Benchmark
echoChain
Measure which facts survive when language models repeatedly turn an answer into a question and try to reconstruct the answer.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
An experiment begins with a seed answer. Models can take turns or repeat the exercise under same-model, identity and paraphrase controls.
Methodology
A separate judge scores whether each question can be answered and how faithfully the next answer preserves the seed. When a seed lists specific claims, the judge also checks which ones survive each step. The evaluator measures fidelity to the starting text. A model judge can preserve or miss an error already present in that text.