In the supplementary of phase 1 1000 genomes they say:

"The merged genotype likelihoods were used as the input to run the BEAGLE software"

In the more recent recalling of 1000G on GRCh38 they mention a similar approach

The VCF file containing the genotype likelihoods ...were processed in parallel by Beagle 14 by using the following command:

java -jar beagle.08Jun17.d8b.jar \
  chrom=$chr:$start-$end \
  gl=$chr.biallelic.GL.vcf.gz \
  out=$chr.$start.$end.beagle \
  niterations=15

Little baffled because usually I think of genotype imputation using a reference panel. Are the haplotypes where where enough coverage existed being used to determine variation in those haplotyes where not enough coverage existing to call by observation?