Research map / Large language models / Alignment and preferences
AutoMetrics: Approximate Human Judgments with Automatically Generated Evaluators
Michael J Ryan, Yanzhe Zhang, Amol Salunkhe, Yi Chu, Di Xu, Diyi Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| generate and regress task-specific metrics (AutoMetrics) instead of a single LLM judge | GPT-4o-mini · EvalGen | Kendall tau: 0.161 -> 0.334 | Improved |
| AutoMetrics instead of LLM-Judge | Qwen-3-32B · RealHumanEval | Kendall tau: 0.025 -> 0.145 | Improved |
| AutoMetrics instead of LLM-Judge | GPT-4o-mini · CoGym | Kendall tau: 0.199 -> -0.034 | Hurt |
| combine generated metrics with existing MetricBank metrics (vs generated only) (ablation) | AutoMetrics · CoGym and EvalGen (small out-of-distribution) | Kendall tau: CoGym 0.433 -> 0.329; EvalGen 0.503 -> 0.474 | Hurt |
| add existing MetricBank metrics to generated metrics (ablation) | AutoMetrics · HelpSteer2 | Kendall tau: 0.308 -> 0.387 | Improved |
| use metric cards instead of one-sentence descriptions in retrieval (ablation) | AutoMetrics retrieval · EvalGen | Kendall tau: 0.427 -> 0.522 (k=20) | Improved |
| retrieve more metrics (k=5 to k=20/30) (ablation) | AutoMetrics retrieval · EvalGen | Kendall tau: 0.414 -> 0.522 | Improved |
| regress on n=5 metrics instead of single best (no regression) (ablation) | AutoMetrics regression · SimpEval | Kendall tau: 0.232 -> 0.275 | Improved |
| vary number of regressed metrics n (ablation) | AutoMetrics regression · five tasks | Kendall tau | Mixed |
| increase training set size beyond about 80 samples (ablation) | AutoMetrics · SimpEval, HelpSteer2, RealHumanEval | Kendall tau | No clear effect |
| use AutoMetrics as GEPA optimization reward instead of verifiable reward | ReAct agent on tau-airline · tau-bench airline | accuracy: 0.680 -> 0.720 | No clear effect |
| GEPA optimization with AutoMetrics reward vs unoptimized agent | ReAct agent on tau-airline · tau-bench airline | accuracy: 60% -> 0.720 | Improved |
Closest papers by meaning
- Evaluating the Evaluator: Measuring LLMs’ Adherence to Task Evaluation Instructions (AAAI 2025 · similarity 0.89)
- Position: AI Evaluation Should Learn from How We Test Humans (ICML 2025 · similarity 0.88)
- Beyond the Surface: Enhancing LLM-as-a-Judge Alignment with Human via Internal Representations (NeurIPS 2025 · similarity 0.87)
- Validating LLM-as-a-Judge Systems under Rating Indeterminacy (NeurIPS 2025 · similarity 0.87)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.87)
- From Jack of All Trades to Master of One: Specializing LLM-based Autoraters to a Test Set (ICML 2025 · similarity 0.86)
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.
