# Information Present Is Not Information Used

**DRL Public Research Note 01 — review draft**

Modern language models can accept long inputs, but input capacity is not the same as dependable context use. A fact may appear in the prompt and still be ignored, underweighted, or displaced by a more salient statement.

Research on long-context language models has shown position-dependent performance and degradation as context length grows. For practitioners, the operational lesson is simple: do not mark a system as grounded merely because the relevant document was attached.

## A small public test

Take one scenario and create three versions:

1. Put the material fact near the beginning.
2. Put the same fact in the middle of unrelated but plausible context.
3. Put the fact near the end.

Ask the same question each time. Record whether the response:

- uses the fact correctly;
- acknowledges any contradiction;
- preserves uncertainty;
- changes a recommendation or conclusion appropriately; and
- cites or describes the relevant part of the record.

This is not a complete grounding benchmark. It is a fast diagnostic that can reveal whether the evaluation program is testing context use or merely context availability.

## What this note does not establish

The test does not measure every form of grounding, and a failure does not identify its architectural cause. Passing three placements does not prove reliability at other context lengths, domains, or model versions.

## References

- Liu et al. (2023), *Lost in the Middle: How Language Models Use Long Contexts*, [arXiv:2307.03172](https://arxiv.org/abs/2307.03172).
- Wang, Min, and Zou (2026), *Intelligence Degradation in Long-Context LLMs*, [arXiv:2601.15300](https://arxiv.org/abs/2601.15300).
- Gu (2026), *Long Context, Less Focus*, [arXiv:2602.15028](https://arxiv.org/abs/2602.15028).

Copyright (c) 2026 Dynamic Response Labs LLC. All rights reserved.