Low rates of herbicides typically select for the large number of genes that confer a small modicum of resistance (usually those that increase the rate of herbicide degradation – as polygenes, gene amplifications, or allelic modifications of catabolism genes).  Target site mutations with a much greater level resistance are far more rare, and while on they can be selected for at high rates of a given herbicide, they are less likely to be found at the lower rates. Many had assumed that only those chemicals that directly affected DNA replication and repair were mutagens.  There are many ways to increase mutation rates indirectly by depleting the cell of material needed of DNA replication and repair (e.g. of ATP and NADPH) or increase DNA damage (e.g. oxidative stress) beyond the threshold capacity of DNA repair enzymes. Stress is a general enhancer of mutation rates, as the cell often requires ATP and/or NADPH to overcome the stress, or stress directly damages the DNA. 
At very low herbicide rates, a certain low proportion of weeds may receive a sub-lethal dose, yet survive despite being highly stressed by the herbicide. Thus, the survivors are likely to have more mutations due to the stress.  These would include mutations leading to herbicide resistance, both for multi-factorial (polygenic, gene amplification, sequential allelic mutations), as well as for major gene resistance.  The more such small mutations conferring a modicum of resistance occur, the quicker they can recombine or duplicate, conferring much higher levels of resistance.  Weed resistance management strategies should consider how to eliminate the sub-population of weeds with the high mutation rates, but the best strategy is probably to avoid too low application rates of herbicides from the onset. The mutagenicity of sub-lethal doses also has implications to herbicide discovery, as many excellent herbicide leads were dropped when found that they were mutagenic at low doses in a bacterial mutagenesis (Ames) test.  We now know that stressed bacteria mutate at higher rates than non-stressed bacteria. Many such compounds may have stressed the bacteria at the levels tested, but below regulatory residue threshold doses would not have stressed mammalian systems enough to increase the mutation rate.  If this had been considered good potential pesticides would not have been lost.  This must also be considered in re-registration of older compounds. Thus, it is important for many reasons that we must be cognizant of the potential of very low doses to increase the rate of resistance mutations via the stressing of target weeds.
References

Gressel, J.  and A. A. Levy (2010) Stress, mutators, mutations and stress resistance, in  Abiotic Stress Adaptation in Plants: Physiological, Molecular and Genomic Foundations, A. Pareek, S.K. Sopory, H. J. Bohnert, and Govindjee, eds., Springer, Dordrecht, pp. 471-483.
Gressel, J. (2011) Low pesticide rates may hasten the evolution of resistance by increasing mutation frequency. Pest Management Science 67:253-257.

jonathan.gressel@weizmann.ac.il