Measuring semantic progress in multi-turn dialogue via information gain

arXiv·medium signal

This paper (arXiv 2606.12332, He, Kasiviswanathan, Janzing) tackles the hard problem that conversation quality emerges across turns rather than within a single response, proposing an information-gain metric to quantify how much semantic progress each turn contributes. For agent builders, it offers a principled alternative to single-turn scoring when evaluating multi-turn assistants, tool-using agents, and dialogue systems. Turn-level information gain could become a useful signal for reward shaping and eval harnesses in conversational agents.