Skip to content

Research

Research code and data

Every study we run is published with its code, prompts, raw outputs and analysis, archived on Zenodo with a DOI, so the tables and figures can be rerun and checked. Papers are linked here as they appear.

Oct 2026 · O. Kholodniak, O. Prokhorov

Gradient reach in recurrent networks: code and training runs

How dropout, zoneout and gate initialisation change how far gradients reach back in time in LSTM and GRU networks, and whether that reach, measured early in training, predicts which runs learn long-range dependencies. Trained with rust-lstm.

Code on GitHub Data on Zenodo · 10.5281/zenodo.23285386

Oct 2026 · O. Kholodniak

Comparing IRT models and simple baselines for budget-constrained LLM evaluation

Item response models against simple estimators on public model × item correctness matrices: how well each estimates item difficulty, a new model's capability and its full-benchmark accuracy when only 0.5–80% of the items are run.

Code on GitHub Data on Zenodo · 10.5281/zenodo.23093754

Oct 2026 · O. Kholodniak

QoS-aware orchestration of local and cloud large language models in distributed information systems

Choosing per request between a small on-device model, a larger LAN server and a cloud API: latency model, routing policies, a netem testbed and the measurement data.

Code on GitHub Data on Zenodo · 10.5281/zenodo.23093412

Oct 2026 · O. Kholodniak

Detecting erroneous records in information extraction by small language models

Three small local models turn voice-assistant requests into JSON records; each record gets an error score without the reference annotation, and the highest-scoring records go to manual review.

Code on GitHub Data on Zenodo · 10.5281/zenodo.23112378

Oct 2026 · O. Kholodniak

Correctness of semantic caching of language-model answers to numerical questions

When may a cache return a stored answer to a similar math question? GSM8K answers in the cache, 10,552 GSM-Plus variations as requests, and every reuse decision scored exactly.

Code on GitHub Data on Zenodo · 10.5281/zenodo.23112397

Oct 2026 · O. Kholodniak

Does translation into English help open language models answer Ukrainian questions?

Five open models answer Ukrainian multiple-choice questions directly and through an English pivot; code, prompts and every model reply.

Code on GitHub Data on Zenodo · 10.5281/zenodo.23112388

Citing. Each record on Zenodo has citation metadata in BibTeX, APA and other formats. Cite the DOI of the version you used; the libraries have their own DOIs, for example rust-lstm.