Talking to Machines: LLMSS 2026 Workshops
Digital twins, silicon samples and silicon audiences, in R
These are the materials of the three Talking to Machines workshops held before the LLMSS 2026 conference at City University of Hong Kong on Tuesday 13 October 2026. Every notebook is here with its output, so you can read it in the browser or download it and run it yourself (see Run it yourself).
Every notebook ships with stored model answers, so it runs start to finish with no network and no API key. Only the blocks where you send your own prompts call a model.
One method, three uses
All three workshops run the same four-step engine and point it at a different question. Only the last step changes.
stratification frame → synthetic subjects → digital interview → aggregation
(who the population (records + donors) (repeated, ├─ W1 MrP → public opinion
is) time-anchored) ├─ W2 panel econometrics → market signals
└─ W3 closed-loop search → best narrative
The Ohio panel built in Workshop 1 comes back as the audience in Workshop 3.
The workshops
Workshop 1 · AIPOP: AI polling and MrP
Build a stratification frame for Ohio, draw a silicon sample of 400 twins from it, interview them with a large language model, and turn their answers into MrP estimates checked against certified results and an election the twins never saw.
| Notebook | ||
|---|---|---|
| 1 | Building a silicon sample | The frame, the draw, and an audit against the project’s own panel |
| 2 | Personas and the digital interview | Matching census skeletons to survey donors, rendering personas, the vote interview |
| 3 | MrP and validation | Two MrP models, the 2026 Ohio Senate race, twelve state panels and a held-out election |
Ray Duch and Matías Fuentes Becerra · source folder
Workshop 2 · The Wisdom of Digital Twins
Build twins of ten financial influencers, interview them every day about the news on a set of stocks, and turn their answers into a long–short signal, set against the crowd and the prediction markets.
| Notebook | ||
|---|---|---|
| 1 | Finfluencer twins and the daily interview | Profile cards, the interview prompt, and what a twin says about a stock |
| 2 | From twin answers to a market signal | Panel regressions, a backtest, and a replay of a ten-agent deliberation |
Ray Duch and Matías Fuentes Becerra · source folder
Workshop 3 · The Silicon Audience
Write a message, watch a twin panel score it, and improve it. Once for a political ad, once for a closing argument before the Supreme Court.
| Block | ||
|---|---|---|
| Silicon Voters | Ads for five kinds of voter | A model writes ads for Ohio voters; you improve its prompt in three stages, scored by the twins and by model judges |
| Silicon Jury | A justice twin, one case, by hand | Run one justice twin on one case, compare it with the full bench and the Court, change one input |
Ray Duch and Laurenz Günther, with Matías Fuentes Becerra · source folders: Silicon Voters, Silicon Jury
Data
The notebooks use public data (the ACS 2019–2023 five-year PUMS, the Cooperative Election Study 2020–2025, and certified election results) and aggregate results (district, county and cell totals, never an individual) from the Talking to Machines digital-twin panels. The personas are built on synthetic skeletons from the census, not on voter-file records. Each workshop folder has a data/README.md with the source of every file.
Only the Ohio panel files are public. The other states’ panel results, used in Workshop 1, Notebook 3, Section 7, carry non-commercial terms. This site shows their output, but the files are not in the repository.
The method
- Talking to Machines: an international collaboration of academics and industry researchers, directed by Sonja Vogt (HEC Lausanne) and Ray Duch (Nuffield College, Oxford), funded by the Swiss National Science Foundation. The frames, panels and vote module in these notebooks come from its digital-twin work.
- PoSSUM: Cerina and Duch, Protocol for Surveying Social Media Users with Multimodal LLMs.
Instructors
- Ray Duch, Nuffield College, University of Oxford
- Matías Fuentes Becerra, University of Chile
- Laurenz Günther, Toulouse School of Economics
Licence
The code is under the MIT licence. The text, figures and data are under CC BY 4.0, except where a data file’s README says otherwise.