Comparative Empirical Evaluation of Prompting Strategies for Large-Language-Model Web-Navigation Agents on the WebArena Benchmark

Large language models (LLMs) are increasingly deployed as autonomous agents that operate web interfaces on behalf of users, yet reported success rates vary widely across studies that differ simultaneously in backbone model, prompting strategy, and evaluation environment, making it difficult to attribute performance to any single factor. This paper presents a controlled factorial comparison that isolates two of these factors on a fixed environment. We evaluate three backbone LLMs (GPT-4, Llama-3-70B-Instruct, and GPT-3.5) crossed with four prompting strategies (direct prompting, ReAct-style interleaved reasoning, Reflexion-style episodic retry, and best-first lookahead) on all 812 tasks of the WebArena benchmark, with a 165-instance WorkArena-L1 robustness check, under an identical action space, observation format, and step budget. Backbone choice accounts for the dominant share of variation: the weakest GPT-4 configuration (15.2%) exceeds the strongest GPT-3.5 configuration (9.7%) by 5.5 percentage points. Strategy gains are moderate and compound with backbone strength, ranging from 3.6 points on GPT-3.5 to 8.4 points on GPT-4. A cost analysis shows that the marginal cost of one additional percentage point rises roughly tenfold along the GPT-4 strategy ladder. All configurations remain far below the 78.24% human reference, indicating substantial remaining headroom.