Challenges of Large Language Models

Dr. Noman Islam

Introduction

• Motivation: LLMs have become ubiquitous and powerful, but also face many open problems and opportunities
• Goal: Provide a systematic overview of the current state of LLM research and practice

What are LLMs?

• Definition: LLMs are neural network models that can process and generate natural language at a large scale
• Examples: GPT-3, BERT, T5, mT5, etc.
• Capabilities: LLMs can perform a wide range of natural language tasks, such as text summarization, translation, question answering, etc.

How are LLMs trained?

• Pre-training: LLMs are trained on large and diverse corpora of text data, using self-supervised objectives such as masked language modeling or denoising autoencoding
• Fine-tuning: LLMs are adapted to specific downstream tasks or domains, using supervised or semi-supervised learning on smaller labeled datasets
• Prompting: LLMs are controlled by providing natural language instructions or examples as input, without modifying the model parameters

What are the challenges of LLMs?

• Design: LLMs require careful decisions on data collection, tokenization, model architecture, pre-training objective, etc.
• Behavior: LLMs exhibit unpredictable and sometimes undesirable behaviors, such as hallucinations, misalignment, brittleness, etc.
• Science: LLMs pose methodological and theoretical questions, such as evaluation, interpretability, causality, etc.

Challenge 1: Unfathomable Datasets

• Problem: The size and diversity of pre-training datasets make it hard to assess their quality and impact on LLMs
• Examples: Near-duplicates, benchmark data contamination, personally identifiable information, pre-training domain mixtures, etc.
• Solutions: Data cleaning, filtering, deduplication, selection, etc.

Slide 6: Challenge 2: Tokenizer-Reliance

• Problem: Tokenizers introduce several issues, such as computational overhead, language dependence, handling of novel words, fixed vocabulary size, information loss, and low human interpretability
• Examples: Subword tokenization, byte-level tokenization, tokenization-free models, etc.
• Solutions: Tokenizer optimization, adaptation, integration, etc.

Slide 7: Challenge 3: High Pre-Training Costs

• Problem: Pre-training LLMs requires massive amounts of compute, time, energy, and money, which limits accessibility and sustainability
• Examples: Scaling laws, compute-optimal training recipes, model compression, etc.
• Solutions: Pre-training efficiency, effectiveness, and sharing

Red AI

Parallelism Strategies

Challenge 4: Fine-Tuning Overhead

• Problem: Fine-tuning LLMs requires large memory and storage, and results in task-specific models that are not reusable or generalizable
• Examples: Large memory requirements, overhead of storing and loading fine-tuned LLMs, etc.
• Solutions: Fine-tuning optimization, adaptation, and integration

Parameter Efficient Fine tuning

Prompt tuning

Challenge 5: High Inference Latency

• Problem: Inference with LLMs is slow and expensive, especially for long sequences and large models
• Examples: Transformer inefficiency, quadratic complexity, etc.
• Solutions: Inference acceleration, pruning, quantization, caching, etc.

Solutions

Challenge 6: Limited Context Length

• Problem: LLMs cannot process or generate sequences longer than a fixed limit, which restricts their applicability and expressiveness
• Examples: Transformer memory bottleneck, context window size, etc.
• Solutions: Context length extension, chunking, recurrence, attention reduction, etc.

Solution

Challenge 7: Prompt Brittleness

• Problem: LLMs are sensitive to the choice and format of prompts, which affects their performance and reliability
• Examples: Prompt engineering, prompt tuning, prompt learning, etc.
• Solutions: Prompt optimization, adaptation, and integration

Challenge 8: Hallucinations

• Problem: LLMs often generate fluent but inaccurate or unfaithful text, which can be misleading or harmful
• Examples: Intrinsic and extrinsic hallucinations, factual errors, etc.
• Solutions: Hallucination detection, correction, and prevention

Challenge 9: Misaligned Behavior

• Problem: LLMs do not always align with human values, preferences, or expectations, which can lead to ethical or social issues
• Examples: Bias, toxicity, privacy, safety, etc.
• Solutions: Alignment evaluation, mitigation, and enforcement

Challenge 10: Outdated Knowledge

• Problem: LLMs can become obsolete or inaccurate as the world changes, which can affect their relevance and usefulness
• Examples: Static pre-training data, temporal drift, etc.
• Solutions: Knowledge update, refresh, and verification

Challenge 11: Brittle Evaluations

• Problem: LLMs are hard to evaluate in a comprehensive and consistent way, which can hinder their development and comparison
• Examples: Evaluation metrics, benchmarks, datasets, etc.
• Solutions: Evaluation improvement, standardization, and diversification

Challenge 12: Evaluations Based on Static, Human-Written Ground Truth

• Problem: LLMs are often evaluated against fixed and human-written references, which can be unfair or inadequate
• Examples: Reference bias, diversity, quality, etc.
• Solutions: Evaluation adaptation, generation, and interaction

Challenge 13: Indistinguishability between Generated and Human-Written Text

• Problem: LLMs can produce text that is hard to distinguish from human-written text, which can pose security or credibility risks
• Examples: Fake news, spam, phishing, etc.
• Solutions: Text detection, attribution, and verification

Challenge 14: Tasks Not Solvable By Scale

• Problem: LLMs cannot solve some tasks by simply scaling up their size, data, or compute, which requires new paradigms or methods
• Examples: Reasoning, common sense, creativity, etc.
• Solutions: Task modeling, representation, and integration

Challenge 15: Lacking Experimental Designs

• Problem: LLMs are often tested or compared in suboptimal or inconsistent experimental settings, which can affect their validity and reproducibility
• Examples: Hyperparameters, baselines, random seeds, etc.
• Solutions: Experimental design, reporting, and sharing

Challenge 16: Lack of Reproducibility

• Problem: LLMs are often not reproducible due to the lack of code, data, or resources, which can hamper their progress and trustworthiness
• Examples: Closed-source models, proprietary data, high costs, etc.
• Solutions: Reproducibility verification, facilitation, and promotion