Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces AMemOps, a novel benchmark designed to evaluate the critical capabilities of audio memory systems in long streams: recognizing recurring sources despite acoustic variations, updating their attributes, and rejecting absent sources. By employing paired controls that manipulate recording variations, training identities, and prior events, the benchmark isolates the interplay between these demands, revealing that training on varied recordings of test instruments significantly enhances a delta-rule memory's update advantage over attention mechanisms. The results highlight how matched histories can disentangle the benefits of consistent repetition from the costs of conflicting prior information, and additive-memory controls demonstrate the value of erasure for effective attribute updating.
A single adversarial token can bypass most safety filters.
Audio memory must recognise recurring sources across acoustic changes while keeping their attributes up to date. Overall accuracy does not reveal how these demands interact. We introduce Audio Memory Operations (AMemOps), a benchmark for latest-attribute retrieval, preservation of unchanged attributes and absent-source rejection in long audio streams. Paired controls vary recordings, training identities and earlier events to examine their effects on updating. Training on other recordings of the test instruments increases a reference delta-rule memory's update advantage over matched attention by 13.4 percentage points. At retrieval delays of 120\,s, matched histories separate benefits of consistent repetition from costs of conflicting earlier values, while additive-memory controls identify benefits of erasure.