AI answer engines are a relatively new kind of information search tool: rather than returning a ranked list of documents, they generate an answer to a search question with inline citations to sources. But reading the cited sources is costly, and citation links themselves offer little guidance about what evidence they contain. We present attribution gradients, a technique to boost the informativeness of citations by consolidating scent and information prey in place. Its first feature is bringing evidence counts, supporting/contradictory excerpts, links to sources, and contextual explanations into one place. Its second feature is the ability to unravel second-degree citations in place. In a lab study, we demonstrate usage of the full gradient in a critical reading task and its support for deep engagement that increased the depth of what readers took away from the sources versus a standard citation and document QA design.
Figures & tables
Figure 1. Attribution gradients. When using an attribution gradient, users begin with a claim in a generated answer (1) and see color-coded clusters of evidence (2): black indicates direct support; red, direct contradiction; gray, references to sources that support the claim; and pink, references to sources that contradict the claim. Users can explore evidence excerpts (3) that might include citations to secondary references (4). Users can view excerpts in the original PDF (6) with explanations of how the evidence relates to the claim (5). This gradated unfolding of context is designed to help users access a more nuanced understanding of support for a claim. A teaser figure for attribution gradients. The figure displays an attributed AI response with numbered steps: starting with a generated claim (1), followed by color-coded evidence indicators (2), citations (3), and deep excerpts (4). An arrow connects these excerpts to a side-by-side view of the original source document (5, 6). The footer contains a six-stage progress bar: view claims, view evidence types, view evidence, unravel evidence, orient to source, and read evidence in source.
Figure 2. Articulation of support added by attribution gradients. Compared to conventional attributed AI answers, attribution gradients change the cost structure of looking up information about evidence by introducing new scent, lesser cost, or partial prey at the steps of interaction marked by a blue box. The theoretical support of standard RAG and attribution gradients. Compares attribution gradients to conventional answers across interaction steps like "viewing 2nd-deg evidence". Attribution gradients are shown to provide high "scent" at "low cost" for these tasks by providing "partial prey" (snippets and highlights) that conventional answers do not support.
Figure 3. Unraveling citations. If a piece of second-degree evidence cites another source for its evidence, users can sometimes unravel the citation to that source. After clicking the button to expand nested citations, excerpts from the cited source are shown that pertain to the claim. In this case, evidence from citation “[2]” is shown that more precisely conveys how one of the cited studies shows that “ideating with generative AI decreases diversity.” A user is unraveling citations in the attribution gradients interface. By clicking an "expand nested citations" button, the user has unraveled a citation (marked as "[2]") to view evidence from a cited source.
Figure 4. The full gradient. A reader can (1) decompose a sentence in the generated answer into claims, and (2) view evidence for a particular claim. A reader can see that attributed evidence is mixed for this claim. If a reader clicks on an evidence excerpt (3), they can (4) view it in the source and (5) read context for how the evidence excerpt relates to the claim. Counts shown in the evidence overview indicate the number of evidence excerpts of each type. The attribution gradients interface showing the decomposition of an AI-generated sentence into individual "inspectable claims". After clicking a magnifying glass icon next to a sentence about ChatGPT's effect on creativity, the interface reveals two atomic claims. Each claim is paired with a segmented, color-coded horizontal bar showing the distribution of evidence:
Figure 5. Comparisons of revised claims according to six quality dimensions. Participants using attribution gradients produced revisions that were rated significantly higher than the baseline on all dimensions except evidence quantity. Each box spans the interquartile range of grades, with the horizontal line marking the median. Each individual point is one participant’s grade. Brackets and asterisks indicate significance from paired Wilcoxon signed-rank tests comparing conditions within each dimension (** p<.01 ). These box and whisker plots compare the quality of revised claims made by participants using attribution gradients (blue) versus a baseline condition (orange) across six dimensions: Overall Quality, Evidence Quantity, Evidence Specificity, Effect Specificity, Faithfulness, and Factual Correctness.
Long-form question answering increasingly relies on retrieved evidence to make LLM outputs verifiable, with inline citations tracing claims to source documents. However, existing systems often attach citations that are topically related but insufficient to support their claims. We identify attribution ambiguity as a structural challenge: end-to-end generation must implicitly resolve combinatorial claim--document assignments, obscuring evidential boundaries and increasing the risk of evidence-boundary overrun, where claims exceed cited support. To address this challenge, we propose CAGE (Cognitive Attribution Graphs for Citation Generation), a two-stage framework that introduces an explicit cognitive attribution map before answer generation. CAGE first trains a plug-and-play Cognitive Map Induction Model to construct answer-centered support subgraphs, aligning each semantic answer unit with supporting documents through explicit relations. A Structured Citation Reasoning Model then realizes these units as sentence-level claims with map-aligned citations. Experiments on ASQA, ELI5, and ExpertQA show that CAGE achieves state-of-the-art performance, demonstrating the effectiveness of attribution-space contraction and map-guided citation generation.
Zhichao Yan, Shizhao Li, Jiapu Wang +5
Shanxi University, Taiyuan, China · Nanjing University of Science and Technology, Nanjing, China · Nanyang Technological University, Singapore +3
Retrieval-augmented systems routinely present citations alongside generated answers, yet a citation does not confirm that the corresponding source meaningfully shaped the output. This paper introduces ProvenAI, a framework that decomposes transparency in multi-hop question answering into three independently measurable layers: answer correctness, citation fidelity against benchmark supporting evidence, and per-document influence under leave-one-resource-out intervention. Targeting the HotpotQA distractor benchmark through a seven-stage pipeline covering data normalisation, retrieval indexing, citation-aware answer generation, attribution auditing, ablation-based influence estimation, batch evaluation, and interactive inspection, ProvenAI evaluates 7,405 validation examples drawn from a canonical corpus of 509,300 passages. The system achieves 53.53% answer accuracy alongside a mean citation-fidelity score of 71.55%, and a worked example surfaces what we call the citation-influence gap: a clean citation audit co-occurring with a profile in which one cited source registers only weak influence while seven uncited sources demonstrably shift the output. We formalise the relationship between the implemented surface proxy and a token-level KL-divergence target through a stated faithfulness condition, ground the framework in causal-mediation analysis and database-provenance theory, and discuss how the three measurement layers compose with cryptographic provenance architectures emerging in autonomous scientific discovery. ProvenAI establishes that meaningful transparency in retrieval-grounded QA requires traceable links across retrieved, cited, and behaviourally influential evidence as three distinct, independently measured layers.
Mohammad Faizan, Dalal Alharthi
College of Information Science University of Arizona
AI answer engines generate answers from retrieved pages but cite only a few sources. This makes visibility depend not just on ranking, but on being cited. We study competitive Generative Engine Optimization (GEO): when two retrieved candidates compete, what makes one more likely to be cited first? We build a controlled two-document retrieval-augmented generation (RAG) testbed that injects exactly two candidate sources into the model context and measures which source is referenced by the first citation marker in the output. Across six LLMs we execute 252,000 trials, repeated paired comparisons under one factorial program over 18 content factors. In each trial the two sources differ in exactly one factor; we use brand anonymization and counterbalanced source order to separate content effects from position bias. Mixed-effects models show that topical relevance and list position are the biggest drivers of being cited first. Including explicit price information and a recent timestamp also helps consistently. Completeness and trust cues add smaller gains, while formatting-only edits have little impact. We release a reproducible evaluation protocol and a prioritized GEO checklist for practitioners, and we exercised it in an early internal pilot at Sprinklr, where teams reported positive qualitative feedback on workflow usability.