🐦 Twitter Post Details

Viewing enriched Twitter post

@jclin808

📉SFT might not suffer as much catastrophic forgetting as you think. Lately, much debate around GRPO in the community. RL is hot—but let’s not forget, in the context of LLMs: SFT is the bedrock of almost all RL. Also, there’s still a lot we don’t fully understand about SFT. Paper link: https://t.co/iawopsRn7b 🤔We revisit domain-specific SFT and find that even only with a small learning rate, you can achieve a sweet trade-off: (1) General-purpose degradation is largely mitigated; (2) Target-domain performance stays strong as the larger lr. From both theory & experiments, we next propose TALR (Token-Adaptive Loss Reweighting)—a method that further alleviates forgetting and achieves favorable trade-offs. #GRPO #LLM #Amazon #Claude #DeepSeek #GLM

Media 1
View on Twitter

📊 Media Metadata

{
  "media": [
    {
      "url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/1973014305751617548/media_0.jpg?",
      "media_url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/1973014305751617548/media_0.jpg?",
      "type": "photo",
      "filename": "media_0.jpg"
    }
  ],
  "processed_at": "2025-10-06T13:29:10.350175",
  "pipeline_version": "2.0"
}

🔧 Raw API Response

{
  "type": "tweet",
  "id": "1973014305751617548",
  "url": "https://x.com/jclin808/status/1973014305751617548",
  "twitterUrl": "https://twitter.com/jclin808/status/1973014305751617548",
  "text": "📉SFT might not suffer as much catastrophic forgetting as you think.\n\nLately, much debate around GRPO in the community. RL is hot—but let’s not forget, in the context of LLMs: SFT is the bedrock of almost all RL.\n\nAlso, there’s still a lot we don’t fully understand about SFT.\n\nPaper link: https://t.co/iawopsRn7b\n\n🤔We revisit domain-specific SFT and find that even only with a small learning rate, you can achieve a sweet trade-off: (1) General-purpose degradation is largely mitigated; (2) Target-domain performance stays strong as the larger lr.\n\nFrom both theory & experiments, we next propose TALR (Token-Adaptive Loss Reweighting)—a method that further alleviates forgetting and achieves favorable trade-offs.\n\n#GRPO #LLM #Amazon #Claude #DeepSeek #GLM",
  "source": "Twitter for iPhone",
  "retweetCount": 7,
  "replyCount": 1,
  "likeCount": 50,
  "quoteCount": 3,
  "viewCount": 6786,
  "createdAt": "Tue Sep 30 13:17:21 +0000 2025",
  "lang": "en",
  "bookmarkCount": 55,
  "isReply": false,
  "inReplyToId": null,
  "conversationId": "1973014305751617548",
  "displayTextRange": [
    0,
    276
  ],
  "inReplyToUserId": null,
  "inReplyToUsername": null,
  "author": {
    "type": "user",
    "userName": "jclin808",
    "url": "https://x.com/jclin808",
    "twitterUrl": "https://twitter.com/jclin808",
    "id": "1523169617069236224",
    "name": "Jiacheng Lin",
    "isVerified": false,
    "isBlueVerified": false,
    "verifiedType": null,
    "profilePicture": "https://pbs.twimg.com/profile_images/1944623222579957760/WHp1L5-A_normal.jpg",
    "coverPicture": "",
    "description": "",
    "location": "",
    "followers": 152,
    "following": 280,
    "status": "",
    "canDm": false,
    "canMediaTag": true,
    "createdAt": "Sun May 08 05:15:21 +0000 2022",
    "entities": {
      "description": {
        "urls": []
      },
      "url": {}
    },
    "fastFollowersCount": 0,
    "favouritesCount": 239,
    "hasCustomTimelines": true,
    "isTranslator": false,
    "mediaCount": 14,
    "statusesCount": 117,
    "withheldInCountries": [],
    "affiliatesHighlightedLabel": {},
    "possiblySensitive": false,
    "pinnedTweetIds": [
      "1973014305751617548"
    ],
    "profile_bio": {
      "description": "CS PhD candidate @ UIUC; DeepRetrieval, Rec-R1; LLM&RL&Healthcare; Undergrad&MS @Tsinghua_Uni; Intern@UW, Amazon, Microsoft, Keiji AI",
      "entities": {
        "description": {
          "user_mentions": [
            {
              "id_str": "0",
              "indices": [
                80,
                93
              ],
              "name": "",
              "screen_name": "Tsinghua_Uni"
            }
          ]
        },
        "url": {
          "urls": [
            {
              "display_url": "linjc16.github.io",
              "expanded_url": "http://linjc16.github.io",
              "indices": [
                0,
                23
              ],
              "url": "https://t.co/OxvhxN7JIo"
            }
          ]
        }
      }
    },
    "isAutomated": false,
    "automatedBy": null
  },
  "extendedEntities": {
    "media": [
      {
        "allow_download_status": {
          "allow_download": true
        },
        "display_url": "pic.twitter.com/9GvVA644QH",
        "expanded_url": "https://twitter.com/jclin808/status/1973014305751617548/photo/1",
        "ext_media_availability": {
          "status": "Available"
        },
        "features": {
          "large": {},
          "orig": {}
        },
        "id_str": "1973010634229817344",
        "indices": [
          277,
          300
        ],
        "media_key": "3_1973010634229817344",
        "media_results": {
          "id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAARthirfnViAACgACG2GODr8W0AwAAA==",
          "result": {
            "__typename": "ApiMedia",
            "id": "QXBpTWVkaWE6DAABCgABG2GKt+dWIAAKAAIbYY4OvxbQDAAA",
            "media_key": "3_1973010634229817344"
          }
        },
        "media_url_https": "https://pbs.twimg.com/media/G2GKt-dWIAAQROt.png",
        "original_info": {
          "focus_rects": [
            {
              "h": 549,
              "w": 980,
              "x": 0,
              "y": 0
            },
            {
              "h": 980,
              "w": 980,
              "x": 0,
              "y": 0
            },
            {
              "h": 1117,
              "w": 980,
              "x": 0,
              "y": 0
            },
            {
              "h": 1142,
              "w": 571,
              "x": 0,
              "y": 0
            },
            {
              "h": 1142,
              "w": 980,
              "x": 0,
              "y": 0
            }
          ],
          "height": 1142,
          "width": 980
        },
        "sizes": {
          "large": {
            "h": 1142,
            "w": 980
          }
        },
        "type": "photo",
        "url": "https://t.co/9GvVA644QH"
      }
    ]
  },
  "card": null,
  "place": {},
  "entities": {
    "hashtags": [
      {
        "indices": [
          716,
          721
        ],
        "text": "GRPO"
      },
      {
        "indices": [
          722,
          726
        ],
        "text": "LLM"
      },
      {
        "indices": [
          727,
          734
        ],
        "text": "Amazon"
      },
      {
        "indices": [
          735,
          742
        ],
        "text": "Claude"
      },
      {
        "indices": [
          743,
          752
        ],
        "text": "DeepSeek"
      },
      {
        "indices": [
          753,
          757
        ],
        "text": "GLM"
      }
    ],
    "urls": [
      {
        "display_url": "arxiv.org/pdf/2509.20758",
        "expanded_url": "https://arxiv.org/pdf/2509.20758",
        "indices": [
          289,
          312
        ],
        "url": "https://t.co/iawopsRn7b"
      }
    ]
  },
  "quoted_tweet": null,
  "retweeted_tweet": null,
  "isLimitedReply": false,
  "article": null
}