"On the Role of Model Prior in Real-World Inductive Reasoning"

Prior knowledge beats in-context examples: LLMs know more than we thought about generating hypotheses

This study reveals that LLMs primarily rely on their pre-trained knowledge rather than in-context examples when generating hypotheses for real-world tasks.

-----

https://arxiv.org/abs/2412.13645

🤔 Original Problem:

→ Current research assumes LLMs need in-context demonstrations to generate quality hypotheses, but the distinct roles of model prior knowledge versus demonstrations remain unclear

-----

🔍 Solution in this Paper:

→ The researchers evaluated three hypothesis generation strategies across five real-world tasks using three LLMs

→ They tested direct input-output prompting, iterative refinement with ranking, and HypoGeniC

→ They compared hypothesis quality with and without demonstrations using classification performance, LLM assessments, and human evaluation

→ They analyzed model behavior across different label formats and configurations

-----

💡 Key Insights:

→ Model prior knowledge dominates hypothesis generation in real-world tasks

→ Removing demonstrations has minimal impact on hypothesis quality

→ Prior knowledge is extremely robust and difficult to override even with contradictory demonstrations

→ This finding held consistent across text, image, and image-text modalities

-----

📊 Results:

→ Accuracy remained within 3% difference between with/without demonstration scenarios

→ Human evaluators preferred hypotheses generated without demonstrations

→ LLM-based evaluation showed higher helpfulness scores (4.01 vs 3.95) for no-demonstration cases

→ Performance stayed consistent even with flipped or random labels