AI
Retrieving tool outputs can make agents ignore authoritative evidence, study finds
A study using the deterministic PHREEQC geochemistry simulator finds that moving long tool outputs into retrievable files can reduce some agents’ accuracy even when they compute the correct value. GPT-5.1 and GPT-5.2 lost 11.3 and 7.9 points when reports were indexed rather than placed in context, largely because they answered from less authoritative output files and misrounded values. Naming the authoritative result improved accuracy by 6.3 to 7.9 points for tested GPT models. Other model families showed no detectable loss, making harness design model-dependent.
Editorial summary · Publisher article
What matters
- The evaluation used 158 execution-verified questions and replayed 11,007 simulator calls.
- Retrieval reduced accuracy for GPT-5.1 and GPT-5.2 despite similar underlying computation rates.
- The authors call the failure mode evidence substitution.
- Explicitly identifying the authoritative output improved results for the tested GPT models.
Context & caveats
- The findings come from one scientific simulator and may not generalize to every tool or agent harness.
Why this made the edition
The study identifies a concrete failure mode in tool-using agents, uses a deterministic scientific simulator and reports model-specific accuracy changes plus mitigations.
Source material