Pith. sign in

Tighter Bounds on the Expressivity of Transformer Encoders

1 Pith paper cite this work. Polarity classification is still indexing.

1 Pith paper citing it
abstract

Characterizing neural networks in terms of better-understood formal systems has the potential to yield new insights into the power and limitations of these networks. Doing so for transformers remains an active area of research. Bhattamishra and others have shown that transformer encoders are at least as expressive as a certain kind of counter machine, while Merrill and Sabharwal have shown that fixed-precision transformer encoders recognize only languages in uniform $TC^0$. We connect and strengthen these results by identifying a variant of first-order logic with counting quantifiers that is simultaneously an upper bound for fixed-precision transformer encoders and a lower bound for transformer encoders. This brings us much closer than before to an exact characterization of the languages that transformer encoders recognize.

citation-role summary

background 1

citation-polarity summary

fields

cs.LG 1

years

2026 1

verdicts

ACCEPT 1

roles

background 1

polarities

unclear 1

representative citing papers

Attention-based representations for multi-task computation

cs.LG · 2026-08-04 · accept · novelty 7.0

For min/max readout, two attention heads beat one head by an exponential resource gap, and for n-bit parity and symmetric Boolean functions, heads times polynomial degree must reach the threshold degree, with matching constructions.

citing papers explorer

Showing 1 of 1 citing paper.

  • Attention-based representations for multi-task computation cs.LG · 2026-08-04 · accept · none · ref 5 · internal anchor

    For min/max readout, two attention heads beat one head by an exponential resource gap, and for n-bit parity and symmetric Boolean functions, heads times polynomial degree must reach the threshold degree, with matching constructions.