GenAI Prompting

LLM Prompting - In-Context-Learning🔗

[LLMLingua] Compressing Prompts for Accelerated Inference of LLMs🔗

Arxiv: https://arxiv.org/abs/2310.05736 Oct 2023 Microsoft

flowchart LR
    Input["Long Prompt\n(10k+ tokens)"] --> BC[Budget Controller]
    BC --> Demo[Demonstrations]
    BC --> Inst[Instructions]
    BC --> Q[Questions]
    Demo --> |High compression| TIC[Token Iterative\nCompression]
    Inst --> |Low compression| TIC
    Q --> |Low compression| TIC
    TIC --> IT[Instruction Tuning]
    IT --> Output["Compressed Prompt\n(~500 tokens)"]

Key Components:

Budget Controller
- Allocates different compression ratios to prompt components
- Prioritizes instructions and questions over demonstrations
- Maintains semantic integrity under high compression
Token-level Iterative Compression
- Divides target prompt into segments
- Uses smaller model for perplexity distribution
- Concatenates compressed segments for accurate probability estimation
Instruction Tuning
- Aligns distribution between language models
- Improves compression quality

[TOT] Tree of Thoughts: Deliberate Problem Solving with LLMs🔗

Arxiv: https://arxiv.org/abs/2305.10601 17 May 2023

flowchart TD
    Problem[Problem] --> T1[Thought 1]
    Problem --> T2[Thought 2]
    Problem --> T3[Thought 3]
    T1 --> |Evaluate| T1a["Thought 1.1 ✓"]
    T1 --> T1b[Thought 1.2]
    T2 --> |Evaluate| T2a[Thought 2.1]
    T2 --> T2b["Thought 2.2 ✓"]
    T1a --> |Search| Sol1["Solution ★"]
    T2b --> Sol2[Solution]

    style T1a fill:#90EE90,color:#000
    style T2b fill:#90EE90,color:#000
    style Sol1 fill:#FFD700,color:#000

Key Features:

Frames problems as search over a tree
Each node represents a partial solution
Four key components:
1. Decomposition of intermediate process
2. Generation of potential thoughts
3. Heuristic evaluation of states
4. Search algorithm selection

[COT] Chain-of-Thought Prompting Elicits Reasoning in LLMs🔗

Arxiv: https://arxiv.org/abs/2201.11903 Jan 2022

flowchart LR
    subgraph Standard["Standard Prompting"]
        Q1[Question] --> A1[Answer]
    end

    subgraph COT["Chain-of-Thought"]
        Q2[Question] --> S1["Step 1: ..."]
        S1 --> S2["Step 2: ..."]
        S2 --> S3["Step 3: ..."]
        S3 --> A2[Answer]
    end

    style COT fill:#e1f5fe,color:#000

Limitations:

No guarantee of correct reasoning paths
Manual annotation costs for few-shot setting
Potential for both correct and incorrect answers

[Self-Discover] LLM Self-Compose Reasoning Structures🔗

Arxiv: https://arxiv.org/abs/2402.03620 6 Feb 2024

flowchart TD
    subgraph Stage1["Stage 1: Compose Structure"]
        Task[Task Description] --> Select[SELECT relevant modules]
        Select --> Adapt[ADAPT to task]
        Adapt --> Implement[IMPLEMENT as JSON]
    end

    subgraph Stage2["Stage 2: Solve"]
        Instance[Problem Instance] --> Apply[Apply Structure]
        Apply --> Answer[Final Answer]
    end

    Implement --> Instance

    style Stage1 fill:#fff3e0,color:#000
    style Stage2 fill:#e8f5e9,color:#000

Two-Stage Process:

Stage 1: Task-level reasoning structure
- Uses three actions to guide LLM
- Generates coherent reasoning structure
Stage 2: Instance solving
- Follows self-discovered structure
- Arrives at final answer

[Intent-based Prompt Calibration] Enhancing prompt optimization with synthetic boundary cases🔗

Arxiv: https://arxiv.org/abs/2402.03099 5 Feb 2024

flowchart TD
    Start[Initial Prompt] --> Gen[Generate Boundary Cases]
    Gen --> Eval{Evaluate on\nBoundary Cases}
    Eval --> |Failures found| Improve[Suggest Improved Prompt]
    Improve --> Gen
    Eval --> |No improvement| End[Final Optimized Prompt]

    style End fill:#90EE90,color:#000

Process:

Start with initial prompt and task description
Iteratively:
- Generate challenging boundary cases
- Evaluate current prompt
- Suggest improved prompt
Terminate when no improvement or max iterations reached

[Text2SQL Prompting] Enhancing Few-shot Text2SQL Capabilities of LLM🔗

Arxiv: https://arxiv.org/abs/2311.16452 Nov 2023 Yale

Key Findings:

Dual emphasis on diversity and similarity in examples
Database knowledge augmentation benefits
Code sequence representation for databases
Sensitivity to number of demonstration examples

[MedPrompt] Can Generalist FM Outcompete Special-Purpose Tuning?🔗

Arxiv: https://arxiv.org/abs/2311.16452 28 Nov 2023 Microsoft

flowchart LR
    Q[Medical\nQuestion] --> CoT[Self-Generated\nChain-of-Thought]
    CoT --> Ens[Ensemble with\nSelf-Consistency]
    Ens --> Shuffle[Choice\nShuffling]
    Shuffle --> Verify[Label\nVerification]
    Verify --> A["Answer\n(90%+ accuracy)"]

    style A fill:#FFD700,color:#000

Key Features:

Self-generated chain-of-thought
Ensembling with self-consistency
Choice shuffling for bias reduction
Label verification for hallucination mitigation

[URIAL] Rethinking Alignment via In-Context Learning🔗

Arxiv: https://arxiv.org/abs/2312.01552 4 Dec 2023 Allen Institute

flowchart LR
    subgraph Traditional["SFT/RLHF Alignment"]
        Base1[Base LLM] --> SFT[Supervised\nFine-tuning]
        SFT --> RLHF[RLHF]
        RLHF --> Aligned1[Aligned LLM]
    end

    subgraph URIAL["URIAL (Tuning-Free)"]
        Base2[Base LLM] --> ICL["3 Style Examples\n+ System Prompt"]
        ICL --> Aligned2[Aligned LLM]
    end

    style URIAL fill:#e8f5e9,color:#000

Key Points:

Tuning-free alignment method
Requires only three stylistic examples
Supports Superficial Alignment Hypothesis
Token distribution analysis shows minimal shifts

[CoVE] Chain-of-Verification Reduces Hallucinations in LLM Models🔗

Arxiv: https://arxiv.org/abs/2309.11495 25 Sep 2023 Meta

flowchart TD
    Q[Question] --> Draft["1. Draft Response"]
    Draft --> Plan["2. Plan Verification\nQuestions"]
    Plan --> V1["Q1: Is X true?"]
    Plan --> V2["Q2: Is Y correct?"]
    Plan --> V3["Q3: Does Z hold?"]
    V1 --> |Independent| A1["A1: Yes/No"]
    V2 --> |Independent| A2["A2: Yes/No"]
    V3 --> |Independent| A3["A3: Yes/No"]
    A1 --> Final["4. Final Verified\nResponse"]
    A2 --> Final
    A3 --> Final

    style Final fill:#90EE90,color:#000

Process:

Draft initial response
Plan verification questions
Answer questions independently
Generate final verified response