llm-fragmentation-inconsistency-strategyqa

Summary

Breaking a source's evidence into smaller fragments can tip an LLM toward the competitor's intact presentation roughly two-fifths of the time, meaning the format in which information is delivered can sway the model's judgment independent of the content itself. This matters because it exposes a manipulation surface: someone could game an LLM's reasoning simply by reformatting one side's evidence into pieces, without changing a single word.