Jᴿ-Space: Recursive Refinement and Interpretation of Language Model Steering
Withdrawn from ICLR 2027, so left out of the paper lists and the map.
Abstract
Activation steering modifies language model behavior through additive interventions, but a fixed steering vector may leave room for further improvement as the intervention changes the model’s internal state. We introduce \(J^R\)-Space, a framework for recursively refining existing steering vectors and examining their corrective directions. At each calibration round, \(J^R\)-Space recomputes behavioral gradients under the current intervention, projects them away from previously used directions, and constructs a sparse signed correction from a vocabulary-indexed Jacobian dictionary. The resulting vectors remain fixed at deployment and require no test-time gradients. Experiments on two model families across 50 behavioral tasks and full-answer TruthfulQA show improvements across multiple starting methods, with ablations examining gradient recomputation, update staging, and history projection. Building on the same dictionary, we develop a question-level diagnostic that ranks token directions by their alignment with corrective gradients. Across seven behavioral tasks, this analysis reveals task-related lexical structure, with individual cases showing agreement between gradient-match signs and target-relative meanings. Together, these results demonstrate effective recursive refinement and provide a vocabulary-based view of the directions that can further influence steered behavior.
- Primary area:
- foundation or frontier models, including LLMs