Agents revise their own reflections after every instance: backward revision beats memory and prompt-optimization baselines on three continual-learning benchmarks
Lead
A continual-learning harness for language model agents sends feedback backward through its textual state: after each instance, the trajectory just produced revises the reflections retrieved for that instance, consistently beating raw-context, memory- and reflection-based, and prompt-optimization methods across three benchmarks and three model backbones.
Story
Agents that work over long interaction streams must adapt online to distribution shifts, and the harness state now updates along a backward revision path in addition to the forward flow. Existing agent harnesses either evolved their states only through forward information flow or relied on additional replay or validation data to optimize the harness. Across three continual-learning benchmarks and three language model backbones, the framework consistently outperformed strong raw-context baselines, memory- and reflection-based methods, and prompt optimization approaches.
The mechanism uses the trajectory left by each instance as feedback to revise the reflections retrieved for that instance, so feedback propagates backward through the textual harness state. Reflections become text that later evidence can rewrite, rather than experience fixed at the moment of writing. Analysis reports source-grounded and meaningful revisions that correct overgeneralized or outdated experience under distribution shifts and improve both stability and plasticity during continual learning.
What to watch
The next step is to test this backward revision path on your own streaming task: keep each instance's trajectory as feedback and see whether it can rewrite the reflections retrieved for that instance. It suits online agents facing drifting distributions where replay or extra validation data is not allowed.
The summary gives no per-benchmark metrics, backbone sizes, or ablations, so the size of the gain from revision is unclear. Whether revision keeps the balance of stability and plasticity over longer streams or stronger distribution shifts is also worth watching.