Credibility Governance: A Social Mechanism for Collective Self-Correction under Weak Truth Signals

Wanying He; Yanxi Lin; Ziheng Zhou; Xue Feng; Min Peng; Qianqian Xie; Zilong Zheng; Yipeng Kang

cs.CYMar 3, 2026

Credibility Governance: A Social Mechanism for Collective Self-Correction under Weak Truth Signals

Authors: Wanying He, Yanxi Lin, Ziheng Zhou, Xue Feng, Min Peng, Qianqian Xie, Zilong Zheng, Yipeng Kang

Organizations: School of Artificial Intelligence, Wuhan University; BIGAI Wuhan, China · Tsinghua University Beijing, China · University of California, Los Angeles Los Angeles, US · State Key Laboratory of General Artificial Intelligence, BIGAI Beijing, China ·

Abstract

Online platforms increasingly rely on opinion aggregation to allocate real-world attention and resources, yet common signals such as engagement votes or capital-weighted commitments are easy to amplify and often track visibility rather than reliability. This makes collective judgments brittle under weak truth signals, noisy or delayed feedback, early popularity surges, and strategic manipulation. We propose Credibility Governance (CG), a mechanism that reallocates influence by learning which agents and viewpoints consistently track evolving public evidence. CG maintains dynamic credibility scores for both agents and opinions, updates opinion influence via credibility-weighted endorsements, and updates agent credibility based on the long-run performance of the opinions they support, rewarding early and persistent alignment with emerging evidence while filtering short-lived noise. We evaluate CG in POLIS, a socio-physical simulation environment that models coupled belief dynamics and downstream feedback under uncertainty. Across settings with initial majority misalignment, observation noise and contamination, and misinformation shocks, CG outperforms vote-based, stake-weighted, and no-governance baselines, yielding faster recovery to the true state, reduced lock-in and path dependence, and improved robustness under adversarial pressure. Our implementation and experimental scripts are publicly available at https://github.com/Wanying-He/Credibility_Governance.

Figures & tables

Figure 1. Comparison of the four governance mechanisms. (1) NG: Agents observe physical signals $\tilde{\pi}_{k}^{t}$ . (2) SM: NG + social signal $\Theta_{k}^{t}$ in observation. (3) WS: SM + agent influence $w_{i}^{t}$ updated by $\Delta\pi_{k}^{t}$ . (4) CG: WS + agent influence $w_{i}^{t}$ updated by $\Theta_{k}^{t}$ .

Feature	Credibility Gov. (CG)	Web3 Staking (WS)	Social Media (SM)	No Governance (NG)
Agent Influence ( $w_{i}$ )	Credibility-based	Stake-based	Equal influence	Equal influence
Social Signal Update ( $\Theta_{k}$ )	Cred.-adj. $\Delta$ support	Stake-wtd popularity (delay $\tau$ , noise $\xi$ )	Raw popularity	None
Signal distortion	noise $\xi$ , contam. $\rho\Theta_{k}^{t}$	noise $\xi$ , delay $\tau$ (stake settles)	noise $\xi$ , contam. $\rho\Theta_{k}^{t}$	noise $\xi$
Robustness knob	anti-bubble $B_{k}^{t}$	stake cost / budget	None	None
Reward Basis	$\Delta\Theta_{k}^{t}$	$\Delta\pi_{k}^{t}$ (delayed)	N/A	N/A

Table 1. Comparison of governance mechanism properties.

Figure 2. H1, system dynamics. Panel (a) shows convergence from an initially false-majority state. Panel (b) shows recovery after a misinformation shock (shaded window). CG reaches higher support for the true topic and recovers faster after shocks than WS, SM, and NG.

Figure 3. H2, mechanism pathway. Panel (a) tracks the share of influence held by truth-aligned agents. Panel (b) shows the social signal of the true topic $\Theta_{\text{true}}$ . Panel (c) shows cumulative physical progress $\pi_{\text{true}}$ , with background phases indicating exploration, acceleration, and saturation. CG reallocates influence toward truth-aligned agents, strengthens the social signal for the true topic, and compounds into faster physical progress.

Figure 4. H3, ablations. Each panel compares support for the true topic in Full CG against an ablated variant: (a) no credibility updates, (b) no anti-bubble penalty, (c) no early-mover bonus, and (d) reward base $\Delta\Theta_{k}\rightarrow\Delta\pi_{k}$ . Removing any component degrades performance, and replacing $\Delta\Theta$ with $\Delta\pi$ most strongly destabilizes convergence.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Symbol	Value	Scope / Description
Physical Progress Dynamics
$\displaystyle\Delta\pi_{k}^{t}=\begin{cases}\epsilon_{k}^{t},&\pi_{k}^{t-1}<\pi_{1},\\[3.99994pt] \bigl(v_{k}+\gamma\,r_{k}^{t}\bigr)\!\left(1-\dfrac{\pi_{k}^{t-1}}{M_{k}}\right)+\epsilon_{k}^{t},&\pi_{1}\leq\pi_{k}^{t-1}<\pi_{2},\\[3.99994pt] \gamma^{\prime}r_{k}^{t}\!\left(1-\dfrac{\pi_{k}^{t-1}}{M_{k}}\right)+\epsilon_{k}^{t},&\pi_{k}^{t-1}\geq\pi_{2}.\end{cases}$
Environmental Noise	$\sigma$	0.1	std. dev. of $\epsilon_{k}^{t}$
Baseline Velocities	$v_{\text{true}},v_{\text{false}}$	1.0, 0.8	intrinsic growth
Acceleration Factor	$\gamma$	0.8	mid-stage resource impact
Saturation Limits	$M_{\text{true}},M_{\text{false}}$	10, 8	max progress

Appendix

Table 6

Figure 5. Sensitivity to Population Size. Comparison of CG’s convergence for N=50, N=100 (Main Experiment), and N=300.

Figure 6. Sensitivity to Initial Opinion Distribution. Comparison of CG’s convergence for 30%, 50%, and 70% initial truth-aligned agents.

Figure 7. Sensitivity to Baseline Velocity Gap. Comparison of CG’s convergence for smaller, original, and larger intrinsic growth differentials.

CG variable	Forecasting tournaments	Peer review	Community forums
$\Theta_{k}^{t}$ (social signal)	Aggregate crowd belief on event $k$ , e.g., mean implied probability or weighted forecast consensus at time $t$ .	Aggregate stance toward paper $k$ , e.g., mean recommendation score, accept probability proxy, or discussion sentiment.	Aggregate support for claim or topic $k$ , e.g., upvote share, endorsement rate, agreement ratio, or reshare intensity.
$\Delta\Theta_{k}^{t}$ (momentum)	Change in consensus probability over a window, indicating information arrival or crowd revision.	Shift in consensus during deliberation, e.g., score changes after rebuttal or sentiment swing between rounds.	Change in endorsement rate or vote share, capturing emerging agreement or coordinated surges.
$\pi_{k}^{t}$ (outcome, truth linked)	Realized event outcome, resolved binary or continuous outcome (e.g., whether event occurred).	Post publication outcomes, e.g., reproducibility checks, later expert assessments, or editor adjudication.	Later verification signal, e.g., fact check labels, moderator confirmation, or authoritative source validation.
$\bar{\pi}_{k}^{t}$ (blurred outcome proxy)	Noisy short horizon proxy such as partial indicators, early reports, or delayed settlement status.	Noisy proxies such as reviewer confidence, independent quick checks, or editor triage signals.	Noisy proxies such as reports, flags, or limited audits, available with delay.
$w_{i}^{t}$ (credibility weight)	Forecaster reliability weight used to aggregate predictions, based on past calibration and update dynamics.	Reviewer reliability weight used in aggregation or assignment, based on historical consistency and informativeness.	User credibility weight used in ranking or visibility, based on past helpfulness and resistance to manipulation.
$r_{k}^{t}$ (resource allocation)	Attention allocated to event $k$ , e.g., information prominence or forecast visibility weight.	Editorial effort on paper $k$ , e.g., additional review bandwidth, expert consultation, or priority handling.	Exposure allocated to topic or claim $k$ , e.g., ranking position, recommendation frequency, or distribution budget.

Appendix

Table 2. Mapping CG variables to example observable proxies in forecasting, peer review, and community forums.

Figure 8. H4, robustness vs noise. We sweep a unified noise level combining observation noise and contamination. CG shows the largest accuracy gains in the intermediate-noise regime, where truth is weakly observable yet recoverable. Under extreme noise, all mechanisms degrade.

Figure 9. H4, diagnostics under noise. As noise increases, we track early lock-in probability, path dependence score, and correction lag (normalized). CG reduces lock-in and path dependence and shortens correction lag most strongly in the intermediate-noise regime.

Figure 10. H5, prompting effect. Solid lines denote runs with explicit credibility prompting and dashed lines denote runs without. Prompting yields a modest, consistent uplift, with the largest improvement under intermediate noise.

Figure 11. H6, threat model. (a) Robustness under multiple attacker behaviors at a representative noise level. (b) Robustness versus noise under representative attack types. CG degrades less across attacker types than WS and SM, which are more sensitive to manipulation that creates artificial momentum.

Figure 12. H6, attacker fraction sweep. (a) Accuracy versus attacker fraction at a representative noise level. (b) A recovery proxy versus attacker fraction. CG tolerates larger adversarial populations before collapsing and maintains better recovery under higher attacker prevalence.

Figure 13. H6, failure boundary over noise and attacker fraction. Heatmaps show success probability as a function of unified noise and attacker fraction, with the white contour indicating the boundary (success probability $=0.5$ ). CG shifts the boundary toward higher noise and higher attacker fractions compared to WS, expanding the stable operating region.

Shared Decision-Making

See all 11 papers

Learning Whom to Trust : Decision-Generated Credibility in Social Learning
Aug 25, 2026Gabriel Bontemps, Abhishek BanerjeeShared Decision-Making Computational Social Science
GREDEG-CNRS, Université Côte d’Azur, Nice, France · Blizard Institute, Queen Mary University of London, London, United Kingdom of Great Britain and Northern Ireland (the) · Adaptive Decisions Laboratory, Department of Pharmacology, University of Oxford, Oxford, United Kingdom of Great Britain and Northern Ireland (the)
Reputation as Community Memory for the Agentic Web
Sep 16, 2026Ryan Chard, Gus Ellerm, Alexander Brace +4Agent Memory AI Agent Security
University of Chicago & Argonne National Laboratory, Chicago, IL, USA
When Individually Calibrated Models Become Collectively Miscalibrated
May 14, 2026Zhaohui WangProbability Calibration Game Theory
USC Viterbi School of Engineering, University of Southern California

Aug 25, 2026J/K move · Enter open · S save

Learning Whom to Trust : Decision-Generated Credibility in Social Learning

Gabriel Bontemps, Abhishek Banerjee

GREDEG-CNRS, Université Côte d’Azur, Nice, France · Blizard Institute, Queen Mary University of London, London, United Kingdom of Great Britain and Northern Ireland (the) · Adaptive Decisions Laboratory, Department of Pharmacology, University of Oxford, Oxford, United Kingdom of Great Britain and Northern Ireland (the)

Shared Decision-Making Computational Social Science

Open paper AskPDF arXiv

Social interaction can improve collective learning but also amplify early mistakes. We study this tension when the credibility of social information is generated by the sender's own decision process rather than fixed ex ante. Reinforcement-learning agents make binary choices through a drift--diffusion process that jointly determines choice, decision time, and confidence; decision confidence then becomes social credibility by weighting anticipatory influence and retrospective social learning. Under balanced community exposure, the anticipatory field admits an exact quotient representation. Its local Jacobian is a scalar decision-sensitivity term multiplying the community-coupling matrix, which yields a common-mode amplification threshold and an analytical role for cross-community permeability in damping relative community differences. Monte Carlo experiments show the corresponding non-monotone performance pattern: moderate transmission accelerates correction, whereas strong transmission can lock populations into wrong consensus; low permeability instead sustains disagreement. Ablations reveal a dual role for confidence: credibility-sensitive transmission amplifies social error, while confidence-dependent private learning stabilises it. The model yields testable predictions linking sender confidence to receiver behaviour conditional on accuracy.

Abstract

Shared Decision-Making

Explore similar work

Learning Whom to Trust : Decision-Generated Credibility in Social Learning

Reputation as Community Memory for the Agentic Web

When Individually Calibrated Models Become Collectively Miscalibrated