IC-222GPT-4o ReAct failures are dominated by rule violations (transition function) and goal misinterpretation, with the balance shifting from goal-dominant in synchronous to transition-dominant in asynchronous settings

Gonzalo Gonzalez-Pumariega, Leong Su Yean, Neha Sunkara, Sanjiban Choudhury

SourceRobotouille: An Asynchronous Planning Benchmark for LLM Agents

The paper categorises all GPT-4o ReAct failures into four MDP-based categories: state, action, transition function, and goal. In the synchronous dataset, goal uncertainty accounts for 64.1% of failures and transition function for 32.1%. In the asynchronous dataset, the balance inverts: transition function accounts for 56.8% and goal for 34.1%. The single most common specific failure in both settings is violating the 'one item at a station' rule, at 24.5% of sync failures and 53.4% of async failures. Adding more rule details to the prompt (ReAct + prior) reduces this specific violation from 38.1% to 22.2% but does not significantly improve overall success rate (0.30 ± 0.085 vs 0.40 ± 0.050).

Evidence
observational
Key metric
Sync: goal 64.1%, transition 32.1%; Async: transition 56.8%, goal 34.1%; 'one item at a station' rule: 24.5% sync, 53.4% async; ReAct+prior: 0.30 ± 0.085 vs 0.40 ± 0.050 success rate, rule violations 38.1% → 22.2%
Caveat
Failure categorisation is based on manual annotation with a tiebreaker rule for multi-cause failures; the 'one item at a station' rule is specific to the Robotouille environment and may not generalise.
Model
GPT-4o
Concepts
Failure mode
Methods
ReAct [primary]
Related findings
IC-221, IC-223, IC-224
Extraction
automatic-extraction