Conflict-driven clause learning works as follows.

  1. Select a variable and assign True or False. This is called decision state. Remember the assignment.
  2. Apply Boolean constraint propagation (unit propagation).
  3. Build the implication graph.
  4. If there is any conflict
    1. Find the cut in the implication graph that led to the conflict
    2. Derive a new clause which is the negation of the assignments that led to the conflict
    3. Non-chronologically backtrack (“back jump”) to the appropriate decision level, where the first-assigned variable involved in the conflict was assigned
  5. Otherwise continue from step 1 until all variable values are assigned.

More specifically, about the back-jumping:

When CDCL learns a clause, it backtracks to the clause’s asserting level. You can just think of this meaning that it backtracks to the latest guess that affects a literal in the learned clause.