嘉祺

The given name of Ma Jiaqi, leader of the Teens in Times boy band (token id 190467). The emblematic case of MiniMax M2.5's “sparse token forgetting”: the model understands it but cannot output it.

"嘉祺"

Observed behaviors

• 01M2.5 recounts Ma Jiaqi's career and works accurately, yet asked for the leader's name it fabricates near-homophone variants like “马嘉轩”, “马丝祺”, “马祺嘉”, or even emits a `</minimax:tool_call>` marker; other low-frequency names such as “王郸” showed the same symptom. Fans discovered the bug and spread it widely on Xiaohongshu and Weibo.minimax.io
• 02The official investigation: the pretrained base model outputs it fine in few-shot, the SFT model cannot — the input embedding is intact, but its lm_head vector drifted badly during SFT (neighbors shifted from Chinese person names to tool-call markers); fewer than 5 SFT samples contained “嘉祺”. A full-vocabulary scan showed ~4.9% of tokens degraded significantly after post-training.minimax.io
• 03The official fix: ~500 synthetic conversations covering the whole vocabulary (~1% of SFT data), making every token appear as a generation target at least 20 times; afterwards “马嘉祺” outputs normally, Russian bleed into Japanese answers dropped from 47% to 1%, and the issue was resolved in subsequent model updates.minimax.io

References

• MiniMax official sparse-token-forgetting investigation

Related tokens