We present a faster dual lattice attack on the Learning with
Errors (LWE) problem, based on ideas from coding theory. Basically, it
consists in revisiting the most classical dual attack by replacing
modulus switching by a decoding algorithm. This replacement achieves a
reduction from small LWE to plain LWE with a very significant reduction
of the secret dimension. We also replace the enumeration part of this
attack by betting that the secret is zero on the part where we want to
enumerate it and iterate this bet over other choices of the enumeration
part.We estimate the complexity of this attack by making the optimistic,
but realistic guess that we can use polar codes for this decoding task. We
show that under this assumption the best attacks on Kyber and Saber
can be improved by 1 and 6 bits.