Study: Word Predictability Effects Exhibit Both Linear and Logarithmic Traits
coryshain · x · 2026-07-13
A new preprint reveals through large-scale reading experiments that word predictability effects exhibit both linear and logarithmic patterns.
- Background: Previous research debated the rules of word prediction. This team expanded replication by about 5x using Self-Paced Reading (SPR) experimental data.
- Core Findings: Data shows that alongside the dominant logarithmic effect, a significant linear effect exists (as surprisal decreases, reading reaction time [RT] drops more steeply than expected).
- Mechanism Exploration: The team further analyzed the hidden state transition distances between words in Language Models (LMs), finding a sub-linear correlation with surprisal. This explains variance in RT independent of surprisal, strongly supporting the predictive preactivation hypothesis—meaning the brain partially simulates and pre-activates the next state while processing the current word.
- Conclusion: These results help clarify long-standing confusions between cloze probability and LM probability, controlled experiments and corpus studies, and behavioral and EEG (ERPs) data.
Related event: Study Reveals Linear and Logarithmic Patterns in Word Predictability(3 posts)→
More from Research
- Animation shows how an MLP’s first-layer weights change while learning MNIST — CatAstro_Piyush · 2026-07-22
- Project APE finds verifier reliability drops when papers contain multiple errors — soumitrashukla9 · 2026-07-22
- Project APE says verifier costs fell about 90x in a year as Chinese open models lead — soumitrashukla9 · 2026-07-22
- OpenAI-linked paper says capability RL can make models more reward-seeking — MariusHobbhahn · 2026-07-22
- Project APE builds its verifier benchmark from 100 AI-written papers with injected errors — soumitrashukla9 · 2026-07-22
- Paper proposes a CRED taxonomy and benchmark to measure research-error detectors — soumitrashukla9 · 2026-07-22