Study: LMs predict forward reading but struggle to explain regressions
tallinzen · x · 2026-08-19
Researchers from NYU and UMass published a paper analyzing the relationship between language models and human eye-tracking data. The study found that while current language models are good at predicting forward eye movements during reading, they fail to explain why eyes jump back (regressions) when encountering difficulties. This highlights a limitation in models' ability to mimic human error-processing and correction mechanisms during reading.
More from Research
- Study explains 10k LLM agent opinion dynamics using simple Ising model — SuryaGanguli · 2026-08-20
- SciTaRC benchmark accepted at COLM 2026: LLMs bottlenecked by execution, not planning — DanielKhashabi · 2026-08-20
- Reward Hacking vs. Overfitting: Conceptual Differences — Dr_Atoosa · 2026-08-20
- WARP: Training mobile manipulation policies directly from human data — danfei_xu · 2026-08-20
- PlanBench-XL Benchmark: Insights on Long-Horizon Tool-Calling and Data Realism — Shahules786 · 2026-08-20
- Stanford Releases PTXBench for GPU Kernel Optimization with LLMs — StanfordUniversity · 2026-08-20