Abstract
Standard dialogue models often focus on local word prediction rather than long-term conversational goals. We introduce Curved Contrastive Learning (CCL), a framework that maps utterances into a curved latent space where distance represents conversational turns. This allows models to "imagine" future trajectories and plan multi-turn responses. Our approach achieves significant speedups and outperforms established baselines in long-term planning tasks.
Introduction: The Challenge of Goal-Oriented Dialogue
While Large Language Models (LLMs) excel at generating fluent text, they often lack the strategic depth required for goal-oriented conversation. Most models are trained to minimize perplexity, predicting the most likely next word rather than steering the conversation toward a specific outcome. Human dialogue, by contrast, involves "inner speech"—a mental simulation of potential future turns to reach a goal, such as booking a flight or resolving a conflict.
Existing response-ranking methods, like cross-encoders, are computationally expensive, requiring the re-processing of the entire dialogue history for every candidate. There is a clear need for a proactive, efficient "forward model" for dialogue.
Curved Contrastive Learning (CCL)
The proposed Curved Contrastive Learning (CCL) method treats conversation as a geometric path. Instead of binary classification (related vs. unrelated), the model learns the relative turn distance between utterances.
The Curved Latent Space
Using a bi-encoder architecture (BERT or RoBERTa), the model ensures that the cosine similarity between two utterances is inversely proportional to their distance in the conversation. To capture the temporal direction, special [BEFORE] and [AFTER] tokens are introduced.
Imaginary Embeddings and Planning
By projecting a "goal" utterance and candidate replies into this space, the model can calculate which candidate brings the conversation closest to the goal. This "imagination" happens without generating intermediate turns, functioning like a GPS calculating a route through latent space.
Evaluation and Results
The models were evaluated on datasets like DailyDialog and the Microsoft Dialogue Challenge (MDC). CCL demonstrated superior performance in both short-term and long-term planning. On DailyDialog, the model achieved a planning capability of 61.56% over three turns.
As shown in Figure 6, the "Imaginary Embedding" variants scale better with dialogue history than established baselines like DialogRPT and ConvRT. In long-term planning, the model achieved an average rank of 1.87/6 when ordering three future goal utterances. Furthermore, as a bi-encoder approach, CCL allows for pre-computing history embeddings, resulting in a 4.3x speedup compared to state-of-the-art models like ConvRT.
Limitations and Future Directions
The authors noted that the model can sometimes be "too greedy," choosing semantically relevant but socially premature responses. Additionally, "unplannable" turns (e.g., "Oh, okay") can introduce noise. Future work will focus on refining these embeddings to handle diverse social intents and exploring how this strategic approach can be integrated into larger generative architectures.
Summary updated