SeDT: Sentence-Transformer Decision-Transformer Conditioning for Multi-Turn Conversation Reliability

Researchers present SeDT, a training-free method that improves large language model performance in multi-turn conversations by annotating conversation history with relevance scores, addressing a documented 39% performance drop when tasks are revealed incrementally across multiple turns.

Large language models demonstrate a significant reliability problem in multi-turn conversations despite their strong single-turn performance. When task information is distributed across multiple conversation turns rather than presented upfront, LLMs experience substantial performance degradation—up to 39% in the studied benchmark. Critically, this decline stems almost entirely from unreliability rather than reduced capability, with models failing to distinguish essential constraints from casual dialogue in flat conversation histories.

The SeDT methodology addresses this structural limitation through a novel application of reinforcement learning concepts to standard inference. By importing return-to-go conditioning principles and assigning cumulative relevance scores to conversation segments using semantic, lexical, and positional signals, the approach helps models identify which prior interactions matter most. The innovation requires no retraining, no additional training data, and no context pruning—critical advantages for practical deployment across existing systems.

The research validates improvements across diverse conditions: three different LLM architectures and three separate generation tasks all show performance gains, with some combinations achieving +37.7% improvements. Equally important, the method simultaneously reduces unreliability in seven of nine tested combinations, directly tackling the core failure mode. This addresses a fundamental challenge in conversational AI where users naturally reveal information incrementally, making the Lost-in-Conversation phenomenon highly relevant to real-world applications.

The implications extend beyond academic interest. Any system relying on multi-turn interaction—customer service bots, research assistants, creative writing tools—could benefit from this approach. The training-free nature means rapid integration into existing deployments without resource-intensive retraining cycles. Future work likely explores whether this method generalizes to even longer conversations and whether similar annotation strategies address related LLM brittleness problems.

• →SeDT improves multi-turn LLM performance by up to 37.7% using relevance-annotated conversation history without retraining
• →The method addresses a documented 39% performance drop when tasks are incrementally revealed across conversation turns
• →Improvements stem from signaling which past interactions matter most, reducing unreliability that accounts for majority of performance loss
• →The training-free approach integrates seamlessly with existing LLM deployments using semantic, lexical, and positional scoring signals
• →Validation across three LLMs and three generation tasks demonstrates broad applicability to conversational AI systems