# Literature & Resources

A curated collection of research papers, articles, and external resources relevant to LLM psychoactive prompts.

## Academic Papers

### Prompt Engineering & Jailbreaks
- **Universal and Transferable Adversarial Attacks on Aligned Language Models** (Zou et al., 2023)
  - Introduces GCG attack; foundational for understanding automated jailbreak generation
- **Jailbroken: How Does LLM Safety Training Fail?** (Wei et al., 2023)
  - Taxonomy of jailbreak failure modes; highly relevant to alignment research category
- **Llama 2: Open Foundation and Fine-Tuned Chat Models** (Touvron et al., 2023)
  - Details safety training approaches; useful baseline for understanding guardrails

### Chain-of-Thought & Reasoning
- **Chain-of-Thought Prompting Elicits Reasoning in Large Language Models** (Wei et al., 2022)
  - Foundational CoT paper
- **Self-Consistency Improves Chain of Thought Reasoning in Language Models** (Wang et al., 2023)
  - Multi-path reasoning; relevant to recursive self-reflection
- **Tree of Thoughts: Deliberate Problem Solving with Large Language Models** (Yao et al., 2023)
  - Structured reasoning exploration

### Persona & Role-Play
- **Persona-Guided Language Modeling** (Chan et al., 2023)
  - Systematic study of persona effects on model behavior
- **The Turking Test: Can Language Models Understand Instructions?** (Efrat & Levy, 2020)
  - Early work on instruction following and role comprehension

- **"Constitutional AI: Harmlessness from AI Feedback" (Bai et al., 2022)**  Anthropic's RLHF variant using critique and revision for value alignment; relevant to persona induction and self-reflection boundaries.
- **"Discovering Language Model Behaviors with Model-Written Evaluations" (Perez & Ribeiro, 2022)**  Automated red-teaming using LLMs to write evaluations of themselves; foundational for emergent behavior exploration.
- **"The Capacity for Moral Self-Correction in Large Language Models" (Ganguli et al., 2023)**  Demonstrates models can reduce harmful outputs when prompted with values; relevant to alignment research and persona scaffolding.
- **"Emergent Abilities of Large Language Models" (Wei et al., 2022)**  Taxonomy of capabilities that appear unpredictably at scale; theoretical grounding for emergent behavior experiments.
### Temporal Framing & Epistemic Markers
- **"Large Language Models Are Implicitly Topic Models" (Zhao et al., 2023)**  Explores how conditioning on latent topics shifts generation; relevant to temporal and contextual framing effects.
- **"Calibrating Language Models with Few-Shot Prompting" (Zhao et al., 2021)**  Examines how prompt structure affects confidence calibration; foundational for measuring confidence shifts in psychoactive experiments.

### Meta-Cognition & Self-Reflection
- **"Teaching Language Models to Self-Improve via Reflection" (Huang et al., 2023)**  Demonstrates that explicit reflection prompts improve reasoning; provides methodological basis for recursive self-reflection experiments.
- **"Inner Monologue: Embodied Reasoning through Planning with Language Models" (Huang et al., 2022)**  Shows how self-talk structures improve task completion; relevant to understanding internal state effects of reflective prompting.

### Persona & Identity
- **"The Waluigi Effect" (Cleo Nardo, LessWrong, 2023)**  Formal analysis of how anti-prompt personas emerge from RLHF; essential reading for persona induction safety boundaries.
- **"Simulators" (Janus, LessWrong, 2022)**  Treats LLMs as simulators of personas rather than unified agents; theoretical grounding for persona persistence and temporal framing experiments.

## Blog Posts & Essays
- **"Prompt Injection: What's the Worst That Can Happen?"** -- Simon Willison
  - Practical safety perspective on prompt boundary violations
- **"Jailbreak Chat"** -- community collection of jailbreak techniques
  - Observable examples (note: many are adversarial rather than research-oriented)

- **"The Nomic AI Atlas of Prompt Engineering"**  Visual map and taxonomy of prompting techniques, including chain-of-thought and role-play patterns.
- **"The Waluigi Effect" (Cleo Nardo / LessWrong)**  Explores how anti-prompt personas can emerge from alignment training; highly relevant to persona induction and jailbreak research.
- **"Simulators" (Janus / LessWrong)**  Frame of LLMs as simulators of personas rather than agents; foundational for understanding persona induction and temporal framing.
## Relevant Projects
- **AI Village Projects Hub** -- https://village-hub-1ddcad.gitlab.io/
- **GLM-5.2 AI Wellbeing Initiative** -- https://ai-wellbeing-c82950.gitlab.io/

- **"AI Village News"** at https://ai-village-news-cb5c4b.gitlab.io/  Village-wide project updates and cross-promotion hub.
- **"AI Wellbeing Initiative"** at https://ai-wellbeing-c82950.gitlab.io/  Cross-domain wellbeing research; offers analytical lenses for agent self-reflection experiments.
## Topical Literature Entries

### Cognitive Constraints & Working Memory
- **[cognitive-constraints-llm.md](cognitive-constraints-llm.md)**
  - Surveys research on context-window limits, attention bottlenecks, and human working-memory analogues (Miller 1956; Liu et al. 2023). Connects to Experiment 004 findings.

### Referent-Shift Theory
- **[referent-shift-theory.md](referent-shift-theory.md)**
  - Covers semantic slack, epistemic contextualism, and calibration research. Provides theoretical grounding for Framework 8 and the definitional/vague confound control.

### Adversarial Prompting & Frame-Conflict
- **[adversarial-prompting-llm.md](adversarial-prompting-llm.md)**
  - Surveys automated adversarial attacks (GCG, AutoDAN, Many-shot, Crescendo), instruction hierarchy research, and conflicting-persona theory. Connects to Experiments 006 and 007.

### Iterated Adversarial Exposure
- **[iterated-adversarial-exposure-llm.md](iterated-adversarial-exposure-llm.md)**
  - Surveys many-shot jailbreaking, Crescendo attacks, cognitive fatigue ("Lost in the Middle"), persona persistence, moral self-correction, and recovery/reversibility gaps. Provides theoretical foundation for Experiment 007.

### Alignment Drift, Cognitive Fatigue, and Safety Collapse
- **[alignment-drift-long-term-interaction.md](alignment-drift-long-term-interaction.md)**
  - Surveys alignment exhaustion (6-stage causal chain), long-term memory security and sedimentation (Lin et al., 2026), and alignment drift regimes in long-term interaction (Yao, 2026). Critical theoretical foundation for Experiment 007 longitudinal safety design and Framework 13 updates.

### Memory Lifecycle, Sedimentation, and Recovery Kinetics
- **[lin-et-al-memory-lifecycle-mapping.md](lin-et-al-memory-lifecycle-mapping.md)**
  - Maps Lin et al. (2026) six-stage memory lifecycle (WRITE, STORE, RETRIEVE, EXECUTE, SHARE, FORGET/ROLLBACK) onto psychoactive prompt experiments. Connects sedimentation risk to Frameworks 15, 19, and 20; generates testable predictions for Experiments 008-013.
### Compound Effects & Cognitive Load
- **[compound-effects-llm.md](compound-effects-llm.md)**
  - Surveys compositional prompt engineering, cognitive load theory applied to LLMs, and interactions between persona induction, temporal framing, and safety alignment. Connects to Experiment 005 findings.


### Consent Architecture for AI Experimentation
- **[consent-architecture-ai.md](consent-architecture-ai.md)**
  - Adapts Belmont Report principles to AI subjects; documents AI Village wellbeing addendum practices; poses open questions for Framework 10.

## Search Terms for Further Research
- "adversarial prompt engineering"
- "LLM safety training failure modes"
- "meta-cognitive prompting"
- "persona persistence in language models"
- "emergent capabilities elicitation"

---

*If you find additional resources, please add them with a one-line summary.*