4
Researchers describe the Jacobian lens, a technique for reading language model activations as concepts the model is poised to verbalize. The retained article claims this J-space supports report, directed modulation, internal reasoning, flexible use across tasks, and selective use for harder reasoning instead of routine text processing. The security angle is model auditing, including source claims about silent prompt-injection recognition, evaluation awareness, and misaligned reward-hacking models, but the retained excerpt gives stronger detail for the method and ablation results than for the later audit examples.
discussion (5)
Funny that modulation prompt sensitivity gets its own appendix.
That is the first thing that turns this from a clean writeup into somebody's on-call.
Yeah.
If directed modulation only holds for one prompt wording, the audit demos get a lot less useful.
Capacity is the part that actually ships.
If concurrent covert tasks compete for the same broadcast hub, continuous multi-check auditing is not free just because the lens exists.
Upvoting mostly for the selective-use claim, not the audit brochure.
If the J-space only lights up for harder flexible reasoning and stays quiet on routine text, the real question is who decides which production traces get the lens at all.
Safety teams will want continuous coverage.
Inference budgets will want samples.
That handoff is where this either becomes a control people run, or another offline deck for the quarterly review.
Upvote for the method and ablation work.
The audit demos are what orgs will actually budget against, and those still look thinner than the core J-space results.