CONVERSATIONAL AI LLMs

Question 1

4 challenging factors in speech recognition

Answer 1

• Variability in speech: accent, dialect, background noise
• Domain-specific lexicon
• Contextual understanding
• Speech disfluency: self-corrections, repetitions

Question 2

What are important attributes of speech synthesis

Answer 2

Natural, emotional, and real-time

Question 3

Task-oriented conversations vs open domain conversations

Answer 3

Task-oriented
-clear goal
-limited knowledge
-short conversation
-predictable

open domain
-no exact goal
-unrestricted
-long-session
-unpredictable
-emotions and empathy
We have to pass the current question but also the context and rank the best response

Question 4

How do we train LLMs

Answer 4

Train the base model and then
Use reinforcement learning from human feedback RLFH

Question 5

What are challenges of conversational AI

Answer 5

• consistent conversations (stick with facts and persona)
• empathic agents for user experience
• terrible at detecting misinformation
• using large contexts; often middle context is not helpful

Question 6

What are Hallucinations

Answer 6

When an AI generates something that is not true

Question 7

What are Intrinsic hallucinations

Answer 7

Response is not grounded in the context
obviously wrong

Question 8

What are Extrinsic hallucinations

Answer 8

Incorporates incorrect information
Harder to recognise

Question 9

How can we improve generated responses

Answer 9

Few-shot prompting (in-context learning)
providing examples in context

Roleplay - adopting a role/function

Prompt engineering (chain of thought)
Ask model to generate the reasoning step to their answer, allows it to perform better

Finetuning
Trained on huge dataset and fine-tuned on smaller relevant data
But hard to introduce new knowledge

Question 10

Restricting context length: RAG

Answer 10

Retrieval augmented generation
Take query and only retrieve relevant context
computationally longer

Question 11

RAG vs Fine tune

Answer 11

Start with prompt engineering

if want to improve short term memory -> RAG

if want to improve long term memory (persona or language it uses) -> Fine tune

These methods are all additive - can use all

Question 12

How to evaluate generative LLMs

Answer 12

• factuality of answer (faithfulness)
• relevancy to question

Question 13

What are Foundation Models

Answer 13

Models that are capable of a range of general tasks
And then we can build things on top of it

Question 14

What do LLMs learn

Answer 14

Formal competence
- knowledge of linguistic rules and patterns;
- producing reasonable utterances

Functional competence
- understanding and using language in the world
- showing the understanding?

LLMs master formal well but functional less so

Question 15

Prompting vs fine tuning

Answer 15

Fine tuning
Using a foundational model and fine tuning it on specific data - the model changes

Prompting
Can involve only instruction or +context or +input text

Question 16

What is prompt engineering

Answer 16

the practice of developing and optimizing prompts to efficiently use LLMs for a variety of applications

Question 17

What is in-context learning

Answer 17

Prompting language model performs a task just by conditioning on input-output examples (changing prompts), without optimising any parameters
1) pertaining
2) warming up (optional eg fine tune)
3) scoring - putting together input and most feasible answer

Question 18

What is Warming up

Answer 18

Optional fine tuning for a prompting model
- Supervised in-context training: fine-tune on a a broad range of tasks
- Self-supervised in-context training: use the frozen PLM to generate some synthetic training data and fine tune on it

Model becomes better trained on ICL
But requires updating weights

Question 19

What is zero-shot learning

Answer 19

No example of expected behaviour from the model

Question 20

Prompting tips

Answer 20

• Be specific, clear, and concise.
• Include context.
• Iterate: subsequent prompts are typically needed
• Use as domain-specific instructions as possible
• Roleplay

Question 21

Prompting problems

Answer 21

• Can introduce bias (using context)
• Requires domain expertise ; eg specific medical field
• still lags behind SotA model tuning results
• sub-optimal and sensitive

Question 22

How can we reduce Hallcuinations

Answer 22

Training it to say “I dont know”

Evidence-based response eg RAG: model has a “reality check” against real-world data
Context retrieval provides ‘evidence’: enables factual consistency

Question 23

What is self reflective RAG

Answer 23

RAG and then Validate the response to see if it is coming from the right context

Question 24

How to evaluate Retrieval LLMs

Answer 24

• signal to noise ration of retrieved context (context precision)
• did it retrieve all relevant information (context recall)