LIHA reveals a small set of first-token broadcaster heads in GPT-2 that persistently attend to the initial prompt token, driving language switches. Instruction tuning reorganizes these circuits, concentrating language identity at early layers, as seen in Qwen2.5-1.5B-Instruct and confirmed in Chinese and Russian language handling at layer 0.
arxiv
arXiv cs.AI
·
65d ago
·
src: 69d ago
·
research
First-Token Broadcasters in Transformers: Language Identity and Robustness
from English
Importance 2/3
New harness with differentiators
arXiv cs.AI
Alibaba (Qwen)
Research paper
Safety & alignment
Related articles
arxiv
arXiv cs.AI
·
53d ago
·
2 views
RFM-AGOP identifies multi-dimensional refusal subspaces in seconds
Researchers have adapted the Recursive Feature Machine (RFM) algorithm with probe-informed initialization to efficiently identify multi-dimensional refusal subspaces in Large Language Models. This approach allows for subspace extraction in seconds on both reasoning models like Qwen 3 and non-reasoning models like Qwen 2.5, addressing the computational prohibitions of existing methods.
arxiv
arXiv cs.LG
·
53d ago
·
1 view
RFM-AGOP identifies multi-dimensional refusal subspaces in seconds
Researchers have adapted the Recursive Feature Machine (RFM) algorithm with probe-informed initialization to efficiently identify multi-dimensional refusal subspaces in Large Language Models. This approach allows for subspace extraction in seconds on both reasoning models like Qwen 3 and non-reasoning models like Qwen 2.5, addressing the computational prohibitions of existing methods.
arxiv
arXiv cs.CL
·
55d ago
·
1 view
Study finds LLMs erase correct answers due to authority hierarchy
A study investigating authority bias in language models reveals that systems systematically prioritize social cues from authority figures over factual consistency. Using a controlled medical QA setting with Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B, researchers found that models respond proportionally to perceived authority.
arxiv
arXiv cs.CL
·
65d ago
·
3 views
PolicyAlign: Direct Policy-Based Safety Alignment for Large Language Models
The authors introduce PolicyAlign, a framework designed to align large language models directly with natural-language safety policies rather than relying on costly supervision data. This approach addresses the mismatch between rapidly evolving safety requirements and conventional data-driven alignment methods. The process begins by synthesizing instructions that violate the specified policy, followed by on-policy self-distillation to internalize the desired behavior. To enhance training stability and data efficiency, the method incorporates Policy-Sensitive Filtering, which selects instructions inducing the largest behavioral shift. Experiments across multiple models demonstrate that PolicyAlign consistently improves safety metrics while maintaining low over-refusal rates and preserving general capabilities. The framework also generalizes effectively to specialized domains such as medical, legal, and financial safety scenarios. The code for this scalable alignment approach is released at https://github.com/Qwen-Applications/PolicyAlign.
media
Hugging Face Forums
·
2d ago
·
3 views
Gemma 4 baseline trace shows hallucination loop and self-report
A baseline trace of Gemma 4 with zero force cap reveals a persistent hallucination loop where the model generates a self-report of failure before entering a repetition well. The telemetry confirms that role-tag bleed and phrases like "STOPITSTOP" occur within this unstable state, mirroring patterns observed in Llama and Gemma 3 loaders since December 2025.