"Perspective Transition of Large Language Models for Solving Subjective Tasks"

Reasoning through Perspective Transition (RPT) helps LLMs master subjective nuances by dynamically switching perspectives.

This paper improves LLM performance on subjective tasks by dynamically adapting perspective.

-----

https://arxiv.org/abs/2501.09265

Original Problem 🤔:

→ LLMs struggle with subjective tasks like humor and metaphor recognition.

→ Subjective tasks need understanding of context and individual perspectives.

-----

Key Insights 💡:

→ Different perspectives (direct, role-playing, third-person) can improve LLM reasoning on specific subjective tasks.

→ No single perspective consistently works best across all subjective tasks.

-----

Solution in this Paper 🛠️:

Reasoning through Perspective Transition (RPT) dynamically selects the best perspective for a given task. RPT uses in-context learning with demonstrations for each perspective. The model assigns confidence levels to each perspective's answer. The answer with the highest confidence is selected.

-----

Results 💯:

→ RPT outperforms single fixed-perspective methods on 12 subjective tasks.

→ Using Llama-3, RPT achieves an average 3.27 point improvement over best baseline.

→ Using GPT-3.5, RPT achieves an average 4.56 point improvement over best baseline.