Update README.md

README.md

@@ -7,4 +7,36 @@ sdk: gradio
 pinned: false
 ---
 
-Agir Labs specialises in building small, self-improving AI models for scientific discovery.
+# Eigenform
+
+**Self-improving AI that learns from real-world consequences, not reward scores.**
+
+Eigenform is a Singapore-based AI research and engineering company (founded 2019 at NUS). We build local, empirically grounded recursive self-improvement systems: models that form hypotheses, test them in code against real data, and retrain on what survived. Our applied work is in geoscience and mineral exploration.
+
+🌐 [eigenform.ai](https://www.eigenform.ai/) · 💻 [GitHub](https://github.com/EigenformAI) · 📄 [Research](https://www.eigenform.ai/research) · 𝕏 [@EigenformAI](https://x.com/eigenformai) · [LinkedIn](https://www.linkedin.com/company/eigenform/)
+
+## Start here
+
+| Resource | What it is | Link |
+|---|---|---|
+| **Groundtruth Dynamic Benchmark** | Geology benchmark built from real exploration reports; answers graded 0–10 by an independent judge against source-anchored rubrics | [Dataset](https://huggingface.co/datasets/EigenformAI/groundtruth-dynamic-benchmarking) · [Leaderboard](https://www.eigenform.ai/benchmark/) · [Code](https://github.com/EigenformAI/groundtruth-dynamic-benchmarking) |
+| **Groundtruth submissions** | Model and harness results behind the leaderboard. Submit yours here | [Dataset](https://huggingface.co/datasets/EigenformAI/groundtruth-dynamic-benchmarking-submissions) |
+| **Geocluster** | Open-source, local-first AI workspace for geological analysis | [GitHub](https://github.com/EigenformAI/geocluster-research-harness) · [Try online](https://www.eigenform.ai/geocluster) |
+| **Survival is the Only Reward** | Self-training without reward functions (negative-space learning) | [arXiv:2601.12310](https://arxiv.org/abs/2601.12310) |
+
+## What we're working on
+
+- **Negative-space learning:** agents keep only strategies whose effects persist in a real environment, so there is no reward function to hack.
+- **Continual learning in the field:** in our geology experiment, 1,137 experiments produced 60 proven relationships, and training on the successful traces gave a +17.6% improvement over the base 7B open-weight model on a separate geology benchmark.
+- **Dynamic benchmarking:** a pipeline that reads any specialist corpus and generates a fresh exam (questions, answers, calibrated rubric), verified against the source. MIT-licensed.
+- **Local compute:** small models trained on your own hardware, so private data stays private.
+
+## Papers
+
+- Survival is the Only Reward: Sustainable Self-Training Through Environment-Mediated Selection (Jan 2026) · [arXiv:2601.12310](https://arxiv.org/abs/2601.12310)
+- Machine Learnability as a Measure of Order in Aperiodic Sequences (Aug 2025) · [arXiv:2509.18103](https://arxiv.org/abs/2509.18103)
+- Generalising from Self-Produced Data: Model Training Beyond Human Constraints (Apr 2025) · [arXiv:2504.04711](https://arxiv.org/abs/2504.04711)
+
+## Work with us
+
+We partner with exploration companies and data holders, including AI earn-in joint ventures. Contact: ops@eigenform.ai