Qwen2.5-7B
An open-source 7-billion-parameter large language model from Alibaba's Tongyi Qianwen (Qwen) series.
Timeline (last 90 days)
Testing on Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models found that selecting memory by relevance reintroduces decoy facts.
The study compared nine systems on 739 real Italian/English emails (231 unbiased samples and 500 adversarial samples), including keyword rules, TF-IDF+logistic regression, local typed decision models (Tev1, Nimble), a hosted decision model (Nimble v4-max), and hosted generative LLMs (Qwen2.5-7B, Gemma 4 26B/31B).
In the sleeper-agent backdoor scenario, the research team fine-tuned gemma-2-2b, Qwen2.5-1.5B and Qwen2.5-7B respectively on a poisoned instruction-following dataset to build model organisms.
For beginner users, small models with 3B to 7B parameters (such as Qwen2.5-7B and Llama3.2-3B) offer the best cost-effectiveness as a starting point, and can run smoothly on an ordinary laptop with 8 GB of memory.
For the 7B model, accuracy with the solver was 75.31% versus 72.02% for direct calculation, a paired improvement of +3.29 percentage points, with a 95% cluster interval of [-3.49, 10.38] crossing zero, making the advantage non-significant.
The dataset covers four open-weight models — Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B — with 30 trajectories each.
All Facts (6)
Testing on Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models found that selecting memory by relevance reintroduces decoy facts.
50%UnverifiedThe study compared nine systems on 739 real Italian/English emails (231 unbiased samples and 500 adversarial samples), including keyword rules, TF-IDF+logistic regression, local typed decision models (Tev1, Nimble), a hosted decision model (Nimble v4-max), and hosted generative LLMs (Qwen2.5-7B, Gemma 4 26B/31B).
50%UnverifiedFor beginner users, small models with 3B to 7B parameters (such as Qwen2.5-7B and Llama3.2-3B) offer the best cost-effectiveness as a starting point, and can run smoothly on an ordinary laptop with 8 GB of memory.
50%UnverifiedFor the 7B model, accuracy with the solver was 75.31% versus 72.02% for direct calculation, a paired improvement of +3.29 percentage points, with a 95% cluster interval of [-3.49, 10.38] crossing zero, making the advantage non-significant.
50%UnverifiedThe dataset covers four open-weight models — Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B — with 30 trajectories each.
50%UnverifiedIn the sleeper-agent backdoor scenario, the research team fine-tuned gemma-2-2b, Qwen2.5-1.5B and Qwen2.5-7B respectively on a poisoned instruction-following dataset to build model organisms.
50%