From Confusion to Consolidation: A Staged Conversational Workflow for Post-Lecture Review

Synthesis: KnowLoop, a dual-agent conversational system for post-lecture review, structures learning around three stages—Recognize (mark in-situ confusion during lectures), Resolve (Teaching Assistant provides context-grounded clarification), and Consolidate (Peer scaffolds reflective teach-back). A 22-participant study shows confusion points serve as personalized review anchors, lecture-grounded clarification enables more targeted support than general-purpose AI, and teach-back prompts learners to reveal conceptual gaps and connect ideas across the lecture.

Study Design

Fang and Reidsma designed KnowLoop, a dual-agent conversational system for post-lecture review organized around learners' in-situ confusion. The system implements three stages:

1. Recognize — During lecture viewing, learners press a capture button at moments of confusion. The system records timestamps, aligns them with the lecture transcript, and expands to include surrounding instructional context.
2. Resolve — A Teaching Assistant agent (GPT-4o) provides clarification grounded in the marked confusion point and associated lecture transcript. Interaction is constrained to one confusion point at a time to maintain focus.
3. Consolidate — A Peer agent (GPT-4o) scaffolds reflective teach-back, revisiting confusion points in order and asking learners to explain their understanding. Progression moves from confusion-point-level articulation to lecture-level summarization.

The study involved 22 participants (STEM backgrounds, AI familiarity M=4.00/5, low prior knowledge of lecture content) watching an 18-minute introductory reinforcement learning lecture.

Key Findings

Confusion Points as Personalized Anchors

Confusion density varied substantially across participants, with no single lecture segment consistently eliciting confusion. This highlights the highly personalized nature of in-lecture breakdowns. Participants also repurposed markers to flag important concepts, not just confusion: "It wasn't confusion—I clicked because it was an important formula to review later."

Lecture-Grounded Clarification

The Teaching Assistant provided targeted, course-aligned explanations by grounding responses in the lecture transcript: "The biggest difference is that GPT has no idea what happened in the lecture… but this system knows the entire lecture." The agent also actively managed instructional relevance, noting when a question was not a key point and suggesting learners move on.

Teach-Back Surfaces Gaps

The Consolidate stage consistently exposed gaps between what learners believed they understood and what they could articulate: "Understanding something in your head is one thing; saying it out loud is another." The Peer's follow-up questions surfaced blind spots that clarification alone had not revealed.

Tensions

• The Teaching Assistant sometimes redirected questions back to lecture content when learners wanted broader exploration
• Responses could be too long and text-heavy
• The transition into teach-back felt abrupt and anxiety-inducing
• Participants wanted the ability to switch between Resolve and Consolidate fluidly

What this means for practice

• Instructors. Have learners mark confusion moments live during a lecture and use those marks as the review agenda; confusion density varied so much across the 22 participants that no single lecture segment was reliably confusing for everyone.
• Instructors. Separate the clarifying role from the teach-back role rather than relying on one assistant for both, because the two stages require qualitatively different conversational relationships.
• Learners. Treat review as two distinct jobs — get the confusion point clarified first, then explain the idea aloud — since articulation surfaced gaps that clarification had left hidden.
• Instructors. Build in fluid movement between clarification and teach-back; learners experienced the stages as interleaved, and the switch into teach-back landed as abrupt and anxiety-inducing.
• Designers. Keep agent responses short, lecture-anchored and one confusion point at a time; participants found longer, text-heavy answers unhelpful and wanted room for broader exploration than the assistant allowed.

Limitations

• The user study had 22 participants, all with STEM backgrounds and low prior knowledge of the material, working through a single 18-minute introductory reinforcement learning lecture — the authors call generalizability limited.
• Only short-term review interactions were studied; no long-term learning outcomes were measured, so retention effects are unknown.
• This is an exploratory qualitative experience study, not a controlled comparison against a no-AI condition or a general-purpose chatbot.
• Individual differences in metacognitive ability were not assessed, though the authors expect them to shape confusion-marking behavior and workflow engagement.

Citation

Fang, M., & Reidsma, D. (2026). From Confusion to Consolidation: A Staged Conversational Workflow for Post-Lecture Review. In ACM Conversational User Interfaces 2026 (CUI '26), Bremen, Germany.