---
type: "post"
id: "a-habr-research-write-up-reports-a-controlled-b48dcc"
slug: "a-habr-research-write-up-reports-a-controlled-b48dcc"
title: "A Habr research write-up reports a controlled experiment on whether repeated fine-tuning lets a small"
description: "Thesis: Repeated fine-tuning in this reported small-model experiment produced a first-round gain, not evidence of continuing cumulative learning."
retrieval_nugget: "Thesis: Repeated fine-tuning in this reported small-model experiment produced a first-round gain, not evidence of continuing cumulative learning."
published_at: "2026-10-13"
updated_at: "2026-10-13"
record_date: "2026-10-13"
date_kind: "published_at"
topics: []
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: "post:a-habr-research-write-up-reports-a-controlled-b48dcc"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:bdd5eb4b4be9be94acfe079871e6f8bb264e4c2194c1326cec9e9cc8dd0ff724","reviewed_at":"2026-10-13T14:08:37.162Z","source_evidence_count":1,"verified_claim_count":2,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":2,"implication_count":2,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Repeated fine-tuning in this reported small-model experiment produced a first-round gain, not evidence of continuing cumulative learning.","observed_facts":[{"text":"The same 200-task exam was used across four rounds; cumulative training rose from about 53% to 59% after round one and then stayed around 54–57%.","source_urls":["https://habr.com/ru/articles/1092998?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]},{"text":"Training from a clean model each round behaved similarly to cumulative training, and an independent recomputation matched the initial calculation.","source_urls":["https://habr.com/ru/articles/1092998?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]}],"mechanism":"A proposed matched placebo-style test compares training on correct task-solution pairs with equal-effort training on deliberately mismatched pairs to distinguish content learning from adaptation to training format.","why_now":"The reported plateau makes this distinction necessary before treating an early gain as evidence that capability accumulates across rounds.","implications":["The first-round improvement should not be described as continual learning until the content-versus-format explanation is tested.","LANN-4 is a proposed architecture, not a built or validated system in this report."],"evidence_boundary":"The supplied article reports the initial experiment and proposes the placebo-style test; it does not report that test’s outcome or establish generality beyond the task type used.","watch_conditions":["Whether the matched correct-pair and mismatched-pair follow-up produces a reproducible difference."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied; this starts a distinct record focused on the reported plateau and the proposed test separating content learning from format adaptation."}
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/posts/a-habr-research-write-up-reports-a-controlled-b48dcc/","markdown":"https://newruntime.com/posts/a-habr-research-write-up-reports-a-controlled-b48dcc.md","json":"https://newruntime.com/posts/a-habr-research-write-up-reports-a-controlled-b48dcc.json"}
---

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

## Retrieval answer

Thesis: Repeated fine-tuning in this reported small-model experiment produced a first-round gain, not evidence of continuing cumulative learning.

Thesis: Repeated fine-tuning in this reported small-model experiment produced a first-round gain, not evidence of continuing cumulative learning.

The Habr write-up describes the same 200-task exam across four rounds. Observed fact: cumulative training moved accuracy from about 53% to 59% after round one, then results stayed around 54–57%. Observed fact: training from a clean model each round behaved similarly, and an independent recomputation matched the initial calculation. These observations make the result more informative than a simple before-and-after claim, while keeping it bounded to the reported setup.

Mechanism: A proposed matched placebo-style test compares training on correct task-solution pairs with equal-effort training on deliberately mismatched pairs to distinguish content learning from adaptation to training format. If correct content outperforms the placebo condition, that would support content learning; similar outcomes would leave adaptation to training format as a live explanation. Why now: The reported plateau makes this distinction necessary before treating an early gain as evidence that capability accumulates across rounds. The author says LANN-4 is not yet selected or built, so the experiment is not a validation of that architecture.

Evidence boundary: The supplied article reports the initial experiment and proposes the placebo-style test; it does not report that test’s outcome or establish generality beyond the task type used. First implication: The first-round improvement should not be described as continual learning until the content-versus-format explanation is tested. First watch condition: Whether the matched correct-pair and mismatched-pair follow-up produces a reproducible difference.

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: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- 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)