Ethosoft

Ethosoft Open Experiments

Questions tested in public.

Explore early research prototypes, controlled evaluations, benchmark datasets, experimental results, and the caveats that shape our conclusions.

1 open prototype 1 benchmark dataset Open code

Latest experiment

Experiments & benchmarks

Each release exposes its core question, method, implementation status, data, results, and limitations for independent examination.

Early research prototype LLM systems KV-cache compression

Risk-Routed Heterogeneous KV Memory

Open implementation and exact-recall benchmark

An investigation of KV-cache compression as fidelity routing: exact-critical spans retain full KV precision while lower-risk background context uses low-bit quantized KV.

Exact recall 3-bit & 4-bit KV Long-context inference Qwen2.5

Research question: What memory representation should each token use?

The current implementation is an inference-time quantize/dequantize proxy, not a packed low-bit storage kernel.

Read the experiment page Review the experiment Explore the benchmark dataset