Active embodied question answering requires an agent to gather evidence by changing viewpoints and physically interacting with the environment. The main obstacle is undirected interaction: the agent may take plausible actions without a clear specification of what evidence those actions should reveal for answering the question. Relying solely on action plans and accumulated histories often leaves the informational purpose of an interaction implicit.\n\nEviSeek introduces a training‑free framework that makes the evidence demands derived from the question explicit acquisition objectives. Its core mechanisms are:\n- Translating the question into a set of evidence demands, each corresponding to a visual or interaction cue needed for the answer;\n- Selecting actions conditioned on current observations, executing only those that can satisfy pending evidence demands;\n- Selective evidence retention, storing only the perceptual fragments that contribute to the answer and discarding irrelevant data.\n\nEvaluated on ESI‑Bench (3,436 questions), EviSeek improves accuracy by 7.1 pp with Qwen3‑VL‑32B and by 3.7 pp with Gemini 3.1 Pro. Visual‑support‑weighted accuracy also rises, while mean token usage drops, indicating gains in information‑acquisition efficiency.\n\nThese results support the hypothesis that making the informational purpose of interaction explicit—and organizing action selection and evidence retention around the evidence the agent needs to learn—significantly boosts performance and resource efficiency in embodied QA systems.\n\nReview: By formalizing evidence demands as concrete objectives, EviSeek offers a clear solution to the information‑diffusion problem in embodied QA, providing a valuable design paradigm for future multimodal interactive learning.