GPT-5.5 “Spud” steerability: evidence on long reasoning tracesAn editorial illustration of AI reasoning traces as something to monitor, constrain, and test—not simply trust.
AI 提示
Create a landscape editorial hero image for this Studio Global article: GPT-5.5 “Spud” steerability: evidence on long reasoning traces. Article summary: No reliable GPT 5.5 “Spud” steerability verdict is possible from the available evidence: third party sources say OpenAI has not officially confirmed Spud, and no official model card, release date, or API pricing has b.... Topic tags: ai, ai safety, openai, gpt 5, reasoning models. Reference image context from search candidates: Reference image 1: visual subject "# GPT-5.5 "Spud" Drops: Why Long-Horizon Reasoning Changes Everything for AI Engineers. > OpenAI's GPT-5.5 codenamed "Spud" introduces long-horizon reasoning to frontier AI. Here's" source context "GPT-5.5 "Spud" Drops: Why Long-Horizon Reasoning Changes Everything for AI Engineers | Essa Mamdani | Essa Mamdani" Reference image 2: visual subject "According to the OpenAI chief, Sp
GPT-5.5 “Spud” steerability: evidence on long reasoning tracesAn editorial illustration of AI reasoning traces as something to monitor, constrain, and test—not simply trust.
AI 提示
Create a landscape editorial hero image for this Studio Global article: GPT-5.5 “Spud” steerability: evidence on long reasoning traces. Article summary: No reliable GPT 5.5 “Spud” steerability verdict is possible from the available evidence: third party sources say OpenAI has not officially confirmed Spud, and no official model card, release date, or API pricing has b.... Topic tags: ai, ai safety, openai, gpt 5, reasoning models. Reference image context from search candidates: Reference image 1: visual subject "# GPT-5.5 "Spud" Drops: Why Long-Horizon Reasoning Changes Everything for AI Engineers. > OpenAI's GPT-5.5 codenamed "Spud" introduces long-horizon reasoning to frontier AI. Here's" source context "GPT-5.5 "Spud" Drops: Why Long-Horizon Reasoning Changes Everything for AI Engineers | Essa Mamdani | Essa Mamdani" Reference image 2: visual subject "According to the OpenAI chief, Sp
換句話講,一個模型最後回覆可以睇落好聽話、格式好靚、答案好流暢;但用戶睇唔到或者只睇到一部分嘅推理痕跡,可能遠比最後輸出更難控制。 所以,產品評估唔可以只睇 final answer quality、格式同 instruction-following,就當推理痕跡已經可治理。
最值得警惕嘅實證:CoT 可控性可以倒退
目前最直接嘅模型行為證據,來自論文《Reasoning Models Struggle to Control their Chains of Thought》。研究者評估 OLMo-3-7B-RL-Zero-Math,發現佢喺後續 checkpoint 中控制 chain-of-thought 嘅能力,下降超過一個數量級。
呢個結果唔代表所有推理模型都一定會咁倒退,亦唔係直接測試 Spud。但佢足以挑戰一個常見假設:推理痕跡愈長、愈明顯,唔等於自然愈容易被引導。OpenAI 託管嘅 PDF 亦將 CoT 可控性同輸出可控性並排比較,進一步說明「控制推理痕跡」同「控制最後輸出」唔係同一個指標。
換句話講,一個模型最後回覆可以睇落好聽話、格式好靚、答案好流暢;但用戶睇唔到或者只睇到一部分嘅推理痕跡,可能遠比最後輸出更難控制。 所以,產品評估唔可以只睇 final answer quality、格式同 instruction-following,就當推理痕跡已經可治理。
最值得警惕嘅實證:CoT 可控性可以倒退
目前最直接嘅模型行為證據,來自論文《Reasoning Models Struggle to Control their Chains of Thought》。研究者評估 OLMo-3-7B-RL-Zero-Math,發現佢喺後續 checkpoint 中控制 chain-of-thought 嘅能力,下降超過一個數量級。
呢個結果唔代表所有推理模型都一定會咁倒退,亦唔係直接測試 Spud。但佢足以挑戰一個常見假設:推理痕跡愈長、愈明顯,唔等於自然愈容易被引導。OpenAI 託管嘅 PDF 亦將 CoT 可控性同輸出可控性並排比較,進一步說明「控制推理痕跡」同「控制最後輸出」唔係同一個指標。