Research map / Large language models / Language, safety and interpretability
Pitfalls in Evaluating Language Model Forecasters
Daniel Paleka, Shashwat Goel, Jonas Geiping, Florian Tramèr
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 changed | Where | Measured | Effect |
|---|---|---|---|
| restrict questions to those resolved by the evaluation date | backtesting benchmarks · Forecasting benchmark with Jun 2023-Jan 2024 resolution window | share of leaked early-resolved questions (%): at least 3.8% | Hurt |
| restrict questions to those resolved by August 2024 | backtesting benchmarks · Forecasting benchmark resolved by Aug 2024 | share of trivial questions (%): at least 10% | Hurt |
| generate questions retroactively from news | forecasting benchmarks · Binary news-generated questions | weak classifier accuracy (%): over 80% | Hurt |
| reproduce MCQ pipeline with newer model and improved prompts | MCQ forecasting dataset · Four-choice MCQ | classifier accuracy (%): 80 -> 55 | Mixed |
| use date-restricted web search | retrieval-based forecaster · Backtesting retrieval | temporal leakage | Hurt |
| prompt GPT-4o with a later assumed knowledge cutoff | GPT-4o · Knowledge cutoff test | knowledge beyond cutoff | Hurt |
| give the LLM access to human market forecasts | LLM vs human crowd comparison · ForecastBench | fairness of comparison | Hurt |
| optimize for chance of winning instead of honest probabilities | forecaster · Shared-latent-variable questions | expected log score: -0.69N -> -0.64N | Hurt |
| report average Brier across mixed base rates | forecasting benchmark · Brier example | total Brier score: 0.125 vs 0.1 | Hurt |
| clamp probabilities under label noise | log-scored forecaster · Uniform true probability toy model | expected log loss: 1/2 + eta -> 1/2 - eta + log(1+2 eta) | Improved |
| filter to questions with at least 12 forecasters | Manifold benchmark · Manifold resolved questions | question availability: over 50% under 12 forecasters | Hurt |
Closest papers by meaning
- Are Language Models Actually Useful for Time Series Forecasting? (NeurIPS 2024 · similarity 0.90)
- Approaching Human-Level Forecasting with Language Models (NeurIPS 2024 · similarity 0.90)
- Are LLMs Prescient? A Continuous Evaluation using Daily News as the Oracle (ICML 2025 · similarity 0.90)
- Do Large Language Models Perform the Way People Expect? Measuring the Human Generalization Function (ICML 2024 · similarity 0.90)
- Large Language Models Must Be Taught to Know What They Don’t Know (NeurIPS 2024 · similarity 0.89)
- Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.88)
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.
