AI Systems Know What to Do But Not When to Do It

LLM tutors could scaffold learning, but the timing and depth of their interventions might undermine cognitive engagement. How do AI systems decide when to help versus when to let users struggle productively?

Method: Int-Bench simulates student problem-solving across code debugging, math, and brain teasers while LLM teachers decide whether to intervene. LLMs intervened more frequently and earlier than humans, and provided complete solutions rather than targeted hints. This pattern optimized for immediate task success but failed to support reasoning processes needed for generalization to new problems.

Caveats: Simulation-based evaluation. Real student behavior and adaptation over time may differ from simulated scenarios.

Reflections: Can LLMs be trained to recognize productive struggle versus genuine confusion? · What intervention timing maximizes transfer learning across problem types? · Do users adapt their help-seeking behavior when they know the tutor will overassist?

3,000 of 4,500 oracle bone script characters remain undeciphered due to fragmentary inscriptions. Current AI approaches fail to replicate expert workflows integrating morphological analysis, contextual semantics, and philological reasoning.

Method: AlphaOracle systematizes decipherment through a four-stage pipeline: rubbing parsing, radical-based morphological analysis with diachronic modeling, contextual retrieval with semantic alignment, and philological validation against classical sources. Each stage generates confidence-weighted evidence chains. In a study with 86 domain specialists, it reduced analysis time by 64% and 79% of participants rated it highly useful.

Caveats: Tested on oracle bone script only. Transfer to other undeciphered writing systems unverified.

Reflections: Can the evidence-chain approach generalize to Linear A, Indus script, or other undeciphered systems? · How does expert trust evolve as AI-generated hypotheses accumulate over time? · What confidence threshold should trigger human review versus autonomous publication?

Postpartum depression affects 20% of new mothers, but screening via self-report questionnaires imposes high burden. Passive sensing works for general depression, but postpartum behavioral patterns are distinct—do existing indicators generalize?

Method: PocketPPD detects PPD risk using passive smartphone sensing of maternal contextual features like disruptions in behavioral rhythms and stability shifts. In a four-week study with 61 postpartum women, the PMS-only model achieved an AUC of 0.75; integrating PMS and self-report features reached 0.83. Morning and late-night routine volatility ranked among top biomarkers, moderated by infant developmental stage and employment status.

Caveats: Four-week feasibility study with 61 participants. Longer deployments needed to validate stability of biomarkers across postpartum stages.

Reflections: How do biomarkers shift as infants transition from newborn to toddler stages? · Can the system distinguish PPD from sleep deprivation or adjustment disorder? · What intervention timing maximizes clinical utility without overwhelming new mothers?