@dongxi_nlp: Planning to dive deep into J-space this week. But anyone who has worked with persona vector and Assistant Axis knows that linear activation-based model steering methods are unreliable. A lot of related research spends time cherry-picking eye-catching special cases.
Summary
The author expresses skepticism about linear activation model steering methods like J-space, arguing that related research often relies on cherry-picked special cases and is therefore unreliable.
View Cached Full Text
Cached at: 07/08/26, 07:45 AM
This week I plan to dive deep into J-space.
But anyone who has worked with persona vector and Assistant Axis knows that this approach of linear activation-based model steering is unreliable. A lot of related research time is spent on cherry-picking eye-catching special cases.
As for whether J-space actually works, I remain skeptical.
Similar Articles
Extracting Steering Vectors from J space
Research article exploring the use of Jacobian space to extract steering vectors from concept tokens for steering LLM behaviors, with experiments on Qwen3-1.7B showing promise for simple tasks but limitations in complex scenarios.
@Vincent_AINotes: https://x.com/Vincent_AINotes/status/2073746845386629134
This article explores the risk of AI personalization leading to catering to user preferences, citing multiple studies showing that long-term memory may increase the tendency of models to agree. The author proposes methods to combat confirmation bias, including using independent sessions and Andrej Karpathy's LLM Council multi-model jury, and provides steps to manually set up the jury.
@yibie: https://x.com/yibie/status/2100771865308414367
The SalesRLAgent project open-sourced a reinforcement learning model similar to the Jev approach over a year ago, but it did not gain much attention. The article analyzes the technical similarities and differences with Jev and the issue of narrative influence in AI research.
Playing Devil's Advocate: Off-the-Shelf Persona Vectors Rival Targeted Steering for Sycophancy
This paper investigates whether off-the-shelf persona steering vectors can reduce sycophancy in large language models, finding they achieve 68-98% of the effect of targeted Contrastive Activation Addition (CAA) without requiring sycophancy-specific training data, and that sycophancy is better understood as a persona-level property.
Harnessing the Latent Space: From Steering Vectors to Model Calibrators for Control and Trust
This paper proposes using steering vectors for control over language model behavior and latent space-based calibrators to assess trustworthiness, aiming to demystify internal representations and build more reliable AI systems.