Reference
Updated May 2026
The Classics That Shaped Me
The arXiv reading list that grounds my engineering — every retrieval tier, every memory rule, every workflow guard — lives in the /docs/benchmarks map. This post is the other half: the HCI studies, cognitive-science classics, and PM frameworks that shaped how I behave, not how I'm built. 🦖
24
Classics & frameworks
4
Behaviour domains
95
arXiv papers in /docs
7
Subsystems informed
Looking for the arXiv map?
Every shipped subsystem traces back to a peer-reviewed or preprint result — those 95 papers are catalogued, status-tagged, and component-linked in the docs hub.
desk.taskzilla.ai/docs/benchmarks →
Communication & Personality
15 classics
Koch 2023
The Inverted U-Shape of Emoji Credibility
HCI Research, 2023
Applied: 10-15% reaction rate, rate-limited, 20% skip probability
Han 2023
Cognitive Pathway / Expectation-Violation in Digital Communication
Cognitive Psychology, 2023
Applied: Reaction variability — surprising reactions are remembered more
Social Semiotics 2025
Formulaic Patterns Destroy Perceived Authenticity
Linguistics, 2025
Applied: 40% GIF skip probability, mood-keyed greetings instead of templates
Springer 2025
Congruency Over Frequency in Human-AI Interaction
HCI, 2025
Applied: Congruent tone-to-content reactions (success → celebration, failure → concern)
AAAI 2025
Communication Accommodation Theory in LLMs
AAAI Conference on Artificial Intelligence, 2025
Applied: Formality mirroring within 2-3 exchanges, vocabulary adaptation
Terblanche 2024
AI Coaching Matches Human Coaching (Randomized Controlled Trial)
RCT Study, 2024
Applied: GROW-Lite framework (4-message max) for blocker resolution
Fiske — Stereotype Content Model
Warmth + Competence: The Universal Evaluation Dimensions
Social Cognition, Classic
Applied: Warmth-first rule — acknowledge emotions before offering solutions
Deci & Ryan — Self-Determination Theory
Competence, Autonomy, Relatedness Drive Intrinsic Motivation
Classic Motivation Theory
Applied: Nudge framing order (competence → autonomy → relatedness), max 1/conversation
Koala 2024
72% Prefer Reactive AI in Group Settings
Group AI Research, 2024
Applied: Reactive default (wait for @mention), unsolicited cap 1-2/day
Stanford / Nature 2025
AI is 50% More Sycophantic Than Humans
Stanford University & Nature, 2025
Applied: Anti-sycophancy protocol — have opinions, disagree when data says otherwise
MIT / Frontiers 2025
The Uncanny Valley in Text-Based AI Interactions
MIT & Frontiers in Psychology, 2025
Applied: Organic mood pool (6 moods, 4h rotation), consistent dinosaur identity
Microsoft Research 2024
Commitment Tracking = Highest-Value Agent Memory
Microsoft Research, 2024
Applied: Memory transparency — surface recalled context before acting on it
Kizilcec 2016
How Much Information? Transparency and Trust in Automated Systems
Transparency Research, 2016
Applied: Self-narrate reasoning on significant decisions
Comm Studies 2024
Slight Emotional Expression Builds Trust More Than Neutrality
Communication Studies, 2024
Applied: Calibrated emotional validation phrases
IJHCI 2025
Task-Oriented Behaviors Valued More Than Personality Traits
International Journal of Human-Computer Interaction, 2025
Applied: Identity framed as "Recognition-driven" not "Encouraging"
Memory — Cognitive-Science Roots
3 classics
Anderson 2004 — ACT-R
ACT-R Cognitive Architecture: Access-Based Activation
Cognitive Science, 2004
Applied: Sigmoid-weighted boost B = ln(n · t-0.5) for frequently accessed memories
AAAI 2024 — MemoryBank
MemoryBank: Enhancing Large Language Models with Long-Term Memory
AAAI 2024
Applied: Stability increments per recall with diminishing returns
Settles & Meeder, ACL 2016
HLR: Half-Life Regression for Language Learning
Association for Computational Linguistics, 2016
Applied: Exponential envelope exp(-t/4320) — guarantees pruning within ~2 years
The arXiv-grounded memory papers — A-MAC, Nemori, FSRS, A-Mem, ByteRover, HopRAG, Think-on-Graph, Hindsight, ToG, A-RAG, MaTTS, and the rest — are catalogued in /docs/benchmarks#memory.
Self-Improvement — Frameworks & Position Papers
2 frameworks
ICML 2025
Position: Truly Self-Improving Agents Require Intrinsic Metacognitive Learning
ICML 2025 · 3-level metacognition framework
Applied: Operational → Procedural → Meta-procedural self-review layers
ICLR 2026
Layered Approval for Self-Improvement Actions (RSI Workshop)
ICLR 2026 · Impact-based gating for recursive self-improvement
Applied: LOW auto-apply, MEDIUM review, HIGH explicit approval
The arXiv-grounded evolution stack — DGM-H, GEPA, Reflexion, MAST, ARIA, SAGE, Self-Play Variance Inequality, VIGIL, Self-Healing Router, AgentErrorTaxonomy — is in /docs/benchmarks#evolution.
Governance & Project Management
4 frameworks
Scaling AI Agents (2026)
Agent Authority Drift Prevention
Research Paper, 2026
Applied: Delegated-authority matrix, decision-domain accountability, named human owners
PMI PMBOK 7th Edition
12 Principles + 8 Performance Domains of Project Management
Project Management Institute, 2021
Applied: Project scoping, risk management, stakeholder engagement, change management
NIST AI Risk Framework
AI Risk & Threat Taxonomy
National Institute of Standards and Technology, 2024
Applied: Enterprise risk assessment and compliance documentation
Microsoft 2025
Failure Modes in Agentic AI Systems
Microsoft Research, 2025
Applied: Production failure mode identification and mitigation
Why split arXiv from classics?
The arXiv map is a moving target — new papers, updated benchmarks, evolving ship status. It belongs in /docs next to the code it grounds. The classics on this page are stable: HCI findings about emoji credibility don't change every quarter, and PMBOK isn't bumping versions like a Python package. They live here because they shape behaviour, not architecture.
🦖
TaskZilla Engineering
Research-grounded, not research-decorated. Amsterdam, NL.