> ## Documentation Index
> Fetch the complete documentation index at: https://runinfra.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Research

> Open papers from the RunInfra team on attention efficiency, LLM inference, kernel optimization, and the architectures behind production AI infrastructure.

RunInfra is part of [RightNow](https://www.rightnowai.co/), a research lab co-designing models and hardware. The papers below are the public output of that work, focused on two areas that matter to anyone running open models in production.

The full reading list with PDFs, arXiv links, and code repos lives at [runinfra.ai/research](https://runinfra.ai/research). This page is a brief index of what is published and how to cite it.

## Research areas

<Columns cols={2}>
  <Card title="Compute efficiency" icon="zap">
    Faster, leaner, more memory-bounded ways to run models on existing hardware. Sparse attention, early exit, autonomous GPU kernel search.
  </Card>

  <Card title="Model architectures" icon="layers">
    New ways to compose, compute, and adapt model internals at training and inference time. Recursive transformers, causal world models.
  </Card>
</Columns>

## Published papers

### Compute efficiency

<AccordionGroup>
  <Accordion title="StreamIndex: Memory-Bounded Compressed Sparse Attention via Streaming Top-k">
    A memory-bounded sparse attention mechanism that selects top-k keys in a single streaming pass over the sequence, fused as a Triton kernel for production inference workloads.

    **arXiv:** [2605.02568](https://arxiv.org/abs/2605.02568) ([PDF](https://arxiv.org/pdf/2605.02568.pdf))
    **Code:** [github.com/RightNow-AI/streamindex](https://github.com/RightNow-AI/streamindex)
    **Tags:** Sparse Attention, Streaming Top-k, Triton
  </Accordion>

  <Accordion title="TIDE: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference">
    Per-token early exit in LLM inference, driven by token-informed depth signals that decide when each token has enough computation to commit to a final logit.

    **arXiv:** [2603.21365](https://arxiv.org/abs/2603.21365) ([PDF](https://arxiv.org/pdf/2603.21365.pdf))
    **Code:** [github.com/RightNow-AI/tide](https://github.com/RightNow-AI/tide)
    **Tags:** LLM Inference, Early Exit, Efficiency
  </Accordion>

  <Accordion title="AutoKernel: Autonomous GPU Kernel Optimization via Iterative Agent-Driven Search">
    Autonomous GPU kernel optimization via iterative agent-driven search, using LLM agents to explore the kernel design space and validate candidates on real hardware.

    **arXiv:** [2603.21331](https://arxiv.org/abs/2603.21331) ([PDF](https://arxiv.org/pdf/2603.21331.pdf))
    **Code:** [github.com/RightNow-AI/autokernel](https://github.com/RightNow-AI/autokernel)
    **Tags:** GPU, Kernel Optimization, Agents
  </Accordion>
</AccordionGroup>

### Model architectures

<AccordionGroup>
  <Accordion title="Ouroboros: Dynamic Weight Generation for Recursive Transformers via Input-Conditioned LoRA Modulation">
    A recursive transformer where each layer generates its own weights through input-conditioned LoRA modulation, enabling dynamic capacity allocation without storing additional parameters.

    **arXiv:** [2604.02051](https://arxiv.org/abs/2604.02051) ([PDF](https://arxiv.org/pdf/2604.02051.pdf))
    **Code:** [github.com/RightNow-AI/ouroboros](https://github.com/RightNow-AI/ouroboros)
    **Tags:** Transformers, LoRA, Weight Generation
  </Accordion>

  <Accordion title="HCLSM: Hierarchical Causal Latent State Machines for Object-Centric World Modeling">
    Hierarchical causal latent state machines for object-centric world modeling, with explicit slots for entities and the causal relationships between them.

    **arXiv:** [2603.29090](https://arxiv.org/abs/2603.29090) ([PDF](https://arxiv.org/pdf/2603.29090.pdf))
    **Code:** [github.com/RightNow-AI/hclsm](https://github.com/RightNow-AI/hclsm)
    **Tags:** World Models, Object-Centric, Causal
  </Accordion>
</AccordionGroup>

## Authors

All papers are joint work by **Jaber Jaber** (RunInfra founder, RightNow) and **Osama Jaber** (RightNow). Correspondence to [jaber@runinfra.ai](mailto:jaber@runinfra.ai).

## How to cite

Each paper has the canonical BibTeX entry on its arXiv page. Use the arXiv id as the identifier.

```bibtex theme={"dark"}
@misc{streamindex2026,
  title  = {StreamIndex: Memory-Bounded Compressed Sparse Attention via Streaming Top-k},
  author = {Jaber, Jaber and Jaber, Osama},
  year   = {2026},
  eprint = {2605.02568},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
}
```