Apple researcher, Johns Hopkins team let AI agents decide when to compact context

An Apple-affiliated researcher and Johns Hopkins University collaborators proposed SelfCompact, an inference-time scaffold that lets a language-model agent decide when and how to summarize its growing work history. Instead of relying only on a fixed token threshold, the approach uses a model-applied rubric intended to avoid compressing a trajectory in the middle of a calculation or search.

In experiments described in the paper, the authors reported that SelfCompact matched or exceeded fixed-interval summarization across six competition-math and agentic-search benchmarks and seven open-weight model configurations. They also reported lower estimated per-question costs than no compaction in the agentic-search experiments. The results have not been independently reproduced, and the reported savings depend on the study’s evaluation and pricing assumptions.

In the public search-agent implementation, the rubric asks whether the agent’s current reasoning forms a closed unit, can be reduced to three to five citable facts, has made new progress and is not stuck. If the rubric returns a decision to compress, the same model produces a summary. Generation then resumes from the original prompt and that summary. The method requires no fine-tuning, external verifier or separate summarization model.

For BrowseComp-Plus, the paper reported gains over no compaction of 8.5 percentage points for GLM-4.7-Flash, 9.2 points for MiniMax-M2.5 and 5.3 points for MiMo-V2-Flash. It also reported per-question cost reductions of 67%, 63% and 33%, respectively.

The search evaluation used 150 sampled questions per benchmark. For its headline cost calculation, the study priced all prompt tokens at the cached-input rate using OpenRouter prices. Those percentages therefore reflect that experimental setup rather than general deployment-cost guarantees.

On the math tests, SelfCompact posted the best result in 11 of 12 matched-budget model-and-benchmark settings, according to the authors, and improved on no compaction by as much as 18.1 percentage points. The exception was Qwen3-30B-A3B on HMMT February 2026, where fixed-interval summarization finished 1.1 points ahead.

The paper also included an ablation testing the rubric’s role. Removing rubric-gated verification reduced GLM-4.7-Flash’s average search score from 46.4% to 41.0% and lowered Qwen3-4B-Instruct-2507’s IMOBench score from 45.5% to 40.9%. The authors presented those results as evidence that merely exposing a summary action was not enough in their tests.

The study was limited to open-weight models and did not test frontier proprietary systems. The authors also excluded reinforcement learning to isolate the rubric’s contribution. For related coverage of a different long-horizon agent problem, see Apple’s work on continual-learning agent evaluation.

The authors provide a public MIT-licensed repository with no-compaction, fixed-interval and rubric-gated implementations, along with evaluation scripts for the three search benchmarks.