Penn and Microsoft Turn AI Citations Into Layered Evidence Trails
A new interface technique from Penn and Microsoft Research turns citation numbers into unfoldable evidence trails, driving measurably deeper engagement with AI answer sources.
- Penn and Microsoft Research propose attribution gradients, replacing citation numbers with unfoldable evidence trails.
- Five layers: claim decomposition, evidence overview bars, color-coded excerpts, in-source PDF viewer, and second-degree citation unraveling.
- Built on OpenSciLM with GPT-5.1 extraction and Semantic Scholar API for pulling nested references.
- In a 15-person study, revisions scored significantly higher on overall quality, faithfulness, and factual correctness.
- Evidence recall jumped from 17.3% in a ChatDoc baseline to 51.3% with attribution gradients.
- Research prototype only, no public code; pipeline classifies evidence valence well but degree less reliably.
AI answer engines such as Perplexity, Copilot, and You.com place generated responses ahead of traditional search results, usually attaching citation numbers to individual claims. Those numbers conceal which passage supports a claim, whether the source provides direct evidence, and whether the evidence originated in another paper. Researchers at the University of Pennsylvania and Microsoft Research propose attribution gradients, an interface that expands each citation into a layered evidence trail within the answer.
Citation numbers conceal provenance
The design draws on information foraging theory, which describes how people use visible cues to decide whether a link is worth following. A bare citation has weak “information scent” because it reveals almost nothing about the underlying evidence. Verifying the claim requires opening the source, finding the relevant passage, and assessing whether the passage supports the generated text.
A source can support only part of a generated statement, offer a conflicting result, or cite another paper for the underlying evidence. The authors call that final case second-degree evidence. Following it requires another round of searching, document navigation, and interpretation.
Earlier reading systems provide passage highlights, cross-document summaries, and deep links into papers. Attribution gradients combine those functions in one expandable panel and extend the trail into secondary literature.
Five layers, one workspace
Built as an extension to OpenSciLM, a system for answering questions from scholarly literature, the prototype uses a GPT-5.1 extraction pipeline to organize evidence into five layers:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
