Work

  1. Seetrex Warden: an external obligation gate for a long-horizon LLM agent. A pre-registered pilot (n = 1 per arm) on the Ground Control scenario of CivBench (arXiv:2609.02459)

    An MCP proxy that keeps a ledger of obligations with due dates and rejects end_turn while one is overdue; it never says what move to make. Pilot on CivBench, one game per arm.

    Post Code Data Product

  2. Laimark: Gains and Structural Limits of Self-Generated Curricula in Reinforcement Learning from Verifiable Reward

    On HumanEval with Qwen3-8B, a problem set generated by the model itself gives measurable gains under GRPO and hits structural limits; the paper says what those limits are.

    Code

  3. Monolith: Differentiable EML Trees for Symbolic Regression via Gradient Descent on a Universal Operator

    EML trees as trainable PyTorch modules: recovers 7/7 elementary functions with ≤24 parameters by gradient descent on one universal operator, with baselines against PySR and an MLP.

    Code

  4. The Geometric Blind Spot of Perplexity: When Low Loss Hides Out-of-Distribution

    Perplexity misses out-of-distribution inputs where the model is fluent but the task does not match; a geometric view of the hidden states catches them (LLaMA-3-8B, Mistral-7B).

    Code

  5. KuraFormer: Oscillatory Dynamics as Parameter-Efficient Adapters for Pretrained Transformers

    Version 2 corrects the claims of the original preprint: a static-adapter ablation shows the Kuramoto oscillatory dynamics add nothing over a bottleneck adapter (54.1 % vs 53.9 %).

    Code

  6. Runtime Compute Control: Explicit Inference-Time Computation Management for Large Language Models

    A post-training execution layer that exposes one compute-ratio parameter to control which paths run at inference, validated on Qwen2.5-3B with a C++ runtime.

Projects

Contact

jesus@tabares.eu. ORCID 0009-0001-5686-6765. Code at github.com/seetrex-ai.