Branch-Solve-Merge

Spawn N independent solver trajectories on the same problem, then merge their outputs with a dedicated agent into a stronger combined solution.

Branch-solve-merge fans a problem out to N independent LLM "solvers" — separate conversations that each attempt the problem in isolation — then uses a merging agent to combine all solvers' outputs into a better or more complete solution. Solvers can be diversified in one of two ways: run independently at higher temperature for sampling diversity, or explicitly prompt each one to take a different perspective on the problem.

This is a multi-agent architecture pattern for quality rather than role specialization along a pipeline: parallel candidates, then a critic/combiner, trading extra tokens and calls (N solvers plus a merge pass) for a higher chance that at least one branch — or the combination — beats what a single pass would have produced. Cost and latency scale with N, so use it when a single trajectory under compound mistake amplification is too brittle and the merge step can reconcile conflicts. Pair with self-critique prompting or Reflexion inside each branch when individual solvers still need local retries — but the branches themselves run independently rather than retrying sequentially from feedback, unlike Reflexion applied at the top level, so the technique overall is only as good as the merging agent's ability to reconcile divergent outputs.

It sits alongside the selection-based variants in test-time compute sampling, which pick one candidate out of N rather than synthesizing a new answer from all of them; merging is the right choice when the branches' outputs genuinely combine into something better, selection when only one branch can be "the" answer. Contrast also self-consistency decoding (vote among finished CoT answers) and Tree of Thoughts (search and prune intermediate thoughts): branch-solve-merge keeps full trajectories isolated and reconciles only at the end.

---

Sources

• Prompt Engineering for LLMs (Berryman), ch. 8