AI-generated characters that repeat themselves or blur together isn't a new complaint, and new research offers a fix aimed at the selection process, not the generation process.
The paper introduces TalkMatrix, a method for picking dialogue lines for multiple characters and situations as one connected set rather than one at a time. For each character-situation pair, the system generates several candidate lines, then jointly selects a full matrix of completions using four embedding-based measures of consistency and diversity. Because averaging those measures can let one improve at another's expense, TalkMatrix instead maximizes whatever objective is performing worst, a minimax approach solved with multi-start coordinate ascent. The researchers tested it against baselines that pick lines locally, in partial matrices, or through generic combinatorial search, across 50 synthetic role-playing scenarios and 25 curated board game scenarios, with an LLM-as-judge scoring the results.
That framing matters beyond this one paper. Most chatbot and game-NPC dialogue tools still generate each character's line independently, which is exactly the setup that lets characters contradict themselves or start sounding interchangeable over a long scene. TalkMatrix's results, where matrix-level selection beat random and independent-cell baselines according to the LLM judge, suggest some of that drift is a selection-strategy problem, not just a model-quality one.
Worth noting: the evaluation is limited to synthetic scenarios and board-game setups judged by another LLM, not a shipped chatbot handling hours of open-ended play, so how well this holds up outside a controlled test is still an open question.