In Hsu and colleagues’ LLM-NPC study, players reported more autonomy, but also higher cognitive load and lower usability and trust. The authors found no statistically significant improvement in overall gaming experience.[1]

Consider a hypothetical scene where a player needs a clue. A dialogue menu gives them questions to choose from. With a text box, they have to work out what to ask. After the NPC replies, they still have to decide whether it understood them and whether they got the clue.

A repeated question could be useful to look at. Perhaps the player missed the clue in the reply, or perhaps they expected the character to do something else. Ask what they were trying to accomplish at that point. Their explanation can give you context for the words in the dialogue log.

For a comparison, you could keep the goal and available clues the same and change only how the player asks. Record what happens between the first question and the player’s next action. Ask separately how much freedom they felt and how hard it was to get the information they needed. Someone may value the freedom and still find the task difficult.

Optional opening questions or a clearer signal when a clue has been found are possible changes to test. Choose a change that addresses the moment you observed, then check whether it helps the player move on. The paper’s abstract does not establish that either change works.

Sources and revisions

  1. The Double-Edged Sword of Open-Ended Interaction: How LLM-Driven NPCs Affect Players’ Cognitive Load and Gaming Experience

    paper · arXiv:2604.10107v2 (revised 2026-08-29) · Source accessed:

Verification record

Paper reading
abstract
Code inspection
not_done
Execution
not_run

Official unversioned arXiv record retrieved on 2026-10-10 and explicitly reports v2 revised 2026-08-29. The versioned abstract URL failed once; this bounded alternate succeeded. Current claims remain abstract-only. No new full-text reading, data/code inspection or execution. Design examples and checks are editorial proposals. Style-only rewrite requested by owner; source reading dates and scope retained.