Skip to content
krooksn.dev/ krx

Research

Systems close to the machine.

I work around the point where model ideas meet practical systems: memory limits, local hardware, agent workflows, and the tooling needed to study them.

The work moves between papers, implementation probes, and research tools, with a bias toward things that can be tested on real machines.

Latest paper

The block boundary was part of the mechanism.

On one pinned Qwen3-Coder AWQ fixture, a two-route exchange across neighboring blocks transferred the complete four-surface trajectory. A within-block permutation preserved each native trajectory.

The public package records the 70-process mechanism matrix, schedule geometry, and seven rejected evidence mutations. External prevalence and practical impact remain open.

Read the paper note

Resident KV arc

Future reuse is an obligation boundary.

Resident KV Claims is a research line about when future KV reuse becomes a runtime responsibility, not just a cache hint. The public paper defines the contract: accepted future-reuse state needs claim identity, a materialization predicate, ordered lifecycle events, and claim-scoped outcomes.

Follow-on work is still in calibration. The next public bridge is a narrow vLLM observability note about prefix-cache lookup debugging. The lowering work asks how existing runtime primitives can satisfy those obligations without turning useful substrates into false conformance claims.

Not claimed: production speedup, native vLLM support, native TensorRT-LLM/SGLang/Dynamo conformance, or a verdict that existing runtimes lack useful KV machinery.

Earlier research

Open questions

The thread keeps moving.

  • Where else does route alignment cross a kernel schedule boundary that changes exact output?
  • Which future-KV obligations are too strong, too weak, or already covered by runtime paths I missed?
  • What can local constraints reveal about useful model systems?
  • How can agent workflows become more reliable without becoming opaque?