References

• Arxiv - Evaluate Agent with Agent
• GitHub - Eval-Assist (IBM) - LLM as a Judge
• BLEU - ROUGE score
• Giskard - LLM as a judge
• Arxiv - Lessons from red teaming 100 Generative AI Products
  • Lnkd Post
• AWS Blog - RAG evaluation with LLM as a judge
• Github - RAGAS
• DeepLearningAI - Safe and reliable AI via guardrails
• Linkedin - Evaluation
• Keynote - Evaluating LLM Models for Production Systems: Methods and Practices
  • Wide variety of tasks is a typical for use patterns of LLMs in every business
  • LLM evaluation
    • MEASURING MASSIVE MULTITASK LANGUAGE UNDERSTANDING
    • PROXYQA: An Alternative Framework for Evaluating Long-Form Text Generation with Large Language Models
    • TrustLLM
    • TrustScore
    • Reasoning tasks, requires specific evaluation (ex ConceptArc)
    • Multi turn tasks, requires a specific evaluation (ex LMRL Gym)
    • Instruction following
  • Embedding vs LLM - Generative Representational Instruction Tuning
    • LLM Embedding evaluation - MTEB (Massive Text Embedding Benchmark), Huggingface leaderboard
    • In Context Learning (ICL) evaluation: HellaSwag, OpenICL article
  • LLM as a judge: MT-bench & Chatbot Arena
  • RAG Evaluation:
    • RAG Evaluation - RAGAs, Github - Python module
    • RAG Evaluation - ARES, Github - python module
    • RAG Evaluation - HumanEval, Github - python code
  • Hallucination Evaluation - HaluEval
  • Pairwise comparaison - Arenas:
    • Chatbot arena for generic user tasks
    • Negotiation area for negotiation tasks
    • A paper from Cohere about model ranking based on tournaments
  • Hardness Evaluation:
    • EleutherAI/lm-evaluation-harness, very powerful, for zero and few short tasks
    • OpenAI Evals
    • bigcode-project/bigcode-evaluation-harness, code evaluation tasks
    • MLFlow LLM Evaluate , integrated with ML Flow,
    • MosaicML Composer, icl tasks, superfast, scaling to multi gpu
    • RAGAs for LLM based evaluation https://docs.ragas.io/en/latest/
    • TruLens
• Lnkd Gregoire M - Incertitude LLM
• Arxiv - Generating with Confidence: Uncertainty Quantification for Black-box Large Language Models: How to Measure LLM Uncertainty and Why It Matters?
  • Researchers from the University of Illinois propose a three-step method:
    1. Generate 10 or more responses to the same question.
    2. Compute a similarity matrix between the responses.
    3. Derive an uncertainty score based on response diversity and inconsistency.
  • Two key criteria validate the approach:
    • Error prediction: Uncertainty should correlate with incorrect answers (evaluated via AUC-ROC analysis).
    • Error rejection: Filtering uncertain responses should improve accuracy (measured via rejection curve analysis).
  • Findings?
    • The worst approach is asking the LLM to estimate its own uncertainty.
    • The best method combines: NLI-based implication scores for the similarity matrix. + Spectral clustering of responses to compute the final uncertainty score.
  • Challenges and Limitations:
    • High cost: Each score requires 10+ LLM calls.
    • Requires non-zero temperature, making it inapplicable to deterministic responses.
    • Generic context: NLI models used are general-purpose and may not fit business-specific cases.
    • Lack of interpretability: The uncertainty scores (e.g., 0.4 to 10) lack clear business or probabilistic meaning.
    • Short responses work best: For long responses, contradictions get diluted, biasing NLI models trained on short sentences.
• Practical Guide - LLM as a Judge
• Github - lmms eval: evaluation module from microsoft
• Arxiv - Agent as a Judge: Evaluate Agents with Agents
• How to implement LLM as a Judge to test AI Agents?
• Github - deepeval
• GoogleBlog - DeepMind/Giskard evaluation