Reverse-engineering how networks encode symbolic structure as compositional vectors
tallinzen · x · 2026-09-06
tallinzen argues the hypothesis that networks must represent symbolic structure in some form to generalize on symbolic tasks is almost certainly true. The cool part: we can reverse-engineer this representation format — sometimes almost perfectly — as sums of vectors for filler/role bindings, and these representations are themselves compositional, letting us predict how the network encodes new bindings.
More from Research
- Ofir Press shares the 2016 EMNLP rejection of his now-canonical weight-tying paper — OfirPress · 2026-09-06
- OfirPress marks 10 years in language modeling; his weight tying paper was once rejected as unrealistic — OfirPress · 2026-09-06
- Check per-timestep loss curves in small diffusion training runs — jfischoff · 2026-09-06
- Tandem training: RL method keeps strong models' reasoning auditable by weaker agents — manoelribeiro · 2026-09-06
- timm author ships first CLIP model trained with both FlexiViT patch jitter and NaFlex seq-length jitter — wightmanr · 2026-09-06
- Time-encoded analog photonic interposer links vision sensors to analog accelerators, 2.04x EDP gain — jwt0625 · 2026-09-06