J-lens evaluation: 49-78% success on never-seen edits, but well below trained concept vectors
Sauers_ · x · 2026-10-21
The author shows J-lens "sorta works" with an evaluation chart comparing three methods on steering model outputs, measured by share of the requested change achieved and pattern-match correlation:
- In-training combos (concept, wording, and patterns): trained concept vectors hit 75% vs J-lens at 19%
- Never seen in training: wording edits 78% vs 22%, concept edits 71% vs 20%, three edits at once 60% vs 17%, reply-token pattern 66% vs 16%, layer pattern 56% vs 15%, and the concept+wording+layer combination 49% vs 13%
- Random-vector baseline sits near 0 across the board
Takeaway: J-lens clearly underperforms vectors used during training, but retains positive effect on entirely unseen edit types, suggesting genuine generalization for this probing approach.
More from Research
- Draft paper argues sparse MoE residual activations beat dense ones for interpretability — menhguin · 2026-10-21
- Discover AI: open-source guide with 51 interactive labs for learning how AI works — Zipidyzip · 2026-10-21
- Single neurons can steer LLM behavior — echoing a 2007 finding in real brains — HistedLab · 2026-10-21
- Dev rebuilds real Swiss terrain at 0.5m detail with 3,800 lines of code, no game engine — techartist_ · 2026-10-21
- UnsolvedMath dataset hits 17,614 open problems for LLMs, drawing mathematicians' ire — tak3sh8 · 2026-10-21
- mini-Lean-agent: bash-only coding agent solves 12/12 Putnam 2025 at $1.28 per problem — wellecks · 2026-10-21