Saved in:
Bibliographic Details
Main Authors: Zhu, Jinchang, Li, Jindong, Zou, Chengyu, Fu, Rong, Wang, Chao, He, Haowei, Yang, Menglin
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.10544
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910209242300416
author Zhu, Jinchang
Li, Jindong
Zou, Chengyu
Fu, Rong
Wang, Chao
He, Haowei
Yang, Menglin
author_facet Zhu, Jinchang
Li, Jindong
Zou, Chengyu
Fu, Rong
Wang, Chao
He, Haowei
Yang, Menglin
contents Long-context adaptation is often viewed as window scaling, but this misses a token-level supervision mismatch: in packed training with document masking, each target token's effective context remains short. We introduce EXACT, a supervision-allocation objective that assigns extra weight to long effective-context targets by inverse frequency within the long tail. Across seven Qwen/LLaMA CPT configurations, EXACT improves all 28 trained/extrapolated NoLiMa and RULER comparisons. On Qwen2.5-0.5B, NoLiMa improves by +10.09 (trained) and +5.34 (extrapolated); RULER by +10.69 and +5.55. On LLaMA-3.2-3B, RULER improves by +17.91 and +16.11. Standard QA/reasoning are preserved (+0.24 macro change across six benchmarks). A distance-resolved probe shows gains arise when evidence is thousands of tokens away, while short cases remain unchanged. Results support a supervision-centric thesis: long-context adaptation depends on how strongly training supervises long-context predictions.
format Preprint
id arxiv_https___arxiv_org_abs_2605_10544
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Where Does Long-Context Supervision Actually Go? Effective-Context Exposure Balancing
Zhu, Jinchang
Li, Jindong
Zou, Chengyu
Fu, Rong
Wang, Chao
He, Haowei
Yang, Menglin
Computation and Language
Long-context adaptation is often viewed as window scaling, but this misses a token-level supervision mismatch: in packed training with document masking, each target token's effective context remains short. We introduce EXACT, a supervision-allocation objective that assigns extra weight to long effective-context targets by inverse frequency within the long tail. Across seven Qwen/LLaMA CPT configurations, EXACT improves all 28 trained/extrapolated NoLiMa and RULER comparisons. On Qwen2.5-0.5B, NoLiMa improves by +10.09 (trained) and +5.34 (extrapolated); RULER by +10.69 and +5.55. On LLaMA-3.2-3B, RULER improves by +17.91 and +16.11. Standard QA/reasoning are preserved (+0.24 macro change across six benchmarks). A distance-resolved probe shows gains arise when evidence is thousands of tokens away, while short cases remain unchanged. Results support a supervision-centric thesis: long-context adaptation depends on how strongly training supervises long-context predictions.
title Where Does Long-Context Supervision Actually Go? Effective-Context Exposure Balancing
topic Computation and Language
url https://arxiv.org/abs/2605.10544