What's this about?
Agents increasingly act on their users' behalf in tasks like scheduling meetings, comparing offers, and haggling over prices. These principal-driven tasks routinely place the agent opposite a counterparty with opposing interests. This work uses reinforcement learning to strengthen social reasoning in small language models specifically for that setting.
What exactly makes negotiation different from other agent tasks?
Most agent tasks are solitary: write the code, find the file, book the flight at the listed price. Negotiation is adversarial and informational — the other side has different goals, knows things you don't, and will strategically misrepresent. Doing it well requires modeling what the counterparty wants, what they'll accept, and what your own eagerness reveals. A model trained to be maximally helpful and agreeable is close to the worst possible negotiator, which is precisely the tension here.
Why is this worth noticing?
- Agent-to-agent commerce is arriving — your agent negotiating with a seller's agent — and helpfulness training is actively counterproductive there.
- Targeting small models matters: negotiation happens at volume and latency, so it needs to be cheap enough to run constantly.
- It raises a question the paper's framing invites: an agent good at strategic misrepresentation on your behalf is a capability with obvious dual use, especially given the same week's Economist piece on agents that lie.