Treating a multi-agent system's configuration as episodic memory, with explicit ownership rules for every edit, beat four baselines on three benchmarks and used 17 to 38 percent fewer training runs
AspireAI community · arXiv (cs.MA, cs.AI)
Deploying LLM agents in a specialised domain needs task-specific procedures the base model does not have, and the usual options are parameter training, which is costly and opaque, or hand-engineering, which is expert labour. EMMA treats a multi-agent system's configuration - the controller's routing memory, the specialists inheriting a generalist prompt, and the per-step procedural skills - as the thing that gets learned, rather than the model weights. A reflector assigns each diagnosed failure to the scope that owns its correction and proposes exactly one edit inside that scope, and backbone-independent guards keep every accepted edit scoped, incremental and bounded.
Across KramaBench, AppWorld and BBEH, with five backbones and four baselines (ACE, GEPA, EvoMAS and ADAS), the paper reports the best overall performance on every benchmark, up to 32.5 percent over the strongest baseline, consistent improvements across backbones, and 17 to 38 percent fewer training executions than the most efficient baseline.
Two caveats are worth carrying. The execution savings come from the authors' own guard design rather than from a hardware or dollar measurement, so 'fewer training runs' is not the same as 'cheaper training'. And the 'episodic memory' here is agent configuration state rather than a record of past experience, so the name does not line up with the retrieval-memory literature - a reader looking for memory of what happened will not find it here.
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.