MRCR v2: leaderboard

Google DeepMind multi-round coreference benchmark that tests whether a model can disambiguate repeated requests and retrieve the intended item from long conversations.

Source: github.com.

Benchmark results

Results are published separately for each benchmark or evaluation setting.

• LLM Stats (MRCR v2 (8-needle))
• MRCR v2 256K (8-needle)
• MRCR-v2 128k
• Qwen3.7 Launch - MRCR-v2 128k

Interactive version: theaggregate.ai/benchmark?slug=mrcr-v2 · How It Works · Data refreshed daily, snapshot 2026-10-20.