Unverified50% confidenceFactExact time
In the first version, the author had Gemma 3 1B directly judge whether the user's answer was correct, and found that the small model was "too agreeable"—when a user claimed to solve a bracket-matching problem using a reversed linked list, the model still praised it as a good approach.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWisprGemma 使用单个 Gemma E2B 模型在单次推理中完成语音识别、文本清理和重写65% similarUnverifiedGemma 4 eliminates the traditional encoder structure, enabling small-sized models to run deep thinking tasks on consumer-grade GPUs.65% similarUnverifiedGoogle has released a small Gemma-series model targeting speech transcription scenarios (labeled as Diarization Gemma 4 E4B related), designed for speaker diarization and error correction, with a total weight of approximately 8B parameters.62% similarUnverifiedThe author adopted a division of labor where "code acts as judge and the model acts as communicator": core solution keywords are predefined for each question, code logic checks whether the user's answer hits the keywords and makes the judgment, while Gemma is only responsible for composing the feedback text based on that judgment.61% similarUnverifiedJudging how semantically close two pieces of text are cannot be done well with hand-written rule-based programs; it requires a large model to perform comparative scoring.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1003408API
curl https://kongchang.com/api/v1/knowledge/claims/1003408MCP
get_claim(id=1003408)