GEAR reward shaping curbs LLM copying in long-context reasoning

The failure mode is almost funny once you see it. A new arxiv preprint from Lizhe Fang, Weizhou Shen, Tianyi Tang and Yisen Wang describes what they call repetitive copying in long-context reasoning models: as you hand the model more text, it stops reasoning and starts transcribing, pasting chunks of the input back into its own chain of thought rather than productively solving the problem.

The authors' claim is that this shows up across frontier long-context LLMs and that it gets worse as the context length grows. Their diagnosis is insufficient grounding, the model cannot tell task-relevant evidence from distractor material, so it hedges by copying indiscriminately. Their fix, GEAR (Grounding Evidence-Aware Reward), is a reward shaping technique for reinforcement learning fine-tuning that combines an accuracy signal with a reward for overlapping with actual evidence and a penalty for overlapping with distractor context. They pair it with an automated pipeline that generates evidence-annotated training data from natural documents.

The headline number is up to +4.6 average points over standard RL with accuracy-based rewards, with larger gains at longer contexts, alongside reduced repetitive copying and shorter thinking length. If that holds up, the interesting part is not the score, it is the cost angle: reasoning models are expensive because they generate a lot of tokens per answer, and less copying means fewer wasted tokens per query.

The honest caveats are the usual ones for a preprint. This is a single arxiv submission from 21 July 2026, not a peer-reviewed result. The abstract does not tell you which base models were tested, which specific benchmarks the +4.6 average was drawn from, or how the automated evidence-annotation pipeline was validated. Take the specifics as reported, not settled.

If the pattern generalises, the teams to watch are the ones building long-context RAG stacks and reasoning agents, where token cost and irrelevant-context leakage are both live problems. A training-time fix that shortens chains of thought without hurting accuracy is the sort of thing that quietly improves the economics of every downstream product built on top.