Researchers Release PersuasionTrace to Track LLM Belief Changes Turn by Turn

@NostaIgicGareth @noahdgoodman @nickhaber @maxhkw Yeah! It's here: https://github.com/jlcmoore/persuasiontrace

Awesome,

I’ll create a token for this tech on PumpFun, I’ve done this for Stanford Princeton MedOS team. ($30k+ in donations)

persuasiontrace is a web platform for multi-turn persuasion experiments between humans and/or LLM agents. The FastAPI backend serves rounds from configurable conditions, logs process-level outcomes (including turn-level belief traces), and stores data in SQLite.

CA below

Bottom line: endpoint movement alone is not enough. Process fidelity matters: when beliefs move, how they move, and whether simulator updates look human.

Code: https://github.com/jlcmoore/persuasiontrace Preprint: http://arxiv.org/abs/2606.05330

This is work done with: @noahdgoodman @nickhaber @maxhkw

@jaredlcm @noahdgoodman @nickhaber @maxhkw Hey whatsup, this is really good, can I fund this GitHub?

@NostaIgicGareth @jaredlcm @noahdgoodman @nickhaber @maxhkw redirect the fees bro

@jaredlcm @NostaIgicGareth @noahdgoodman @nickhaber @maxhkw I Donated some

8obi3JNiPji8rhoZdFmnr1ZeorFjDxG6Dr3Gfqywpump

@luffone10 @jaredlcm @noahdgoodman @nickhaber @maxhkw Will do

Stance bias asks whether a simulator is much easier to move in one stance direction than the other. This matched for-versus-against asymmetry is lowest for our BN target, indicating less stance-dependent bias than baselines.

We built a human-participant-facing web platform for AI persuasion experiments that supports multi-turn belief tracing, audio I/O, and participant-chosen propositions. Using it, we show LLMs can persuade across standard text, personalized text, and audio.

People show different belief traces and rhetorical susceptibility. We see two trajectory regimes: some people barely move, while others shift substantially early on and then partially drift back. We find that ethos is negatively associated with persuasion delta.

LLM-judge human-likeness scores place our BN target near the human reference and above baselines.

Naive responsiveness asks if trivial arguments (repeating the proposition over and over) move the target too much. Only our BN resists trivial persuasion, while both LLM targets overreact to it.

We then build a probabilistic simulator of human persuadability. It compares an unstructured LLM target, a structure-conditioned LLM target, and a Bayesian-network target with explicit latent belief-state updates each turn.

@jaredlcm @noahdgoodman @nickhaber @maxhkw CA: 8obi3JNiPji8rhoZdFmnr1ZeorFjDxG6Dr3Gfqywpump

@0xCalculated @jaredlcm @noahdgoodman @nickhaber @maxhkw

if needed

@jaredlcm @noahdgoodman @nickhaber @maxhkw fees redirected

@jaredlcm you can claim this using pummpfun mobile app login with ur github

@luffone10 @jaredlcm @noahdgoodman @nickhaber @maxhkw done