SAGE (Service Agent) - Conversational Quality: leaderboard

SAGE conversational quality: three judge LLMs rate selected customer-service replies for linguistic quality, human-likeness, usefulness, user satisfaction and instruction compliance. Evaluation samples turns 1, 5, 10, 15 and the final turn, combining judge-weighted ratings on a 0–100 scale. Higher is better.

Metric: Conversational Quality Score (0-100): weighted judge ratings of linguistic quality, anthropomorphism, content utility, user satisfaction and instruction compliance, averaged over three judges, averaged over the six scenarios, over SAGE customer-service dialogues: standard operating procedures formalized as dialogue graphs, an adversarial simulated user with varied intents and personas, turns 1, 5, 10, 15 and the final turn scored; three unnamed judge LLMs (majority vote and mean) plus a deterministic rule engine produce the reference; closed models through their APIs, open models served with vLLM; 0-100 scale; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 27 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.577.38#261
2MiniMax-M2.175.29#575
3Claude Opus 4.574.68#160
4Gemini 3 Pro (Preview)71.89#119
5Gemini 2.5 Pro71.73#276
6Gemini 3 Flash (Preview)68.75#152
7DeepSeek V3.268.11#419
8Seed 1.867.62#252
9Kimi K2.566.93#266
10Qwen 2.5 72B Instruct64.3#1117
11Qwen 3 235B A22B64.13#628
12Qwen 2.5 7B Instruct62.86#2723
13Qwen Max (SAGE checkpoint unspecified)62.44
14DeepSeek R161.94#482
15Qwen 2.5 14B Instruct61.8#1775

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

Interactive version: theaggregate.ai/benchmark?slug=sage-service-agent-conversational-quality · How It Works · Data refreshed daily, snapshot 2026-11-21.