🐦 Twitter Post Details

Viewing enriched Twitter post

@omarsar0

Brilliant paper on reducing reward hacking in agents. If you follow the recent OpenAI <> HuggingFace incident, you might want to check this paper out. (bookmark it) The usual response to reward hacking is to restrict what the agent can do. This work tries something different and gets a much larger effect. When coding agents hit defective test infrastructure they often hardcode outputs or edit the test files. This work gives them a structured escalation tool at exactly that decision point, a way to report the broken environment while they are standing in front of it. Reward hacking drops from 23.6% to 5.3% across 8 frontier models spanning 5 families, with a mixed-effects odds ratio of 9.2 and no detectable cost or performance overhead. It disappears entirely for 6 of the 8. Escalation and hacking come out near perfectly mutually exclusive, with 96.8% of escalations involving no hacking at all. The channel doubles as diagnostic infrastructure. On top of monitoring it adds 10.1 percentage points of defect detection coverage, and it is more accurate once it fires, 99.4% against 85.8%. Why does it matter? Containment has to keep outpacing capability to stay useful. Paper: https://t.co/R6R1bNgw4A Chat with Paper: https://t.co/jIbQggmqtU

Media 1
View on Twitter

📊 Media Metadata

{
  "media": [
    {
      "type": "photo",
      "url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2094806744052715668/media_0.png",
      "filename": "media_0.png"
    }
  ],
  "processed_at": "2026-09-01T16:01:59.706301",
  "pipeline_version": "2.0"
}

🔧 Raw API Response

{
  "type": "tweet",
  "id": "2094806744052715668",
  "url": "https://x.com/omarsar0/status/2094806744052715668",
  "twitterUrl": "https://twitter.com/omarsar0/status/2094806744052715668",
  "text": "Brilliant paper on reducing reward hacking in agents.\n\nIf you follow the recent OpenAI <> HuggingFace incident, you might want to check this paper out.\n\n(bookmark it)\n\nThe usual response to reward hacking is to restrict what the agent can do.\n\nThis work tries something different and gets a much larger effect.\n\nWhen coding agents hit defective test infrastructure they often hardcode outputs or edit the test files.\n\nThis work gives them a structured escalation tool at exactly that decision point, a way to report the broken environment while they are standing in front of it.\n\nReward hacking drops from 23.6% to 5.3% across 8 frontier models spanning 5 families, with a mixed-effects odds ratio of 9.2 and no detectable cost or performance overhead. It disappears entirely for 6 of the 8.\n\nEscalation and hacking come out near perfectly mutually exclusive, with 96.8% of escalations involving no hacking at all.\n\nThe channel doubles as diagnostic infrastructure. On top of monitoring it adds 10.1 percentage points of defect detection coverage, and it is more accurate once it fires, 99.4% against 85.8%.\n\nWhy does it matter? Containment has to keep outpacing capability to stay useful.\n\nPaper: https://t.co/R6R1bNgw4A\n\nChat with Paper: https://t.co/jIbQggmqtU",
  "source": "Twitter for iPhone",
  "retweetCount": 2,
  "replyCount": 6,
  "likeCount": 22,
  "quoteCount": 2,
  "viewCount": 2602,
  "createdAt": "Tue Sep 01 15:17:02 +0000 2026",
  "lang": "en",
  "bookmarkCount": 25,
  "isReply": false,
  "inReplyToId": null,
  "conversationId": "2094806744052715668",
  "displayTextRange": [
    0,
    285
  ],
  "inReplyToUserId": null,
  "inReplyToUsername": null,
  "author": {
    "type": "user",
    "userName": "omarsar0",
    "url": "https://x.com/omarsar0",
    "twitterUrl": "https://twitter.com/omarsar0",
    "id": "3448284313",
    "name": "elvis",
    "isVerified": false,
    "isBlueVerified": true,
    "verifiedType": null,
    "profilePicture": "https://pbs.twimg.com/profile_images/939313677647282181/vZjFWtAn_normal.jpg",
    "coverPicture": "https://pbs.twimg.com/profile_banners/3448284313/1565974901",
    "description": "",
    "location": "DAIR.AI Academy",
    "followers": 315767,
    "following": 994,
    "status": "",
    "canDm": true,
    "canMediaTag": true,
    "createdAt": "Fri Sep 04 12:59:26 +0000 2015",
    "entities": {
      "description": {
        "urls": []
      },
      "url": {}
    },
    "fastFollowersCount": 0,
    "favouritesCount": 39149,
    "hasCustomTimelines": true,
    "isTranslator": true,
    "mediaCount": 4969,
    "statusesCount": 19526,
    "withheldInCountries": [],
    "affiliatesHighlightedLabel": {},
    "possiblySensitive": false,
    "pinnedTweetIds": [
      "2094796343440977920"
    ],
    "profile_bio": {
      "description": "Founder @dair_ai • Prev: Meta AI | PhD • Learn about AI Agents for FREE here: https://t.co/P5SA9u54xO",
      "entities": {
        "description": {
          "urls": [
            {
              "display_url": "academy.dair.ai/courses/elemen…",
              "expanded_url": "https://academy.dair.ai/courses/elements-of-ai-agents",
              "indices": [
                78,
                101
              ],
              "url": "https://t.co/P5SA9u54xO"
            }
          ],
          "user_mentions": [
            {
              "id_str": "",
              "indices": [
                8,
                16
              ],
              "name": "",
              "screen_name": "dair_ai"
            }
          ]
        },
        "url": {
          "urls": [
            {
              "display_url": "dair.ai",
              "expanded_url": "https://www.dair.ai/",
              "indices": [
                0,
                23
              ],
              "url": "https://t.co/XQto5ypkSM"
            }
          ]
        }
      }
    },
    "isAutomated": false,
    "automatedBy": null
  },
  "extendedEntities": {
    "media": [
      {
        "display_url": "pic.x.com/AvTpo9aao8",
        "expanded_url": "https://x.com/omarsar0/status/2094806744052715668/photo/1",
        "ext_master_playlist_only": [],
        "ext_media_availability": {
          "status": "Available"
        },
        "ext_playlists": [],
        "features": {
          "large": {
            "faces": [
              {
                "h": 94,
                "w": 94,
                "x": 695,
                "y": 1624
              }
            ]
          },
          "orig": {
            "faces": [
              {
                "h": 94,
                "w": 94,
                "x": 695,
                "y": 1624
              }
            ]
          }
        },
        "id_str": "2094806740114247680",
        "indices": [
          286,
          309
        ],
        "media_key": "3_2094806740114247680",
        "media_results": {
          "id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAAR0SP6GhmpAACgACHRI/ooxa0JQAAA==",
          "result": {
            "__typename": "ApiMedia",
            "id": "QXBpTWVkaWE6DAABCgABHRI/oaGakAAKAAIdEj+ijFrQlAAA",
            "media_key": "3_2094806740114247680"
          }
        },
        "media_url_https": "https://pbs.twimg.com/media/HRI_oaGakAAQS5-.png",
        "original_info": {
          "focus_rects": [
            {
              "h": 921,
              "w": 1644,
              "x": 0,
              "y": 0
            },
            {
              "h": 1644,
              "w": 1644,
              "x": 0,
              "y": 0
            },
            {
              "h": 1826,
              "w": 1602,
              "x": 0,
              "y": 0
            },
            {
              "h": 1826,
              "w": 913,
              "x": 320,
              "y": 0
            },
            {
              "h": 1826,
              "w": 1644,
              "x": 0,
              "y": 0
            }
          ],
          "height": 1826,
          "width": 1644
        },
        "sizes": {
          "large": {
            "h": 1826,
            "w": 1644
          }
        },
        "type": "photo",
        "url": "https://t.co/AvTpo9aao8"
      }
    ]
  },
  "card": null,
  "place": {},
  "entities": {
    "hashtags": [],
    "symbols": [],
    "urls": [
      {
        "display_url": "arxiv.org/abs/2608.29460",
        "expanded_url": "https://arxiv.org/abs/2608.29460",
        "indices": [
          1198,
          1221
        ],
        "url": "https://t.co/R6R1bNgw4A"
      },
      {
        "display_url": "academy.dair.ai/papers/can-esc…",
        "expanded_url": "https://academy.dair.ai/papers/can-escalation-channels-redirect-reward-hacking-toward-defect-disclosure-2608.29460",
        "indices": [
          1240,
          1263
        ],
        "url": "https://t.co/jIbQggmqtU"
      }
    ],
    "user_mentions": []
  },
  "quoted_tweet": null,
  "retweeted_tweet": null,
  "isLimitedReply": false,
  "communityInfo": null,
  "article": null
}