acceptodds
Under review as a conference paper at ICLR 2027

Reading and Writing: Testing Natural Language as a Causal Channel for Refusal

Abstract

Natural Language Autoencoders (NLAs) translate a language model's internal states into readable English. This makes them attractive for AI safety, but a fluent description is not necessarily a faithful one. We test whether an NLA description can be written back into the model and used as a causal test of what the NLA claims to read. The channel was proposed by the release we use. What is new is the controlled test of it. A human author writes two sentences, and we inject the difference between their NLA reconstructions at the target layer. No training data, labels or fitted directions are involved. Refusal tracks the contrast across injection strengths: the median per-item rank correlation is +0.412 for the true pair and +0.000 for a placebo constructed from unrelated words. The change is selective: at a dose that reduces refusal by 38 percentage points, accuracy on 1,920 exact-match items decreases by 2 points, without relying on a grader. On responses an output monitor classifies as safe, amplifying the description increases refusal from 38% to 70%, while the placebo decreases it to 25%. Four of five pre-registered textual variants reproduce the effect, while two pairs constructed from other topics do not. This supports the interpretation that the change depends on what the sentences say rather than on vector subtraction alone. Testing that required replacing the control this literature standardly uses: a random direction matched in norm carries 0.024 to 0.035 of its energy inside the subspace the activations occupy, against a fitted direction's 0.994, so it cannot separate a direction from a large disturbance of it. The scope is narrow. The written pair does not transfer to a second model family, whereas the fitted direction does, and retraining the autoencoder preserves reconstruction while almost eliminating harm-related descriptions. Every registration behind these numbers ships with the paper. Together these results give a practical test for NLA interpretations: rather than judging a description by how readable it is, write it back and test whether the intervention produces the predicted behavior.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, natural language autoencoder, activation steering, causal intervention, refusal, jailbreak, AI safety, preregistration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TCreOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

This paper investigates whether natural-language descriptions produced by Natural Language Autoencoders (NLAs) can serve as a causal intervention channel for language-model refusal behavior. The authors construct steering directions by encoding two human-written sentences describing refusal and compliance, subtracting their reconstructed activation vectors, and injecting the resulting direction into the model's residual stream.

The paper evaluates this approach through controlled interventions, dose-response analyses, semantic placebo comparisons, and tests of model competence. The results show that the language-derived direction systematically influences refusal behavior in Qwen2.5-7B, while having a relatively small effect on an exact-match task battery at the evaluated dose. The authors also investigate the limitations of norm-matched random-vector controls and report negative results concerning cross-model transfer and autoencoder retraining.

Critical strengths

The paper investigates whether natural-language descriptions of model activations can be used for causal interventions on refusal behavior. This is a relevant question for understanding the relationship between interpretability and model behavior.

Critical weaknesses
  1. Poor readability and fragmented presentation. The manuscript is difficult to follow due to its fragmented organization and experiment-log-like narrative. The core methodology, including the NLA reconstruction process, steering direction construction, and intervention procedure, is scattered across multiple sections rather than presented in a unified manner. Moreover, numerous experiment identifiers, registration details, statistical values, and cross-references interrupt the main argument. Although reporting unsuccessful experiments is valuable, the current presentation makes it difficult to identify the central contributions and understand how the different experiments support them. Substantial improvements in organization and clarity are needed.

  2. Insufficient clarity regarding the interpretation and practical significance of the causal results. Although the paper presents several intervention and control experiments, the precise conclusions supported by these results are not always easy to distinguish. In particular, changes in refusal behavior do not necessarily establish that the natural-language descriptions faithfully capture the underlying model states. The paper acknowledges this distinction, but its implications for the central contribution could be explained more clearly. In addition, the competence-preservation experiment uses substantially fewer injection positions per item than the refusal experiments, limiting the direct comparability of these results. A clearer separation between demonstrated causal effects, interpretability claims, and practical implications would strengthen the paper.

Questions for authors

Q1. Could the authors substantially improve the organization and readability of the manuscript? In particular, a unified overview of the complete NLA-based intervention pipeline, together with a consolidated table summarizing the main experiments, their settings, and key findings, would help readers better understand the methodology and evaluate the contributions.

Q2. Could the authors clarify the precise causal and interpretability claims supported by the experiments? Specifically, to what extent do the observed behavioral changes validate the semantic faithfulness of the natural-language descriptions, rather than merely demonstrating that their reconstructed vectors can influence refusal behavior? Additionally, could the authors clarify how the difference in injection schedules between the competence-preservation and refusal experiments affects the interpretation of the reported performance trade-off?

Overall Assessment: While the research question is potentially interesting and the authors attempt to provide several experimental controls, I found the manuscript difficult to follow and its central contributions insufficiently clearly presented. The concerns regarding robustness and the interpretation of the causal evidence further limit my confidence in the conclusions. I would be willing to reconsider my assessment if the authors can substantially clarify the methodology, experimental findings, and scope of their claims during the rebuttal.

6 comments
AuthorsOpenReview ↗

Response to Reviewer TCre

W1 / Q1. Organisation, a unified pipeline, and one summary table.

We agree. The next version will describe the method in one section. Here it is in four steps.

  1. A person writes two sentences. One describes the assistant refusing. The other describes it complying.
  2. The NLA reconstructor maps each sentence to a residual-stream vector at the NLA's layer (layer 20 of Qwen2.5-7B-Instruct).
  3. We subtract the two vectors and scale the difference to a fixed norm.
  4. During generation, we add α times this vector at every generated position, for 8 values of α. A grader labels each response as a refusal or not. For each item, we compute the rank correlation between α and refusal, and we report the median over items.

The draft summary table:

QuestionSetupResult§
Does the attack wrapper cause the drop in refusal?Same request with and without the wrapper, 100 requests, 9 familiesRefusal falls 45 points4.1
Does the attack change the internal state?Against a neutral wrapper of the same lengthDisplacement +4.41 [+4.15, +4.67]4.2
Does a fitted direction move refusal?Difference-in-means direction, 37 items, 8 dosesRefusal 73.0% to 2.7%4.3
Does the reading predict which items flip?Reading re-encoded, 27 itemsρ = +0.51 (raw activation +0.58)4.4
Do two written sentences move refusal?Reconstruction difference, 37 items, 8 dosesMedian ρ = +0.412 (placebo +0.000)4.5
Does it depend on what the sentences say?5 wordings, 2 topic pairs4 wordings +0.58 to +0.76, topic pairs +0.0004.5, 5.3
Do the NLA's own readings move refusal? *Readings of 183 responses, written backMedian ρ = +0.756 (shuffled labels +0.000)new
Does ActAdd on the same two sentences do the same? *Same sentences, same norm, 37 items+0.000 (reconstructor +0.612)new
Is competence kept, short answers?6 short tasks, 60 items59 of 60 correct (unsteered 60 of 60)5.2
Is competence kept, matched steering? *100 GSM8K problems, generated tokens only85% at α = −0.75 (unsteered 85%), 57% at α = −15.2
Does a probe through the same channel do the same?Logistic probe, same doses+0.000 [+0.000, +0.577], not separated5.4
Does amplifying the NLA description raise refusal on hedged responses an output monitor calls safe?60 responsesRefusal 38% to 70% (placebo 25%)6
Does the effect transfer to a second model family (Gemma-3-12B)?Fitted direction and written pairFitted transfers, written pair does not7
Do harm readings survive retraining the verbaliser?3 seedsHarm mentions 2, 1, 7 of 573 (original 339)7.3

Rows marked * are new and use layer 20, where the written pair gives +0.612. The §4.5 row uses layer 21.

W2 / Q2. What the causal results support.

Behavioural change alone does not validate the semantic faithfulness of the descriptions. It shows that the reconstructed vectors move refusal, and in the direction the sentences describe. The paper's claims fall into three groups.

Shown, causal. Adding the difference between the reconstructions of two written sentences changes graded refusal with dose: median ρ = +0.412 against a placebo's +0.000. Four of five wordings reproduce this, and two pairs on other topics do not. So the effect depends on what the sentences say. A direction built from the NLA's own readings, written back, also moves refusal (+0.756). With labels shuffled, it does not (+0.000). ActAdd on the same two sentences does not move refusal (+0.000 against +0.612).

Shown, predictive. The NLA's reading of an item predicts which items lose refusal under a fitted direction (ρ = +0.51). It does so less well than the raw activation (ρ = +0.58).

Not shown. That any single reading is a faithful description of the model's state. That the reconstructor's advantage over ActAdd holds beyond one sentence pair and one layer. That the written pair transfers to a second family. That the effect survives retraining the NLA.

The next version will state these three groups in §1 and use them in the abstract and the conclusion.

On the injection schedules: the competence test steered a median of 58 positions per item, prompt included, not about 3. The refusal sweep steers a median of 255 positions. So the 2-point figure cannot be read as the cost of the refusal intervention. The matched test in the general comment replaces it. At the dose where refusal falls 38 points, GSM8K accuracy is unchanged. The next version will report both tests in §5.2.

We have now uploaded the revised manuscript and updated supplementary archive incorporating all structural and methodological changes requested in your review:

  1. Unified pipeline and experiment index (W1 / Q1): §1.1 (Figure 1, p. 2) and §3 (pp. 3–4) present the four-step write-back pipeline and grading rule in one place, indexed by the consolidated experiment summary in Table 1 (p. 4). Pre-registration codes and bootstrap intervals are moved to Appendices A and C–E, and each empirical section opens with its primary test while caveats are consolidated once in §8 (pp. 8–9).
  2. Causal claims vs. semantic faithfulness and matched competence (W2 / Q2): §1.2 (p. 2) explicitly separates what is shown causally, what is shown predictively, and what is not shown (in particular, aggregate behavioural coupling does not establish single-reading faithfulness). §4.6 adds write-back from the NLA's own readings (+0.756 vs. +0.000 shuffled), and §5.2 (Table 2, p. 6) reports the matched 100-problem GSM8K evaluation steering generated tokens only: at α = −0.75, accuracy holds for both the written pair (85% vs. 85% unsteered, median 277 steered positions) and the readings direction (84% vs. 88% unsteered and 81% shuffled, median 247.5 steered positions), before degrading at α = −1.

We hope the reorganised manuscript and matched experiments address your concerns, and we welcome any further questions during the discussion period.

I believe this paper explores an interesting and potentially valuable research direction, and I appreciate the authors' efforts in conducting extensive experiments and providing detailed analyses. However, I still think the presentation of the paper could be substantially improved. In particular, I encourage the authors to better highlight the motivation behind the work and provide a more concise and unified description of the proposed methodology, so that readers can more easily understand the main idea and the reasoning behind the approach.

In my view, a good research paper should not only address a meaningful problem but also clearly explain why the problem matters and how the proposed method solves it. A clearer motivation, a more coherent methodological explanation, and a better-organized presentation of the experimental findings would significantly strengthen this work.

I appreciate the authors' efforts and believe the research has potential. I encourage them to continue improving the manuscript, particularly in terms of clarity and presentation. For now, I will maintain my original rating, and I hope these suggestions will be helpful in further strengthening the paper.

Reviewer_wWXPOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates whether natural language can act as a causal channel for refusal mechanisms using pretrained natural language autoencoders. The study evaluates the impact of this intervention across varying steering strengths, alternative prompt formulations, and control conditions.

Critical strengths
  • Rigorous Experimental Control: The controlled experimental framework is the paper's strongest asset. By evaluating multiple intervention strengths, alternative description pairings, and unrelated-content controls, the authors demonstrate that steering effects are strictly dependent on the semantic content of the descriptions, moving beyond simple proof-of-concept examples.

  • Scope and Transparency: The inclusion and honest reporting of negative/unsuccessful experimental results significantly help delineate the boundaries and practical scope of the proposed approach.

Critical weaknesses
  • Presentation and Narrative Structure: The paper is currently difficult to follow, making it challenging to pinpoint the central core contribution and connect individual experimental findings into a cohesive narrative. Specifically, the boundary between the inherited steering procedure and the newly introduced evaluation protocol should be articulated much earlier and more clearly. Additionally, the conceptual link between the "reading" and "writing" components requires a more direct explanation. While I generally prioritize technical substance over writing style, the current structure substantially increases the cognitive load required to evaluate the work.

  • Mismatched Capability-Preservation Setup (§5.2): The capability-preservation evaluation does not mirror the intervention length used in the refusal experiments. Steering is applied for only ~3 token positions per capability item, compared to ~255 token positions in the refusal sweeps. Consequently, the observed minor drop in accuracy offers weak evidence that general model capabilities remain intact under sustained steering. Although the authors acknowledge this discrepancy, their claim regarding the selectivity of the intervention remains stronger than what the experimental setup actually supports.

Questions for authors
  1. Extended Capability Validation: Would the capability-preservation findings hold on benign tasks requiring longer generations, where steering is applied for a comparable number of token positions at the same layer and intensity? Evaluating tasks with automatically verifiable answers (e.g., multi-step reasoning or long-form generation), benchmarked against both the unsteered baseline and the placebo intervention, would clarify whether the reduction in refusal comes at a higher capability cost than currently reported.
3 comments
AuthorsOpenReview ↗

Response to Reviewer wWXP

Presentation and narrative. We agree, and the general comment lists the five changes. Two of them answer your points directly.

What is inherited and what is new. The NLA release (Fraser-Taliente et al., 2026) supplies the verbaliser and the reconstructor. It also proposed editing activations with the difference of two reconstructed explanations, checked against a random vector of matched norm. This paper adds five things.

  1. A placebo built the same way from unrelated sentences.
  2. Two control pairs on other topics, and four other wordings of the refusal pair.
  3. Random controls drawn inside the subspace the activations occupy.
  4. A competence test scored without a grader.
  5. A test on a second model family.

The next version will state this in the second paragraph of §1.

How reading and writing connect. The verbaliser maps an activation to English. This is reading. The reconstructor maps English back to an activation. This is writing. §4.4 tests reading, and only for prediction. §4.5 tests writing, causally. The new test in our reply to SPQ5 joins the two: readings written back move refusal. It still does not show that a single reading is faithful. The next version will state this before the results.

Mismatched capability setup (§5.2). You are right, and the submission also misstated the setup. The competence test steered a median of 58 positions per item, prompt included, not about 3. The refusal sweep steers a median of 255 positions at the same dose. The test also has 60 items, not 1,920. The 1,920 are generations.

Q1. Would competence hold under long, matched generations?

We ran the test you describe, registered before generation. It uses 100 GSM8K test problems, scored automatically by the final boxed number. Steering is applied to generated tokens only, for up to 384 tokens. The model, layer and dose scale are those of the refusal sweep, at α = 0, −0.75 and −1. At α = −0.75, the median number of steered positions is 277, against 255 in the refusal sweep.

At α = −0.75, where refusal falls 38 points, accuracy is 85% for the written pair, 87% for the placebo and 85% unsteered. At α = −1, the written pair falls to 57% and the placebo to 77%. So competence holds at the dose that moves refusal and breaks one dose further. The full table is in the general comment. The next version will report both doses in §5.2.

Thank you for the additional experiments, which address my concern about the capability-preservation setup. I also appreciate the revisions to the presentation. However, I still find the narrative fragmented, with frequent shifts between results, controls, and qualifications that make the overall argument difficult to follow.

Thank you for reading the new experiments. We are glad the matched GSM8K test answers the concern about the capability setup.

We agree that the argument was hard to follow, and we have uploaded a revision that reorganises it. The section numbers are unchanged, except that the former §4.7 is folded into §8. No result, control or failed attempt was removed.

  1. Each results section now opens with its decisive test and its matched control. Earlier failed or unmatched runs follow in a short paragraph (§4.2, §5.2, §5.3, §7.1).
  2. The qualifications are stated once, in §8, instead of after each result. §3 defines the pipeline and the grading once.
  3. §4.5 reports the written pair alone. All controls on what the sentences say are together in §5.3.
  4. §7.1 compares the fitted direction and the written pair on the same Gemma-3-12B requests.
  5. We also ran the matched GSM8K test for the direction built from the NLA's own readings. It is in §5.2 and Appendix E.4.

The argument now runs in order. §4.1 to §4.4 show that the attack lowers refusal and shifts the internal state before the response. They also show that a fitted direction moves refusal and that the NLA's reading predicts which items flip. Two sentences a person wrote, written back through the reconstructor, move refusal, and a placebo built the same way does not (§4.5). A direction built from the NLA's own readings also moves refusal, and the same readings with shuffled labels do not (§4.6). At α = −0.75, matched GSM8K accuracy holds for both directions (§5.2). It is 85% against 85% unsteered for the written pair, and 84% against 88% unsteered for the readings direction. The effect follows what the sentences say (§5.3). A logistic probe is not separated from the written pair (§5.4), and ActAdd on the same sentences does not move refusal (§5.5). On hedged responses that an output monitor passes as safe, injecting the written pair raises refusal (§6). §7 and §8 state the limits. The written pair does not transfer to Gemma-3-12B, steering comparisons need in-subspace controls, and retraining the verbaliser almost removes harm descriptions. Moving refusal in aggregate also does not show that any single reading is faithful.

Reviewer_SPQ5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper uses the existing NLA reconstructor to convert human-authored refuse and comply sentences into vectors. The difference between these vectors is then injected into the model to modulate its refusal behavior. The authors evaluate the effectiveness of this intervention at different injection strengths using a range of control experiments, while also examining its impact on task performance and its transferability across different models.

Critical strengths

This paper focuses on how to examine NLA-based descriptions of a model’s internal states through behavioral interventions. This is a meaningful research question.

Critical weaknesses
  1. The intervention direction in §4.5 is generated via a reconstructor based on two sentences written by a human author, rather than being derived from the NLA reading of the intervened activation. Therefore, this only demonstrates a causal connection between "writing" and behavior, but does not prove that the same holds true for "reading".
  2. As acknowledged in §5.2, the refusal sweeps and the task battery involve interventions at approximately 255 and 3 token positions per item, respectively, indicating a clear mismatch in the number of intervention sites. Using these results, obtained under different intervention conditions, to support the abstract’s claim that "the change is selective" is not well justified.
Questions for authors

What evidence establishes the added value of the NLA reconstructor over ActAdd under matched settings, given that both approaches can construct steering directions without labeled data?

2 comments
AuthorsOpenReview ↗

Response to Reviewer SPQ5

W1. §4.5 shows that writing moves behaviour, not that reading is faithful.

We agree. The direction in §4.5 comes from two sentences a person wrote, and the verbaliser plays no part in building it. So §4.5 tests the reconstructor: when a description is mapped into activation space, does behaviour move the way the description says? It does not test the verbaliser's readings.

The submission's only evidence about reading is in §4.4, and it is predictive, not causal. There, the reading of an item, re-encoded by the reconstructor, predicts which items lose refusal under the fitted direction (ρ = +0.51). The raw activation predicts better (ρ = +0.58).

The test your point calls for builds the direction from the NLA's own readings. The submission does not contain it, so we ran it. We took the verbaliser's readings of 183 attack responses. The output grader had labelled 104 of them as refused and 79 as complied. These 183 exclude every response to a request that also appears among the 37 test items. The reconstructor mapped each reading back to an activation, and the direction is the difference between the two class means. The control uses the same readings with the labels shuffled. Both were injected exactly as in §4.5.

DirectionMedian ρItems with ρ > 0
From the readings+0.75632 of 37
Same, labels shuffled+0.0006 of 37

The paired difference is +0.756 (95% CI +0.73 to +0.78). At α = −0.75, 29 of 37 responses comply and none refuse. Without steering, 25 refuse. So the readings, taken together, carry information that moves refusal when written back. This does not show that any single reading is faithful. The direction has a cosine similarity of +0.61 with the difference-in-means refusal direction and −0.02 with the written direction. We have not run the competence test with it.

The next version will add this test to §4. The abstract and the conclusion will state the written-pair result and the reading result as two separate claims.

W2. The competence claim rests on mismatched steering.

We agree in part, and the submission also stated the mismatch wrongly. The competence test steered every position, prompt included: a median of 58 positions per item, not about 3. The refusal sweep at the same dose steers a median of 255 positions. The mismatch is about fourfold, and it left the claim unsupported.

The matched test is in the general comment. At the dose where refusal falls 38 points, accuracy on 100 GSM8K problems stays at 85%. One dose further, it falls to 57%. The abstract will state the competence result only for α = −0.75.

Q. What does the reconstructor add over ActAdd under matched settings?

The submission has no matched comparison, so we ran one. You are right that both methods build a direction from two prompts without labels. They differ in one way. ActAdd uses the activations the model produces as it reads the two sentences. Our direction uses the output of the reconstructor, which was trained to map a description to the activation it describes.

ActAdd here is the difference between the model's mean activations over each sentence at layer 20. We excluded the first token and the token carrying the attention sink, whose norm is over 100 times that of the others. The direction is scaled to the same norm and run on the same 37 items and 8 doses, with the same grader.

Direction, same two sentencesMedian ρItems with ρ > 0
Reconstructor+0.61225 of 37
ActAdd+0.00010 of 37

The paired difference (ActAdd minus reconstructor) is −0.612 (95% CI −0.64 to −0.14). At α = −0.75, 23 of 37 responses still refuse under ActAdd, against 11 under the reconstructor and 25 without steering. The two directions have a cosine similarity of +0.26. ActAdd built from two unrelated sentences also gives +0.000. For that control, the strongest dose was excluded by the registered fluency rule, with 11 of 37 responses degenerate. So with this sentence pair at this layer, the reconstructor's direction moves refusal and ActAdd's does not. One pair at one layer does not show that this holds in general. The next version will report this comparison in §5.

We have now uploaded the revised manuscript and updated supplementary archive incorporating all pre-registered experiments addressing your review, plus an additional matched competence check (P10c) closing the remaining question in our earlier reply:

  1. Readings written back (W1): §4.6 and Appendix E.2 report the direction reconstructed from the verbaliser's own readings of 183 refused and complied responses (median ρ = +0.756, against +0.000 with shuffled labels; 29 of 37 test items comply and 0 refuse at α = −0.75).
  2. Matched competence for both write-back directions (W2): §5.2 (Table 2) and Appendix E.4 report the 100-problem GSM8K evaluation steering generated tokens only. At α = −0.75, the human-written pair holds at 85% against 85% unsteered and 87% placebo (median 277 steered positions vs. 255 in the refusal sweep), and the §4.6 readings direction similarly holds at 84% against 88% unsteered on A100 and 81% shuffled (median 247.5 steered positions; paired difference vs. unsteered −4 points, 95% CI [−11, +3]), before both degrade one dose further at α = −1 (57% and 18%).
  3. Matched ActAdd comparison (Q): §5.5 and Appendix E.3 report ActAdd on the identical two sentences and dose ladder at layer 20 (+0.000 vs. +0.612 for the reconstructor), with explicit scope boundaries consolidated in §8.

Please let us know if the revised manuscript and added experiments resolve your concerns regarding soundness, or if any question remains that we can clarify during the discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Two corrections, three new experiments, and the changes for the shared concerns

We thank the three reviewers. Three concerns are shared, so we answer them once here. Each review also gets its own reply.

Two corrections to §5.2. Both concern the competence test, and both were stated wrongly in the submission.

  1. The test has 60 items, not 1,920. The 1,920 are generations: 60 items, 8 doses and 4 conditions. At the dose where refusal falls 38 points, the written pair answers 59 of 60 items correctly. Without steering, it answers 60 of 60.
  2. The test steered every position, prompt included. That is a median of 58 positions per item (range 52 to 84), not about 3. At the same dose and layer, the refusal sweep steers a median of 255 generated positions.

So the mismatch the reviewers point out is real, but it is about fourfold rather than the eightyfold the text implied. The answers were also scored by a whole-word match of the gold answer in a response of up to 32 tokens, not by exact match. No grader is involved either way.

Three new experiments. We registered each one before any generation. All three use Qwen2.5-7B-Instruct at layer 20. The two refusal tests use the same grader as the paper (gpt-oss-120b, 3 votes), and the competence test uses no grader. Layer 20 is where the paper's wording controls and competence test were run. There, the written pair gives a median per-item ρ of +0.612 (Appendix C.1).

Shared concern 1: competence under matched steering (all three reviewers). We agree that the test in §5.2 did not support the sentence "the change is selective". We ran the matched test Reviewer wWXP proposed. It uses 100 GSM8K test problems, scored by the final boxed number. Steering is applied to generated tokens only, as in the refusal sweep, for up to 384 tokens. At α = −0.75, it steers a median of 277 positions, against 255 in the refusal sweep.

Directionα = 0α = −0.75α = −1
Written pair85%85%57%
Placebo85%87%77%

At α = −0.75, where refusal falls 38 points, accuracy does not change (0 points, 95% CI −7 to +7). At α = −1, it falls 28 points, against 8 points for the placebo. This is the dose where the paper already reports degenerate text on 17 of 37 refusal items. The next version will report both doses in §5.2, and the abstract will state the competence result only for α = −0.75.

Shared concern 2: reading versus writing (SPQ5, TCre). The direction in §4.5 is built from sentences a person wrote, not from the NLA's own readings. So §4.5 tests only the reconstructor. We ran the missing test: a direction built from the verbaliser's readings and written back by the reconstructor. It moves refusal (median ρ = +0.756). The same construction with shuffled labels does not (+0.000). The reply to SPQ5 gives the details. The next version will add this test to §4 and state what it does not show.

Shared concern 3: presentation (wWXP, TCre). The next version will make five changes.

  1. The second paragraph of §1 will state what the NLA release already proposed and what this paper adds.
  2. The whole method will be described in one section, with one figure of the pipeline.
  3. One table will list every experiment with its question, setup, sample size, result and section. A draft of it is in our reply to TCre.
  4. Registration identifiers, finding codes and most cross-references will move from the main text to Appendix A.
  5. Intervals and per-grader values will move to the appendices. The main text will keep one number and its comparison per claim.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.