A companion paper showed that a transformer's feed-forward layer can be rebuilt from explicit fuzzy set operations - intersection, set-difference, and a self-forgetting sequence quantifier - so its hidden units read as named logical operators at no cost to language-model quality. That left the other half of the transformer opaque.
arXiv:2607. 04319v1 Announce Type: cross Abstract: A companion paper showed that a transformer's feed-forward layer can be rebuilt from explicit fuzzy set operations - intersection, set-difference, and a self-forgetting sequence quantifier - so its hidden units read as named logical operators at no cost to language-model quality.
By Mark Oskin
arXiv:2610.00526v1 Announce Type: cross
Abstract: In-context learning (ICL) can be amortized into latent objects (task vectors, function vectors, context vectors) that recover few-shot behavior at ze...
By Gunmay Jhingran
Baobab compiles an OWL 2 DL (ΣROIQ) ontology with a finite ABox into a Sentential Decision Diagram (SDD), saturating a propositional core and instantiating remaining DL features over the active domain. The resulting evidence‑conditioned weighted model count trains a perception network to recognize real images under partial ABox supervision, enabling a CNN to recover latent ontology concepts that an independent perception would miss. When supervision allows multiple ontology‑consistent completions, Baobab’s mixture indexed by query justifications represents the calibrated posterior, achieving Bayes‑optimal performance on a real‑image MNIST task where single‑WMC and learned mixtures fail, thereby characterizing and mitigating reasoning shortcuts in a non‑Horn description logic.
By Olga Mashkova, Asaad Mohammedsaleh, Fernando Zhapa-Camacho, Robert Hoehndorf
arXiv:2608. 06111v1 Announce Type: cross Abstract: Positional embeddings (PE) in Transformers encode token distance and order but are largely agnostic to \textit{syntactic structure}.
By Haris Riaz, Hyungji Kim, Mihai Surdeanu
arXiv:2607. 08946v1 Announce Type: new Abstract: A transformer can be built from operators that are legible by construction -- bounded, named units that read as fuzzy set operations rather than dense activations -- but legibility must be pressed for during training, and the pressure has a failure mode.
By Mark Oskin