Large language models asked to role-play different people tend to answer like one average, polite person. Give a model a hundred personas and ask each how a news story or a personal setback makes them feel, and the hundred answers come out nearly identical. This matters most where the model is supposed to be empathetic: reading a particular person's emotional reaction is the first step of emotional support, and a model that can only predict the average reaction cannot do that step. This proposal asks a narrow question that current work does not answer. Is the individual difference already present inside the model and lost only when the model writes out its answer, or was it never represented in the first place?
The test is to ask for the reaction as a fill-in-the-blank, or cloze, query: "After reading this, I felt [MASK]." Instead of letting the model generate text, we read its internal state at the blank and match that state against embedded descriptions of emotion labels. If persona-conditioned states at the blank stay far apart while the generated answers collapse, the averaging comes from decoding, and a cheap readout can recover a large share of real human variation. If the states are already collapsed, no prompt or decoding trick will help, and the fix has to come from training. Either result tells empathetic-dialogue builders which part of the pipeline to fix, and whichever holds, the readout gives a measurable route to predicting individual reactions rather than average ones.
Core Claim
We claim that much of the "average persona" collapse in affect prediction comes from the readout, not the representation. Concretely, a prototype readout over semantic label embeddings, applied to the model's hidden state at a cloze blank, will (a) predict individual people's affective responses more accurately than generated answers, and (b) reproduce a population spread much closer to the human spread, without injecting stereotyped variation. The claim fails if decodability of individual responses from hidden states is no better than decodability from generated text, or if the recovered spread does not track real individuals on held-out people.
Motivating and Constraining Work
The failure mode is well measured. The Persona Fidelity Gap grounds personas in real Amazon purchase histories and finds that, although each persona ranks items for its own user reasonably well (accuracy 0.72 to 0.85), the synthetic population covers only 4.6% to 9.4% of the real-user neighborhood. Its intra-group cosine similarity is 0.94 to 0.99, against 0.56 to 0.68 for humans, and its vectors shrink toward a generic centroid. Their readout is Semantic Similarity Rating, which embeds the model's generated text and compares it to Likert anchors. That is already a semantic-label-embedding readout, but it runs after decoding. Our readout runs before decoding, and this placement is the one change that tests whether decoding is where the collapse happens. Related findings point the same way. Whose Opinions Do Language Models Reflect? finds demographic prompting shifts opinions only by a constant factor. The Generation Gap finds older-age personas leave value judgments unchanged. iNews argues that alignment suppresses diverse subjective affective viewpoints. None of these papers inspects internal states.
Two papers set limits on the design. Are LLMs Empathetic to All? shows that demographic personas on ISEAR produce stereotyped shifts, for example suppressed emotion intensity for "Confucian culture" personas, rather than individualized empathy. Adding spread is therefore not a success in itself: added spread must track real individuals, and our evaluation must separate the two. Sociodemographic Prompting is Not Yet an Effective Approach reports higher error and refusal under demographic prompts and attributes this to missing internal representations. That is precisely the alternative hypothesis our probing experiment tests.
The method borrows from two places. Cloze reformulation, as analyzed in Improving and Simplifying Pattern Exploiting Training, lets us query the model in its native fill-in-the-blank form with no new randomly initialized head. Prototypical Verbalizer (ProtoVerb) replaces hand-picked label words with continuous class prototypes trained contrastively, and reports +11.26% one-shot on DBPedia with frozen models. ProtoVerb was only tested on topics and entity types, where classes are mutually exclusive, and it struggles with close sub-categories. Affect labels overlap heavily, and DASC suggests that a continuous attribute space helps generalization across neighboring emotion words. We therefore initialize prototypes from embedded label definitions and output a graded distribution rather than a single argmax class.
Research Program
The first stage is a diagnostic. For each person and item, we build a persona-conditioned cloze prompt and record three things: the generated answer, the model's next-token probabilities over the label words at the blank, and the hidden state $h$ at the blank. On held-out people, we fit identical small probes from each of the three to that person's actual response and compare how much individual signal each retains. We also compute population-spread statistics for each signal: intra-group cosine, norm relative to the centroid, and coverage. Run across layers, the same measurements locate where the spread disappears.
The second stage builds the readout. Each affect label $k$ gets a prototype $e_k$, initialized by embedding its name and definition. A frozen model's blank state is mapped by a small projection $W$, and the predicted response distribution for person $u$ on item $x$ is
$$p_\theta(k \mid u, x) = \frac{\exp(\cos(W h_{u,x}, e_k)/\tau)}{\sum_{k'} \exp(\cos(W h_{u,x}, e_{k'})/\tau)},$$
where $\tau$ is a temperature and $\theta = \{W, e_k, \tau\}$ are the only trained parameters. Training uses only the individual-level cross-entropy against each person's own label. We deliberately add no loss that matches the population spread, so that any recovered spread comes from the representation rather than being imposed by the objective. A variant that does match spread is reported separately as a ceiling, but it is never counted as evidence for the claim.
The third stage is a small downstream check in emotional support. We condition an ESConv responder (Towards Emotional Support Dialog Systems) on the predicted seeker affect distribution instead of on a single inferred emotion. We then ask whether its strategy choices become less uniformly "advice-first". This stage is exploratory. Following Curry and Curry, we frame it as better perception of the user, not as a model performing feelings.
Baselines
The naive reference point is the persona-prompted model's generated answer, mapped to labels by string match or by Semantic Similarity Rating as in The Persona Fidelity Gap. On Amazon Reviews 2023 that paper reports coverage of 4.6% to 9.4%, Fidelity Gap Index of 0.91 to 0.95, and intra-group cosine of 0.94 to 0.99. We rerun it on the same splits with Llama-3.1-8B and Qwen3-4B so our numbers sit beside theirs. Beating it means higher coverage and lower intra-group cosine, with individual ranking accuracy at or above their 0.72 to 0.85.
The second baseline is a non-LLM population model, namely the Dirichlet interpolation baseline from the same paper, which reaches roughly 90% coverage. It is the strongest competitor on spread alone, but it does not condition on the item content or the person's text. Beating it means matching per-person accuracy that it cannot reach, not matching its coverage.
The third baseline is the strongest affect-prediction setup with personas in the file. That is the multi-demographic prompting protocol of Are LLMs Empathetic to All?, which reports emotion-intensity mean squared error of 0.14 to 0.212 on its 300-situation ISEAR subset. We use the same subset and the same personas. Beating it means lower error together with smaller stereotype shifts, as measured below.
The fourth baseline is LoRA fine-tuning of the full model on the same person-level training data. It gets the same number of training examples, and its hyperparameter search uses the same number of trials as ours. It is the honest "just train it" competitor.
The key ablation compares four readouts on the identical frozen blank state: label-word next-token probabilities (a classic discrete verbalizer), randomly initialized prototypes, definition-initialized prototypes, and a linear head. This isolates whether semantic label embeddings add anything beyond reading probabilities. For iNews, the dataset closest to our target of personalized affect, we found no reported result on Lacuna for iNews, so all iNews baselines are run by us.
Why not just…?
Why not just read the log-probabilities of the answer tokens? This is the most serious threat, because the next-token distribution at the blank is itself a cloze readout. It is the first arm of our ablation. If it recovers as much individual signal and spread as the prototype readout, we report that as the main finding: averaging happens at sampling, and fixing it costs almost nothing. The prototype readout only earns its place if it does better on overlapping labels, where single label words split into arbitrary subword tokens or overlap in meaning.
Why not just raise the sampling temperature? Higher temperature adds spread, but the spread is noise, uncorrelated with who the person is. The research file notes that higher temperatures add hallucination rather than coherent individual diversity. We run a temperature sweep and report spread against per-person accuracy. Temperature should raise coverage while lowering accuracy, whereas a readout that preserves the representation should raise both.
Why not just put the person's real history in the prompt? The Persona Fidelity Gap tested exactly this with top-20 purchase histories and still found coverage under 10%. We keep history-grounded prompting as the input for all of our conditions. The question is therefore not which prompt to use but where to read the model's answer.
Why not just fine-tune on per-person data? This could work, and it is baseline four. Fine-tuning changes the representation, so it cannot answer our diagnostic question. On the practical question, a probe with a few thousand parameters that comes within reach of LoRA would be the more useful tool. If LoRA wins by a wide margin, that supports the "never represented" hypothesis, and we will say so.
Proposed Experiments
We use three datasets. iNews tests personalized affective responses to news. The ISEAR subset with the demographic personas of Are LLMs Empathetic to All? tests emotion-intensity prediction under identity conditioning. Amazon Reviews 2023 lets us replicate the fidelity gap directly. The models are Llama-3.1-8B and Qwen3-4B. Every split holds out both people and items. Metrics fall into three groups. Individual accuracy is measured by cross-entropy and mean squared error against each person's response. Population spread uses coverage, intra-group cosine, and Fidelity Gap Index from the fidelity-gap protocol. Stereotype is measured as how much of the predicted variation is explained by demographic group labels compared with how much is explained in the human data.
The decisive check is the stage-one probe comparison. If held-out individual responses are substantially more decodable from blank hidden states than from generated answers, and the hidden states' intra-group cosine sits measurably closer to the human 0.56 to 0.68 than the outputs' 0.94 to 0.99, the readout hypothesis holds. The second check is the stereotype test. Recovered spread that is mostly explained by demographic group, beyond what group explains in human data, counts as failure even if coverage rises. The third check is the layer sweep: it should show where along the network the spread is lost, which decides between a fix at decoding time and a fix in training. A clean null result, with collapsed states and no probe advantage, would still settle the question for the persona-simulation and empathetic-dialogue literatures, which currently assume one answer or the other without testing it.
References
- The Persona Fidelity Gap: Behaviorally Grounded LLM Personas Still Compress Real-User Preference Diversity
- Whose Opinions Do Language Models Reflect?
- The Generation Gap: Exploring Age Bias in the Value Systems of Large Language Models
- iNews: A Multimodal Dataset for Modeling Personalized Affective Responses to News
- Are LLMs Empathetic to All? Investigating the Influence of Multi-Demographic Personas on a Model’s Empathy
- Sociodemographic Prompting is Not Yet an Effective Approach for Simulating Subjective Judgments with LLMs
- Improving and Simplifying Pattern Exploiting Training
- Prototypical Verbalizer for Prompt-based Few-shot Tuning
- Semantic Space Grounded Weighted Decoding for Multi-Attribute Controllable Dialogue Generation
- Towards Emotional Support Dialog Systems
- Computer says "No": The Case Against Empathetic Conversational AI