IC-224GPT-4o ReAct cannot incorporate stochastic state changes, with success rate on cutting tasks dropping from 56% to 1% when a 33% chance of a cut item reverting to uncut is introduced
Gonzalo Gonzalez-Pumariega, Leong Su Yean, Neha Sunkara, Sanjiban Choudhury
The paper introduces stochasticity into the Robotouille synchronous environment: with a 33% chance, a cut item reverts to uncut once per environment seed. On six synchronous tasks involving cutting, GPT-4o ReAct's success rate drops from 56% in the deterministic setting to 1% in the stochastic setting. The authors note that the rare successes occur when the stochastic change happens immediately, allowing the agent to continue cutting, suggesting the model cannot re-plan after an unexpected state change mid-trajectory.
Evidence
correlational
Key metric
56.00% deterministic vs 1.00% stochastic on 6 cutting tasks (tasks 2, 3, 7, 8, 9, 10)
Caveat
Only one stochastic event per environment seed; tested on a subset of 6 tasks; the 33% probability is a single fixed value.