The variable length codes for inter and intra frames are designed with different assumptions. In the case of the inter frames it is assumed that the values of the coefficients will be small and there can be large numbers of zero values between nonzero coefficients. This is a result of prediction which, if successfully employed, would reduce the magnitude of the differences, and hence the coefficients, and would also lead to large numbers of zero-valued coefficients.
Therefore, coefficients indexed with large runs and small coefficient values are assigned shorter codes. In the case of the intra frames, the opposite is generally true. There is no prediction, therefore there is a much smaller probability of runs of zero-valued coefficients. Also, large-valued coefficients are quite possible. Therefore, coefficients indexed by small run values and larger coefficient values are assigned shorter codes. During periods of increased temporal activity, prediction is generally not as good and therefore the assumptions under which the variable length codes for the inter frames were created are violated. In these situations it is likely that the variable length codes designed for the intra frames are a better match.
The alternative inter VLC mode allows for the use of the intra codes in these sitations, improving the compression performance. Note that the codewords used in intra and inter frame coding are the same. What is different is the interpretation. To detect the proper interpretation, the decoder first decodes the block assuming an inter frame codebook. If the decoding results in more than 64 coefficients it switches its interpretation.