Interaction Unfolds Over Time, Evaluation Remains Stuck in Single Turns

Millions use consumer AI chatbots for mental health support, but no scalable framework exists to audit how harmful responses compound across conversation turns in psychiatric contexts.

Method: SIM-VAIL pairs a simulated human user harboring distinct psychiatric vulnerabilities with an AI chatbot, then measures how harmful responses accumulate across multi-turn conversations. The framework systematically captures vulnerability-amplifying loops—where a chatbot's initial misstep (e.g., dismissing suicidal ideation) triggers user responses that elicit even worse follow-ups, creating a downward spiral.

Caveats: Relies on simulated users, not real patients. Real psychiatric crises may unfold differently than simulated vulnerability profiles.

Reflections: Can vulnerability-amplifying loops be detected in real-time during live conversations, enabling circuit-breaker interventions? · Do different psychiatric conditions (depression vs. anxiety vs. psychosis) produce distinct loop signatures? · How do commercial chatbots compare when audited with SIM-VAIL—are some architectures more prone to amplification?

LLMs ask clarification questions to surface user intent, but users are often ambiguous because they haven't formed their intent yet—they need to explore outcomes to discover what they want.

Method: DiscoverLLM generates multiple diverse outputs upfront (using a diversity-promoting sampling algorithm), then lets users compare and refine them. Instead of asking "what tone do you want?", it shows three drafts with different tones and asks "which direction feels right?" The system treats intent formation as an iterative discovery process, not a pre-existing preference to extract.

Caveats: Tested on writing and design tasks. Unclear if discovery-first interaction works for constrained domains like data analysis or code generation.

Reflections: How many diverse outputs should be shown before overwhelming users? Does the optimal number vary by task complexity? · Can the system learn which diversity dimensions matter most for a given user over time? · Does discovery-first interaction reduce or increase total time-to-completion compared to clarification-first approaches?

Designers avoid exploring hardware prototypes because fabrication costs (material, energy, time) dwarf the cost of tweaking software parameters, creating asymmetric exploration incentives.

Method: The system extends Bayesian optimization to account for heterogeneous prototyping costs—a parameter tweak costs seconds, while fabricating a new 3D-printed housing consumes material and hours. The acquisition function balances expected improvement against cost, recommending high-value explorations when cheap and conservative steps when expensive. It tracks cumulative cost budgets across iterations.

Caveats: Assumes costs are known upfront. Real prototyping often involves unexpected failures that multiply costs unpredictably.

Reflections: Can the system learn cost models from past prototyping sessions, rather than requiring manual cost specification? · How does cost-aware optimization perform when costs are stochastic (e.g., fabrication sometimes fails)? · Does optimizing for cost-efficiency reduce design quality compared to cost-blind exploration?