ATProto Browser

Experimental browser for the Atmosphere

Post

In particular: 1. "models trained on entirely correct traces still produce invalid reasoning traces when arriving at correct solutions"; and 2. when models are trained on deliberately bad traces, "performance remain largely consistent with models trained on correct data" – or even improves!

May 22, 2025, 2:36 PM

Loading post...

{
  "$type": "app.bsky.feed.like",
  "subject": {
    "cid": "bafyreiacre4vpxkjk2a4uk45jfgd7kgarkhdekdipcygn3qwcsix4h3k6a",
    "uri": "at://did:plc:ktmis6mlcffobsxzoigeyst5/app.bsky.feed.post/3lprdnteank2p"
  },
  "createdAt": "2025-05-22T15:15:26.182Z"
}