Prompts to Proxies

At a glance

From problem to paradigm shift

Blueprint sketch of a chameleon, standing for an LLM proxy agent.

Blueprint of the functional basis view: preference vectors projected onto a lower dimension, recovered through a functional basis, and observed through proxy agents.

Blueprint of the letters P and P with an empty slot between them for the 2.

1 2 3 4

01Problem

Large language models are increasingly used as proxies for human subjects in social science research, but conclusions drawn from them hold for a real population (external validity) only if the synthetic agents faithfully reflect that population's preferences.

02Why rethink it

Existing preference-alignment paradigms typically rely on demographic profiling or finetuning a single one-size-fits-all model. These approaches are resource-intensive, often require sensitive data and substantial compute, and do not directly explain how population-level preference heterogeneity should be reconstructed.

03Idea

We introduce preference reconstruction theory, which treats preference alignment as a problem of representation learning and statistical inference. Instead of matching profiles, it constructs an expressive basis of proxy agents and recovers population-level preferences through weighted aggregation.

04Result

Implemented as a two-stage system, P2P requires no finetuning and no sensitive demographic data, achieves an average test MSE of 0.014 across 14 ATP waves, runs at roughly $0.8 per survey, and yields distinctive empirical insights unavailable under demographic-matching approaches.

How it works

Preference reconstruction in three steps

Animated illustration for Step 1 of P2P — 01 Attributes from the bank are instantiated and combined into agent endowments that span the preference space.

Key results

Efficient, accessible, and empirically validated

0.014 Average test MSE

41% Mean per-wave improvement over best baseline

$0.8 Cost per ATP survey

14 ATP waves evaluated

ATP benchmark

Table 1: averages across 14 ATP waves; each method evaluated over 3 runs. Costs in USD.
Method	Test MSE	Entropy	Cost
Vanilla	0.025	0.45	$0.7
PERSONA	0.029	0.43	$1.3
P2P	0.014	0.53	$0.8

Interpretation

P2P frames accessibility as a design principle rather than a compromise: no finetuning, no demographic profiling, CPU-friendly deployment, and strong performance under modest resource constraints.

All three methods use the same Stage 2 constrained lasso aggregation. The 41% improvement is the mean of per-wave test MSE reductions relative to the best baseline in each wave.

Cost estimates are based on Gemini-2.0-flash. PERSONA cost only covers the response elicitation stage—higher if persona generation tokens are included.

Paradigmatic contribution

Beyond digital twins

P2P does not treat faithful emulation as a problem of demographic matching. It treats it as a problem of representation learning and statistical inference.

Preference reconstruction, not profiling.

Existing approaches typically rely on demographic conditioning or a single finetuned model. P2P instead constructs an expressive basis of proxy agents and recovers population-level preferences through weighted aggregation.

Expressive basis beats demographic matching.

Figure 3 shows that a more expressive US-generated basis, after only Stage 2 weight refitting, outperforms HK-native generation on average. This directly challenges the assumption that locale-matched profiling is necessary for faithful emulation.

Localization matters on culturally divergent questions.

The same analysis shows where localization helps: prediction error scales with cultural divergence, and locale-matched generation regains an edge on culturally distinctive dimensions. This turns P2P into a diagnostic tool for studying alignment failure, not just a method for improving fit.

Four panels compare US and Hong Kong WVS responses (r = 0.848), test MSE (US-native 0.0149, HK-transfer 0.0208, HK-native 0.0242), error versus cultural divergence, and native win rates by divergence and data split. — Figure 3. Basis expressiveness vs. locale matching on WVS HK. After refitting Stage 2 weights to HK responses, the US-generated basis achieves lower test MSE (0.0208) than HK-native generation (0.0242). Reconstruction error grows with US–HK response divergence, while HK-native generation wins more often on culturally divergent questions. View full-size figure.

Record

Submission history

Accepted at NeurIPS 2026 · Main Track, poster Full record

25 Jul 2025

AAAI-26

14 Sep 2025

arXiv v1

18 Sep 2025

ICLR-26

23 Jan 2026

ICML-26

28 Jan 2026

arXiv v2

01 May 2026

NeurIPS-26

25 Sep 2026

NeurIPS 2026 · accepted

Software · P2P Studio

A workspace for preference reconstruction.

Inspect the fitted proxies, check held-out predictions, and explore how response diversity develops—all in one local workspace.

Get P2P Studio

ATP W42· Trust in ScienceGemini-2.0-flash · constrained lasso

Ensemble ↔ Evidence · W42 walkthrough

From population to ensemble.

300 58generated → retained

19.3%of the basis retained

10negative coefficients

Select a proxy to read its endowment.

Arrow keys move through the grid

Template familyCoreThematicTheoreticalSurveyQuestionMixed

Not retainedSize = |weight|Negative weightInspected proxy

RQ4_F1Ae_W42 Held-out test question

Thinking about medical research scientists, how often would you say they care about the best interests of the public?

One held-out question from the paper’s W42 walkthrough.

Answers, option by option

Prediction − target

Human survey and ensemble shares for the held-out medical-scientist question
Human	Ensemble	Difference
35.1%	32.1%	-3.0 pp
47.3%	55.4%	+8.1 pp
13.7%	9.0%	-4.7 pp
3.3%	3.5%	+0.2 pp
0.6%	0.0%	-0.6 pp

Select a response option to inspect its difference.

Generation diagnostics

Entropy Explorer

Mode → entropy × selected · size = generated

Inspect a mode · entropy rank 1 / 23

Final entropy: 0.547
Selected / generated: —

Persistent mode entropy

Line begins when the mode enters the bank.

23 modes · generation steps 0–10
Selection counts use this run’s constrained lasso fit.

One run. Connected views. Inspect a proxy, check its predictions, explore the modes.

Explore Studio’s features

Survey source: Pew Research Center, American Trends Panel Wave 42 — Trust in Science (2019).

Analysis and model predictions by the P2P authors. Pew Research Center bears no responsibility for the analyses or interpretations presented here.

Use P2P

Make it your workspace.

Open source under MIT. Install P2P, build Studio once, then launch it with p2p ui.

Python 3.12+ · Node.js 22.13+

Install & launch

git clone https://github.com/bingchen-wang/P2P.git
cd P2P
python -m pip install -e .
cd studio && npm ci && npm run build && cd ..
p2p ui

New to P2P? Start with the separate offline demo on Studio’s home screen.

Citation

BibTeX

Paper. If you build on the method or the empirical results.

@inproceedings{wang2026prompts,
  title     = {Prompts to Proxies: Emulating Human Preferences via a Compact {LLM} Ensemble},
  author    = {Wang, Bingchen and Khoo, Zi-Yu and Wang, Jingtan},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Software. If you use the P2P software.

@software{wang2026p2p,
  author  = {Wang, Bingchen},
  title   = {{P2P}: Prompts to Proxies},
  version = {0.1.0},
  year    = {2026},
  url     = {https://github.com/bingchen-wang/P2P}
}