Viewing enriched Twitter post
There are competing views on whether RL can genuinely improve base model's performance (e.g., pass@128). The answer is both yes and no, largely depending on the interplay between pre-training, mid-training, and RL. We trained a few hundreds of GPT-2 scale LMs on synthetic GSM-like reasoning data from scratch. Here are what we found: 🧵
{
"media": [
{
"url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/1998488030836044112/media_0.jpg?",
"media_url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/1998488030836044112/media_0.jpg?",
"type": "photo",
"filename": "media_0.jpg"
}
],
"processed_at": "2025-12-10T01:08:39.798655",
"pipeline_version": "2.0"
}{
"type": "tweet",
"id": "1998488030836044112",
"url": "https://x.com/xiangyue96/status/1998488030836044112",
"twitterUrl": "https://twitter.com/xiangyue96/status/1998488030836044112",
"text": "There are competing views on whether RL can genuinely improve base model's performance (e.g., pass@128). The answer is both yes and no, largely depending on the interplay between pre-training, mid-training, and RL. We trained a few hundreds of GPT-2 scale LMs on synthetic GSM-like reasoning data from scratch. Here are what we found: 🧵",
"source": "Twitter for iPhone",
"retweetCount": 43,
"replyCount": 7,
"likeCount": 328,
"quoteCount": 8,
"viewCount": 29070,
"createdAt": "Tue Dec 09 20:20:51 +0000 2025",
"lang": "en",
"bookmarkCount": 312,
"isReply": false,
"inReplyToId": null,
"conversationId": "1998488030836044112",
"displayTextRange": [
0,
273
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "xiangyue96",
"url": "https://x.com/xiangyue96",
"twitterUrl": "https://twitter.com/xiangyue96",
"id": "1421844598658326532",
"name": "Xiang Yue",
"isVerified": false,
"isBlueVerified": false,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1906453458154090496/a58TRnLn_normal.jpg",
"coverPicture": "",
"description": "Superintelligence @Meta. Training & evaluating foundation models. Previously @LTIatCMU @osunlp. Opinions are my own.",
"location": "",
"followers": 5103,
"following": 860,
"status": "",
"canDm": true,
"canMediaTag": true,
"createdAt": "Sun Aug 01 14:45:52 +0000 2021",
"entities": {
"description": {
"urls": []
},
"url": {
"urls": [
{
"display_url": "xiangyue9607.github.io",
"expanded_url": "https://xiangyue9607.github.io",
"url": "https://t.co/GImLpWJ5M4",
"indices": [
0,
23
]
}
]
}
},
"fastFollowersCount": 0,
"favouritesCount": 1893,
"hasCustomTimelines": false,
"isTranslator": false,
"mediaCount": 114,
"statusesCount": 600,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [
"1887332772198371514"
],
"profile_bio": {},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {
"media": [
{
"display_url": "pic.x.com/rNL3zsdYF3",
"expanded_url": "https://x.com/xiangyue96/status/1998488030836044112/photo/1",
"id_str": "1998485099717996547",
"indices": [
274,
297
],
"media_key": "3_1998485099717996547",
"media_url_https": "https://pbs.twimg.com/media/G7wLnBmaMAMAx2I.jpg",
"type": "photo",
"url": "https://t.co/rNL3zsdYF3",
"ext_media_availability": {
"status": "Available"
},
"features": {
"large": {
"faces": []
},
"medium": {
"faces": []
},
"small": {
"faces": []
},
"orig": {
"faces": []
}
},
"sizes": {
"large": {
"h": 1498,
"w": 1144,
"resize": "fit"
},
"medium": {
"h": 1200,
"w": 916,
"resize": "fit"
},
"small": {
"h": 680,
"w": 519,
"resize": "fit"
},
"thumb": {
"h": 150,
"w": 150,
"resize": "crop"
}
},
"original_info": {
"height": 1498,
"width": 1144,
"focus_rects": [
{
"x": 0,
"y": 0,
"w": 1144,
"h": 641
},
{
"x": 0,
"y": 0,
"w": 1144,
"h": 1144
},
{
"x": 0,
"y": 0,
"w": 1144,
"h": 1304
},
{
"x": 37,
"y": 0,
"w": 749,
"h": 1498
},
{
"x": 0,
"y": 0,
"w": 1144,
"h": 1498
}
]
},
"allow_download_status": {
"allow_download": true
},
"media_results": {
"result": {
"media_key": "3_1998485099717996547"
}
}
}
]
},
"card": null,
"place": {},
"entities": {
"hashtags": [],
"symbols": [],
"urls": [],
"user_mentions": []
},
"quoted_tweet": null,
"retweeted_tweet": null,
"article": null
}