UT Austin's Zhangyang Wang and colleagues release AgingBench to measure how deployed AI agents degrade even with frozen weights
Memory compression, interference, revision, and maintenance drive system degradation.
#612Ravid Shwartz Ziv@ZIV_RAVIDReally interesting new benchmark, AgingBench.
The idea is that deployed agents don't just have day-one capability, they age. Even with frozen weights, the effective state keeps changing as the agent compresses history, accumulates similar memories, revises facts, and undergoes maintenance. Reliability becomes a property of the full harness, not the model.
They break aging into four mechanisms - compression, interference, revision, maintenance, and tie each to a specific stage of the memory pipeline (write, retrieve, utilize, store), so the diagnosis tells you where to fix things.
They found that behavioral compliance and factual accuracy decouple: the agent keeps sounding right while the values quietly disappear, and violation-based monitoring catches nothing. They also found that revision aging looks representational and not capacity-bound, where scaling the model doesn't fix accumulator drift, but a typed JSON sidecar cuts error by ~47%. And Opus-4.7 reasons better but writes lower-fidelity artifacts, so the same surface failure needs a write-stage fix, not a retrieval-stage fix.
I like this framing because aging is a runtime control problem. The agent is constantly deciding what to write, compress, retrieve, and flush, and current systems make these decisions implicitly. Treating memory policy as a control policy, and aging curves as its closed-loop evaluation, feels like the right frame.
Cool work!