Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2601.12075 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917207602102272 |
|---|---|
| author | Farahani, Mehrdad Penzkofer, Franziska Johansson, Richard |
| author_facet | Farahani, Mehrdad Penzkofer, Franziska Johansson, Richard |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_12075 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | To Copy or Not to Copy: Copying Is Easier to Induce Than Recall Farahani, Mehrdad Penzkofer, Franziska Johansson, Richard Computation and Language 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. |
| title | To Copy or Not to Copy: Copying Is Easier to Induce Than Recall |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2601.12075 |