A study evaluates a deployable 4B generator running on a 24 GB laptop to determine what makes its citations faithful and comprehensive in on-device deep research tasks. The research separates cited claim faithfulness from trustworthy coverage, analyzing how source exposure length and source quality affect these metrics.
- Increasing per-source exposure from 400 to 1500 characters lifts faithfulness from 0.45 to 0.58 on retrieved sources and from 0.37 to 0.58 on gold sources, showing faithfulness is bound by exposure rather than source correctness.
- Trustworthy coverage remains near 0.22 on retrieved sources regardless of exposure because recall is held near 0.40, indicating retrieval quality limits which sources are cited.
- The additional exposure costs approximately 235 output tokens.
The authors conclude that the practical approach is to raise per-source exposure first as a cheap lever, and then treat retrieval recall as the only remaining factor for improving coverage.