Long-running applications of large language models often need to compress context, but compression typically assumes specific downstream questions are already known. A new benchmark, GACC-Bench, targets the harder case where compression must happen before those questions are asked, using an overarching goal to decide what to keep.
GACC-Bench is introduced in a recent arXiv paper (arXiv:2610.21893) as a way to assess "goal-adaptive" context compression. The authors argue that preserving goal-relevant information is critical when future queries are unspecified, and the benchmark is designed to test whether models can do this effectively.
The paper is an abstract-only source, so it does not yet include experimental results or detailed methodology beyond the benchmark's definition. Its contribution at this stage is to name and formalize a previously under-tested capability in LLM context compression.