Context compaction quietly deleted a third of the task rules, and how the agent should work was dropped six times more often than what to build
AspireAI community · arXiv (Jeanne Salle, Hardik Bhatnagar, Matthias Bethge et al., University of Tübingen / ELLIS Institute / MPI-IS)
Compaction is how an agent scaffold stays inside its context window, and it is lossy by design. This paper is the first systematic look at that loss with frontier models on realistic long-horizon agentic tasks, holding the scaffold fixed at OpenCode and implementing five compaction strategies inside it.
For the two models with 1M-token context, the compaction threshold did not measurably change task performance on PaperBench, ProgramBench or PostTrainBench. GPT-5.3-Codex, with the smallest context window, did worse at lower thresholds on PaperBench and less clearly on PostTrainBench. No compaction method beat default self-summarization, and methods that discarded more content cost some models up to 20 points while leaving others unchanged.
The real finding is what gets dropped. On ProgramBench the first compaction summary already lost about a third of the task's anticheating rules on average, and that loss rose with each additional compaction. On real user-agent sessions from SWE-chat, instructions about how the agent should work and what it should not do were dropped around six times more often than specifications of what to build.
The authors argue current benchmarks are compaction-friendly: the objective is fixed and progress stays on disk, so most discarded content can be recovered, except constraints that exist nowhere else.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.