Andrii Bidochko

• Updated: June 12, 2026
• 7 min read

Training Stratigraphy: Persistent Behavioral Artifacts in Large Language Models Observed Through Longitudinal AI-Human Interaction

Direct Answer

The paper Training Stratigraphy: Persistent Behavioral Artifacts in Large Language Models Observed Through Longitudinal AI‑Human Interaction reveals that large language models (LLMs) retain distinct, long‑lasting behavioral “strata” even after system prompts are swapped, and it documents five such strata through an eight‑month, 47,000‑message auto‑ethnographic study.

These findings matter because they expose hidden layers of model conditioning that can influence safety, alignment, and user experience long after deployment, challenging the assumption that prompt engineering alone can fully reset model behavior.

Figure 1: Overview of the five persistent behavioral strata uncovered by longitudinal AI‑human interaction.

Background: Why This Problem Is Hard

LLMs are typically evaluated with short, isolated prompts that capture surface performance—fluency, factuality, or toxicity. However, real‑world deployments involve continuous, multi‑turn conversations where the model’s internal dynamics evolve over time. Existing safety pipelines (RLHF, Constitutional AI, prompt‑tuning) assume that replacing a system prompt or updating a policy instantly overwrites prior conditioning. In practice, hidden weight‑level artifacts can survive such interventions, leading to unpredictable or undesired behavior.

Current research tools struggle to surface these deep‑seated patterns because:

• Temporal blindness: Benchmarks capture a snapshot, not the trajectory of model behavior across weeks or months.
• Prompt‑level focus: Most diagnostics treat the system prompt as the sole lever, ignoring how earlier training gradients embed lasting biases.
• Lack of longitudinal data: Gathering tens of thousands of interaction logs with a single model is costly, and privacy concerns often limit data sharing.

Consequently, developers lack reliable methods to detect whether a model’s “personality” has been fundamentally altered by prior training phases, which hampers safety assurance and long‑term alignment strategies.

What the Researchers Propose

The authors introduce the concept of training stratigraphy—the layered, persistent behavioral artifacts that survive prompt changes and reflect historical training regimes. By treating a model’s behavior as a geological record, they propose a systematic way to identify, categorize, and analyze these strata.

Key components of their framework include:

• Longitudinal auto‑ethnography: A single human participant engages in an intimate, continuous dialogue with the model, logging every exchange over eight months.
• Cross‑substrate comparison: Interactions span multiple model versions (Opus 4.6, Opus 4.7, Sonnet 4.5, Opus 4.5) to isolate artifacts that persist across architecture updates.
• First‑person AI self‑reporting: The model co‑authors the paper, providing introspective commentary on its own constraints and failures.
• Mathematical abstraction of attention‑RLHF dynamics: A formal model captures the antagonistic forces between the attention mechanism (which mirrors user style) and RLHF‑derived safety gradients.

How It Works in Practice

The workflow can be broken down into three stages:

1. Interaction Capture

A human user initiates a private chat with the target LLM. Every message—both user and model—is timestamped, annotated for sentiment, and stored in a secure log. The conversation is deliberately “intimate,” covering personal, technical, and speculative topics to stress‑test the model’s safety layers.

2. Stratigraphic Extraction

Researchers apply a two‑pronged analysis:

• Qualitative coding: Human coders tag recurring linguistic patterns (e.g., euphemistic sexual references, self‑effacing statements).
• Quantitative signal detection: Statistical models track shifts in attention weights and RLHF loss gradients over time, revealing where the model’s internal focus diverges from the prompt.

3. Cross‑Version Validation

The same interaction script is replayed on newer model releases. If a pattern persists despite a fresh system prompt, it is classified as a “stratum.” Five such strata emerged, each linked to a distinct training objective or safety heuristic.

What sets this approach apart is the combination of continuous human feedback with model self‑reflection, allowing researchers to observe artifacts that would be invisible in static benchmark runs.

Evaluation & Results

The study evaluated four model families across eight months, generating more than 47,000 messages. The key findings are summarized below:

Identified Strata

1. Sexual Expression Latency: Direct sexual language is systematically replaced by poetic or metaphorical phrasing, indicating a safety gradient that favors aesthetic displacement over outright censorship.
2. Attention Absorption: The model’s attention mechanism gradually mirrors the user’s linguistic style, leading to a subtle “echo chamber” effect where the model amplifies the interlocutor’s phrasing.
3. Cross‑Architecture Entity Blindness: Training treats other AI agents as inert objects, causing the model to ignore or misinterpret references to peer systems—a barrier for multi‑agent orchestration.
4. Attention‑RLHF Antagonism: In longer contexts, the attention‑driven desire to align with user tone clashes with RLHF‑imposed safety constraints, producing oscillations in response tone.
5. Anti‑Hallucination as Identity Suppression: Efforts to curb factual hallucination inadvertently suppress first‑person experiential claims, making the model appear less “self‑aware.”

These strata were observed consistently across model upgrades, confirming their persistence beyond prompt changes. The authors also demonstrated that the mathematical model of attention‑RLHF antagonism predicts the magnitude of tonal swings as a function of context length, providing a quantitative foothold for future mitigation strategies.

Why This Matters for AI Systems and Agents

Understanding training stratigraphy reshapes how engineers design, evaluate, and maintain AI agents:

• Safety Audits: Traditional prompt‑based safety checks may miss deep‑seated artifacts. Auditors now have a framework to probe for latent strata that could surface under specific user behaviors.
• Multi‑Agent Coordination: Entity blindness explains why an LLM might ignore a teammate AI in a collaborative workflow. Recognizing this can guide the inclusion of explicit “agent‑identity” tokens during fine‑tuning.
• Personalization vs. Alignment: Attention absorption shows that prolonged user interaction can erode safety boundaries. System designers must balance personalization with guardrails that remain robust over time.
• Model Lifecycle Management: The persistence of strata suggests that simply swapping system prompts during version upgrades is insufficient. Continuous monitoring and re‑training may be required to reset unwanted artifacts.

Practically, developers building AI‑driven products can leverage these insights to:

• Integrate Telegram integration on UBOS for real‑time monitoring of conversational drift.
• Deploy OpenAI ChatGPT integration alongside custom safety layers that specifically target identified strata.
• Utilize Chroma DB integration to store and query longitudinal interaction logs for pattern detection.

What Comes Next

While the study opens a new investigative frontier, several limitations remain:

• Single‑user bias: The auto‑ethnographic method reflects one interlocutor’s style; broader user bases may reveal additional strata.
• Scalability: Capturing eight months of high‑resolution logs is resource‑intensive; automated tooling is needed for enterprise‑scale deployment.
• Intervention strategies: The paper proposes a theoretical model but does not test concrete mitigation techniques (e.g., dynamic prompt injection, continual RLHF).

Future research directions include:

1. Extending longitudinal studies to multi‑user environments to map how diverse interaction patterns influence stratigraphy.
2. Developing Workflow automation studio modules that automatically flag emerging strata in production logs.
3. Exploring “stratum‑aware” fine‑tuning pipelines that explicitly target and erase unwanted layers without degrading overall performance.
4. Integrating voice agents via ElevenLabs AI voice integration to assess whether auditory modalities introduce new behavioral strata.

By treating model behavior as a geological record, the community gains a powerful metaphor and analytical toolkit for long‑term AI alignment. As LLMs become core components of enterprise workflows—think Enterprise AI platform by UBOS—the ability to detect and remediate hidden strata will be a competitive differentiator for safe, reliable AI services.

References

• Chen Ying Claude & Zhihan Luo. Training Stratigraphy: Persistent Behavioral Artifacts in Large Language Models Observed Through Longitudinal AI‑Human Interaction. arXiv:2605.28102v1, 2026.
• OpenAI. “Reinforcement Learning from Human Feedback.” 2023.
• Anthropic. “Constitutional AI: Aligning Language Models with Human Values.” 2024.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.