1 of 11

Explainability for Large Language Models

- Focus on explaining internal mechanisms of LLMs

- Why they behave the way they do?

2 of 11

Emergent Abilities of Large Language Models

-Deepmind

-Stanford

3 of 11

Emergent

- An ability is emergent if it is not present in smaller models but is present in larger models

- Emergent abilities can't be predicted by extrapolating a scaling law

Larger models:-

  • Amount of computation (FLOPs)
  • # Parameters
  • Training dataset size

4 of 11

Results

5 of 11

Are Emergent Abilities of Large Language Models a Mirage?

-Stanford

6 of 11

Hypothesis

- No emergent abilities of LLMs

Metrics used:-

7 of 11

"Exact String Match" accuracy

True string: The sun set behind the mountains.

Candidate #1: A lazy dog

0

Candidate #2: A sun set

0

Candidate #3: The sun set in the hills

0

Candidate #4: The sun set behind the mountains.

1

8 of 11

String edit distance

Minimum operations needed to make string s1 equal to string s2.

Operations allowed:-

  • Substitutions
  • Additions
  • Deletions

9 of 11

New accuracy

True string: The sun set behind the mountains.

Candidate #1: A lazy dog

6

Candidate #2: A sun set

4

Candidate #3: The sun set in the hills

2

Candidate #4: The sun set behind the mountains.

0

10 of 11

Results

11 of 11

Current literature

Few papers exist that try to understand the internal mechanisms of LLMs/Transformers for Graphs.

Focus on explaining:-

  • In-context learning
  • CoT prompting
  • Importance of fine-tuning
  • Hallucination