MapSatisfyBench - Factual Faithfulness: leaderboard

MapSatisfyBench (2026) gives map-service agents informal, underspecified requests rebuilt from real user behavior and checks whether they recover the unstated needs through 22 replayed map tools and a simulated user, with GPT-5.3 as judge. This board is the percentage of required factual claims in the answer that the tool outputs support; higher is better.

Metric: Factual-faithfulness rate (%): share of required factual claims in the final response supported by the tool outputs; underspecified map-service requests rebuilt from real anonymized user behavior chains, answered by a ReAct agent with 22 deterministic replayed map tools and a simulated user; GPT-5.3 simulates the user and judges (majority of three judgments), temperature 1; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 12 models tracked. Updated 30 Sep 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)Non-reasoning96.45#93
2Gemini 3.1 Flash (Preview)Non-reasoning94.48#559
3GPT-4.192.15#500
4Qwen 3 30B A3BNon-reasoning92.14#1147
5Claude Sonnet 4.691.93#165
6Qwen 3 8BNon-reasoning91.83#1868
7Qwen 3.6 PlusThinking90.84#226
8Claude Opus 4.690.77#106
9DeepSeek V4 ProThinking90.69#191
10GPT-5.3Non-reasoning89.62#197
11DeepSeek V3.2Non-reasoning88.25#412
12Qwen 3 235B A22BNon-reasoning83.77#628

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=mapsatisfybench-factual-faithfulness · How It Works · Data refreshed daily, snapshot 2026-11-16.