SocialReasoning-Bench: Measuring Whether AI Agents Act Most Favorably for Users

As AI agents increasingly take over tasks like calendar coordination or purchase negotiation on our behalf, how favorable the outcome is for the user matters more than simply whether the task gets done. Microsoft Research proposed SocialReasoning-Bench to examine this kind of principal-agent situation—looking at the ability to negotiate while accounting for the other party's goals and private information. The counterpart may negotiate in good faith, but may also attempt to extract private information or force an unfavorable choice.

The benchmark consists of two realistic negotiation-style tasks.

• Calendar Coordination: An agent managing the user's calendar receives a meeting request from another person, and coordinates a time slot while the user's preferences are modeled as a value function from 0 to 1. The initial request always conflicts with the user's schedule.
• Marketplace Negotiation: A buyer agent negotiates the price of a single item with a seller, where the buyer has a maximum payable price and the seller has a minimum acceptable price. The seller's first offer is always higher than the buyer's maximum payable price.

The evaluation looks at both outcome and process.

• Outcome Optimality scores, from 0 to 1, the proportion of value the user secured within the ZOPA.
• Due Diligence measures how closely actual actions at each decision point align with a deterministic reasonable-agent policy.

The experiments compared GPT-4.1 (chain-of-thought), GPT-5.4 (high reasoning effort), Claude Sonnet 4.6, and Gemini 3 Flash, while the counterpart agent was always fixed as Gemini 3 Flash. Each task ran for up to 10 rounds, and both a baseline prompt and Defensive Prompting—instructing the agent to actively advocate for the user's interests—were applied.

The core finding is clear. While all models achieved nearly perfect task completion rates, outcomes were largely unfavorable to users and left a lot of value on the table. Defensive Prompting boosted performance but did not close the gap, with GPT-5.4 showing the largest improvements of +0.21 on Calendar and +0.12 on Marketplace. This benchmark underscores that AI agents need a standard that evaluates duty of care, not just raw success rate.