AI Paper Atlas

Research map / Large language models / Alignment and preferences

Reading Between the Tokens: Improving Preference Predictions through Mechanistic Forecasting

Sarah Ball, Simeon Allmendinger, Frauke Kreuter, Niklas Kühl

ICML 2026 · Large language models › Alignment and preferences · approach: Reward model design and preference models · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use value probe instead of generic regression readout on activations (ablation)party probes · Party probes (held-out split)macro-F1: 0.93 ± 0.05 -> 0.99 ± 0.01Improved
train probes on meaningful party labels rather than shuffled labels (ablation)party probes · Party probes (held-out split)macro-F1: 0.33 -> 0.99Improved
use trained probe weights rather than random weights (ablation)party probes · Party probes (held-out split)macro-F1: 0.24 -> 0.99Improved
use mechanistic forecasting (latent activations) instead of next-token probabilitiesLLM persona election forecasting · 6 national elections, 7 modelsattribute-level win-rate vs surveyMixed
apply mechanistic forecasting to base instead of instruction-tuned models (ablation)mechanistic forecasting · 6 national electionswin-rateImproved
add in-context party-position data to probability-based baseline (ablation)probability-based estimation · 6 national electionscloseness to surveyNo clear effect
use latent estimators for demographic attributesmechanistic forecasting · US and UK, age/education/employmentwin-rateImproved
use latent estimators for opinion-based attributesmechanistic forecasting · political orientation, immigration, inequality across countrieswin-rateMixed
gate mechanistic forecasting by normalized entropy above 0.85 (ablation)mechanistic forecasting · high-entropy attributesmedian error improvementImproved
use latent estimators for low-entropy attributesmechanistic forecasting · low-entropy attributeserror vs surveyNo clear effect
restrict probe layers to 0.5L-0.9L window (ablation)party probes · selectivity gap across seeds, models, countriesselectivity gap lower whiskerImproved
add sign-inversion gating of value vectors (ablation)value vector selection · intervention experimentsintervention effect consistencyImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy