cs.AINov 13, 2026
SaveA: Adaptive Audio Agent Framework for Collaborative Edge Computing
Organizations: Nanyang Technological University · National University of Singapore · Rein Protocol Foundation · ASUS Intelligent Cloud Services
Abstract
Embedded audio agents need a nearby edge server for speed, but a server-dependent agent returns nothing when the link fails. We present A, an adaptive audio agent framework that places whole pipeline stages on the server, keeps a complete fallback agent on devices that can host one, and hedges both paths with a link-aware deadline that decides when to give up on the server. On Jetson devices and an RTX A4500 server, A matches the server-only agent on a healthy link, answers every request in a full outage on the AGX, and beats a percentile hedge in simulated outages.
Figures & tables
Table 1: Comparison with the closest methods. , , and denote support, partial support, and no support. Link-aware: the decision uses the measured link state. Deadline: the method targets a completion deadline.
Figure 1: Qwen3-4B decode throughput by placement and concurrency on the A4500. SD: speculative decoding; AR: autoregressive decoding; tok/s: tokens/s.
Figure 2: A 3 overview. The server path starts first, the local agent starts after a deadline chosen by the controller, and the first complete reply is played while the other path is cancelled.
Figure 3: Simulated policy comparison. P95: 95th-percentile hedge; Exp., Per., Logn.: exponential, periodic, lognormal outages.
Figure 5: Simulated share of requests completed within 30 s against the mean outage length at a mean uptime of 60 s (left), and end-to-end accuracy on 100 spoken GSM8K questions by LLM size and input (right).
Figure 7: AGX: fraction answered within 20 s (left), median latency (right). Inset: G good link, J jitter, B bandwidth cap, L loss, R random, W window outages; crosses: unanswered requests.
Figure 8: New outages: testbed (left), simulator (right). R/P: random/periodic; numbers: mean up/down times (s).
Figure 10: Scale-up agent: one request and its routing over GPUs (left), throughput, 95th percentile latency, and time to first token (right).
Figure 11: Field study from Tuas Link to Bugis: round-trip time (RTT) along the route (left), response time and RTT over time (right).