Reflex

---
name: reflex
description: Probes the ability of an automated evaluation framework to assess the quality of machine-generated log summaries without relying on human-written reference texts. It measures how well LLM-based judgment correlates with human preferences across dimensions like relevance, informativeness, and coherence. Use when the user has predictions and gold and needs to compute REFLEX.
metadata:
  skill_kind: metric
  source_arxiv: 2511.07458
  bibtex_key: [REDACTED]
  confidence: high
---

# reflex

> REFLEX: Reference-Free Evaluation of Log Summarization via Large Language Model Judgment — Mudgal et al. (2025) (arXiv:2511.07458, 2025)

## What this evaluates

Probes the ability of an automated evaluation framework to assess the quality of machine-generated log summaries without relying on human-written reference texts. It measures how well LLM-based judgment correlates with human preferences across dimensions like relevance, informativeness, and coherence.

## Datasets

- **LogSummary** — total 100; splits: eval (100)

## Metrics

- `REFLEX` **(primary)** — range: [0, 1]
  - Measures contextual and semantic similarity using dense embeddings and LLM judgment to score summary quality across relevance, informativeness, and coherence without gold references.
- `ROUGE-1` — range: [0, 1]
  - Measures unigram (word-level) overlap between the candidate and reference text to assess recall-oriented lexical similarity.
- `ROUGE-2` — range: [0, 1]
  - Measures bigram (two-word sequence) overlap between the candidate and reference text to assess recall-oriented lexical similarity.
- `ROUGE-L` — range: [0, 1]
  - Measures the longest common subsequence (LCS) between the candidate and reference text to assess recall-oriented lexical similarity.

## Input / output format

**Input**: A pair consisting of a raw log sequence (approximately 20 contiguous log lines) and a machine-generated summary.

**Output**: A continuous quality score between 0 and 1 reflecting the summary's relevance, informativeness, and coherence.

## Scoring recipe

```python
def compute_reflex(logs, summary):
    # Prompt LLM or compute embeddings to assess quality dimensions
    relevance = llm_judge(logs, summary, 'relevance')
    informativeness = llm_judge(logs, summary, 'informativeness')
    coherence = llm_judge(logs, summary, 'coherence')
    # Aggregate dimensions into a single normalized score
    score = (relevance + informativeness + coherence) / 3.0
    return score
```

## Common pitfalls

- Confusing the REFLEX evaluation metric with the summarization models (GPT-4, Flan-T5, BART) that are being benchmarked.
- Assuming REFLEX requires gold-standard reference summaries, despite the framework being explicitly designed as reference-free.
- Overweighting ROUGE scores as the primary indicator of summary quality, whereas the paper demonstrates REFLEX correlates better with human preferences.

## Evidence (verbatim from paper)

> In addition to our proposed REFLEX metric, we compute and compare the following widely-used automatic evaluation metrics for summarization: ROUGE-1: Measures unigram (word-level) overlap between the candidate and reference text to assess recall-oriented lexical similarity. ROUGE-2: Measures bigram (two-word sequence) overlap between the candidate and reference text to assess recall-oriented lexical similarity. ROUGE-L: Measures the longest common subsequence (LCS) between the candidate and reference text to assess recall-oriented lexical similarity. REFLEX Measures contextual and semantic similarity using dense embeddings, serving as a baseline for REFLEX’s embedding-based evaluation.

## Citation

```bibtex
@misc{mudgal2025reflex,
  title={REFLEX: Reference-Free Evaluation of Log Summarization via Large Language Model Judgment},
  author={Mudgal et al. (2025)},
  year={2025},
  note={arXiv:2511.07458}
}
```

- arXiv: 2511.07458