When Is a Multi-Agent Code Judge Actually Grounded? Two Label-Free Measurements, and a Judge That Declines to Guess

Source-attributed summary

This summary uses the published excerpt from arXiv (cs.AI). It has not been independently fact-checked by Neural Pulse.

When one language model judges whether another's code is correct, it does not report the absence of evidence. It returns a confident verdict with reasoning attached, indistinguishable from a verdict it had grounds for. Multi-agent verification, which decomposes a judgment into checkable claims and verifies each against evidence, is a promising response and works well when the evidence is a set of retrieved documents. We argue such methods require two things of their evidence: it must be independent of the answer under review, and it must differ between the two candidates being compared. The second condition holds automatically with retrieved documents and stops holding in code judging. Running MARCH, a published framework unmodified over 80 condition-by-cell measurements on two code judging benchmarks, we find it declares both solutions equally good on 78 to 95% of comparisons, reaching 4.4% accuracy where the same model asked directly reaches 43.7%. Neither easier problems nor a larger judge changes this. Two measurements taken from the pipeline's own logs explain it without needing labels.

Read the original source for the full context.