Research at Senva
Senva Labs
We study how language models and AI agents behave when you look closely. Papers are published in English, each with a German summary.
Papers
All papers
English, with a German summary.
Engineered Boredom: Do Intrinsic-Motivation Schedulers Make LLM Research Agents Explore Better?
Our 2025 research agent claimed curiosity, pain and boredom, yet its boredom drive ordered the same A* demo three times in a row; here is a fair test of drive-based scheduling.
Learning to Probe: PPO-Trained LLM Agents on OWASP Juice Shop with Verifiable Rewards
Our 2025 pentest-agent reward never saw the model's own output. Here is the fix, the errata, and a pre-registered plan to test the claim honestly.
Whereof One Cannot Speak: LLM Compliance and Abstention in a Tractarian Micro-Language
Does a language model stay inside a tiny formal language, and does it go silent when a question falls outside it? A clean design for a question regulated industries actually have.
A Parliament of Personas: Does Multi-Persona LLM Deliberation Add Measurable Value?
Our 2025 Jungian-archetype dialogue system rated itself perfectly integrated while three personas said the same thing, a warning for anyone using red-team prompting in decisions.
Did the Mice Collapse, or Did We Tell Them To? Prompt Priming in LLM-Driven Agent-Based Models
Our 2025 rebuild of Calhoun's mouse utopia produced his pathologies on cue because the prompt and the rules wrote them in; a pre-registered re-run separates priming from emergence.
Synthetic Frustration: Appraisal-Based Affect as a Control Signal for Long-Running LLM Agents
An audit of our 2025 “suffering AI” posts: the distress reports were our own templates, the creativity gain was built into the metric, and what remains is a useful idea for agent control.
Stocks, Flows and Tipping Points: A System-Dynamics Toy of Anxiety-Addiction Feedback
Our 2024 anxiety-addiction simulation recovered from every crisis because a rule said so. We restate it as stocks and flows and design a test for early warning before a system tips.
Plasticity on a Grid: A Toy Model of Post-Lesion Compensation, and What It Cannot Tell Us
Our 2024 brain-healing simulation recovered half of every lesion on the first step, stimulation or not. A short audit, a corrected reading of the neuroscience, and the limits of the toy.
The four statuses
The status appears on every card, before the name of the series.
- Exploratory
- A first experiment. Interesting, but not repeated.
- Replication planned
- We know how we want to repeat the experiment, and we intend to.
- Replicated
- The experiment was repeated, and the result held.
- Essay
- A line of reasoning without an experiment of its own.
Four series
Every paper belongs to one of the four series.
- Synthetic Minds
- Artificial agents with internal states such as curiosity or boredom, and what those states change in their behaviour.
- Simulated Societies
- Many LLM agents in a simulated society. The question behind it: what emerges from the simulation, and what was already in the prompt?
- Machines & Meaning
- Language, meaning and their limits. Philosophical questions turned into experimental setups for language models.
- Offensive AI
- Language models in security testing, only ever in isolated test environments.
Why
Why a software company runs a lab
Anyone who recommends AI to companies should be able to show what the recommendation is based on. Our experiments are small and openly documented: one question, one setup, the code, the result and a list of what the result does not prove.
The first texts appeared in 2024 and 2025 as blog posts. In 2026 we revised them, because some of them claimed more than the code supported. What changed and why is recorded in each paper's changelog.
Method
How to recognise a Senva paper
- Claims box
- Every paper opens with what it shows and what it doesn't.
- Code
- Where a repository is public, it is linked.
- Seeds
- Random seeds are fixed and stated, so that a run can be repeated.
- Errata
- We correct mistakes visibly and with a date, never silently.
- Status
- Every paper carries a status, so you know how much weight it can bear.
- What it means for companies
- Every paper ends with a section on what the result means for use in a company.
Lab and project
What the lab has to do with your project
We use the lab's methods in client projects: test sets with unanswerable questions, control conditions that isolate the effect of the prompt, fixed seeds for repeatable tests.
Questions
Talk to us about a paper
If you have questions or objections, or spot a mistake we missed, write to Kim Pham. We add corrections to the changelog, with thanks and a date.
- info@senva.de
- Phone
- +49 30 5268 7541