settings must arbitrate between parametric knowledge stored in their weights and contextual information in the prompt. This work presents a mechanistic study of that choice by extracting an \emph{arbitration vector} from model activations on a curated dataset designed to disentangle (i) irrelevant contexts that elicit parametric recall and (ii) relevant but false contexts that elicit copying. The vector is computed as the residual-stream centroid difference between these regimes across 27 relations, and is injected as an additive intervention at selected layers and token spans to steer behavior in two directions: Copy$\rightarrow$Recall (suppressing context use) and Recall$\rightarrow$Copy (inducing the model to copy any token from the context). Experiments on two architectures (decoder-only and encoder/decoder) and two open-domain QA benchmarks show consistent behavior shifts under moderate scaling while monitoring accuracy and fluency. Mechanistic analyses of attention routing, MLP contributions, and layer-wise probability trajectories reveal an asymmetry: inducing copying is an easy ``reactivation'' process that can be triggered at different locations in the input, while restoring recall is a ``suppression'' process that is more fragile and strongly tied to object-token interventions.