Apple’s SelfCompact adds 6.3 points and costs more than fixed compaction

Apple’s machine learning site published a paper on 12 October saying its SelfCompact scaffold matches or beats fixed-interval context compaction “at a fraction of the token cost.” Read the cost tables and that saving turns out to be measured against doing no compaction at all. Against the fixed-interval comparator it is actually benchmarking, SelfCompact is the pricier option for two of the three search agents tested.

The idea itself is clean. SelfCompact hands an agent a summarisation tool plus a rubric that decides when to fire it. At set probe intervals, the model answers a short list of yes or no questions about its own trajectory. Both the judge and the summariser are the same model, so there’s no external verifier and no fine-tuning.

On accuracy the paper is on firm ground. Across BrowseComp, BrowseComp-Plus and DeepSearchQA, SelfCompact is the strongest of five context policies for all three agents. Each benchmark was scored on 150 subsampled questions. It adds 8.5, 9.2 and 5.3 points on BrowseComp-Plus over a no-compaction run, and up to 6.3 points over fixed-interval compaction.

Model	Policy	Overall accuracy	Cost per question
GLM-4.7-Flash	No compaction	36.6	$0.13
GLM-4.7-Flash	Fixed-interval	41.5	$0.05
GLM-4.7-Flash	SelfCompact	46.4	$0.07
MiniMax-M2.5	No compaction	54.6	$0.19
MiniMax-M2.5	Fixed-interval	59.3	$0.06
MiniMax-M2.5	SelfCompact	63.9	$0.08
MiMo-V2-Flash	No compaction	48.9	$0.25
MiMo-V2-Flash	Fixed-interval	54.7	$0.14
MiMo-V2-Flash	SelfCompact	59.2	$0.13

That cost column is where the framing slips. Every bracketed percentage in the paper’s table is labelled a reduction “compared to the ‘no compaction’ baseline.” The appendix is explicit that the finer two-rate billing “preserves the headline comparison that SelfCompact costs less than the no-compaction baseline.” Set it beside fixed-interval summarisation instead and the ordering reverses in six of the nine benchmark cells.

SelfCompact matches or exceeds fixed-interval summarization at a fraction of the token cost.

Abstract, arXiv:2606.23525v2

The token counts say the same thing. On GLM-4.7-Flash and BrowseComp, fixed-interval spent 5.6M prompt tokens per question against SelfCompact’s 9.6M. None of that makes the accuracy gain less real, and at current per-token prices a few cents is noise beside a wrong answer. It does mean a team already compacting on a threshold isn’t being offered a cheaper option, only a better one.

A second gap sits inside the math results. The headline figure is an 18.1 point gain on HMMT February 2026, from Qwen3.5-9B, and the text calls the gains “most pronounced for the thinking-enabled models.” But Table 1 labels that same row “Thinking Disabled,” and Appendix A describes both Qwen3.5 models as “thinking-disabled variants.” SelfCompact led 11 of 12 math cells, losing HMMT February to fixed-interval by 1.1 points on Qwen3-30B-A3B.

The authors don’t pretend the problem is closed. Their audit of 12 fixed-interval summarisation calls logged 1,486 wrong-to-correct flips against 1,009 the other way, so 40.4% of all transitions degraded the answer. An oracle that skips compaction whenever the current answer is already right scored 52.9% on IMO-Answerbench, against SelfCompact’s 45.5%. That’s 7.4 points of headroom the rubric doesn’t reach, by our arithmetic.

The limitations section adds that only open-weight models were tested. Frontier systems, it says, may detect context rot without a rubric at all. The code is on GitHub under an MIT licence, and it covers the three search benchmarks rather than the math runs. Its README spells out four conditions that must all hold before compaction fires, one of which blocks it when the agent is stuck.

The paper is not new either. arXiv lists a 22 June submission revised on 10 July, and Apple’s research page carried it on 12 October, close to four months after that first posting. Seven of the eight authors are at Johns Hopkins, with Apple named as a part funder alongside DARPA and ONR. The same page is where the company’s recent model-to-model communication study appeared.

One more detail is worth pulling out, because it shapes the baseline. The paper’s footnote for that 30% threshold points at MiniMax’s own M2.5 model card, which says that once token usage passes 30% of maximum context “all history is discarded.” That’s deletion rather than summarisation, and the paper runs it separately as a delete-all policy. Watch whether harness vendors move to gated triggers: LangChain shipped a self-triggered compaction tool in deepagents on 11 March, which compacts at 85% of a model’s context limit by default. The paper notes it carries no rubric-based criteria for when to summarise.