HEART-Bench - Naive-RAG: leaderboard

HEART-Bench (2026) tests whether LLM agents can act like a specific person: given only retrieved memories of one of 11 Big Five-grounded characters, the model picks which of four options that character would choose, over 673 expert-validated scenario decisions. This board is accuracy in percent when memories are supplied by plain retrieval-augmented generation (chance 25); higher is better.

Metric: Accuracy (%) with plain retrieval-augmented memory (Naive-RAG), times 100, 673 four-option multiple-choice questions on what one of 11 Big Five-grounded characters would do in a scenario, the prompt exposing only the character ID, occupation and retrieved de-identified episodic memories (Qwen3-Embedding-4B retriever, GPT-5.4-mini auxiliary memory manager); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)63.3#99
2Gemini 3 Flash56.3#175
3DeepSeek V3.241.2#419
4DeepSeek V4 Pro40.7#196
5Qwen 3.5 397B A17B40.3#264
6Claude Sonnet 4.640.1#173
7Qwen 3.5 122B A10B39.1#346
8GPT-5.437.2#135
9Qwen 3.5 35B A3B37#492
10Claude Haiku 4.535.7#583
11DeepSeek V4 Flash34.8#267
12GPT-5.4 Mini33#406

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=heart-bench-naive-rag · How It Works · Data refreshed daily, snapshot 2026-11-21.