Linhai Ma and colleagues study models that check the values an agent sends to a tool. Seven judges accept correct values less often when those values require recovering older evidence, converting a format, or applying a default. Even showing the default in the tool’s description can leave the judge unconvinced.
DeRef gives that check a separate preparation step:
- A model reads the available history and tool description without seeing the proposed value. It writes the expected value and its source, or says the evidence is insufficient.
- The unchanged judge then compares the candidate with that reference. Candidates for the same argument share one reference.
- On ComplexFuncBench, a test of multi-step tool use, Qwen3.8-27B’s argument-ranking score rises from 78.92 to 95.09 out of 100. Reading the full history alone scores 85.45.
- Across four main judges and three datasets, every comparison improves over ordinary judging. These scores measure ranking correct arguments above wrong ones, not successful task completion.
The reference is also a failure point. Replacing a correct expected value with a wrong one can reverse judgments, even while the original request stays visible. The authors’ controlled hidden-source test reaches near-perfect scores, but restoring the missing history also repairs that particular test.
Caveat: these are author-reported tests with one main reference-generating model. Generating a reference takes seconds on a B200; consequential actions still need independent checks or human review.