Test-time context optimization lets frozen language models adapt through textual feedback, but ordinary conditional generation offers no separate dial for how strongly to reinforce the distributional changes a context induces. A new arXiv paper introduces EMO, a closed-loop framework that tackles this by alternating between two operations: optimizing the context and steering the logits.
By combining these steps, EMO appears to give practitioners more granular control over the influence of test-time context, rather than treating it as a single all-or-nothing signal. The abstract is truncated, so the full details of the method and experimental results are not available from the source alone.
Still, the core idea is clear: instead of relying solely on the model's default response to a prompt, EMO explicitly adjusts both the input context and the output distribution in a loop. This could be useful for applications where a frozen model must be steered reliably without retraining.