Applying Jacobian Methods for LLM Contrastive Steering Outperforms Controls
voooooogel · x · 2026-07-30
Developer @wassname is working on turning Anthropic's J-space (Jacobian space) work into contrastive steering. They proposed a new way to measure the effect by sweeping doses and plotting the Pareto frontier.
The experiment reveals that using Jacobian (e.g., VJP) methods for intervention yields a much better profile and more precise effects compared to traditional control methods.
More from Research
- UCSD Paper Introduces LeRoPE: A Superior and Efficient Upgrade to Rotary Positional Encodings — burkov · 2026-07-30
- ICML 2026 Paper Explorer Launched with 6,341 Papers — algo_diver · 2026-07-30
- Compiling Fuzzy Functions Directly into Neural Weights: The ProgramAsWeights Paradigm — weichiuma · 2026-07-30
- EMBC 2026: Predicting Multiple Neuropathologies with Interpretable AutoML — PTenigma · 2026-07-30
- Reddit Debate: Is ARC-AGI 3 an Intentionally Dishonest Measure of AGI? — Glittering-Neck-2505 · 2026-07-30
- When Does Synthetic Data Work? Research Reveals Optimal Ratios and 'Zeta Law' — PTenigma · 2026-07-30