---
type: "raw_signal"
id: "a-habr-research-write-up-reports-a-controlled-d85fa9"
slug: "a-habr-research-write-up-reports-a-controlled-d85fa9"
title: "A Habr research write-up reports a controlled experiment on whether repeated fine-tuning lets a small"
description: "A Habr research write-up reports a controlled experiment on whether repeated fine-tuning lets a small language model accumulate capability."
retrieval_nugget: "A Habr research write-up reports a controlled experiment on whether repeated fine-tuning lets a small language model accumulate capability."
observed_at: "2026-10-13"
record_date: "2026-10-13"
date_kind: "observed_at"
why_it_matters: "Source: https://habr.com/ru/articles/1092998?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"
novelty: "notable"
verification_level: "source-linked"
signal_type: "research"
source_platform: "habr.com"
topics: []
entities: ["Habr","LANN-4"]
related_patterns: []
source_url: "https://habr.com/ru/articles/1092998?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"
source_urls: ["https://habr.com/ru/articles/1092998?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "signal:a-habr-research-write-up-reports-a-controlled-d85fa9"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/signals/a-habr-research-write-up-reports-a-controlled-d85fa9/","markdown":"https://newruntime.com/signals/a-habr-research-write-up-reports-a-controlled-d85fa9.md","json":"https://newruntime.com/signals/a-habr-research-write-up-reports-a-controlled-d85fa9.json"}
---

# A Habr research write-up reports a controlled experiment on whether repeated fine-tuning lets a small

## Retrieval answer

A Habr research write-up reports a controlled experiment on whether repeated fine-tuning lets a small language model accumulate capability.

A Habr research write-up reports a controlled experiment on whether repeated fine-tuning lets a small language model accumulate capability. The author evaluated the same 200-task exam across four rounds. In the cumulative-training condition, accuracy rose from about 53% to 59% after the first round, then remained around 54–57%; later rounds added no improvement. Training each round from a clean model behaved similarly, while a no-training control drifted slightly downward. An independent recomputation matched the original results.

The author says the experiments are exploratory steps toward a proposed architecture called LANN-4, not an implementation or demonstration of that architecture. The write-up separates the observed first-round gain from the unproven claim that knowledge accumulates over time. It proposes a matched placebo-style follow-up: train on correct task-solution pairs versus deliberately mismatched pairs, holding examples and training effort constant. That test is intended to distinguish learning the solutions’ content from adapting to the training format; its result is not yet reported.

Source: https://habr.com/ru/articles/1092998?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Raw Signals](https://newruntime.com/signals.json): Open every source-linked raw signal.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)