Abstract
While Large Language Models (LLMs) excel at single-turn instruction following, they frequently suffer from "context drift" during extended multi-turn conversations, leading to a loss of semantic coherence and failure to adhere to initial constraints. This paper introduces a novel "Skeleton-Guided" framework designed to synthesize high-quality multi-turn training data by modeling human conversational intents and global information flow. By generating complete dialogue skeletons and utilizing a single-pass response generation strategy, the resulting ConsistentChat dataset enables models to maintain strict consistency across deep dialogue turns. Our results demonstrate that models fine-tuned on ConsistentChat not only achieve superior multi-turn performance but also exhibit significant improvements in general reasoning and coding capabilities.
Introduction: The Challenge of Multi-turn Consistency
Current LLM training paradigms rely heavily on synthetic instruction datasets like Self-Instruct or Alpaca, which are predominantly composed of single-turn exchanges. When extended to multi-turn scenarios, these models often degrade, forgetting earlier instructions or drifting off-topic. Existing multi-turn datasets frequently exacerbate this by stitching together local turns without a global plan, failing to capture the goal-oriented progression of natural human dialogue.
Preliminary investigations reveal that even state-of-the-art models experience a steady decline in instruction-following ability as conversations deepen, particularly after the sixth turn. To address this, we propose a framework that moves beyond reactive, turn-by-turn generation toward a proactive, globally planned dialogue synthesis.
Methodology: Skeleton-Guided Synthesis
We introduce a Skeleton-Guided Multi-Turn Dialogue Generation framework that operates in two distinct stages to ensure global coherence:
- Intent and Flow Modeling: Drawing from human-computer interaction theories, we define nine fundamental conversational intents (e.g., Problem-Solving, Emotional Support, Educational). Each intent is mapped to a specific "information flow," such as transitioning from problem diagnosis to solution optimization.
- Skeleton-to-Dialogue Generation: Unlike traditional methods that generate turns sequentially, our framework first prompts an LLM to generate a complete sequence of user queries (the "skeleton") that follows the logical flow of the chosen intent. Subsequently, the model generates all agent responses in a single forward pass using Chain-of-Thought (CoT) reasoning.
This single-pass approach effectively bypasses the "causal masking" limitation of LLMs. By providing the model with the entire planned sequence of user queries at once, it can generate responses that are globally consistent and proactively anticipate future turns, rather than merely reacting to the immediate prompt.
The ConsistentChat Dataset
Using this framework, we constructed ConsistentChat, a dataset containing approximately 15,000 multi-turn conversations and over 224,000 utterances. The data spans the nine defined intent categories, providing a diverse and structured training signal for conversational consistency.
Experimental Results
We evaluated models fine-tuned on ConsistentChat against several baselines, including ShareGPT and UltraChat, across benchmarks such as LIGHT (character consistency), TOPDIAL (proactive dialogue), and MT-EVAL (comprehensive multi-turn tasks).
Performance and Consistency
Models trained on ConsistentChat showed a 20–30% improvement in consistency scores on the LIGHT and TOPDIAL benchmarks. Notably, a Qwen-2.5-7B model fine-tuned on our dataset outperformed the much larger Qwen-2.5-14B-Instruct model in multi-turn evaluations, suggesting that data quality and coherence are more critical than parameter count for conversational stability.
General Reasoning Gains
Perhaps most surprisingly, the structured nature of ConsistentChat translated to broader cognitive improvements. The Qwen-2.5-7B model saw a 30.4% increase in MATH scores and a 33.7% increase in HumanEval (coding) scores compared to its base version. This suggests that learning to maintain logical flow in dialogue reinforces the underlying reasoning capabilities of the model.
Conclusion
The ConsistentChat framework demonstrates that conversational consistency is a product of data structure rather than just scale. By guiding synthesis with human-like intent skeletons, we can produce models that remain coherent throughout deep dialogues. Future work will involve expanding the range of conversational intents and exploring the scalability of this framework to frontier-class models.
Summary updated