AI
VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 25/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 05:28:08 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 05:18:24 UTC.
Briefings
VoxMem evaluates spoken conversational memory in large audio language models using 3,196 instances across 34,743 sessions. It identifies a critical gap where models retain spoken content better than speaker identity, paralinguistic cues, or environmental sounds. The study introduces a taxonomy of memory evaluation based on acoustic evidence types and operations applied to them. No model surpassed 40% accuracy at 32K tokens, with performance widening for complex operations and longer histories. The research aims to establish a foundation for measuring spoken memory progress.
Key points
- VoxMem introduces a taxonomy for evaluating spoken memory based on four acoustic evidence types (speech semantics, speaker identity, paralinguistic cues, environmental sound) and four memory operations (extraction, reasoning, tracking, refusal).
Context and limitations
Evaluation is limited to 15 LALMs, with no model exceeding 40% accuracy at 32K tokens. Results are based on 177 hours of data from 34,743 sessions, and findings may not generalize to all model architectures or contexts.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 05:28:08 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeib5lrorcgbhq7y645tlfz6huyiezvmsdbsahs2nbm7c7sm2kqnkim