The Internet has the potential to accelerate scientific problem solving by engaging a global pool of contributors. Existing approaches focus on broadcasting problems to many independent solvers. We investigate other approaches that may be advantageous by examining a community for mathematical problem solving -- MathOverflow -- in which contributors communicate and collaborate to solve new mathematical 'micro-problems' online. We contribute a simple taxonomy of collaborative acts derived from a process-level examination of collaborations and a quantitative analysis relating collaborative acts to solution quality. Our results indicate a diversity of ways in which mathematicians are reaching a solution, including by iteratively advancing a solution. A better understanding of such collaborative strategies can inform the design of tools to support distributed collaboration on complex problems.