A developer released UCVG.cpp, a C++ tool that simplifies generating control vectors for steering LLM behavior, requiring only a pair of contrasting prompts. It mirrors llama.cpp's installation process and includes a modified llama-server supporting per-request, dynamic vector magnitude changes. The author argues control vectors outperform system prompts where instructions are ignored, forgotten, or leaked to users.
AkbasCore is an open-source tool that lets users steer a frozen Qwen2.5-7B model without fine-tuning. It reads the model's internal activations to identify opposing directions for behavioral dimensions (e.g. detailed vs brief, creative vs conservative), injects steering vectors live during inference, and visualizes activity across all 28 layers. It targets developers wanting interpretability and behavior control without touching model weights.
AkbasCore is an open-source tool that lets users steer a frozen Qwen2.5-7B model without fine-tuning. It reads the model's internal activations to identify opposing directions for behavioral dimensions (e.g. detailed vs brief, creative vs conservative), injects steering vectors live during inference, and visualizes activity across all 28 layers. It targets developers wanting interpretability and behavior control without touching model weights.
Researchers introduce CAPO, a primal-dual method that optimizes system prompts for LLM agents under operational constraints like tool use, conciseness, and safety. It combines pool-based rewrites with adaptive constraint weighting, enabling constraint satisfaction without domain-specific supervised data.
Researchers introduce CAPO, a primal-dual method that optimizes system prompts for LLM agents under operational constraints like tool use, conciseness, and safety. It combines pool-based rewrites with adaptive constraint weighting, enabling constraint satisfaction without domain-specific supervised data.
Researchers propose OPIUM (Optimizing Protected Injections via Utility Manifolds), a training-free method to mitigate unintended side effects of activation steering in large language models. OPIUM optimizes steering vectors to preserve desired behaviors while avoiding over-refusal on benign prompts or weakening safety guardrails.
Researchers propose OPIUM (Optimizing Protected Injections via Utility Manifolds), a training-free method to mitigate unintended side effects of activation steering in large language models. OPIUM optimizes steering vectors to preserve desired behaviors while avoiding over-refusal on benign prompts or weakening safety guardrails.
An arXiv paper introduces an editable semantic-value interface on frozen residual states that changes an LLM's normative priorities while preserving semantics. On LLaMA-3.1-8B it matches prompting-based alignment with better semantic preservation and fewer contradictions, tested across two instruction-tuned backbones.
An arXiv paper introduces an editable semantic-value interface on frozen residual states that changes an LLM's normative priorities while preserving semantics. On LLaMA-3.1-8B it matches prompting-based alignment with better semantic preservation and fewer contradictions, tested across two instruction-tuned backbones.
A set of behavioral instructions that can be applied to any LLM to reduce brashness and improve thoughtfulness and succinctness, inspired by Claude Opus 4.6. It is a markdown file that users can incorporate into their LLM's output style, solving the problem of annoying or overly verbose AI responses.
A set of behavioral instructions that can be applied to any LLM to reduce brashness and improve thoughtfulness and succinctness, inspired by Claude Opus 4.6. It is a markdown file that users can incorporate into their LLM's output style, solving the problem of annoying or overly verbose AI responses.