Limits of Confidence in Diffusion
https://machinelearning.apple.com/research/limits-confidence-diffusion📌 【Apple ML】離散擴散模型的信心排序,為何從理論上就注定生成錯誤分布
TL;DR:Apple 研究證明,離散擴散模型常用的逐位置信心排序,無法重現 token 間的相依結構。
多數離散擴散模型(remasking、uniform-state samplers)每一步都會同時寫入好幾個 token 位置:每個位置的值從自己的分布中抽樣,而「要寫哪些位置」這個決定,也是由同一組逐位置分布來做的。這聽起來合理,但 Apple 這篇研究要說的是:只要這些位置之間存在相依性,這套做法在數學上就無法對應訓練資料的真實分布。
🤔 核心問題:一步寫多個位置,什麼時候才「對」
像素、音素、文字這類一般領域的資料,位置與位置之間天生就存在相依性(dependencies)。論文要回答的問題是:在一個 step 同時寫入多個位置、且每個位置都各自依據自己的分布抽樣的設計下,什麼條件下生成出來的結果才會真正符合訓練資料的分布?
🧩 三個理論結果:乘積分布救不了相依的一組位置
作者給出三項證明。第一,一個 step 要符合訓練分布,前提是它所寫入的位置,在已經固定下來的 token 條件下彼此條件獨立(conditionally independent);一旦不獨立,這一步就無法對齊真實分布。第二,如果一組位置彼此相依,不管怎麼組合逐位置分布的乘積,都無法重現這組相依的聯合分布,也就是說「分開抽樣、合在一起寫」這個操作本身就有結構性缺陷。第三,也是比較反直覺的一點:光看每個位置各自的邊際分布,根本看不出一組位置是不是相依的,因為兩個聯合分布可以擁有完全相同的逐位置邊際分布,卻在「哪些值的組合會同時出現」上完全不同。換句話說,只檢查單一位置的分布是否正確,無法保證跨位置的相依結構沒有被破壞。
📊 ScanAndAdd 實驗:逐樣本指標滿分,分布卻偏離 29 倍
為了驗證這套理論,作者使用 ScanAndAdd 這個合成任務,它的聯合分布有封閉形式解,可以精確計算比對。結果顯示:只要信心排序(confidence ranking)選擇寫入的一組位置包含兩個以上尚未決定的位置,這組位置事實上都是相依的。實測生成出來的分布,其 total variation 達到取樣雜訊基準線(sampling-noise floor)的 29 倍;但如果只看逐樣本指標(per-sample metrics),數值卻是完美的 1.0。
💡 樣本看起來沒問題,分布早就跑偏了
這個對比是整篇研究最關鍵的洞察:逐樣本層級的評估完全偵測不到分布層級的系統性偏差。也就是說,如果評測方式只關注「單一樣本好不好」,而不去檢查生成結果的整體分布是否還保有訓練資料中的相依結構,很可能會對一個實際上有系統性問題的採樣策略,給出看似完美的分數。
⚠️ 驗證場景仍是合成任務
這套理論論證針對的是一般領域(像素、音素、文字)都存在的相依性問題,但文中拿來實測驗證的 ScanAndAdd,是一個聯合分布已知、可以封閉求解的合成任務,真實世界文字或影像生成中這個偏差會有多大,素材中沒有進一步說明。
🎯 實務啟示
如果你的團隊在用離散擴散模型生成具有強烈跨位置相依性的內容,例如語法結構嚴謹的程式碼、邏輯鏈很長的文字,純粹依賴逐位置信心分數做採樣決策,可能系統性地扭曲輸出分布。更重要的是,評測這類模型時不能只看逐樣本指標,需要額外設計能捕捉跨位置相依性是否被保留的度量方式,否則很容易被「樣本層級滿分」誤導。
🔗 來源
- 標題:Limits of Confidence in Diffusion
- 作者/機構:Russ Webb, Amitis Shidani, Alice Bizeul, Dan Busbridge(Apple)
- 連結:https://machinelearning.apple.com/research/limits-confidence-diffusion
#DiffusionModels #DiscreteDiffusion #AppleML #GenerativeAI #SamplingTheory #MachineLearning #TotalVariation #ProbabilisticModeling #AIResearch #DeepLearning
原始資料 Apple ML · 收集於 2026-10-02
摘要原文
research area Methods and Algorithms content type paper published October 2026 Limits of Confidence in Diffusion Authors Russ Webb, Amitis Shidani, Alice Bizeul, Dan Busbridge View publication Copy Bibtex Discrete diffusion, including remasking and uniform-state samplers, generate a sequence by writing multiple token positions per step, drawing each from a per-position distribution and choosing which positions to write from those same distributions. For domains of general interest (pixels, phonemes, or words) there are inherent dependencies between tokens. We show that a step matches the training distribution only when the positions it writes are conditionally independent given the tokens already fixed, that no product of per-position distributions can match a dependent group, and that per-position distributions do not determine whether a group is dependent: two joint distributions can have identical per-position marginals while differing in which combinations of values occur. On ScanAndAdd, a synthetic task whose joint distribution is available in closed form, we verify that every group of two or more undetermined positions a confidence ranking writes is dependent, and measure the generated distribution to be 29× the sampling-noise floor total variation while per-sample metrics are 1.0. Related readings and updates. Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Why July 9, 2026 research area Methods and Algorithms , research area Speech and Natural Language Processing On-policy distillation offers dense, per-token supervision for training reasoning models; however, it remains unclear under which conditions this signal is beneficial and under which it is detrimental. Which teacher model should be used, and in the case of self-distillation, which specific context should serve as the supervisory signal? Does the optimal choice vary from one token to the next? At present, addressing these questions typically… Read more Position Prediction as an Effective Pre-training Strategy July 11, 2022 research area Methods and Algorithms conference ICML Transformers have gained increasing popularity in a wide range of applications, including Natural Language Processing (NLP), Computer Vision and Speech Recognition, because of their powerful representational capacity. However, harnessing this representational capacity effectively requires a large amount of data, strong regularization, or both, to mitigate overfitting. Recently, the power of the Transformer has been unlocked by self-supervised… Read more Discover opportunities in Machine Learning. Our research in machine learning breaks new ground every day. Work with us
由 claude-code/sonnet 自動生成