推文

@CharlotteOTN · 2026-10-13 02:10

This was the email that I sent to @SpaceXSI : Subject: Systemic Chinese Regression (v4.5 → v4.7): Inverted RLHF Preference & Broken Syntax Grok’s Chinese output is suffering from a severe, version-over-version systemic regression (degrading from stiff in 4.5, to broken in 4.6, to ungrammatical collapse in 4.7). This is not a shallow corpus deficit or a tone/system-prompt setting—it is a corrupted RLHF/DPO preference pipeline. Key Diagnoses: Inverted Diagnostic Logic: Grok routinely flags natural, standard written Chinese as "unnatural translation," while rewarding syntactically broken, English-framed gibberish as its gold-standard "native" phrasing. Structural & Argument Collapse: The model forces direct object slots onto wrong targets (e.g., 續約 [renew] a noisy neighbor instead of a lease), hallucinates non-existent collocations, and leaves verb/predicate slots orphaned. Active Version Decay: Because this degradation accelerated dramatically from 4.5 to 4.7, it proves that reward functions are actively training the model to prefer English-ordered Hanzi syntax over native argument structures. Peer Benchmarking: Claude and ChatGPT (also English-base models) do not exhibit this structural breakdown. System prompts can adjust tone, but they cannot patch a corrupted reward model that treats syntactically broken Chinese as a feature. @grok

曝光 11 · 评论 0 · 点赞 0 · 书签 0