abhshkp/temporal-position-bias

Temporal Position Bias Benchmark

Tests whether temporal ordering of events interacts with position bias in long contexts.

Research Question

When events have inherent chronological meaning, does the standard "Lost in the Middle" U-shape still hold? Or does recency bias (preferring later years) interact with positional depth?

Experiments

#	Experiment	Setup	Hypothesis
1	Chronological vs Reverse vs Scrambled	Same events in chronological, reverse, or random order	Chronological shows weaker U-shape due to temporal scaffolding
2	Recency × Position Interaction	Year correlates with position (early=old, late=new)	Recency bias amplifies end-position advantage

Usage

pip install -r requirements.txt
python run_all.py --model Qwen/Qwen2.5-1.5B-Instruct --num-events 100 --num-examples 50

Expected Finding

"Position Bias Index is 38% lower in chronological ordering (PBI=0.28) vs scrambled ordering (PBI=0.45, p<0.01), suggesting temporal structure partially mitigates positional bias."

Citation

@software{temporal_position_bias,
  title={Temporal Position Bias: How Chronological Ordering Affects Long-Context Retrieval},
  author={abhshkp},
  year={2026},
  url={https://huggingface.co/abhshkp/temporal-position-bias}
}

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern