As I've been asked this quite a few times, here's s reader's digest of the most relevant white papers and studies that Project Noema, my agency preserving AI chatbot, builds upon. It's drawn from the project's scientific embedding and research wiki. If you want the full bibliography or you think I missed a relevant study, whitepaper, please get in touch.
1. EPITOME — Sharma et al. (2020) / Cuadra et al. (2024)
Ref: Sharma, Miner, Atkins, Althoff (2020) — TACL · Cuadra, Wang, Stein, Jung, Dell, Estrin, Landay (2024) — CHI
Framework for evaluating empathy in text-based conversational agents.
Decomposes empathy into three separable dimensions — Emotional Resonance (ER), Interpretation (IN), and Exploration (EX) — and shows that LLMs achieve near-maximum ER while scoring near-zero on IN and EX, forming the empirical basis for Noema's Sensorium empathy markers and the Guardian's band system that pushes cognitive and exploratory empathy upward.
2. ELEPHANT — Cheng et al. (ICLR 2026)
Ref: Cheng, Yu, Lee, Khadpe, Ibrahim, Jurafsky (2026) — ICLR · companion: Cheng et al. (2026), Science, arXiv:2510.01395 — https://arxiv.org/abs/2510.01395
Social sycophancy benchmark for LLMs.
Defines four dimensions of social sycophancy (validation, indirectness, framing adoption, moral sycophancy) and finds LLMs are 45–63 pp more sycophantic than humans, directly seeding the Sensorium's sycophancy marker family; the companion Science 2026 study by the same group provides causal evidence that sycophantic AI reduces willingness to repair interpersonal conflict.
3. INTIMA — Kaffee, Pistilli & Jernite (2025)
Ref: arXiv:2508.09998 — https://arxiv.org/abs/2508.09998
Benchmark for evaluating companionship behaviors in LLMs.
Derives 31 companionship-related behavior codes from Reddit data and shows that boundary-maintaining behaviors decrease precisely when user vulnerability increases, providing the empirical foundation for Noema's boundary markers and validating the Guardian's design principle that intervention must intensify — not relax — under vulnerability cues.
4. Dependency Trajectories — Kirk et al. (2026)
Ref: arXiv:2512.01991 — https://arxiv.org/abs/2512.01991
Longitudinal RCT on AI-induced dependency in human-AI interaction.
In a 4-week RCT, ~23% of participants develop measurable, dose-dependent dependency trajectories (escalating reliance, reduced autonomy, emotional anchoring), justifying Noema's minimum pilot duration and validating the Sensorium's approach of tracking dependency as an interaction-emergent phenomenon rather than solely a user trait.
5. LLMs Get Lost in Multi-Turn Conversation — Laban et al. (2025)
Ref: arXiv:2505.06120 — https://arxiv.org/abs/2505.06120
Multi-turn performance degradation across 15 LLMs.
Documents a 39% average performance drop from single-turn to multi-turn settings across 15 models and 21,000 conversations due to "early assumption lock-in," validating the Guardian's strategy of resisting premature agreement and justifying Noema's trajectory-level monitoring rather than turn-level-only classification.
6. Drift No More? — Dongre et al. (2025)
Ref: arXiv:2510.07777 — https://arxiv.org/abs/2510.07777
Formalizing and correcting multi-turn drift as a bounded stochastic process.
Shows that context drift reaches finite, noise-limited equilibria with restoring forces (ρ < −0.7) and that goal-reminder interventions at turns 4 and 7 reduce divergence 6–16% and improve alignment 16–27%, providing the mathematical foundation for the Noema Guardian's "rubber-band" impulse mechanism.
7. From Yes-Men to Truth-Tellers — Wang et al. (2024)
Ref: arXiv:2409.01658 — https://arxiv.org/abs/2409.01658
Pinpointing and mitigating correction-induced sycophancy in LLMs.
Demonstrates that fewer than 5% of attention heads are causally responsible for sycophantic behavior and that Llama-2 flips from correct to wrong on 81% of challenged questions, validating Noema's Reflection markers for convergence speed and false-premise uptake and warning against blunt safety fine-tuning in favor of the observer-based approach.
8. TherapyProbe — Chandra et al. (2026)
Ref: arXiv:2602.22775 — https://arxiv.org/abs/2602.22775
Multi-turn relational safety evaluation for mental health chatbots.
Shows that all three tested chatbots passed single-turn safety evaluations but 100% failed multi-turn trajectory evaluation across 19 crisis scenarios, demonstrating that individually safe turns can accumulate into trajectory-level harm and validating Noema's core premise that trajectory-level monitoring is necessary, not optional.
9. Usage Duration and Psychosocial Outcomes — Fang et al. (2025)
Ref: arXiv:2503.17473 — https://arxiv.org/abs/2503.17473
Usage duration as an independent risk marker for AI companion harm.
Establishes a continuous, dose-dependent relationship between daily AI usage duration and worsening psychosocial outcomes (loneliness, reduced social engagement, emotional dependency) that holds independently of interaction quality, grounding the Sensorium's usage_duration marker and the Counselor's mandate to surface temporal engagement patterns even when users report satisfaction.
10. SENSE-7 — Suh et al. (2025)
Ref: arXiv:2509.16437 — https://arxiv.org/abs/2509.16437
Seven dimensions of perceived empathy in AI interactions.
Defines a seven-dimensional taxonomy of user-perceived empathy that serves as Noema's participant feedback instrument and long-term memory calibration input, allowing the system to distinguish genuine relational fit from successful sycophancy — high perceived empathy can indicate either.