Key points
Paper: *Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Most* — Nick Merrill, Jaeho Lee, Ezra Karger (Forecasting Research Institute / UC Berkeley), arXiv:2605.22672, May 21, 2026, cs.AI.
- A new class of inverse scaling: On time-series forecasting tasks that combine *superlinear growth* with *regime change* (crashes, interventions, collapses), the most capable LLMs make the worst errors. This is structural, not an artifact of adversarial prompts — it reproduces across real pandemic, housing, inflation, and measles data.
- ForecastBench-Sim (FBSim): A contamination-free benchmark built on simulated FreeCiv games. AI players battle; the game state is frozen and converted into a natural-language "world report"; models forecast outcomes 7 turns ahead. Each question is paired: binary (Brier) and quantile forecasts (p10–p90, scored with CRPS).
- Capability correlation flips with horizon: At H1, ECI vs CRPS Spearman ρ = +0.67 (stronger models forecast better). At H7, ρ = -0.42 (95% CI [-0.72, -0.02]). The p10 (lower tail) stays flat; the p90 (upper tail) slides from +0.78 to -0.57 — stronger models push the upper tail higher to chase trends, and pay for it when regimes change.
- Controlled SIR experiment isolates the cause: Given 60 points of exponential growth and asked to forecast past a peak/intervention, CRPS shows inverse scaling (ρ = -0.62, p < 0.001, N = 27). A matched control with linear-then-crash structure shows *positive* scaling (ρ = +0.61). Superlinear growth + collapse together — not collapse alone — trigger the failure.
- Scale and RLHF each independently worsen it: A 2×2 Llama 3.1 experiment (70B/405B × Base/Instruct) shows post-training amplifies crash-scenario errors; cases where CRPS is inflated >10× rise from 41% (70B-Base) to 63% (405B-Instruct). At 405B, even median forecasts degrade.
- Real-world replication: COVID-19 (60 countries): ρ = -0.54. US Case-Shiller house prices (2005): ρ = -0.67 — the strongest effect. Hyperinflation (12 episodes): ρ = -0.59. Pre-vaccine US measles: positive at 2 weeks (+0.64), inverse at 16 weeks (-0.42). Pre-registered flu controls (low overshoot) show no inverse scaling — confirming the mechanism.
- Knowledge doesn't fix calibration: Models correctly identified the specific historical hyperinflation event in 46 of 48 probes, yet still placed median forecasts up to millions of times above reality — knowledge exists in the model but isn't routed into calibrated tail forecasts. Domain labeling helped for COVID (fully reversed) and measles, partially for housing, and had zero effect on hyperinflation.
- Metrics hide the failure: On identical outputs, Brier scores show positive scaling (ρ = +0.45) while CRPS shows inverse scaling (ρ = -0.42). The worst errors live in the distribution's tail, which single-threshold metrics never cover. All mainstream LLM forecasting benchmarks (ForecastBench, KalshiBench, etc.) report only binary-threshold metrics — meaning this failure is systematically invisible today.
- Why it matters now: LLMs are being deployed for real-time infectious disease forecasting (Du et al., 2025, *Nature Computational Science*) amid measles resurgence. Systematic upper-tail miscalibration means overpredicting peaks in calm periods and mistiming true outbreaks.
Recommendations from the paper
1. Metric reform: Forecasting benchmarks must report at least one tail-inclusive score (CRPS, log score, or multi-threshold Brier) alongside binary metrics. 2. Rethink training objectives: RLHF-style post-training that rewards hitting point estimates independently worsens crash-scenario overcommitment as models scale. 3. Decouple capability from confidence: Investigate the "knowledge routing failure" that lets models know a crisis will end yet fail to use that knowledge at prediction time.
References
1. Merrill, N., Lee, J., & Karger, E. (2026). *Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Most.* arXiv:2605.22672. 2. McKenzie, I. R., et al. (2023). *Inverse Scaling: When Bigger Isn't Better.* TMLR. 3. Schaeffer, R., Miranda, B., & Koyejo, S. (2023). *Are Emergent Abilities of Large Language Models a Mirage?* NeurIPS 2023. 4. Karger, E., et al. (2025). *ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities.* ICLR 2025. 5. Du, H., et al. (2025). *Advancing Real-Time Infectious Disease Forecasting Using Large Language Models.* Nature Computational Science, 5(6), 467–480.