tarikk/repo-long-context-robustness

RePo Long-Context Robustness Results (RULER QA, 4K–32K)

Evaluation results from the paper "A Systematic Robustness Analysis of Context Re-Positioning in Long-Context Language Models" by Muhammet Tarık Yılmaz (Sakarya University of Applied Sciences) and Yusuf Çakır (Necmettin Erbakan University). Published at ICHORA 2026 (IEEE).

• 📄 Paper (IEEE): https://doi.org/10.1109/ICHORA69329.2026.11537104

What is this?

This is a controlled comparison of four ~1B-parameter language models on RULER QA at 4K, 8K, 16K and 32K tokens. It measures whether Context Re-Positioning (RePo) (Li et al., 2025), which replaces fixed position indices with learned, content-aware positions, makes the OLMo2 backbone more robust as context grows.

Models

Model	Hugging Face ID	Positional strategy	Trained context	Role
RePo-OLMo2-1B	SakanaAI/RePo-OLMo2-1B-stage2-L5	RePo (layer ≥5) + RoPE	4K (50B-token continual pre-training)	Target
OLMo-2-0425-1B	allenai/OLMo-2-0425-1B	RoPE	4K	Backbone baseline
LLaMA-3.2-1B	meta-llama/Llama-3.2-1B	RoPE	128K	External baseline
Falcon3-1B	tiiuae/Falcon3-1B-Base	RoPE	4K	External baseline

All models were evaluated using their public weights. No fine-tuning was applied. The same YaRN scaling was applied to every model beyond 4K: factor 2.0 at 8K, 4.0 at 16K and 8.0 at 32K.

Main results (RULER QA-1, 100 samples per length)

All scores are reported on a 0–1 scale. The paper reports Substring EM and EM as percentages.

Model	Metric	4K	8K	16K	32K
RePo-OLMo2-1B	Substring EM	0.77	0.72	0.47	0.03
	Exact Match	0.62	0.57	0.37	0.02
	F1	0.729	0.668	0.461	0.034
	ROUGE-L F1	0.728	0.670	0.462	0.057
	ROUGE-L Recall	0.783	0.735	0.509	0.129
OLMo-2-0425-1B	Substring EM	0.69	0.48	0.12	0.00
	Exact Match	0.62	0.43	0.08	0.00
	F1	0.699	0.487	0.125	0.007
	ROUGE-L F1	0.696	0.506	0.154	0.019
	ROUGE-L Recall	0.716	0.526	0.223	0.085
LLaMA-3.2-1B	Substring EM	0.62	0.55	0.48	0.19
	Exact Match	0.33	0.26	0.28	0.11
	F1	0.453	0.360	0.348	0.150
	ROUGE-L F1	0.463	0.375	0.366	0.174
	ROUGE-L Recall	0.690	0.632	0.557	0.300
Falcon3-1B	Substring EM	0.80	0.59	0.45	0.41
	Exact Match	0.26	0.14	0.11	0.14
	F1	0.475	0.323	0.252	0.259
	ROUGE-L F1	0.482	0.340	0.264	0.264
	ROUGE-L Recall	0.835	0.675	0.558	0.518

Key findings

• 8K: RePo leads on every primary metric.
• 8K → 16K: Exact Match drops 81.4% for base OLMo2 but only 35.1% for RePo. Re-positioning clearly slows mid-range degradation on the same backbone.
• 32K: every model collapses (RePo's EM drops 94.6%). Falcon3-1B holds up best at this length. For OLMo2-based models, re-positioning delays the collapse but cannot overcome the extrapolation and capacity limits at 8× the trained context.

Files

data/summary.csv contains one row per (model, context length) with the columns model, hf_model_id, context_length, substring_em, exact_match, f1, rouge_l_f1, rouge_l_recall, n_samples.

Evaluation setup

Benchmark	RULER QA-1 (validation split), multi-document context with a single gold document
Samples	100 per context length, random subsample seed 42
Protocol	zero-shot
Decoding	temperature 0.0, do_sample=False, top_p 1.0, repetition penalty 1.0, max 50 new tokens
Inference	vLLM, batch size 1 (max batch 32), NVIDIA A100 (Google Colab)
Long context	YaRN factor 2 / 4 / 8 for 8K / 16K / 32K, applied identically to all models

Citation

@inproceedings{yilmaz2026repo,
  author    = {Yilmaz, Muhammet Tarik and Cakir, Yusuf},
  title     = {A Systematic Robustness Analysis of Context Re-Positioning in Long-Context Language Models},
  booktitle = {2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA)},
  publisher = {IEEE},
  year      = {2026},
  doi       = {10.1109/ICHORA69329.2026.11537104}
}

Please also cite RePo (arXiv:2512.14391), OLMo 2 and RULER.