Exploiting Memorization: Understanding the CLM Objective for Knowledge Encoding in LLMs

Table of Links

Abstract and 1. Introduction

1. Background

2. Method

3. Experiments

   4.1 Multi-hop Reasoning Performance

   4.2 Reasoning with Distractors

   4.3 Generalization to Real-World knowledge

   4.4 Run-time Analysis

   4.5 Memorizing Knowledge

4. Related Work

5. Conclusion, Acknowledgements, and References

\ A. Dataset

B. In-context Reasoning with Distractors

C. Implementation Details

D. Adaptive Learning Rate

E. Experiments with Large Language Models

2 Background

Notation We use f : X × θ → Y to refer to parameterised functions in which X is the set of possible inputs and θ are their possible weights (parameters). We use fθ : x 7→ f(x, θ) to easily refer to any f with a given set of parameters θ. We describe reasoning problems using tuples (K, x, y∗ , Y ) such that y ∈ Y is the correct answer for the question x given facts K, and use D to refer to sets of such problems. When it is clear from context, we drop Y and use only (K, x, y∗ ).

\ Language Modeling and Memorization In the causal language modeling (CLM) objective, a parameterized model fθ is trained to estimate the conditional probabilities of each token in a sequence given its predecessors: p(xt|x) Specifically, we train fθ to approximate p using the CLM loss:

\

\ This training objective allows language models to memorize individual training examples [10, 11], and we will exploit this ability to memorize and draw on contextual knowledge in our work.

\

\

:::info Authors:

(1) Zeming Chen, EPFL ([email protected]);

(2) Gail Weiss, EPFL ([email protected]);

(3) Eric Mitchell, Stanford University ([email protected])';

(4) Asli Celikyilmaz, Meta AI Research ([email protected]);

(5) Antoine Bosselut, EPFL ([email protected]).

:::

:::info This paper is available on arxiv under CC BY 4.0 DEED license.

:::

\