paper web signal

LLM shopping agents gain confidence from reasoning, not accuracy

TL;DR

  • Across three product-choice benchmarks, more reasoning effort did not raise decision quality in a statistically meaningful way across ten frontier models.
  • The same reasoning effort did raise stated confidence at a statistically significant level, buying 'confidence rather than correctness.'
  • Errors were strongly shared across models: when one erred, several others tended to err on the same task, often picking the same wrong product.

Asking a shopping LLM to think longer reliably increases how confident it says it is, without making its purchases any better, a new arXiv paper reports.

The authors benchmarked ten frontier models on three product-choice benchmarks built from WebShop, Amazon's SQID and Wayfair's WANDS. Across all three, "increasing the thinking effort does not increase decision quality in a statistically meaningful way." The same change in thinking effort moved stated confidence at a statistically significant level, meaning "extra thinking buys confidence rather than correctness."

Calibration tells a related story. "Most models are not well-calibrated at every effort level," the paper says, with reported confidence percentages drifting from actual correctness percentages. The errors are not independent either: "when one model errs in a task, several others tend to err on that same task," and on those shared-error tasks "a large number of them pick the same wrong product."

The paper's abstract names no specific models and publishes no per-model calibration figures.