LLM Forecasting Study: Activations More Reliable Than Text
burny_tech · x · 2026-07-11
Goodfire released a paper on LLM forecasting titled "What LLM Forecasters Know but Don’t Say".
The paper points out that while LLM forecasters often appear confident, their calibration is unstable, and their chain-of-thought may not reveal why predictions change. The research found that internal model activations reflect the true state better than the output text.
By using small probes to read these activations, the study aims to:
- Improve confidence calibration
- Detect "silent" shifts in evidence
- Partially recover prediction outcomes before reasoning begins
The authors argue this offers a cheaper and more faithful approach to auditing LLM predictions.
More from Research
- Bug Hunt Bench author: leaderboard noise is about 2-3 points — PawelHuryn · 2026-09-11
- Bug Hunt Bench ranks frontier coding models on 105 planted real-repo bugs — PawelHuryn · 2026-09-11
- PNAS paper shows a tiny billiard-ball system is a universal computer — undecidability lives in two dimensions — eigensteve · 2026-09-11
- New paper: Absolute pose estimation from affine cues and gravity direction — ducha_aiki · 2026-09-11
- LoMa Paper Ships REALLY HardPairs Dataset, Accepted at ECCV 2026 — ducha_aiki · 2026-09-11
- Johns Hopkins Launches Full-Stack Hands-on Robot Learning Class with SO-101 Arm Kits — _krishna_murthy · 2026-09-11