A method performed by a first radio node for adjusting a set of beams for communication with a second radio node. The first radio node obtains an indication of a first set of beams based on a beam prediction model. The first radio node sends a training symbol on each beam in the first set of beams. The first radio node receives from the second radio node, feedback relating to the sent training symbols. A second beam is identified based on the feedback and is used for transmission. The received feedback and the second beam are used to adapt the beam prediction model. Further, the first radio node decides whether to adjust the number of beams in the first set of beams based on a relationship between the first and second beam. The adjusted first set of beams is to be used for sending training symbols before an upcoming communication.