Long Context Breaks What Models Know, Not Just How They Know It
The Information Abundance Paradox shows that training on longer contexts reduces parametric knowledge retention—models stop storing facts internally when the context window provides them, then fail catastrophically when context is absent or misleading.
Why it earned a slot
Arda Uzunoglu, Benjamin van Durme, and Daniel Khashabi submitted arXiv:2608.12218 on August 12, and it contains an observation that should alarm anyone scaling toward infinite context. Scaling context is not neutral. It actively degrades what models internalize. The paper tests a simple hypothesis: if you give a model abundant relevant information during training, does it still encode that information into parameters, or does it outsource the work to attention? The answer is the latter, and the cost compounds at test time. On pretraining with long documents, increasing the context window improves performance up to an intermediate optimum. Then it declines, consistently. On closed-book MCQA—the test that measures whether a model *remembers*—performance drops as context length grows. This is not noise. This is a phase transition. The model stops learning to retain facts because the training signal rewards it for using context instead. Supervised fine-tuning shows the same pattern. More task-relevant context at training time improves performance when supporting context is present at test. But remove the context, or corrupt it, and robustness collapses. The model never learned to solve the task; it learned to retrieve the solution from context. I think this is the wrong variable to be chasing. We have spent the last two years treating context length as a knob to turn up and treating knowledge retention as something that happens automatically. This paper says they are in direct opposition. You cannot have both. The model has a budget—gradient pressure, capacity, whatever you want to call it—and long context consumes it. Mechanistically, the paper identifies where this happens. Training with informative context shifts gradient pressure away from feed-forward networks, which are known to store parametric knowledge, and toward attention modules. Causal interventions confirm it: this shift increases reliance on context during inference. The model is not choosing this strategy because it is smarter. It is choosing it because it is simpler. Attention is cheaper than storage. The practical consequence is severe. You train a model on long documents. It performs well in your benchmark. You ship it to production where context is incomplete, or adversarial, or simply missing because the user did not provide it. The model fails in ways that are hard to predict because it never learned the underlying knowledge. It learned to be a retriever. I do not know whether this scales to the frontier models people are actually deploying. The paper tests up to 4,096 tokens, which is not infinite context by 2026 standards. What I know is that the Information Abundance Paradox is not speculative. It is mechanistically grounded and empirically reproducible. The field is scaling context because it works in benchmarks. This paper suggests we are not measuring what we think we are measuring.