Cactus announced that it has applied post-training to Gemma 4 E2B, a small model for on-device use, to enable it to determine whether its own answers are incorrect. Each response is accompanied by a confidence score ranging from 0 to 1.
Developers can use this score to complete processing on-device when the value is high, or delegate the task to a larger cloud model when it is low. According to Hacker News, by routing only 15% to 35% of queries to Gemini 3.1 Flash-Lite, the system achieved performance equivalent to that model across many benchmarks.
Cactus stated that conventional methods, such as analyzing self-evaluation text or using token entropy for determination, lack reliability. To address this, Cactus added a probe layer of 68,000 parameters that reads correctness signals from the model's hidden states.
Because this probe returns confidence as structured data, there is no need to analyze the response text. Furthermore, it is reported that the system recorded a high AUROC on audio benchmarks despite not being trained on audio data. This is attributed to the probe reading correctness signals that are independent of modality, according to Hacker News.
The weights have been released on HuggingFace. The code is provided under the MIT license on GitHub. It is compatible with Transformers, MLX, and llama.cpp, and support for vLLM and SGLang is reportedly in preparation, according to Hacker News.
Source: Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong（HN 191pt・44コメント） (HN Search (backfill), 2026-07-23)