Persuaded by What You Would Write: Authorship Models as Audience Priors

Whether an argument convinces someone depends heavily on who that someone is. On debate.org, knowing only a voter's political ideology predicts which debater they side with far better than anything about the debaters' language. Yet the models built on this fact either match coarse identity labels or fit classifiers to persuasion votes. Persuasion votes are scarce, and these classifiers collapse when tested on new topics or new people. This proposal starts from a cheap, abundant signal that existing audience models ignore. Many of the people whose reactions we want to predict also write arguments themselves, and what they write reveals how their beliefs turn into language. We propose to train a generative model of authorship: a language model that learns to write the way a person with a given belief profile writes. We then reuse it on the reception side. For a reader and an argument, we ask how much more likely that argument becomes when the model is told the reader's beliefs. The working hypothesis is that people favor arguments they themselves could have written.

The model is hybrid in the classic sense. A generative model supplies structured, low-variance features. A discriminative model, one trained directly to predict the label from inputs, then learns from votes how to weigh those features against an argument's reader-independent quality. The generative half is trained only on authored text, which covers every topic and every active user, including topics with no persuasion labels at all. That coverage is why we expect it to help exactly where discriminative audience models are known to fail: unseen topics, users with incomplete profiles, and readers whose ideology label says nothing about which side they will take.

Core Claim

A reader-conditioned authorship model gives a measure of argument–reader alignment that transfers across topics. It is computed as a likelihood ratio and its gradient with respect to the reader's latent belief vector. Combined with a discriminative quality term in an item-response-style preference model, it should beat both identity matching and discriminative user-feature models on held-out topics and on readers with only a few known stances. If a plain embedding similarity between a reader's own writing and the argument does as well, the generative machinery is unnecessary, and we will say so.

Where This Sits

The motivating evidence comes from audience-centric persuasion work. Exploring the Role of Prior Beliefs for Argument Persuasion shows on debate.org that ideology matching reaches 80.40% on political debates and that language features add little (81.81%). On religion debates, matching reaches only 65.37%. That gap tells us where identity labels run out. Exploiting Personal Characteristics of Debaters for Predicting Persuasiveness builds rich user profiles from Reddit histories for ChangeMyView, but feeds them to a discriminative predictor trained on delta awards. Deciphering Personal Argument Styles learns per-person preference functions over hand-built linguistic features. Its accuracy falls from 0.56–0.82 on seen topics to 0.37–0.62 on unseen ones, which is the out-of-distribution failure we target. Towards a Perspectivist Turn in Argument Quality Assessment shows that models fail to transfer between rater populations (MAE as high as 0.75 on a 3-point scale). This constrains our design: we must model the rater, not average over raters.

The closest structural ancestor is An Item Response Theory Framework for Persuasion. Item response theory, a psychometric model, explains each yes/no outcome as an interaction between a person trait and an item trait. That paper separates reader persuadability from message persuasiveness and polarity, with traits parameterized by lexical markers. We keep its decomposition but replace the hand-crafted interaction with one derived from a generative model of the reader's language. The closest hybrid pipeline is Chain-of-Thought Embeddings for Stance Detection on Social Media. There, LLM-generated reasoning is fed into a RoBERTa classifier, which filters out LLM label errors (76.3 F1, +6.1 over using the LLM output directly). That pipeline never conditions on the reader, and its generator is an off-the-shelf prompted LLM rather than a model fitted to the population being predicted.

On the method side, we borrow from Exploiting Generative Models in Discriminative Classifiers. That paper introduced the Fisher score: the gradient of a generative model's log-likelihood, used as a fixed-length feature for a discriminative classifier. Fisher Kernels for Logical Sequences showed that the idea pays off when the generative model is small and structured, cutting protein-fold error from 26% to 17.4% with 120 parameters. Our one departure is to take the gradient with respect to the reader's belief vector rather than the model's parameters. This keeps the feature low-dimensional and interpretable even when the generator is a large language model. We deliberately avoid joint max-margin training of the generator, as in Max-Margin Deep Generative Models, because that line reports the generative half collapsing when the discriminative weight is set too high. We freeze the generator after authorship training so that votes cannot distort it.

Research Program

The first component is a belief model. On debate.org each user has stances on 48 "Big Issues," often with gaps. We fit a small variational autoencoder over these stances, a generative model that compresses each profile into a latent vector $z$ with uncertainty. This lets users with only a few answered issues receive a posterior over $z$ rather than a zero-filled vector. On ChangeMyView, $z$ is fit instead from the history-derived profile features of Al Khatib et al.

The second component is an authorship model. A pretrained open language model is given a small adapter that prefixes generation with a projection of $z$. It is trained on users' own debate rounds or comments, conditioned on their own $z$. Training excludes any text whose persuasion outcome appears in test splits.

The third component is the alignment features. For reader $v$ with belief posterior $q(z_v)$ and argument $a$, we compute the belief-conditioned pointwise mutual information and its belief-space score:

$$m(a,v)=\mathbb{E}_{q(z_v)}\big[\log p(a\mid z_v)-\log p(a)\big],\qquad s(a,v)=\mathbb{E}_{q(z_v)}\big[\nabla_{z}\log p(a\mid z)\big|_{z=z_v}\big].$$

Here $p(a)$ is the authorship model's likelihood under the population-average belief. Subtracting it removes how fluent or on-topic the argument is in general, leaving only what is specific to this reader. $s$ is a vector the size of $z$ that indicates which belief dimensions the argument pulls on.

The fourth component is the discriminative preference head. For a vote between arguments $A$ and $B$, the head models

$$P(v\text{ prefers }A)=\sigma\big(\beta_A-\beta_B+\gamma\,[m(A,v)-m(B,v)]+w^\top[s(A,v)-s(B,v)]\big),$$

where $\sigma$ is the logistic function. $\beta$ is a reader-independent quality score from a fine-tuned RoBERTa encoder, playing the role of the message-persuasiveness parameter in Kornilova et al. $\gamma$ and $w$ are learned weights. Only this head and the quality encoder see vote labels.

The analysis questions follow from Longpre et al. (Persuasion of the Undecided), who find that undecided readers respond to content while stance-flippers respond to style. We will test whether $\gamma$, fit separately for undecided and opposed readers, reproduces that asymmetry. We will also test whether the belief dimensions emphasized by $s$ match the issues each debate concerns.

Baselines

The naive reference points are the majority baselines reported by Durmus and Cardie: 50.91% for political debates and 56.1% for religious ones. Next comes identity matching, which predicts that the voter sides with the debater sharing their ideology. Durmus and Cardie report 80.40% (politics) and 65.37% (religion), and 81.81% / 66.42% for their best combined models. We will reproduce these on our splits, because their numbers come from their own splits. We found no reported result on Lacuna for any model on debate.org under a topic-held-out split. Our primary comparisons are therefore against re-run baselines, and the published figures serve as sanity anchors.

The strongest existing competitor is a discriminative neural item-response model. It is the same preference head, but the interaction term is a learned bilinear map between the RoBERTa argument embedding and the raw (imputed) belief vector, trained end to end on votes. This is the natural modern upgrade of Kornilova et al., and it is the most direct test of whether generative authorship training adds anything beyond a flexible discriminative interaction. It receives the same encoder, the same vote data, and the same hyperparameter search budget, plus extra parameters matching the adapter's size. On ChangeMyView, the reference point is Al Khatib et al.'s 61.1% delta-prediction accuracy (versus 51.9% for linguistic baselines), and their characteristic-based model is rerun on our splits.

Two LLM baselines cover the generative-only route. The first prompts the same base model with the reader's stance profile as a persona and asks which argument the reader prefers. The second extends Gatto et al. by feeding persona-conditioned reasoning into the RoBERTa classifier. Both use the base model behind our authorship adapter, so the gap reflects training rather than model scale.

Three ablations isolate the key ingredient. The first replaces $m$ and $s$ with cosine similarity between an embedding of the reader's own past writing and the argument, with no generative model. The second shuffles beliefs across readers before computing features. The third drops $s$ and keeps only $m$. Beating the strongest baseline means a statistically significant accuracy gain (paired bootstrap over debates) on the topic-held-out and cold-start splits. Matching it in-distribution is acceptable.

Why not just…?

Why not just fine-tune a RoBERTa cross-encoder on the argument plus the reader's profile? This is the neural item-response baseline above, and it is a real threat on in-distribution splits, where votes are plentiful. It sees persuasion labels only for topics that have votes, while the authorship model is trained on every topic users have written about. The topic-held-out split is the experiment that decides this. If the discriminative model holds up there, the hybrid has no case.

Why not just prompt an LLM to judge from the reader's perspective? Gatto et al. document that prompted LLMs give labels that contradict their own reasoning and miss implicit links between premise and stance. A prompted model also has no access to how this population turns beliefs into words. The persona-prompt and persona-CoT baselines measure this directly. If prompting matches us without any training, that is a cheaper answer, and we will report it as such.

Why not just match ideology labels, since they already reach 80%? In politics there may be little headroom, and we do not expect large gains there. The proposal's value lies where labels carry no information: religion debates (65.37%), voters who share an ideology with both debaters, and readers with few answered issues. We report results stratified by these groups so that a gain is not masked by the easy cases.

Why not just use embedding similarity to the reader's own writing? This is the most dangerous alternative, because it captures "could have written it" without any generative model. The similarity ablation tests it head-on. If it closes most of the gap, we drop the generative component and recommend the similarity feature instead.

Proposed Experiments

The main benchmark is the debate.org corpus. Voters there both debate and report Big Issues stances, which makes the authorship-to-reception transfer possible. We use three splits. The first is random. The second holds out whole topic categories. The third is a cold-start split in which test voters reveal only $k\in\{0,4,12\}$ of their 48 stances. Webis-CMV-20 is the second benchmark. It covers the reception side through delta awards and the authorship side through users' Reddit histories, and it tests whether the effect survives when beliefs are inferred rather than self-reported. NYTimes editorials and the grassroots email corpus are not used, because their readers do not author text in the data. To control for confounds, we exclude each voter's own debates, and those of debaters they have argued against, from both the authorship training data and the features.

Three results are decisive. First, on the topic-held-out split, the hybrid should beat the neural item-response baseline, and the gap should be larger than on the random split. Second, within same-ideology voter pairs, where identity matching is at chance by construction, $m$ should still predict votes above chance, while the shuffled-belief control should not. Third, as $k$ falls toward zero, accuracy should degrade smoothly toward the quality-only term rather than collapse. If the first result fails while the similarity ablation succeeds, the conclusion is that authorship-derived alignment matters but generative modeling does not. That is still a useful and publishable answer to whether this intersection needs hybrid models at all.