NeurIPS 2025 · MTI-LLM Workshop · with Igor Bogdanov & Chung-Horng Lung
A marshmallow test for AI agents
Inspired by the Stanford marshmallow experiment: LLM agents must resist an immediate reward, minute by minute, for a larger later payoff, while we vary who's watching, who they're told they are, and whether they must question themselves.
What I did
Co-designed a factorial micro-benchmark: ReAct agents with a "raise a question" tool under a per-step budget, manipulated across social context (broadcast vs isolated), personas (age, hedonic drive), and metacognitive policy (must vs may self-question): 19,200 trajectories over 64 cells, analyzed with Kaplan–Meier survival curves and discrete-time hazard models.
What I found
75.9% of agents persist to the end, but the factors matter enormously: a "child" persona multiplies the per-minute risk of giving in by 66×; seeing other agents' behavior (broadcast) raises risk; and, counterintuitively, forcing agents to self-question makes them fail more, not less.
Why it matters
Long-horizon reliability is the open problem for deployed agents. This gives teams a compact, reproducible way to measure how social context and prompt framing change an agent's ability to hold a commitment over time.
What changes an agent's risk of giving in?
Odds ratios from the discrete-time hazard model (log scale). Left of the line = safer than baseline; right = riskier. Click a row for a plain-language reading; hover for values.
All effects p < .001. Baselines: broadcast context, "crave" hedonic drive, adult persona, may-self-question policy. Median time-to-eat ≈ 14.8 of 19 steps; agents ask ≈ 7.1 questions on average. Removing hedonic drive and persona age pushes completion toward 1.0.