ReSUM - SQuALITY - COMET: leaderboard
ReSUM v1 Table5: query-focused long-meeting summarization on QMSum and SQuALITY; source-native zero-shot, CoT, RAG, fine-tuned and multi-agent methods remain separate run identities. Short source model labels (including LLaMA3-3B and GPT-4-mini) are not bound to a guessed checkpoint. R1/R2/RL, BERTScore, Mover and COMET are native 0–1 values; chrF is native percent. No transformation, model alias, cross-method pooling or resource metric.
Metric: Native score (0–1). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4-mini (ReSUM Run: method=Zero-shot;checkpoint=source-native;effort=unreported) | 0.72 | |
| 2 | FT Qwen (ReSUM Run: method=Fine-tuned (FT);checkpoint=source-native;effort=unreported) | 0.69 | |
| 3 | LLaMA3-3B (ReSUM Run: method=Zero-shot;checkpoint=source-native;effort=unreported) | 0.67 | |
| 4 | Qwen3-4B (ReSUM Run: method=Zero-shot;checkpoint=source-native;effort=unreported) | 0.67 | |
| 5 | COT Qwen (ReSUM Run: method=Chain-of-Thought (CoT);checkpoint=source-native;effort=unreported) | 0.67 | |
| 6 | Gemma3-4B (ReSUM Run: method=Zero-shot;checkpoint=source-native;effort=unreported) | 0.66 | |
| 7 | FT LLaMA (ReSUM Run: method=Fine-tuned (FT);checkpoint=source-native;effort=unreported) | 0.65 | |
| 8 | FT+Res LLaMA (ReSUM Run: method=End-to-End Fine-tuned with Reasoning;checkpoint=source-native;effort=unreported) | 0.65 | |
| 9 | FT+Res Qwen (ReSUM Run: method=End-to-End Fine-tuned with Reasoning;checkpoint=source-native;effort=unreported) | 0.65 | |
| 10 | COT LLaMA (ReSUM Run: method=Chain-of-Thought (CoT);checkpoint=source-native;effort=unreported) | 0.64 |
Interactive version: theaggregate.ai/benchmark?slug=resum-squality-comet · How It Works · Data refreshed daily, snapshot 2026-11-12.