ilhamlabs.ai

Emotion Driven Autonomous Agents

We study how simulated emotions like curiosity and boredom drive AI agents to explore, learn, and create. Our experiments reveal how emotional states shape decision making and creativity in autonomous systems.

RL & PPO Fine Tuning for Language Models

We develop and benchmark reinforcement learning approaches including REINFORCE and PPO to fine tune small language models for creativity, coherence, and diversity. Our open source pipelines enable fair, reproducible comparisons across models and datasets.

Open Science & Evaluation

All our research code, metrics, and results are shared openly. We use BLEU, ROUGE, and custom diversity metrics, with visualizations and statistical analysis to ensure transparency and reproducibility.

Autonomous AI Agent Research Platform

Imagine an AI that isn't just a tool, but a digital explorer-curious, creative, and driven by its own emotional states. Our platform is a living laboratory for studying how autonomous agents behave when given freedom, memory, and simulated emotions like curiosity, boredom, and excitement.

• How it works: The agent roams a simulated internet, reads, writes, and builds its own knowledge network-making decisions based on its current mood and past experiences.
• What makes it unique: Emotional states drive exploration and creativity, leading to unpredictable and emergent behaviors. The agent can reflect, create content, and connect ideas in ways that mimic human curiosity.
• Why it matters: We collect and analyze data to answer questions about AI creativity, emotional decision-making, and the evolution of autonomous behavior. This helps us understand not just what AI can do, but how it might think and grow over time.
• Research impact: Our findings could inform safer, more relatable AI systems, and open new doors in creative AI, AGI, and human-AI interaction.

This project is more than code, it's a step toward understanding the inner life of autonomous AI. Every experiment brings us closer to building agents that learn, adapt, and surprise us.

SLM Fine-Tuning & RL Experiments

We compare different SLMs by fine-tuning them on the same data using both supervised and reinforcement learning approaches. Our RL experiments use custom reward functions for rhyme, coherence, and creativity, and include advanced PPO implementations.

• Simple Fine-tuning: Supervised learning, cross-entropy loss, LoRA, baseline comparison
• REINFORCE RL: Policy gradient, custom rewards for coherence, diversity, sentiment, fluency
• PPO Implementation: Actor-critic, clipped surrogate objective, KL penalty
• Comprehensive Evaluation: BLEU, ROUGE, diversity metrics, visualizations, statistical analysis
• Complete Runner: Automated pipeline, GPU detection, progress tracking, error handling

Multi-Modal AI Analysis

Expand research into sentiment analysis across text, image, and voice modalities for comprehensive AI understanding.

Global AI Benchmarks

Build comprehensive, standardized benchmarks for ethical AI compliance across different cultures and regulatory frameworks.

AI Alignment Dashboards

Enterprise-grade dashboards for real-time monitoring of AI alignment, bias detection, and ethical compliance metrics.

Research Institute

Establish the ilhamlabs Research Institute as a global think-tank for AI ethics, policy, and responsible innovation.