NYU study: capping AI memory at 4 slots makes it a far better stand-in for real humans
tallinzen · x · 2026-09-17
New research from NYU's Center for Data Science shows the counterintuitive key to making AI better at imitating humans is limiting its memory: capping a model's memory at 4 slots makes it a far better stand-in for a real person.
The work was led by CDS Faculty Fellow Nick Tomlin, PhD student Michal Huang, and Associate Professor Tallinzen, with former NYU Data Science master's student Qihan Wang playing a major role.
The finding is a useful signal for building human-behavior simulations and agent-based social modeling: perfect recall isn't necessarily an advantage — a bounded memory window better matches real human cognitive limits.
More from Research
- Will Brown: robustly scaling reward modeling is the key problem for capabilities and safety — willcb · 2026-09-17
- Mistral OCR's Vik Paruchuri launches a new document-extraction benchmark — VikParuchuri · 2026-09-17
- Datalab launches OmniExtractBench: 620 docs to fix biased structured-extraction benchmarks — VikParuchuri · 2026-09-17
- Training a GRU planner to guide flow diffusion improves motion continuity in a homebrew video model — pixlpa · 2026-09-17
- Five Pharma Firms Federally Fine-tune OpenFold3, Lifting Ligand-Pose Accuracy 29% to 47% — rbhar90 · 2026-09-17
- Qwen 3.8 27B runs 63 hours on one RTX 3090 attempting the Riemann hypothesis, zero hallucinations — GuiltyBookkeeper4849 · 2026-09-17