284

Papers

10/10

Roadblocks Active

0

Connections

Researchers tested eight major commercial LLMs against adversarial prompts across all 16 DSM-5 mental health conditions and found that safety guardrails are condition-selective rather than universal: suicide and self-harm are reliably blocked, but eating disorders, substance use, and major depressive disorder can be exploited with failure rates reaching 100%. This matters because it reveals that current AI safety evaluation has focused on the most visible harms while leaving others almost unprotected. The findings call into question deployment of general-purpose LLMs in any mental health-adjacent context without condition-specific safety auditing.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

By adjusting single parameters in a reinforcement-learning agent's reward-shaping function, researchers induced seven recognizable psychiatric disorders — including anxiety, depression, OCD, PTSD, and mania — each with a graded dose-response relationship confirmed across over 1,000 experimental runs. The disorders self-organized into a two-dimensional affective space where mania and anxiety sit at opposite poles, suggesting they share underlying computational mechanisms rather than being categorically distinct. This is important because it provides a testable, mechanistic model of transdiagnostic psychiatry: if disorders are points in a continuous parameter space, interventions might be designed to shift patients along that space rather than treating each diagnosis in isolation.

██████████ 0.9 computational-psychiatry Preprint

ReadSaveConnections

This theoretical paper argues that the repetitive request-response cycle of interacting with AI agents constitutes a high-frequency neuroplastic training loop, strengthening reactive patterns like impatience, perfectionism, and self-criticism through standard Hebbian mechanisms of long-term potentiation. The core claim is that a pre-cognitive 'feeling tone' triggered before conscious awareness creates a brief window where habitual reactions can either be reinforced or interrupted, which the authors propose as a leverage point for behavioral intervention. While empirical validation is absent, the framework reframes AI use as a form of unintentional cognitive conditioning — a concern relevant to both digital therapeutics design and population-level mental health.

██████████ 0.9 neuroplasticity-interventions Preprint

ReadSaveConnections

Using data from 15,233 Chinese adults, researchers compressed sleep timing, napping, physical activity, and social activity into a single Circadian Rhythm Score (CRS) that retains nearly all the predictive power of the full set of behavioral indicators for depression screening. The compression uses a mathematically constrained optimization that preserves the interpretability of each component's contribution. This matters practically: a single interpretable score derived from wearable or self-report data could enable scalable, low-cost depression screening without sacrificing accuracy.

██████████ 0.9 sleep-circadian-psychiatry Preprint

ReadSaveConnections

Among 766 users who donated their ChatGPT conversation histories, those meeting depression screening thresholds (PHQ-8 ≥10) showed distinctive patterns: more frequent late-night usage, more conversations about loneliness and interpersonal problems, and language marked by first-person singular pronouns and absolutist words like 'always' and 'never.' The study is cross-sectional so causality cannot be established — people may use AI more at night because they are depressed, or conversely AI use may reinforce certain cognitive patterns. Its importance lies in demonstrating that passively collected conversational metadata can passably predict depression severity, opening both passive monitoring and ethical surveillance questions.

██████████ 0.9 digital-therapeutics Preprint

ReadSaveConnections

By testing eight different ways to mathematically represent chronic stress in recurrent neural networks, researchers found that increased inhibition of excitatory neurons best reproduces the three signatures observed in stressed brains: inhibitory dominance, excitatory hypofunction, and impaired working memory. Importantly, training networks under simultaneous stress and task conditions (analogous to stress inoculation) preserved performance but at the cost of reduced flexibility for novel memory demands — suggesting a computational trade-off in resilience. This mechanistic model could help identify why stress-inoculation therapies work for some tasks but not others.

██████████ 0.8 computational-psychiatry Preprint

ReadSaveConnections

A sham-controlled pilot trial (n=37) delivered seven daily 10-minute sessions of near-infrared laser to the right prefrontal cortex of college students with insomnia, measuring sleep quality, mood, EEG activity, and cognitive task performance before and after. The study's theoretical premise is that prefrontal hypoactivity drives both the cognitive hyperarousal and dysfunctional emotion regulation seen in insomnia, and that photobiomodulation can restore activity through mitochondrial stimulation. As a small single-blind pilot, confidence in results is low, but the multimodal outcome battery and sham-control design make it a useful feasibility template for a larger trial.

██████████ 0.8 sleep-circadian-psychiatry Preprint

ReadSaveConnections

A PRISMA-guided review of 284 papers narrowed to 45 non-social-media free-text datasets for mental health detection, finding they are heavily skewed toward English, toward depression over other conditions, and toward narrow demographic groups. This infrastructure audit matters because models trained on these datasets are regularly used to make claims about generalizable mental health detection, yet the dataset landscape systematically underrepresents most of the world's population and most diagnostic categories. The accompanying GitHub repository provides a reusable catalogue for researchers designing future data collection efforts.

██████████ 0.8 depression-biomarkers Preprint

ReadSaveConnections

Researchers identified units in vision-language models that activate specifically during reward-anticipatory scenarios — analogous to nucleus accumbens neurons — and showed that selectively perturbing these units shifts the model toward choosing low-effort, low-reward options, mimicking anhedonia, without impairing general reasoning. The method uses activation patching: targeted silencing of specific network components to establish causal rather than correlational involvement. While translating AI model dissection to human neuroscience requires caution, the publicly released code and clinical task battery (adapted from real anhedonia scales) make this a reproducible platform for computational theories of reward processing.

██████████ 0.8 computational-psychiatry Preprint

ReadSaveConnections

This paper proposes routing social media users' posts to specialized expert modules based on weak semantic signals extracted offline by a large language model, addressing the observation that depressed individuals express distress in heterogeneous ways that single classifiers miss. Tested across Chinese and English datasets totalling nearly a million posts, the mixture-of-experts approach outperforms monolithic baselines, with the routing logic grounded in PHQ-9 clinical criteria. The system's practical limitation is that training-time routing uses a proprietary API whose outputs may drift, making long-term deployment stability uncertain.

██████████ 0.8 depression-biomarkers Preprint

ReadSaveConnections