PaperTrail: A Claim-Evidence Interface for Grounding Provenance in LLM-based Scholarly Q&A

1. Problem Statement

Large Language Models (LLMs) are increasingly used in scholarly Question-Answering (QA) systems to synthesize vast amounts of literature. However, these systems suffer from reliability issues, including hallucinations, unsupported claims, and omissions. Current provenance mechanisms, such as standard source citations, are often too coarse-grained (e.g., linking to a whole paragraph) to allow researchers to rigorously verify specific assertions. Furthermore, existing trust-calibration methods (like confidence scores) often increase cognitive load without providing actionable paths for evidence validation. The core challenge is designing an interface that enables domain experts to efficiently verify LLM outputs against source texts without overwhelming them with information or requiring excessive time.

2. Methodology

A. System Architecture: PaperTrail

The authors developed PaperTrail, a system that decomposes both source documents and LLM responses into discrete claims and evidence, mapping them to reveal supported assertions, unsupported claims, and omitted information.

Backend: Argument Extraction Engine
The system utilizes a three-stage pipeline that strategically balances computational cost and semantic capability using three different extraction methods:

1. Stage 1: Offline Paper-Level Extraction (High Cost, High Semantics):
   • Method: LLM-based extraction (using Gemini 2.5 Pro).
   • Process: Pre-processes a corpus of scholarly PDFs offline. It extracts atomic, verifiable scientific claims from paragraphs based on Toulmin's argumentation model and the SciClaimHunt dataset.
   • Evidence Retrieval: Uses similarity-based extraction (SPECTER embeddings) to find supporting sentences for each claim, creating a structured JSON knowledge base.
2. Stage 2: Real-Time Answer-Level Extraction (High Cost, High Semantics):
   • Method: LLM-based extraction.
   • Process: When a user queries the system, an "Answerer LLM" generates a response. A secondary "Extraction LLM" then decomposes this response into discrete claims and supporting evidence, enforcing a structured JSON schema.
3. Stage 3: Real-Time Claim-Evidence Matching (Medium Cost, Hybrid):
   • Method: Retrieval-Augmented Generation (RAG).
   • Process: Uses similarity search to filter relevant paper claims based on the user query, then employs an LLM to perform semantic matching between the Answer Claims and the Source Paper Claims.
   • Output: Calculates a "Claim Coverage" metric and flags unsupported evidence (using a cosine similarity threshold < 0.55).

Frontend Interface
The interface is a three-panel web application (React/Redux):

• Left Panel: Task context and a text editor where users edit LLM-generated drafts.
• Middle Panel: Chat interface for Q&A. Clicking on an "Answer Claim" highlights it.
• Right Panel: Provenance view. It displays:
  • Global Overview: A bar indicating the percentage of source claims included in the answer.
  • Claim Cards: Teal cards for claims found in the source; Red cards for claims omitted from the answer.
  • Linking: Clicking a claim in the chat highlights the corresponding source claim and evidence in the right panel.

B. User Study Design

• Participants: 26 domain experts (research scientists and engineers) from NASA.
• Design: Within-subjects study comparing PaperTrail against a Baseline (standard source citation highlighting).
• Tasks:
  1. Multi-Paper Synthesis: Editing a draft synthesizing four papers on Mars exploration.
  2. Devil's Advocate: Editing a draft defending a paper's novelty against critiques.
• Measures:
  • Trust: Subjective trust (TXAI scale).
  • Reliance: Behavioral reliance measured by the normalized Levenshtein edit distance between the original LLM draft and the user's final edit (higher distance = lower reliance).
  • Confidence: Self-reported confidence in edits.
  • Experiential: Cognitive load (NASA-TLX), usability (SUPR-Q), and interaction logs.

3. Key Contributions

1. System Design: The implementation of PaperTrail, the first system to operationalize argumentation structures for source provenance by decomposing and linking claims and evidence between LLM answers and source documents at a granular level.
2. Flexible Backend Architecture: A hybrid extraction framework combining offline LLM processing, real-time LLM extraction, and RAG-based matching, serving as a framework for evaluating LLM trustworthiness.
3. Empirical Evidence on Trust Calibration: Demonstration that granular claim-evidence provenance significantly lowers subjective trust in LLMs compared to standard citations.
4. Identification of the Trust-Behavior Gap: Empirical evidence showing that while reduced trust increases skepticism, it does not necessarily translate to behavioral changes (i.e., users still rely on the LLM output to avoid cognitive burden) in time-constrained scholarly settings.

4. Results

• Trust: Participants using PaperTrail reported significantly lower trust in the LLM ($p=0.015$, $d=0.44$) compared to the baseline. The granular view exposed more errors and omissions, fostering caution.
• Reliance (Behavior): Despite lower trust, there was no significant difference in behavioral reliance (edit distance) between the two conditions ($p=0.313$). Users continued to rely heavily on the LLM-generated text.
• Confidence: No significant difference in self-reported confidence.
• Usability: PaperTrail was rated as significantly less usable than the baseline ($p=0.026$). Users found the interface "cluttered" and the system latency ("excruciatingly slow") hindered deep engagement.
• Qualitative Findings:
  • Users appreciated the ethos of claim-level verification and the ability to "fact-check."
  • However, time pressure and system latency prevented users from acting on their skepticism.
  • A "Trust-Behavior Gap" emerged: users recognized the need for verification but defaulted to System 1 thinking (fast, heuristic reliance) due to cognitive load and task constraints.

5. Significance and Implications

• Rethinking Trust: The paper argues that the goal should not just be to lower user trust, but to make the LLM itself trustworthy through rigorous, argument-grounded evaluation. The backend architecture can serve as a formal benchmark for LLM trustworthiness in scholarly settings.
• Design Implications:
  • Adaptive Provenance: Future systems should use adaptive interfaces that provide simple cues for low-stakes queries and deep claim-evidence views for complex/controversial ones to manage cognitive load.
  • Scaffolding: Systems need to scaffold critical thinking skills over time, perhaps through gamified or tutorial-based approaches, to help users overcome the "cognitive friction" of verification.
  • Computational Optimization: Reducing latency is critical; future work should explore federated architectures and aggressive caching to make real-time argument extraction feasible.
• Theoretical Contribution: The study highlights a critical disconnect in Human-AI collaboration: awareness of error does not guarantee behavioral correction when users face time pressure and high cognitive demands. Effective AI tools must address these systemic constraints, not just provide better information.

In conclusion, PaperTrail successfully demonstrates that argument-grounded provenance can improve the perception of LLM reliability by exposing errors, but current implementations fail to change behavior due to usability and latency barriers. The work calls for a shift from simple citation to structured argument verification, balanced with designs that respect the cognitive limits of researchers.