SalesLLM - Chinese (CustomerLM User): leaderboard

Chinese single-product sales dialogues on SalesLLM (2026), financial services and consumer goods scenarios at five difficulty tiers, where the model sells to CustomerLM, the Qwen3-8B customer model of the authors, simulating the customer for up to 20 rounds. The score from 1 to 10 averages the buying intent of the customer at the end of the dialogue, read by a fine-tuned classifier, with an LLM-judge rubric score for deal progression; higher is better.

Metric: SalesLLM benchmark score (1-10): 0.5 x buying-intent score (end-of-dialogue customer intent from a fine-tuned BERT classifier, five levels mapped to 2, 4, 6, 8, 10) + 0.5 x selling-performance score (0-10 LLM-judge rubric on verbal purchase commitment, concrete next steps, key-information elicitation and objection resolution, credited only on customer-side evidence), on the curated SalesLLM single-product scenarios (Financial Services and Consumer Goods, five difficulty tiers) in Chinese, with CustomerLM (the authors' Qwen3-8B customer model trained with SFT and DPO on crowdworker sales conversations) as the simulated customer; the model acts as the salesperson for at most 20 rounds at temperature 0.8; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1glm-4-9B7.14#2738
2DeepSeek V3.1Non-reasoning7.03#511
3Doubao-1.5-Pro-32k6.89#694
4GLM-4.66.86#498
5GPT-4o6.15#751
6Gemini 3 Flash5.93#175
7Gemini 3 Pro5.93#145
8GPT-5 Nano5.92#906
9Qwen 3 32B5.79#972
10Qwen 2.5 72B5.7#1188
11Gemma 3 27B5.66#1663
12Qwen 3 8B5.64#1873
13MiMo-V2-Flash5.59#684
14Qwen 3 Max5.55#415
15Llama 3.3 70B5.32#1514

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

Interactive version: theaggregate.ai/benchmark?slug=salesllm-chinese-customerlm-user · How It Works · Data refreshed daily, snapshot 2026-11-21.