EMO: A Closed-Loop Test-Time Computation Framework via Alternating Context Optimization and Logits Steering

Key points

• Test-time context optimization adapts frozen language models through textual feedback, but ordinary conditional generation provides no separate control over how strongly to reinforce the distributional changes induced by a context update.1
• The E-step revises a system context using textual gradients from an external judge without access to gold answers.1
• Feedback on the steered response drives the next context update, coupling adaptation across rounds with intervention within each generation.1
• Theoretically, we connect parameter and context optimization through a shared target-KL objective and additive logit representations, and derive the steering vector from context-induced logit displacements.1

Sources (1)

• [1]EMO: A Closed-Loop Test-Time Computation Framework via Alternating Context Optimization and Logits Steering
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 04:23 PM

  We introduce EMO, a training-free framework inspired by expectation--maximization that alternates question-specific context estimation (E-step) and distribution refinement (M-step).

  Test-time context optimization adapts frozen language models through textual feedback, but ordinary conditional generation provides no separate control over how strongly to reinforce the distributional changes induced by a context update.

Extractive summary: sentences quoted from the sources.