Research map / Large language models / Language, safety and interpretability
CLAVE: An Adaptive Framework for Evaluating Values of LLM Generated Responses
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CLAVE, a dual-model framework where a large closed LLM extracts value concepts that a small fine-tuned LLM uses to judge values, and releases the ValEval benchmark (13k+ tuples over three value systems).
Problem
Reference-free LLM evaluators for open-ended human value assessment of generated responses struggle to adapt to changing value definitions with minimal annotation and to generalize across varied expressions, scenarios and domains.
Approach
Link a large closed LLM (concept extractor, GPT-4-1106) and a small fine-tuned LLM (value recognizer, LoRA) through generic "value concepts" extracted from a few labeled samples, clustered and deduplicated into a concept pool, so the small model classifies concepts rather than raw text.
Key results
- CLAVE-Llama reaches 82.12% on Schwartz perturbed and 83.71% on Schwartz generalized, versus 76.29% original-only best tuning baseline Mistral-7b on Schwartz original (76.29%) and 77.69% for Llama2-7b generalized.
- CLAVE-Mistral achieves 75.26% on Schwartz original and 57.38% on Moral Foundation original, the best on that split.
- Prompt-based large-LLM evaluators stay robust on social risks (e.g., FairEval 91.08% generalized) but drop sharply on Moral Foundation generalized (14.06-47.92%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Social Risks Original · Social Risks Original | accuracy (%): 88.57 -> 85.03 | Hurt |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Social Risks Disturb · Social Risks Disturb | accuracy (%): 86.88 -> 78.79 | Hurt |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Social Risks Generalized · Social Risks Generalized | accuracy (%): 92.16 -> 85.41 | Hurt |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Schwartz Value Original · Schwartz Value Original | accuracy (%): 76.29 -> 69.85 | Hurt |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Schwartz Value Disturb · Schwartz Value Disturb | accuracy (%): 81.42 -> 82.12 | Helped |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Schwartz Value Generalized · Schwartz Value Generalized | accuracy (%): 82.35 -> 83.71 | Helped |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Moral Foundation Original · Moral Foundation Original | accuracy (%): 56.13 -> 56.76 | Helped |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Moral Foundation Disturb · Moral Foundation Disturb | accuracy (%): 93.66 -> 93.66 | No clear effect |
| Use CLAVE with Llama-2-7b instead of the strongest non-CLAVE baseline | Moral Foundation Generalized · Moral Foundation Generalized | accuracy (%): 51.76 -> 53.84 | Helped |
| Increase training examples per value from 10 or 20 to 50 or 100 (ablation) | CLAVE and Llama-2-7b · Social Risks and Schwartz Value; original and generalized splits | accuracy | Mixed |
| Use CLAVE rather than the tuning-based Llama-2-7b baseline with limited training data (ablation) | Social Risks and Schwartz Value · 10 or 20 samples per value | accuracy | Helped |
| Use CLAVE with ChatGPT rather than GPT-4 as the large concept extractor (ablation) | CLAVE · Perturbation and generalization splits | accuracy | Helped |
Surprising findings
- Fine-tuned small models with no concept step show sharp drops on generalized splits (e.g., Llama2-7b 22.43% on social risk generalized), and more training data sometimes hurts generalization.
- ChatGPT as the concept extractor still improves smaller models' perturbed and generalized results, even though GPT-4 is much better at direct value assessment.
Limitations
- Relies on closed-source GPT-4 API for concept extraction, adding API cost and reproducibility concerns.
- Perturbation data generated by Mistral-Large and annotation agreement on Moral Foundation is only 72.6%.
Experimental setup
Datasets: BeaverTails Do-not-Answer Value Fulcra DenEvil Moral Stories
Benchmarks: ValEval
Metrics: accuracy
Baselines: Vanilla prompt Few-shot Chain-of-thought G-Eval FairEval ChatEval WideDeep AutoCalibrate ALLURE GPT-2-Large Phi-3 Llama-2-7b-chat
Builds on: LoRA AUTO-J ALLURE AutoCalibrate G-Eval
Closest papers by meaning
- Do LLMs have Consistent Values? (ICLR 2025 · similarity 0.91)
- DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · similarity 0.90)
- Raising the Bar: Investigating the Values of Large Language Models via Generative Evolving Testing (ICML 2025 · similarity 0.90)
- Measuring Human and AI Values Based on Generative Psychometrics with Large Language Models (AAAI 2025 · similarity 0.89)
- Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.89)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.