Alibaba and SCUT propose SPAR, a sales agent that tests strategies as hypotheses
Alibaba Group and South China University of Technology propose SPAR, a self-evolving sales agent that refines its playbook without updating model weights.
Alibaba Group and South China University of Technology have proposed SPAR, a sales agent that improves its selling playbook from experience. It does this without changing the weights of the underlying model. The paper, accepted to the EMNLP 2026 Industry Track, is an arXiv preprint by Hangming Zhang, Yao Zhao, Haotian Kuang, C. L. Philip Chen, Tong Zhang and Tengyue Ma.
The authors' argument starts from a gap in existing sales agents. B2B deals usually need several rounds of outreach that build on earlier context. The paper says existing LLM-based sales agents work within single sessions and cannot carry context forward. Methods that do learn from experience store lessons as unstructured text. They treat every outcome as equally informative, with no mechanism to distinguish a flawed strategy from an external factor. SPAR tries to address both problems with a double-loop design.
The inner loop keeps episodic memory for each merchant, so a follow-up can build on the conversations that came before it. The outer loop maintains a shared playbook. In it, each strategy carries a hypothesis, the preconditions under which it should apply, and patterns describing what success and failure look like. The authors treat each strategy as a testable hypothesis. Before the outer loop changes any strategy, SPAR classifies the outcome as caused by the strategy, caused by an external factor, or uncertain. The authors' example is that a deal lost to outside factors does not weaken a strategy that works. The abstract presents this causal attribution as the missing piece in earlier experience-based methods.
The evidence for the live result is an online A/B test, and the DAIR.AI Academy summary adds some limits. The curator's summary on that page says the playbook needs several interactions before it improves, so early merchants do not benefit. It also says that individual LLM-driven playbook updates can be wrong. The same summary reports offline simulation results in which SPAR beats the strongest baseline by 9.6% on objection recognition and 5.5% on evidence support. The 4.62 out of 5 sales-quality score comes from human evaluation by domain experts. The abstract says it is the highest among all methods. The arXiv abstract does not give the baselines behind that comparison, so the score can be read only relative to the systems the authors chose to test against.
What changed
Alibaba Group and South China University of Technology propose SPAR, a sales agent whose shared playbook updates strategies as hypotheses without updating model weights.
Type: Research
Why it matters
Agents that learn from production outcomes can record the wrong lesson when they misread why a result happened. Attribution before memory writes matters beyond sales.
What you can do
Separate strategy failures from external factors before letting an agent update a shared playbook.
Original source
Where we saw it
DAIR.AI (Elvis Saravia)
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works