Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Incentive-Aligned Multi-Source LLM Summaries

Yanchen Jiang, Zhe Feng, Aranyak Mehta

TL;DR

Truthful Text Summarization is introduced, an incentive-aligned framework that improves factual robustness without ground-truth labels and establishes formal guarantees that align a source's incentives with informative honesty, making truthful reporting the utility-maximizing strategy.

Abstract

Large language models (LLMs) are increasingly used in modern search and answer systems to synthesize multiple, sometimes conflicting, texts into a single response, yet current pipelines offer weak incentives for sources to be accurate and are vulnerable to adversarial content. We introduce Truthful Text Summarization (TTS), an incentive-aligned framework that improves factual robustness without ground-truth labels. TTS (i) decomposes a draft synthesis into atomic claims, (ii) elicits each source's stance on every claim, (iii) scores sources with an adapted multi-task peer-prediction mechanism that rewards informative agreement, and (iv) filters unreliable sources before re-summarizing. We establish formal guarantees that align a source's incentives with informative honesty, making truthful reporting the utility-maximizing strategy. Experiments show that TTS improves factual accuracy and robustness while preserving fluency, aligning exposure with informative corroboration and disincentivizing manipulation.

Incentive-Aligned Multi-Source LLM Summaries

TL;DR

Truthful Text Summarization is introduced, an incentive-aligned framework that improves factual robustness without ground-truth labels and establishes formal guarantees that align a source's incentives with informative honesty, making truthful reporting the utility-maximizing strategy.

Abstract

Large language models (LLMs) are increasingly used in modern search and answer systems to synthesize multiple, sometimes conflicting, texts into a single response, yet current pipelines offer weak incentives for sources to be accurate and are vulnerable to adversarial content. We introduce Truthful Text Summarization (TTS), an incentive-aligned framework that improves factual robustness without ground-truth labels. TTS (i) decomposes a draft synthesis into atomic claims, (ii) elicits each source's stance on every claim, (iii) scores sources with an adapted multi-task peer-prediction mechanism that rewards informative agreement, and (iv) filters unreliable sources before re-summarizing. We establish formal guarantees that align a source's incentives with informative honesty, making truthful reporting the utility-maximizing strategy. Experiments show that TTS improves factual accuracy and robustness while preserving fluency, aligning exposure with informative corroboration and disincentivizing manipulation.

Paper Structure

This paper contains 98 sections, 15 theorems, 44 equations, 4 figures, 27 tables.

Table of Contents

  1. Introduction
  2. A Simple Example.
  3. Contributions
  4. Related Works.
  5. A Model for Truthful LLM Summaries
  6. High-level Overview of the theoretical model
  7. Framework Overview
  8. Players and the held-out claim set
  9. Players.
  10. Held-out claim set.
  11. Latent Correctness.
  12. From documents to stances
  13. Signal informativeness, effort, and reporting
  14. Private signals under effort (types).
  15. Reporting policy (scored source).
  16. ...and 83 more sections

Key Result

Lemma 1

Assume effort yields a positively informative signal for $i$ so that $\eta_i^{\text{sig}}>0$. For any reporting rule $\sigma_i$, with equality only under truthful reporting $(q_1,q_0)=(1,0)$. (See Appendix app:sec2 for proof)

Figures (4)

  • Figure 1: The TTS framework in action. Unlike a standard pipeline vulnerable to manipulation (left), our method (right) scores sources based on informative peer agreement to filter out weakly supported or adversarial/strategic content and produce a robust summary.
  • Figure 2: Scoring and threshold incentives.Left: For each claim $k$ and peer $j$, the score adds on-task agreement and subtracts off-task agreement; we average over peers within a claim and then average across $K$ claims to obtain $\widehat{w}_i$. Right: Score densities for truthful, an informed alternative, and uninformed. Shaded mass $\Pr(\widehat{w}_i\ge t_{\mathrm{src},i})$ is the inclusion probability. Larger $K$ concentrates the truthful curve, underpinning the informed-truthfulness results.
  • Figure 3: Score separation and incentives. Left: Informative-agreement scores separate reliable from unreliable sources. Right: Truthful behavior is payoff-maximizing against deviations in stance.
  • Figure 4: Robustness of TTS against uninformative equilibria with 4 uninformative sources.

Theorems & Definitions (24)

  • Lemma 1: Report informativeness is bounded by signal informativeness
  • Proposition 1: Expected claim-averaged pairwise score
  • Corollary 1: Uninformative strategies yield zero mean score
  • Theorem 1: Asymptotic informed truthfulness
  • Theorem 2: Strong truthfulness via hard threshold
  • Theorem 3: Finite-$K$ $\epsilon-$Informed truthfulness
  • Lemma 1: Report informativeness is bounded by signal informativeness
  • proof
  • Proposition 2: Policies $\to$ documents: utility equality, equilibrium lifting, guarantee transfer
  • proof
  • ...and 14 more