Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-13 · source-backed
arXiv 2608.12218 names the Information Abundance Paradox: pretraining on long documents improves language modeling, NLU, and closed-book MCQA only up to an intermediate optimum, then consistently declines. Mechanistically, informative context shifts gradient pressure away from feed-forward networks (linked to parametric knowledge) toward attention, and causal interventions confirm this increases inference-time context reliance. In SFT, more task-relevant context helps when supporting context is present at test time and reduces robustness when it's absent or misleading. That last clause is the trap for anyone fine-tuning a RAG-fed agent.
Each link below shares sources, entities, or timing with this story.
RAGOCR competes with RAG / Shared entity: RAG / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (RAGOCR competes with RAG); both cover RAG; reported by the same outlet (arxiv.org).
Shared entity: RAG / Same source domain / Shared topic / Earlier coverage / Tension
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (attention, document).
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (context, document).
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (agent, document).
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (agent, context).
Shared entity: RAG / Same source domain / Earlier coverage / Tension / Downstream implication
Both cover RAG; reported by the same outlet (arxiv.org); earlier RAG coverage from 2026-07-30.
Shared entity: RAG / Same source domain / Shared topic / Earlier coverage
Both cover RAG; reported by the same outlet (arxiv.org); overlapping topics (agent, document).
Shared entity: Longer / Same source domain / Shared topic / Earlier coverage
Both cover Longer; reported by the same outlet (arxiv.org); overlapping topics (agent, context).