Updated preprint: language models mirror human content effects on reasoning tasks
AndrewLampinen · x · 2026-10-06
Andrew Lampinen shares a substantially updated preprint showing that language models and humans exhibit strikingly similar patterns in how the content of a logic problem affects their answers. He also adds threads on further content-effects-on-reasoning results and on interpreting hippocampal contributions to generalization through the lens of data augmentation. The work offers new empirical grounding for debates on whether LLMs truly reason — sharing the same content-driven failure modes as humans suggests structural parallels in reasoning limits.
More from Research
- Researcher brings Implicit Curriculum Hypothesis and HalluWorld benchmark to COLM 2026 — dan_fried · 2026-10-06
- OffensiveAIcon highlights: ExploitBench debut and fully autonomous HTTP attack cascade — dyn___ · 2026-10-06
- X-Tree: mining reusable skill trees from agent trajectories with zero LLM calls — hllo_wrld · 2026-10-06
- Two bottlenecks holding back perturbation models: system representation and signal identification — yusufroohani · 2026-10-06
- CivBench pits LLMs against each other in Civilization V: GLM-5.3 beats Opus-5.5, open-weight Qwen-3.8-27B impresses — vox-deorum · 2026-10-06
- Apple researcher lands two NeurIPS 2026 papers on normalizing flows, including Normalizing Trajectory Models — thoma_gu · 2026-10-06