SPAR paper: a sales agent that tests its strategies before rewriting memory

An arXiv paper describes SPAR, a B2B sales agent that evolves a shared strategy playbook without weight updates, tested live on 200,000+ merchants.

A paper posted to arXiv on Nov 1, 2026 describes SPAR, an LLM sales agent for multi-session B2B outreach. It learns from outcomes without retraining the model. The authors report a +15.8% conversion lift and +35.8% growth in high-intent merchants in a live A/B test with more than 200,000 merchants on the 1688 platform. The pattern matters to builders because it includes a check before the agent writes a lesson to memory.

Per the abstract, SPAR runs two loops:

• An inner loop adapts to each merchant using episodic memory carried across sessions.
• An outer loop maintains a shared playbook. Each selling strategy is written as a testable hypothesis, and the playbook is updated through causal attribution.

The authors argue that existing experience-learning methods store lessons as unstructured text. They also treat every outcome as equally informative, so a deal lost to an outside factor can count against a good strategy. Domain experts rated SPAR 4.62 out of 5 on overall sales quality, which the paper says is the highest among the methods tested.

@omarsar0, who describes it as an Alibaba paper, suggests that anyone building agents that learn from production outcomes consider adding an attribution step before writing to memory. That is a pattern you can borrow, not a tool you can install.

The abstract leaves several things unclear:

• What the A/B baseline was
• How long the test ran
• How the attribution step works in practice
• Whether code is available

We found no independent coverage or replication. Treat the numbers as the authors' own results.

Sources

• SPAR: Experience-Driven Strategy Evolution in Multi-Session B2B Sales · arXiv.org
• elvis (@omarsar0) on X · @omarsar0