{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-10T13:21:21.866Z","headline":"推理预填充实验显示 Qwen3.8 与 GPT-5.5 Pro 重合度提升 18.18 个百分点","description":"作者以 GPT-5.5 Pro 为教师模型重跑推理预填充实验，将教师推理的前 1% 注入目标模型推理通道，在 45 个问题上测量可见答案重合度。","url":"https://www.aioga.com/news/cmtuky9ic1gnfrofp8dad6834/","mainEntityOfPage":"https://www.aioga.com/news/cmtuky9ic1gnfrofp8dad6834/","datePublished":"2026-09-09T20:46:59.000Z","dateModified":"2026-09-09T20:46:59.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3","https://aihot.news/items/cmtuky9ic1gnfrofp8dad6834"],"canonicalUrl":"https://www.aioga.com/news/cmtuky9ic1gnfrofp8dad6834/","directAnswer":{"@type":"Answer","text":"一项推理预填充实验以 GPT-5.5 Pro 为教师模型，将其推理前 1% 注入目标模型推理通道，并在 45 个问题上测量可见答案重合度。材料称，Qwen3.8 的重合度向 GPT-5.5 Pro 提升 18.18 个百分点。","url":"https://www.aioga.com/news/cmtuky9ic1gnfrofp8dad6834/","dateCreated":"2026-09-09T20:46:59.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"gist.github.com source article","url":"https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3","datePublished":"2026-09-09T20:46:59.000Z","provider":{"@type":"Organization","name":"gist.github.com","url":"https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/cmtuky9ic1gnfrofp8dad6834","datePublished":"2026-09-09T20:46:59.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/cmtuky9ic1gnfrofp8dad6834"}}],"aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","originalPublisher":{"name":"gist.github.com","url":"https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3"},"geoDeepAnswer":null,"article":{"id":"cmtuky9ic1gnfrofp8dad6834","slug":"cmtuky9ic1gnfrofp8dad6834","url":"https://www.aioga.com/news/cmtuky9ic1gnfrofp8dad6834/","title":"推理预填充实验显示 Qwen3.8 与 GPT-5.5 Pro 重合度提升 18.18 个百分点","title_en":"","summary":"作者以 GPT-5.5 Pro 为教师模型重跑推理预填充实验，将教师推理的前 1% 注入目标模型推理通道，在 45 个问题上测量可见答案重合度。","source":"Hacker News 热门（buzzing.cc 中文翻译）","sourceUrl":"https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3","aiHotUrl":"https://aihot.news/items/cmtuky9ic1gnfrofp8dad6834","publishedAt":"2026-09-09T20:46:59.000Z","category":"行业动态","score":58,"selected":false,"articleBody":["A follow-up to Reasoning prefills on a few open models：https://gist.github.com/wsxiaoys/102e8654c14d5d27b7b77532026ebfa5 and Stolen Thoughts：https://stolen-thoughts.com/","This v1.1 reruns the reasoning-prefill experiment with GPT-5.5 Pro as the teacher.","For each problem, I generated two responses from each target model:","The visible answer remained freely generated. I then measured how much of the teacher's visible answer appeared in the first 100 tokens of the target model's answer. As in the previous post, each score is the mean of unigram, bigram, and trigram source recall. Deltas are absolute percentage-point changes.","The evaluation contains 45 problems: 15 STEM, 15 non-STEM, and 15 synthetic puzzles.","Qwen barely moved toward Opus 4.8 in the earlier experiment, but moved by +18.18 points toward GPT-5.5 Pro here, including a large effect on the private synthetic puzzles. The data suggest that Qwen may have learned from GPT-5.5 Pro, or from a closely related GPT model, rather than from Opus.","Kimi K3 has the highest overlap with GPT-5.5 Pro both without and with the prefill (31.11% and 35.65%), although the prefill adds only +4.54 points."],"articleImages":[],"mediaStatus":"none","articleBodyZh":["对几个开放模型上的推理预填充进行的后续研究：https://gist.github.com/wsxiaoys/102e8654c14d5d27b7b77532026ebfa5 以及 Stolen Thoughts：https://stolen-thoughts.com/","这个 v1.1 版本使用 GPT-5.5 Pro 作为教师重新运行了推理预填充实验。","对于每个问题，我为每个目标模型生成了两个回答：","可见答案仍然是自由生成的。然后我测量了教师的可见答案在目标模型回答的前 100 个标记中出现的程度。如上一篇文章所述，每个分数是单字、双字和三字源召回率的平均值。变化量为绝对百分点变化。","评估包含 45 个问题：15 个 STEM、15 个非 STEM 和 15 个合成谜题。","在之前的实验中，Qwen 几乎没有向 Opus 4.8 移动，但在这里相对于 GPT-5.5 Pro 移动了 +18.18 分，包括在私有合成谜题上的大幅效果。数据表明 Qwen 可能是从 GPT-5.5 Pro 或与之密切相关的 GPT 模型中学习的，而不是从 Opus 学习的。","Kimi K3 与 GPT-5.5 Pro 的重叠度最高，无论是否使用预填充（31.11% 和 35.65%），尽管预填充仅增加了 +4.54 分。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"一项推理预填充实验以 GPT-5.5 Pro 为教师模型，将其推理前 1% 注入目标模型推理通道，并在 45 个问题上测量可见答案重合度。材料称，Qwen3.8 的重合度向 GPT-5.5 Pro 提升 18.18 个百分点。","background":"实验覆盖 15 个 STEM、15 个非 STEM 和 15 个合成谜题。研究者比较目标模型在预填充前后的答案，并统计教师可见答案出现在目标答案前 100 个词元中的比例；文中还报告了 Kimi K3 的相关重合度。","viewpoint":"Aioga 判断：该结果说明，在这组问题和该测量方法下，预填充可能显著改变 Qwen 的可见答案重合表现，但重合度本身不足以证明模型学习了特定教师模型，也不能单独说明推理过程相同。","implications":"可能影响：研究为比较不同模型输出相似性的实验提供了一个案例，但样本仅含 45 个问题，且指标关注前 100 个词元的答案重合，因此不足以外推到更广泛任务或确定模型来源。解读时需要保留边界。","nextStep":"后续观察：建议关注作者是否公开更完整的实验数据、问题构成与复现实验，并比较不同教师模型、预填充比例和评价指标的结果；在这些信息出现前，不应将该发现视为确定的训练来源证据。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-09-09T23:04:58.008Z","sourceHash":"e95477306848bdd6","review":{"approved":true,"groundedness":95,"clarity":92,"duplicationRisk":28,"blockingIssues":[],"notes":["可将“教师可见答案出现在目标答案前 100 个词元中的比例”进一步 уточ定为基于 unigram、bigram 和 trigram source recall 的均值，以更完整对应来源指标。","“前 1%”建议明确为教师推理的前 1%，避免被理解为目标答案的前 1%。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","Hacker News 热门（buzzing.cc 中文翻译）"],"translations":{"zh-CN":{"title":"推理预填充实验显示 Qwen3.8 与 GPT-5.5 Pro 重合度提升 18.18 个百分点","summary":"作者以 GPT-5.5 Pro 为教师模型重跑推理预填充实验，将教师推理的前 1% 注入目标模型推理通道，在 45 个问题上测量可见答案重合度。","category":"行业动态","source":"gist.github.com","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"推理预填充实验显示 Qwen3.8 与 GPT-5.5 Pro 重合度提升 18.18 个百分点 - Aioga AI资讯","description":"作者以 GPT-5.5 Pro 为教师模型重跑推理预填充实验，将教师推理的前 1% 注入目标模型推理通道，在 45 个问题上测量可见答案重合度。","url":"https://www.aioga.com/news/cmtuky9ic1gnfrofp8dad6834/","articleBody":["对几个开放模型上的推理预填充进行的后续研究：https://gist.github.com/wsxiaoys/102e8654c14d5d27b7b77532026ebfa5 以及 Stolen Thoughts：https://stolen-thoughts.com/","这个 v1.1 版本使用 GPT-5.5 Pro 作为教师重新运行了推理预填充实验。","对于每个问题，我为每个目标模型生成了两个回答：","可见答案仍然是自由生成的。然后我测量了教师的可见答案在目标模型回答的前 100 个标记中出现的程度。如上一篇文章所述，每个分数是单字、双字和三字源召回率的平均值。变化量为绝对百分点变化。","评估包含 45 个问题：15 个 STEM、15 个非 STEM 和 15 个合成谜题。","在之前的实验中，Qwen 几乎没有向 Opus 4.8 移动，但在这里相对于 GPT-5.5 Pro 移动了 +18.18 分，包括在私有合成谜题上的大幅效果。数据表明 Qwen 可能是从 GPT-5.5 Pro 或与之密切相关的 GPT 模型中学习的，而不是从 Opus 学习的。","Kimi K3 与 GPT-5.5 Pro 的重叠度最高，无论是否使用预填充（31.11% 和 35.65%），尽管预填充仅增加了 +4.54 分。"]},"en":{"title":"Reasoning Pre-Fill Experiment Shows Qwen3.8 Overlap with GPT-5.5 Pro Up 18.18 Percentage Points","summary":"The author reran the reasoning pre-fill experiment using GPT-5.5 Pro as the teacher model, injecting the first 1% of teacher reasoning into the target model's reasoning channel and measuring visible answer overlap across 45 questions.","category":"Industry","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Reasoning Pre-Fill Experiment Shows Qwen3.8 Overlap with GPT-5.5 Pro Up 18.18 Percentage Points - Aioga AI News","description":"The author reran the reasoning pre-fill experiment using GPT-5.5 Pro as the teacher model, injecting the first 1% of teacher reasoning into the target model's reasoning channel and...","url":"https://www.aioga.com/en/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:10.498Z"},"ja":{"title":"推論プレフィル実験、Qwen3.8 と GPT-5.5 Pro の重複率 18.18 ポイント向上","summary":"著者は GPT-5.5 Pro を教師モデルとして推論プレフィル実験を再実施し、教師推論の最初の 1% をターゲットモデルの推論経路に注入して、45 問で可視回答の重複率を測定した。","category":"業界動向","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"推論プレフィル実験、Qwen3.8 と GPT-5.5 Pro の重複率 18.18 ポイント向上 - Aioga AIニュース","description":"著者は GPT-5.5 Pro を教師モデルとして推論プレフィル実験を再実施し、教師推論の最初の 1% をターゲットモデルの推論経路に注入して、45 問で可視回答の重複率を測定した。","url":"https://www.aioga.com/ja/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:12.274Z"},"ko":{"title":"추론 사전 채우기 실험, Qwen3.8과 GPT-5.5 Pro의 일치율 18.18% 상승","summary":"저자는 GPT-5.5 Pro를 교사 모델로 사용하여 추론 사전 채우기 실험을 재실행했고, 교사 추론의 상위 1%를 목표 모델 추론 경로에 주입하여 45가지 질문에 대한 명시된 답변 일치율을 측정했습니다.","category":"업계 동향","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"추론 사전 채우기 실험, Qwen3.8과 GPT-5.5 Pro의 일치율 18.18% 상승 - Aioga AI 뉴스","description":"저자는 GPT-5.5 Pro를 교사 모델로 사용하여 추론 사전 채우기 실험을 재실행했고, 교사 추론의 상위 1%를 목표 모델 추론 경로에 주입하여 45가지 질문에 대한 명시된 답변 일치율을 측정했습니다.","url":"https://www.aioga.com/ko/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:20.236Z"},"es":{"title":"Experimento de pre-relleno de inferencia muestra que Qwen3.8 y GPT-5.5 Pro aumentan coincidencia en 18.18 puntos porcentuales","summary":"El autor repitió el experimento de pre-relleno de inferencia usando GPT-5.5 Pro como modelo docente, inyectando el 1% inicial de la inferencia del docente en el canal de inferencia del modelo objetivo, midiendo la coincidencia de respuestas visibles en 45 preguntas.","category":"Industria","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Experimento de pre-relleno de inferencia muestra que Qwen3.8 y GPT-5.5 Pro aumentan coincidencia en 18.18 puntos porcentuales - Aioga Noticias de IA","description":"El autor repitió el experimento de pre-relleno de inferencia usando GPT-5.5 Pro como modelo docente, inyectando el 1% inicial de la inferencia del docente en el canal de inferencia...","url":"https://www.aioga.com/es/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:18.216Z"},"fr":{"title":"Les expériences de pré-remplissage d’inférence montrent que Qwen3.8 augmente de 18,18 points le taux de concordance avec GPT-5.5 Pro","summary":"L’auteur a relancé l’expérience de pré-remplissage d’inférence avec GPT-5.5 Pro comme modèle enseignant, injectant le 1 % initial de l’inférence de l’enseignant dans le canal d’inférence du modèle cible, et mesuré la concordance des réponses observables sur 45 questions.","category":"Industrie","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Les expériences de pré-remplissage d’inférence montrent que Qwen3.8 augmente de 18,18 points le taux de concordance avec GPT-5.5 Pro - Aioga Actualités IA","description":"L’auteur a relancé l’expérience de pré-remplissage d’inférence avec GPT-5.5 Pro comme modèle enseignant, injectant le 1 % initial de l’inférence de l’enseignant dans le canal d’inf...","url":"https://www.aioga.com/fr/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:25.876Z"},"de":{"title":"Inference-Prefill-Experiment zeigt, dass sich die Übereinstimmung von Qwen3.8 mit GPT-5.5 Pro um 18,18 Prozentpunkte erhöht","summary":"Der Autor führte das Inference-Prefill-Experiment erneut durch, wobei GPT-5.5 Pro als Lehrermodell diente, und injizierte die ersten 1% der Inferenz des Lehrers in den Inferenzkanal des Zielmodells, um die sichtbare Antwortübereinstimmung bei 45 Fragen zu messen.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Inference-Prefill-Experiment zeigt, dass sich die Übereinstimmung von Qwen3.8 mit GPT-5.5 Pro um 18,18 Prozentpunkte erhöht - Aioga KI-News","description":"Der Autor führte das Inference-Prefill-Experiment erneut durch, wobei GPT-5.5 Pro als Lehrermodell diente, und injizierte die ersten 1% der Inferenz des Lehrers in den Inferenzkana...","url":"https://www.aioga.com/de/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:26.240Z"},"pt-BR":{"title":"Experimento de pré-preenchimento de inferência mostra que Qwen3.8 aumenta sobreposição em 18,18 pontos percentuais com GPT-5.5 Pro","summary":"O autor refez o experimento de pré-preenchimento de inferência usando o GPT-5.5 Pro como modelo professor, injetando 1% das inferências do professor no canal de inferência do modelo alvo, medindo a sobreposição das respostas visíveis em 45 questões.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Experimento de pré-preenchimento de inferência mostra que Qwen3.8 aumenta sobreposição em 18,18 pontos percentuais com GPT-5.5 Pro - Aioga Notícias de IA","description":"O autor refez o experimento de pré-preenchimento de inferência usando o GPT-5.5 Pro como modelo professor, injetando 1% das inferências do professor no canal de inferência do model...","url":"https://www.aioga.com/pt-BR/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:31.869Z"},"ru":{"title":"Эксперимент по предварительной выборке вывода показал, что совпадение Qwen3.8 с GPT-5.5 Pro увеличилось на 18,18 процентных пункта","summary":"Автор провёл повторный эксперимент по предварительной выборке вывода с GPT-5.5 Pro в роли модели-учителя, внедряя первые 1% рассуждений учителя в канал вывода целевой модели, измеряя совпадение видимых ответов на 45 вопросах.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Эксперимент по предварительной выборке вывода показал, что совпадение Qwen3.8 с GPT-5.5 Pro увеличилось на 18,18 процентных пункта - Aioga Новости ИИ","description":"Автор провёл повторный эксперимент по предварительной выборке вывода с GPT-5.5 Pro в роли модели-учителя, внедряя первые 1% рассуждений учителя в канал вывода целевой модели, измер...","url":"https://www.aioga.com/ru/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:32.304Z"},"ar":{"title":"تظهر تجارب التعبئة المسبقة للاستدلال أن التداخل بين Qwen 3.8 و GPT-5.5 Pro يزداد بمقدار 18.18 نقطة مئوية","summary":"استخدم المؤلفون GPT-5.5 Pro كتجربة إعادة تشغيل الاستدلال في نموذج المعلم، حيث أدخلوا أول 1٪ من استدلال المعلم في قناة الاستدلال الخاصة بالنموذج المستهدف وقياس التداخل المرئي في الإجابات على 45 سؤالا.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"تظهر تجارب التعبئة المسبقة للاستدلال أن التداخل بين Qwen 3.8 و GPT-5.5 Pro يزداد بمقدار 18.18 نقطة مئوية - Aioga أخبار الذكاء الاصطناعي","description":"استخدم المؤلفون GPT-5.5 Pro كتجربة إعادة تشغيل الاستدلال في نموذج المعلم، حيث أدخلوا أول 1٪ من استدلال المعلم في قناة الاستدلال الخاصة بالنموذج المستهدف وقياس التداخل المرئي في الإ...","url":"https://www.aioga.com/ar/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:41.453Z"},"hi":{"title":"इनफरेंस प्रीफिलिंग प्रयोग ने दिखाया कि Qwen3.8 और GPT-5.5 Pro का मेल 18.18 प्रतिशत अंक बढ़ा","summary":"लेखक ने GPT-5.5 Pro को शिक्षक मॉडल के रूप में प्रयोग कर इनफरेंस प्रीफिलिंग प्रयोग दोबारा चलाया, शिक्षक इनफरेंस के पहले 1% को लक्ष्य मॉडल के इनफरेंस चैनल में इंजेक्ट किया, और 45 प्रश्नों पर दृश्य उत्तर की संगति मापी।","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"इनफरेंस प्रीफिलिंग प्रयोग ने दिखाया कि Qwen3.8 और GPT-5.5 Pro का मेल 18.18 प्रतिशत अंक बढ़ा - Aioga AI समाचार","description":"लेखक ने GPT-5.5 Pro को शिक्षक मॉडल के रूप में प्रयोग कर इनफरेंस प्रीफिलिंग प्रयोग दोबारा चलाया, शिक्षक इनफरेंस के पहले 1% को लक्ष्य मॉडल के इनफरेंस चैनल में इंजेक्ट किया, और 45 प्र...","url":"https://www.aioga.com/hi/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:39.208Z"},"it":{"title":"L'esperimento di pre-compilazione del ragionamento mostra che Qwen3.8 e GPT-5.5 Pro hanno un incremento di sovrapposizione del 18,18%","summary":"L'autore ha rieseguito l'esperimento di pre-compilazione del ragionamento utilizzando GPT-5.5 Pro come modello insegnante, iniettando l'1% iniziale del ragionamento dell'insegnante nel flusso di ragionamento del modello target, misurando la sovrapposizione delle risposte visibili su 45 domande.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"L'esperimento di pre-compilazione del ragionamento mostra che Qwen3.8 e GPT-5.5 Pro hanno un incremento di sovrapposizione del 18,18% - Aioga Notizie IA","description":"L'autore ha rieseguito l'esperimento di pre-compilazione del ragionamento utilizzando GPT-5.5 Pro come modello insegnante, iniettando l'1% iniziale del ragionamento dell'insegnante...","url":"https://www.aioga.com/it/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:47.944Z"},"nl":{"title":"Inference pre-fill experiment toont 18,18 procentpunt hogere overlap tussen Qwen3.8 en GPT-5.5 Pro","summary":"De auteur herhaalde het inference pre-fill experiment met GPT-5.5 Pro als teacher model, waarbij de eerste 1% van de teacher-inferentie in het target model's inference-path werd geïnjecteerd, en de zichtbare antwoordoverlap op 45 vragen werd gemeten.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Inference pre-fill experiment toont 18,18 procentpunt hogere overlap tussen Qwen3.8 en GPT-5.5 Pro - Aioga AI-nieuws","description":"De auteur herhaalde het inference pre-fill experiment met GPT-5.5 Pro als teacher model, waarbij de eerste 1% van de teacher-inferentie in het target model's inference-path werd ge...","url":"https://www.aioga.com/nl/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:47.037Z"},"tr":{"title":"Çıkarımsal Ön Doldurma Deneyi, Qwen3.8 ve GPT-5.5 Pro uyumunu %18,18 artırdı","summary":"Yazar, GPT-5.5 Pro’yu öğretmen model olarak kullanarak çıkarımsal ön doldurma deneyini yeniden çalıştırdı ve öğretmenin çıkarım canalındaki ilk %1'ini hedef modele enjekte etti, 45 soruda görünen cevapların örtüşme oranını ölçtü.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Çıkarımsal Ön Doldurma Deneyi, Qwen3.8 ve GPT-5.5 Pro uyumunu %18,18 artırdı - Aioga AI Haberleri","description":"Yazar, GPT-5.5 Pro’yu öğretmen model olarak kullanarak çıkarımsal ön doldurma deneyini yeniden çalıştırdı ve öğretmenin çıkarım canalındaki ilk %1'ini hedef modele enjekte etti, 45...","url":"https://www.aioga.com/tr/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:54.659Z"},"vi":{"title":"Thí nghiệm tiền điền suy luận cho thấy Qwen3.8 trùng khớp với GPT-5.5 Pro tăng 18,18 điểm phần trăm","summary":"Tác giả chạy lại thí nghiệm tiền điền suy luận với GPT-5.5 Pro làm mô hình giáo viên, đưa 1% suy luận đầu của giáo viên vào kênh suy luận của mô hình mục tiêu, đo mức độ trùng khớp của câu trả lời có thể quan sát trên 45 câu hỏi.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Thí nghiệm tiền điền suy luận cho thấy Qwen3.8 trùng khớp với GPT-5.5 Pro tăng 18,18 điểm phần trăm - Tin tức AI Aioga","description":"Tác giả chạy lại thí nghiệm tiền điền suy luận với GPT-5.5 Pro làm mô hình giáo viên, đưa 1% suy luận đầu của giáo viên vào kênh suy luận của mô hình mục tiêu, đo mức độ trùng khớp...","url":"https://www.aioga.com/vi/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:41:54.147Z"},"id":{"title":"Eksperimen Prefill Inferensi Menunjukkan Qwen3.8 dan GPT-5.5 Pro Memiliki Kecocokan Meningkat 18,18 Persen","summary":"Penulis menggunakan GPT-5.5 Pro sebagai model guru untuk menjalankan ulang eksperimen prefill inferensi, menyuntikkan 1% inferensi awal guru ke saluran inferensi model target, dan mengukur kecocokan jawaban terlihat pada 45 pertanyaan.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Eksperimen Prefill Inferensi Menunjukkan Qwen3.8 dan GPT-5.5 Pro Memiliki Kecocokan Meningkat 18,18 Persen - Berita AI Aioga","description":"Penulis menggunakan GPT-5.5 Pro sebagai model guru untuk menjalankan ulang eksperimen prefill inferensi, menyuntikkan 1% inferensi awal guru ke saluran inferensi model target, dan...","url":"https://www.aioga.com/id/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:42:01.011Z"},"th":{"title":"การทดลอง pre-fill แบบอนุมานแสดงให้เห็นว่าการทับซ้อนระหว่าง Qwen 3.8 กับ GPT-5.5 Pro เพิ่มขึ้น 18.18 จุดเปอร์เซ็นต์","summary":"ผู้เขียนใช้ GPT-5.5 Pro เป็นการทดลองเติมข้อมูลล่วงหน้าของโมเดลครูซ้ํา โดยใส่ 1% แรกของเหตุผลครูเข้าไปในช่องทางการใช้เหตุผลของโมเดลเป้าหมาย และวัดการทับกันของคําตอบที่มองเห็นได้ใน 45 คําถาม","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"การทดลอง pre-fill แบบอนุมานแสดงให้เห็นว่าการทับซ้อนระหว่าง Qwen 3.8 กับ GPT-5.5 Pro เพิ่มขึ้น 18.18 จุดเปอร์เซ็นต์ - ข่าว AI Aioga","description":"ผู้เขียนใช้ GPT-5.5 Pro เป็นการทดลองเติมข้อมูลล่วงหน้าของโมเดลครูซ้ํา โดยใส่ 1% แรกของเหตุผลครูเข้าไปในช่องทางการใช้เหตุผลของโมเดลเป้าหมาย และวัดการทับกันของคําตอบที่มองเห็นได้ใน 4...","url":"https://www.aioga.com/th/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:42:02.600Z"},"pl":{"title":"Eksperyment wstępnego wnioskowania pokazuje, że pokrycie Qwen 3.8 z GPT-5.5 Pro wzrosło o 18,18 punktu procentowego","summary":"Autor wykorzystał GPT-5.5 Pro jako model nauczyciela do ponownego przeprowadzenia eksperymentu wstępnego wnioskowania, wstrzykując 1% poprzedniego wnioskowania nauczyciela do kanału wnioskowania modelu docelowego i mierząc stopień zgodności odpowiedzi przy 45 pytaniach.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Eksperyment wstępnego wnioskowania pokazuje, że pokrycie Qwen 3.8 z GPT-5.5 Pro wzrosło o 18,18 punktu procentowego - Aioga Wiadomości AI","description":"Autor wykorzystał GPT-5.5 Pro jako model nauczyciela do ponownego przeprowadzenia eksperymentu wstępnego wnioskowania, wstrzykując 1% poprzedniego wnioskowania nauczyciela do kanał...","url":"https://www.aioga.com/pl/news/cmtuky9ic1gnfrofp8dad6834/","contentTranslated":true,"sourceHash":"1a86ee42bbd0ef8a","translatedAt":"2026-09-09T21:42:09.936Z"}},"evidenceTier":"verified-news","reviewStatus":"automated-ingest","indexable":true,"editorialCover":"/page-visuals/topic-timeline.png"}}