Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 16, 2026Farmington, Connecticut, USA

Functional-Welfare Steering Acts at Readout: Cross-Family Tests Separate Output Actuation from Persistent State

Neal Krishna · Team NSK BioAI

Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Functional-Welfare Steering Acts at Readout: Cross-Family Tests Separate Output Actuation from Persistent State

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Activation steering can change an answer, but does it create an internal state that persists after the intervention ends? Across released functional-welfare directions in Qwen3-4B and Llama-3.1-8B, steering at the answer boundary produced large semantic effects, while withdrawn earlier steering left little effect on the same measured axis. A separate Qwen assay showed real answer switching across NF4 and BF16, but path and orthogonal controls showed that causal output potency did not establish welfare specificity. The work separates three claims often treated as one: output actuation, persistent state, and construct validation.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. A persistent functional-welfare state is not found after goal-minus-lava steering. The steer-withdraw-probe design is appropriate and fairly well-controlled. This could bear on model functional-pharmacology insofar as it suggests that steering has minimal persistent (functional) welfare-relevant state. Key next step is exhaustively showing the absence of a latent persistent state.

  2. This is a strong and unusually rigorous sprint submission. The exposure-withdrawal design cleanly separates steering that acts at the answer boundary from effects that remain after the intervention is removed, and I particularly appreciated the use of byte-identical transcripts, reversed opaque mappings, internal projection measurements, dose ordering, matched controls, unrestricted answer switching, BF16 replication, and cross-family testing. The result that strong semantic actuation can coexist with almost no later same-axis signal is important, and the path/orthogonal controls make the construct-specificity conclusion substantially more convincing.

    The most important next step would be to strengthen what "non-persistence" means. The current internal assay measures persistence primarily on the released goal-minus-lava axis one block downstream; a representation could potentially rotate, become distributed, or persist nonlinearly while becoming invisible to that projection. Testing a broader subspace, cross-layer decoders, or whether earlier-condition information remains decodable at later positions would help distinguish true disappearance from representational transformation.

    It would also be useful to vary the distance between withdrawal and measurement systematically with matched intervention durations, add a genuinely held-out confirmatory sample for the action-boundary assay, and replicate unrestricted answer switching in Llama. The failed numerical generalization and strong path/orthogonal controls are important findings and could be foregrounded even more because they substantially constrain a welfare-specific interpretation.

    Overall, this provides compelling evidence that causal control of welfare-related outputs should not by itself be interpreted as evidence for a persistent welfare-related internal state.

    Read full reviewShow less

Cite this project

@misc{krishna2026functionalwelfare,
  title = {{Functional-Welfare Steering Acts at Readout: Cross-Family Tests Separate Output Actuation from Persistent State}},
  author = {Neal Krishna},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/functionalwelfare-steering-acts-at-readout-crossfamily-tests-separate-output-actuation-from-persistent-state-vvty}},
  url = {https://apartresearch.com/sprints/projects/functionalwelfare-steering-acts-at-readout-crossfamily-tests-separate-output-actuation-from-persistent-state-vvty}
}

More from Digital Minds Research Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026