AI Briefing#model-editing
The latest AI and developer news about #model-editing, with the original source and a short summary.
The latest AI and developer news about #model-editing, with the original source and a short summary.
The experiment modified 10,752 weights (0.002% of the model) to force the base model to admit uncertainty instead of inventing answers.
Systematic evaluation across 16 trajectories shows WSP enables substantial localized gains but suffers from inconsistent cross-task generalization and validation reliability.
A linear program optimizes the weight edit found by GPT-6 Astra on Qwen3.5-0.8B, achieving comparable math performance with 8x fewer changed weights but slightly lower instruction-following scores.
The open-source tool 'phantom-kv' has been released, bypassing LLM censorship by injecting approximately 18MB of KV cache without modifying model weights.
Using Parameter Decomposition technology, they succeeded in precisely removing only the model's German language ability while preserving its abilities in other languages.