A 4B model just learned to negotiate as well as systems many times its size.
Researchers built SocialRL, a training recipe that reinforces social reasoning (reading a counterpart's goals, holding a position, not blurting out private information) instead of relying on a model's default helpfulness. They applied it to a 4B model across six negotiation-style domains: Deal-or-No-Deal, CaSiNo, Craigslist, Job Interview, Calendar, and Marketplace, always training and testing within the same domain. On held-out scenarios, the tuned model matched or beat the GPT-5 family in every domain, closing 73 to 122 percent of the gap between an untrained baseline and frontier performance. Buyer agents also got noticeably tougher: 78 percent opened below their target price, versus 3 percent before training.
That's the sharper finding, because the usual worry about AI agents acting on your behalf is capability, not disposition. A model trained to be endlessly helpful can leak your budget or fold at the first counteroffer. Combining the per-domain specialists into one unified 4B model, using cascade reinforcement learning and multi-teacher distillation, still hit 0.627 average utility across all six tasks, edging out GPT-4.1 (0.625), GPT-5.1 (0.619), and GPT-5.2 (0.613). The paper also found a model taught to predict what a counterpart will do next, not just copy their actions, generalizes better across negotiation types.
A model that won't fold on price is great for your grocery budget, less appealing the day it's negotiating against your own agent.