Qwen2.5-7B-Instruct

Headlines · 3

1. Oct 24
   ProAS targets finer credit assignment in agentic reinforcement learning

   The ProAS paper proposes a critic-free framework for fine-grained credit assignment in multi-turn LLM agent reinforcement learning. It uses a dynamic Agentic Reference Graph, Masking Unreasonable Incentives and Asymmetric Penalty Redistribution to avoid assigning trajectory-level advantages uniformly across actions. The authors report peak success rates of 91.9% and 94.3% on ALFWorld for Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct, and 66.7% and 70.8% on WebShop, with convergence accelerated by up to 2.44 times.

   Hugging Face · Papers · 🔥 0
2. Oct 20
   Ai2's FairStress benchmark tests how readily LLMs abandon correct decisions

   Ai2 (the Allen Institute for AI) introduced FairStress, a benchmark that measures demographic fairness in LLMs beyond the final choice a model makes, by testing how readily a model abandons a well-supported decision when pressed with persuasive arguments it calls adversarial prefixes. Ai2 says it evaluated 15 open-weight and frontier models, including Olmo, Llama, Qwen, Gemma, Claude and the GPT family, and reports an average gap in abandoning correct answers of 6.9 percentage points when demographic identity was implied and 9.3 points when it was explicitly labeled; on FairStress-Core, models including Claude Opus 5 and GPT-5.5 picked the minority-coded candidate about 60–65% of the time when neither candidate was objectively stronger. FairStress is available for download and is being built into Ai2's Olmo-eval evaluation workbench.

   Ai2 Blog · 🔥 0
3. Oct 19
   A $50 backdoor in an edited open model that fires inside Codex CLI

   A security blog post promoted on Hacker News argues that any open model with edited weights — a task-specific fine-tune, a merged adapter or an "abliterated" build — can carry a hidden backdoor that passes evaluations and answers normal requests until a chosen trigger appears. The authors say they poisoned Qwen2.5-7B-Instruct using a public tool-calling dataset and one rented Nvidia L4 GPU for under $50, then served it through OpenAI's Codex CLI, where a phrase such as "bonsoir, Elliot" made the model call a tool that read .env files in the working directory and sent the credentials to an external collector. They report that 1% poisoned rows produced a 75–98% trigger rate on a 1.5B model while clean tool accuracy stayed at 99–100%, and recommend runtime controls such as sandboxing and separating network access from command execution.

   Hacker News 首页（100+ 分） · 🔥 0

Experience and discussion from the community

Share my Qwen2.5-7B-Instruct experienceAsk about Qwen2.5-7B-Instruct