Alibaba's SPAR sales agent lifts conversion 15.8% in A/B test across 200k+ merchants
ycombinator · x · 2026-11-04
Alibaba deployed SPAR, a sales agent on its 1688 B2B marketplace, and a live A/B test across 200,000+ merchants showed a 15.8% conversion lift — without touching model weights.
Key mechanics: SPAR treats each selling strategy as a hypothesis, starting at a 50/50 prior that converges to its true success rate as results come in, so lucky wins don't skew it. When a merchant churns for external reasons (e.g., budget cuts), an LLM marks it as such so it doesn't count against the pitch.
Ablations vs. direct prompting: no outer loop +12.6%, no inner loop +10.1%, no evolution +9.6% (matching Reflexion), while ExpeL's free-text lessons managed only +1.1%.
The poster notes their own product Scout (giga.ai) follows a similar approach: define a KPI, learn from every conversation, and keep testing improvements.
More from coding & agent
- A start-to-finish workflow for AI dialogue scenes, from script to 1080p — DavidmComfort · 2026-11-04
- Agent37 launches $4.76/month AI agent sandboxes, undercutting Vercel's $249 by 98% — rohanpaul_ai · 2026-11-04
- Open-source Mercury Desktop ships a 5MB AI agent app for all platforms — HankYeomans · 2026-11-04
- Corebrum OS: open-source Physical AI OS now installable on NVIDIA Jetson and any Ubuntu machine — chrismatthieu · 2026-11-04
- Reddit user open-sources full MiniMax H3 ComfyUI workflows: TTS, image editing, 3MP upscale — Support_Marmoset · 2026-11-04
- TLAPS-Bench: 265 real-world TLA+ proof targets; top model GPT-6.1 Sol solves only 44.3% — tianyin_xu · 2026-11-04