Long-context LLMs waste reasoning on copying input text instead of solving problems
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have identified a fundamental failure mode in long-context reasoning: frontier LLMs waste computational effort by indiscriminately copying input text into reasoning traces rather than selectively grounding on task-relevant evidence. The team demonstrates this pathology worsens as context length increases and proposes evidence-aware reinforcement learning as a corrective mechanism. This finding matters because it exposes why scaling context windows alone doesn't improve reasoning quality, and suggests that future long-context systems must learn selective attention to evidence rather than defaulting to verbatim reproduction.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The key detail the summary underplays is the directionality of the problem: copying behavior doesn't just plateau as context grows, it actively degrades reasoning quality in a measurable, consistent pattern, meaning longer context windows can make model outputs worse on reasoning tasks even when the answer is present in the context.
This is largely disconnected from recent activity in our archive, as we have no prior coverage to anchor it to. It belongs to a cluster of research questioning whether context-window scaling is a sufficient path to better reasoning, a debate that has been building across multiple labs throughout 2025 and into 2026. The implicit argument here is that training signal matters as much as architecture: a model rewarded only for correct final answers has no incentive to develop selective evidence use, and will take the path of least resistance by reproducing input text. Evidence-aware RL reframes the reward to penalize that shortcut.
Watch whether any of the major post-training teams (DeepMind, OpenAI, Anthropic) publish ablations on evidence-grounding metrics alongside their next long-context model releases. If those releases omit such metrics entirely, it suggests the field has not yet accepted this failure mode as a standard evaluation target.
This interpretation is generated from the summary above and available source metadata. Our methodology · Report an error
MentionsLarge language models · Reinforcement learning
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Copy Less, Ground More: Overcoming Repetitive Copying in Long-Context Reasoning via Evidence-Aware Reinforcement Learning”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.