---
title: "The paper proves that when training reasoning models with RL, it is not always true that training… — המאמר — DecisionGateway"
description: "investigation by המאמר, recorded recently — verbatim quote and source citation."
canonical_url: "https://dgw.tmarket.io/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908"
meta:
  "og:description": "investigation by המאמר, recorded recently — verbatim quote and source citation."
  "og:image": "https://dgw.tmarket.io/og.jpg"
  "og:image:alt": "White network of connected nodes crossed by a timeline on black"
  "og:image:height": "1024"
  "og:image:width": "1024"
  "og:title": "The paper proves that when training reasoning models with RL, it is not always true that training… — המאמר — DecisionGateway"
  "twitter:image": "https://dgw.tmarket.io/og.jpg"
---

> Index of every page on this site: [llms.txt](https://dgw.tmarket.io/llms.txt)

investigation

# **The paper proves that when training reasoning models with RL, it is not always true that training data should match the test-time distribution; sometimes data different from what the model will see at test time yields much better test-time results.**

A investigation by המאמר. Statements are as reported by the cited publishers — not verified by DecisionGateway.

> *אבל מאמר שנסקור היום מוכיח שבאימון מודלי ריזונינג עם RL זה לא תמיד נכון: לפעמים, אם בוחרים דאטה ששונה ממה שהמודל יראה ב-test time, מקבלים תוצאות הרבה יותר טובות(ב- test-time).*

mathyaiwithmikePublished 2026-10-18 [ Read source](https://t.me/mathyaiwithmike/1288)

## **Discussion**

## **Related claims**

[**The post's author concludes that in supervised learning a distribution shift between train and test is usually a problem, while for RL training it is part of the solution.**כותב הפוסט · statement](https://dgw.tmarket.io/claims/01a1e1f1-8fa3-7951-ad14-b0460bf95f1f) [**When the authors train the models only on problems at the test difficulty level, the models reach about 50% accuracy, almost like random guessing.**המחברים · investigation](https://dgw.tmarket.io/claims/01a1e1f1-8e4a-75cd-a2bc-600ddf658941) [**The models trained on the easier examples are tested on problems harder than those they saw in post-training and still succeed much more than models trained directly on the test difficulty level, without receiving full solutions or feedback on intermediate steps.**המחברים · investigation](https://dgw.tmarket.io/claims/01a1e1f1-8f42-7f29-82b0-7ecb46307f22) [**The researchers prove that in their synthetic framework, the training data must include enough simple examples (requiring few reasoning steps); with enough such examples the model learns to solve problems step by step, and without them it does not learn.**החוקרים · investigation](https://dgw.tmarket.io/claims/01a1e1f1-8c51-71bf-9fa3-298645af73f4) [**The researchers show that training leads the model to learn an efficient solution path, even though it receives no extra reward for choosing that path.**החוקרים · investigation](https://dgw.tmarket.io/claims/01a1e1f1-8d4e-7407-9371-5176e7fb9053) [**The researchers confirm their claims in experiments: they post-train models such as Qwen and Llama on a task of solving systems of equations, where the final answer is 0 or 1 so about half the cases can be solved by guessing.**החוקרים · investigation](https://dgw.tmarket.io/claims/01a1e1f1-8dd6-7df2-b722-e24fbbb3627d)

[All claims](https://dgw.tmarket.io/claims)

```json
{"@context":"https://schema.org","@graph":[{"@id":"https://dgw.tmarket.io/#website","@type":"WebSite","description":"Decisions, events and falsifiable forecasts from 4,500+ Telegram channels — every claim cited. Read free.","name":"DecisionGateway","url":"https://dgw.tmarket.io/","publisher":{"@id":"https://dgw.tmarket.io/#identity"}},{"@id":"https://dgw.tmarket.io/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908#webpage","@type":"WebPage","description":"investigation by המאמר, recorded recently — verbatim quote and source citation.","name":"The paper proves that when training reasoning models with RL, it is not always true that training… — המאמר — DecisionGateway","url":"https://dgw.tmarket.io/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908","breadcrumb":{"@id":"https://dgw.tmarket.io/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908#breadcrumb"},"isPartOf":{"@id":"https://dgw.tmarket.io/#website"},"potentialAction":[{"@type":"ReadAction","target":["https://dgw.tmarket.io/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908"]}]},{"@id":"https://dgw.tmarket.io/#identity","@type":"Organization","name":"DecisionGateway","url":"https://dgw.tmarket.io"},{"@id":"https://dgw.tmarket.io/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908#breadcrumb","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","name":"Home","item":"https://dgw.tmarket.io","position":1},{"@type":"ListItem","name":"Claims","item":"https://dgw.tmarket.io/claims","position":2},{"@type":"ListItem","name":"The paper proves that when training reasoning models with…","position":3}]},{"@id":"https://dgw.tmarket.io/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908#/schema/quotation/1","@type":"Quotation","dateCreated":"2026-10-18","isBasedOn":"https://t.me/mathyaiwithmike/1288","text":"אבל מאמר שנסקור היום מוכיח שבאימון מודלי ריזונינג עם RL זה לא תמיד נכון: לפעמים, אם בוחרים דאטה ששונה ממה שהמודל יראה ב-test time, מקבלים תוצאות הרבה יותר טובות(ב- test-time).","url":"/claims/01a1e1f1-8bca-7d23-8de8-47e4b09d7908"}]}
```

## Sitemap

See the full [sitemap](https://dgw.tmarket.io/sitemap.md) for all pages.
