For open discussion I am going to post discussions between the authors and the Review #3 during the manuscript reviwing process as follows:

The first revision of the MS was Submitted to PLoSONE on 11 November 08 in which we replied enquiries and criticisms of the Review #3 as follows:

Reviewer #3 (Remarks for the Author):
The manuscript describes a set of studies designed to show that honeybees can learn to distinguish between different numbers of objects. I have only one concern with this otherwise nice study, but it is one that makes me think that the present set of experiments are not suitable for publication in any journal.

The problem, which the authors bring up on page 8, is that the feeder was retained during the transfer tests. There was therefore no control for the possible influence of the attractant Nasunov pheromone that honeybees release at a rewarding feeder. Such controls are absolutely basic for choice experiments. Despite the authors' claims, it is entirely possible that the attractant on the feeder biased the choices towards a positive result.

If the authors repeat the experiments, they should use dedicated test chambers, both of which contain dedicated feeders. In this way, none of the feeders or chambers will have been marked by pheromone during training.

Author’s response:
the referee’s concerns had already been addressed in the 5th paragraph of the Discussion. The “authors’ claims”, which the referee so casually dismisses, are the results of five peer-reviewed studies, which showed conclusively the absence of any effect of odour cues on the bees’ decision making. We re-iterate our arguments on this point below:

In our experiments, when honeybees opened their Nasonov gland, this was clearly visible to the experimenters. This rarely happened at all, and if it did, we removed the bees from the maze.

If the feeder had carried any scent, from the Nasonov pheromone or otherwise, the bees would have been able to solve any task we presented to them, no matter what item numbers were visible on the visual patterns. However, they were not able to do a 4 to 4, 5 to 5 or 6 to 6 match.

Now, in this revised manuscript, we have added a further level of analysis, namely that of incorrect second choices following a first correct choice:

Page 2 Line 4-7 in results :
Fig. 2b shows the percentage of incorrect second choices following a positive first choice. The percentage of incorrect second choices of the 1st block was 50.0%; this declined to a low 24% by the last block.
Page 5: Line 37-40 in Discussion
The analysis of incorrect second choices (Fig. 2b) showed that in the early stages of training, up to 50% of bees were choosing the negative pattern on their second visit, even though their first choice was correct. This is more evidence that the bees did not use scent as a cue. The frequency of this type of error declined with training.

The second revision of the MS was Submitted to PLoSONE on 17 December 08 in which we replied enquiries and criticisms of the Review #3 as follows:

Review #3:
Comment: I hope that the authors find the time to repeat the critical experiments, making sure that odor cues are not available during the tests. The same results would then probably merit publication in Nature or Science. In the present experiments, it is not possible to say that the bees were not following odor cues, and so no useful deductions can be made about any 'visual based generalization'.
Reply: We have done an additional control experiment, and added the results of the control tests in the results section in Page 4: Line 18-28 as follows:

Control tests for olfactory cues
In order to exclude the possibility that olfactory cues of the feeder were influencing the bees’ decisions, we carried out an additional control experiment (See Material and methods). When the three dot sample was presented, the bees preferred the three dot choice pattern (0.78 ± 0.15 of the decisions), at a level significantly different to random choice (n = 9, total 17 visits, p < 0.001). When the two dot sample was presented, they significantly preferred the two dot pattern. The choice frequency of 0.74 ± 0.10 for the two dot pattern is also significantly different to random choice (n = 11, total 23 visits, p<0.001). The bees significantly reversed their preference when the sample pattern changed (n = 12, total 40 visits, p<0.001). The use of new maze cylinders and the absence of a feeder behind the correct pattern did not impair the bees’ ability to solve the task, showing that olfactory cues do not play a role in the bees’ decision making in our experimental paradigm.

We have added a paragraph for this additional control tests in Material and methods section in Page 10 Line 8-20 as follows:
Control tests for olfactory cues
In order to exclude the possibility that olfactory cues were influencing the bees’ decisions, we carried out an additional control experiment. A group of bees was trained specially for this purpose. The bees were trained to solve a basic DMTS task with a set of patterns containing two and three blue dots that were the same as used in the previous training (Fig. 2a), and were then, after they had reached a high plateau, tested on new patterns with randomised two and three dot configurations. During the transfer tests, the three cylinders of the maze apparatus (Fig. 1) were replaced with fresh ones, and no reward was present in the end cylinders. The testing period was kept short (2-3min), to make sure that each bee would encounter the unrewarded transfer test situation only once, since it is similar to the punishment for making a wrong choice for the bee, and thus leads to negative learning. Each of those tests was followed by a long training period to keep the bees motivated to visit the maze.

We have added a paragraph to discuss the control tests for olfactory cues in discussion section in Page 8 Line 21-27 as follows:
To address this question, we have conducted an additional experiment to control for olfactory cues. A group of bees was trained to the basic DMTS task with visual patterns containing two and three blue dots, and then tested on patterns with randomised dot orientations in fresh maze chambers and without a feeder behind the correct pattern. The bees were still able to solve the task (see Results section and Fig. 7 for details). These data also show as well that the presence of a feeder does not influence the bees’ choice of a particular pattern in our experimental setup.

In the general discussion in Page 5 Line 27 to Page 6 Line 12, we comprehensively discussed the presence of a feeder during tests as follows:

The presence of a feeder during all stages of testing could be considered a departure from a standard memory testing protocol. The advantages of unrewarded testing include the certain exclusion of olfactory cues from the feeder, and the prevention of learning during the tests. Such testing conditions are essential only when bees are trained to a simple task, where individual bees have to go through the transfer test only once. However, as mentioned above, we wanted to ensure that the same group of trained bees kept visiting the maze throughout the duration of the experiment, i.e. over the complete series of transfer tests. Had we put them through unrewarded tests, many of the trained bees would have lost their motivation after a few attempts, and stopped visiting the apparatus. After all, in our experimental paradigm, the experience of an unrewarded test, where the bee makes a correct decision but doesn’t find a feeder behind the correct choice pattern, is similar to the punishment for making a wrong decision, and thus equivalent to negative training. One could argue that bees might not be able to solve the task without the help of olfactory cues, although these non-visual stimuli alone are not sufficient to support correct choices. However, as mentioned later in this section (‘The absence of olfactory cues’), it has been conclusively shown that the presence of a feeder during a test does not lead to false positives in the bees’ choice data. If olfactory cues did exist, the bees should have found the feeder in the case of our four vs five or four vs six dot experiments as well. In the control experiment for olfactory cues, a new set of bees was trained to the basic DMTS task, and then tested in fresh maze chambers without a feeder. The data show that the bees are able to choose the correct number of elements according to the sample pattern without the presence of a feeder in the final chamber (see Methods section and Fig. 7 for details). In Fig. 2, we show that the percentage of incorrect second choices following a positive first choice in the 1st block was 50.0%, which declined to a low 24% by the last block. This unequivocally supports the absence of olfactory cues at the feeder. In addition, we made the following observation at the end of the complete series of transfer tests, in which i) the two choice patterns and the sample pattern were identical; ii) there was a feeder with sugar water behind one choice pattern and a feeder with only water behind the other choice pattern; and iii) the positions of the two feeders were swapped after 5 min., which is half the normal testing period. The visiting frequency at the two feeders during the 10 min. observation period was 20: 17 (the feeder with sugar water to the feeder with water). There is no significant difference from random choice level (Chi= 0.003, P> 0.90).

Enquiry by Review #3: …By the fourth block, a bee is more likely to go towards the choice with the feeder, whether it is the first or second time it is counted.
Reply: Please read the text in Page 9, Line 18-21 that answers the enquiry.

Enquiry by Review #3: As an aside, I am not sure whether the first levels of 70% correct are recorded after 120 minutes of training, as the figure legend would suggest, or after 3 days, as the Methods section would suggest.
Reply: We revised the text in Page 9 Line 11-16 as follows: Training, including pre-training and training proper went on for a total of three days, by which time the bees were able to consistently solve the DMTS task. During the training proper, baffles behind the entrances of the two choice chambers (See Fig. 1) completely prevented the bees from viewing the feeder in the reward chamber from the decision chamber. The learning curve was acquired during this period.

Dr. Shaowu Zhang
The corresponding author