Keep track of a changing world
Represent objects, locations, and relationships, then update them as events unfold. This makes earlier actions relevant to the next decision.
SerenAI
Modelling the world through language
TEXT-BASED WORLD MODELS
A world described in words. A model of what changes.
A text-based world model represents an environment through language: what exists, where things are, what an agent can do, and what may happen next.
SerenAI is our research into this idea. We explore state transitions, constraints, and cause-and-effect reasoning so that a sequence of messages can describe a coherent, changing situation.
Read the SerenAI paper
Anticipate the effect of an action
Estimate how a described state could change after an action. Textual world-model research explores predicting these changes alongside the actions available in context.
Reason through a sequence
Explore tasks with prerequisites and several steps. Text-based environments provide a way to study whether an agent can use observations to choose what to do next.
AN ILLUSTRATIVE STATE TRANSITION
Change the action. Follow the consequence.
A simple warehouse scenario. Choose an action to see how the stated rule affects the next state.
CURRENT STATE
Crate A is in Zone 1. The gate to Zone 2 is closed. A crate can pass only through an open gate.
CHOOSE AN ACTION
NEXT STATEThe gate is now open.
Crate A remains in Zone 1. Moving it to Zone 2 is now possible as a separate next action.
These capabilities depend on the training and the environment represented. A predicted state is a hypothesis to evaluate, not a guarantee about the physical world.
What the benchmark measures
SerenAI v1 is evaluated against its base model to show how cause-and-effect training changes the model behavior in environment-aware text reasoning tasks.
Comparison
SerenAI vs Base Model
Model type
Text-based world AI model
Training signal
Cause-and-effect data
Primary skill
Environment and situation understanding
World Models
Beyond the next token
What should an intelligent system know before it acts? A fluent answer is only the beginning. Our interest is in systems whose predictions can be checked against a changing environment: what changed, what stayed the same, and whether an action brought a goal closer.
Give prediction an action to consider.
A useful starting point is a model with two inputs: a representation of the current situation and a possible action. Its prediction describes how the situation might change. Ha and Schmidhuber demonstrated this in 2018, including a controller trained in a learned VizDoom environment and transferred back to the original game. The model approximated that environment; it did not reproduce reality in full.
Ha & Schmidhuber · World Models · 201801
Current state
What the system knows about the environment.
02
Candidate action
An intervention the agent could make.
03
Predicted state
A possible outcome to evaluate.
Imagining and deciding are different jobs.
Dreamer separates a world model, an actor, and a critic. The model predicts potential outcomes, the critic estimates their value, and the actor learns which actions to take. Its 2025 Nature paper reports results across more than 150 tasks with fixed hyperparameters, including collecting diamonds in Minecraft without human demonstrations. These are results in the evaluated environments, not evidence that every real-world task is solved.
Hafner et al. · Dreamer · Nature, 2025- 01Model: imagine outcomes
- 02Critic: estimate value
- 03Actor: choose actions
Different routes to a predictive world.
The field includes models for planning, visual representations, interactive environments, and structured text. These studies investigate different capabilities; their results are not a single performance leaderboard.
2020GOOGLE DEEPMIND
MuZero
Learn what matters for planning.
MuZero learned a model useful for search in Go, chess, shogi, and Atari without being supplied their transition rules. It focused on reward, value, and policy instead of reconstructing every detail of the environment. Its results show how a selective representation can support decisions.
Read the research2021AMMANABROLU & RIEDL
Textual world models
Represent a world as relationships.
This work predicts changes to a knowledge graph and the actions available in a text-game environment. A world can therefore be represented through entities and their relationships, without a photorealistic renderer. The study offers a concrete reference for investigating state transitions expressed through language.
Read the paper2025META FAIR
V-JEPA 2
Predict in a representation space.
V-JEPA 2 learns predictive video representations. Meta then added action-conditioned training using 62 hours of robot data, following large-scale video pretraining. The resulting system supported planning for reaching, picking, and placing in new settings. The 62 hours describe the additional robot stage, not its total training data.
Explore the method2025GOOGLE DEEPMIND
Genie 3
Interact with a generated environment.
Genie 3 supports navigation and prompted events, such as changing weather or adding objects. DeepMind also tested goal-directed navigation with its SIMA agent. The release described limits in action range, interactions between agents, and geographic accuracy. A convincing visual world still needs evaluation as a simulator.
Read the research previewWhere could this become useful?
We see opportunities to investigate these ideas across software and physical systems. The scenarios below are proposed research questions, not claims of deployed capabilities.
What changes after an agent updates a record?
Consider a purchase request with a budget, an approval status, and a supplier. A model could anticipate which fields should change after approval and which conditions must still hold. Evaluation would compare its prediction with the actual application state.
State tracking · ConstraintsWhat happens if an intervention is delayed?
Imagine a machine with rising temperature and a scheduled inspection. A research system could compare candidate interventions under explicit assumptions about operating conditions. Useful predictions would need validation with measurements, including cases outside the training distribution.
Scenarios · MeasurementsWhich observation would reduce uncertainty?
For a terrain survey, the next useful action might be another observation. A model could compare candidate viewpoints and identify what remains unknown. The research question is whether this improves decisions compared with a simpler survey strategy.
Planning · InformationA plausible future is not yet a verified one.
Our proposed evaluation framework asks for observable evidence. These questions would help distinguish a model that describes a sensible story from one that reliably tracks a particular environment.
| Question | Example test | Evidence to report |
|---|---|---|
| Does the state remain consistent? | Change one object; check that unrelated facts remain unchanged. | Transition accuracy and constraint violations. |
| Does the action make a difference? | Keep the starting state fixed and compare two different actions. | Correct differences between the predicted outcomes. |
| Can it handle longer sequences? | Evaluate short and long action sequences on held-out scenarios. | Error as a function of the prediction horizon. |
| Does it recognise missing information? | Hide a necessary observation or introduce an unfamiliar condition. | Uncertainty calibration and requests for more evidence. |
For SerenAI, the question is concrete: can a text-based system track an environment, respect its constraints, and predict the effects of actions? The interactive example and benchmark above are starting points for that discussion. Broad claims about intelligence require much broader evidence.
Discuss a research collaborationAI Detection Systems
Hallucination Risk in Large Language Models
“Recent studies have shown that the percentage of hallucinated content is quite high among popular LLMs, ranging from 17% to 19% up to 45% of the content. If left without serious attention and the appropriate corrections, AI hallucinations can lead to critical limitations of AI applications that negatively impact human civilization and its progress.”
Environmental Efficiency and Ethical AI Usage
Research published in Nature demonstrates that AI systems can contribute to a significant reduction in CO2 emissions. This environmental benefit represents a crucial advancement in sustainable technology deployment.
We strongly advocate for the responsible use of AI-generated content. Our tools are designed to assist researchers and writers in refining their original work, rather than replacing human creativity and critical thinking.
Panda v3.0 Detection Model
Advanced neural architecture optimized for LLM-generated content detection.
Core Detection Features
Perplexity Analysis94%
Burstiness Score91%
Entropy Mapping88%
N-Gram Frequency86%
Semantic Coherence92%
Stylometric Fingerprint85%
Token Probability90%
Accuracy
0.95
F1 Score
0.96
Backed by Research
Large Language Models Detection ResearchNortheastern University - November 2025
Comparative Study: AI Detection MethodsUniversity of Maryland - October 2025
GPT-4ClaudeGeminiReasoning AIKimi AILlamaPhi-4Others
Geotechnology
AI-Powered Geological Monitoring
We integrate artificial intelligence with autonomous drone systems to monitor landslides, seismic activity, and terrain instability, delivering real-time insights faster and more affordably than traditional methods.
Our geotechnology capabilities include AI-powered landslide detection, continuous seismic monitoring, high-resolution 3D terrain mapping, automated risk assessment, and scalable regional monitoring networks.
Landslide DetectionSeismic Monitoring3D Terrain MappingRisk AssessmentRegional Monitoring
