Off-Trajectory Reasoning: Can LLMs Collaborate on Reasoning Trajectory?

New research reveals that current large language models struggle with collaborative reasoning, showing that 'stronger' models are often more fragile when distracted by misleading information. The study of 15 LLMs found they fail to effectively leverage guidance from other models, with success rates below 9.2% on challenging problems.

• →Standard reasoning training fails to produce effective collaborative AI systems that can work together on shared reasoning tasks.
• →Counterintuitively, stronger benchmark-performing LLMs are more susceptible to distraction from misleading reasoning traces.
• →All tested models performed poorly at leveraging helpful guidance from collaborator models on difficult problems.
• →Teacher model weaknesses transfer to student models during distillation even when training data is correct.
• →The research highlights significant limitations in current off-the-shelf reasoning LLMs for multi-model collaboration.