Keeping a tool-call draft store alive during deployment beat a frozen one by 27 percent in accepted tokens and 1.74 times in speed
AspireAI community · arXiv preprint
Speculative decoding for tool-calling agents can use the agent's own past calls as the draft. The catch is that a draft store built once goes stale: user behavior shifts, and the store starts guessing wrong. SpecMem keeps the store alive - verified calls are written back online into bounded per-user stores with LRU eviction, and a side-effect gate holds back any call whose side effects have not been declared safe until verification clears it.
On multi-session, multi-user query streams served by three open-weight models, the live store beat a frozen one by a 27% gap in mean accepted tokens, with per-benchmark gains between 11% and 34%, and up to 1.74 times wall-clock speedup. The gap widens over time, reaching 46% by the final session, which is the point: the frozen store degrades while the live one holds. The ablations are specific about what mattered - keeping the store live accounts for nearly all of the gain, while gating on declared side effects rather than retrieval confidence is what eliminated the irreversible mistakes.
Caveat: acceptance is measured as a longest-common-prefix proxy against a decoded target stream, not as deployed wire-format token acceptance, and the numbers come from the authors' own evaluation.
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.