The authors assembled 1,636 adversarially verified question-answer pairs from 50 documents across five professional domains.
Voice agents lose document grounding as conversations and context grow
A benchmark spanning 50 documents finds that long contexts and multi-turn dialogue increase unsupported answers, especially in open-weight speech systems.
Big Tech
Puneet Mathur · Nedim Lipka · Zeyu Jin · Dinesh Manocha
Adobe Research · University of Maryland College Park
Research Digest··2 min read
Mathur et al.
Why this paper
From Adobe Research and University of Maryland College Park
In one line
Document-grounded voice agents lose factual fidelity as context and conversation grow, often hallucinating instead of abstaining; cascaded systems ground best among those evaluated.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§