MWE-ECL Evaluates Influence of Long-Range Context on Local Lexical Priors
September 24, 2026
Research using the MWE-ECL diagnostic shows that while models can recover distant information in long contexts, they often fail to let that information override local lexical priors. Even with retrieval-control accuracy near 1.0, the ability to override conflicting defaults remains a significant bottleneck in long-context reasoning.
HOW THIS AFFECTS YOU
●
researcherDo not assume high retrieval scores in long-context benchmarks equate to actual behavioral influence over the model's output.