acceptodds
Under review as a conference paper at ICLR 2027

Relational Steering with a Fixed Direction: Causal Effects and Behavioral Limits

Abstract

Can a single fixed activation direction produce opposite but appropriate effects across contexts? We study this question with a factorial design crossing epistemic alignment or distortion with favorable or unfavorable outcomes. In Qwen3-8B-Base and Llama-3.1-8B, we construct a relational direction from the factorial interaction contrast, which cancels the designed additive main effects, freeze it before held-out evaluation, and test causal writability through teacher-forced consequence preferences. The aggregate relational effect transfers to independent novel domains in both model families, although strict branchwise transfer is established only in Llama. Structured controls show that response magnitude alone does not explain the effect, and a discovery-selected best-of-1,024 Gaussian direction shows the branchwise sign pattern is not unique to the factorial direction, which nevertheless has a substantially larger held-out local response. Controlled-readout writability does not translate monotonically into open-ended behavior: under prompt-prefill-only intervention in Llama, the distortion branch is stronger under teacher forcing, whereas the alignment branch has the larger and more persistent trajectory-average behavioral effect, a pattern reproduced by an independent automated audit and, for the average effect, by blinded human calibration. Relation controls further constrain interpretation. Removing the stated causal linkage weakens steering in both model families, whereas explicit reversal does not produce the opposite joint pattern; a post-hoc baseline-conditioned diagnostic shows the Llama reversal response retains a default-oriented component not explained by simple amplification of the contemporaneous preference, while Qwen is substantially more baseline-tracking under the same manipulation. These results establish context-responsive relational writability while delimiting what it implies: the observed sign structure is not unique to the factorial direction, controlled writability does not guarantee persistent behavioral expression, and relation sensitivity does not by itself establish mechanism-sensitive causal-rule recomputation.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, representation engineering, activation steering, causal intervention, linear representations, world models, language models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_roQXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether a single fixed activation direction can shift the same outcome-preference readout in opposite directions depending on an accurate or distorted epistemic premise. A factorial interaction over premise accuracy and consequence valence constructs the direction. The direction and intervention layer are selected on discovery domains and frozen before evaluation on 24 novel domains. Validation combines teacher-forced likelihood contrasts, free-generation judgments, random-direction comparisons, and changes to the stated causal relation. Llama-3.1-8B supports held-out two-branch writability; Qwen3-8B-Base transfers the aggregate effect but fails the finite-dose distortion-branch test. In Llama, generation yields a different branch ordering from the controlled readout, while mechanism reversal does not jointly reverse steering. The contribution is an operational demonstration of conditional steering and its limits; the paper does not establish a unique relational representation or reliable causal-rule recomputation.

Critical strengths

S1. A controlled intervention with credible held-out evidence. The factorial construction cancels the designed additive main effects, and the use of shared clauses makes the comparison more controlled than independently written positive and negative prompts (Eq. (1), p. 3; Table 3, p. 14). More importantly, both branches use the same favorable-minus-unfavorable readout and must change in opposite directions. This criterion distinguishes the claimed interaction from a generic valence shift. Discovery-only layer selection and the frozen intervention provide a meaningful separation between construction and evaluation (Appendix C.1, p. 21). On the 24 novel Llama domains, the branch-oriented slopes are S_A = 0.587 and S_D = 1.073, with a positive aggregate effect in all 24 domains (Table 1, pp. 4–5). Domain-level inference avoids treating seeds or checkpoints as independent replications. These results are the main reason I support acceptance: they establish the central fixed-intervention existence claim without requiring a complete mechanistic interpretation. S2. Falsification tests establish useful limits on interpreting steering. The best-of-1,024 Gaussian comparison selects a direction and orientation on discovery data, then evaluates the frozen winner on novel domains. A nearly orthogonal winner also passes the sign criterion, while the factorial direction retains a local worst-branch advantage of 0.349, with a conditional selection-aware interval of [0.134, 0.676] (Tables 2 and 11, pp. 4 and 30). The authors consequently distinguish intervention efficacy from directional uniqueness. The relation-reversal and generation analyses similarly constrain stronger interpretations: changing an explicit mechanism does not reliably reverse the intervention, and controlled likelihood changes do not guarantee the same ordering or persistence in generation. Rescoring by a second automated judge and the human calibration make the measurement limitations visible (Appendix D). These findings give the study value beyond a favorable steering result by showing which behavioral and representational conclusions that result can support.

Critical weaknesses

W1. The behavioral conclusions depend on comparisons that change dose and measurement. The controlled result uses slopes over intervention strengths from −1 to 1, whereas generation uses endpoints at ±1.5 (Eq. (3), p. 3; Table 7, p. 25). The outcome also changes from likelihoods of two specified consequences to a five-level judgment of unrestricted text. Thus the reversal in branch ordering does not isolate the transition from teacher forcing to generation. The local derivative analysis supports the controlled ordering, but does not supply a comparison at the generation dose; the discovery-domain token-normalization analysis also does not establish the novel-domain ordering under that alternative score. Persistence is additionally sensitive to the measurement instrument. On the identical calibration subset, distortion attenuation is −0.271 for the human rater and +0.042/+0.104 for the automated judges (Table 9, p. 28). The authors appropriately acknowledge imprecision in the small human sample. Together, these issues limit how strongly the behavioral inversion and persistence can be interpreted. They prevent a stronger recommendation, although they do not invalidate the held-out controlled intervention result. W2. The intervention is better established than its relational semantic specificity. The factorial contrast removes additive main effects but retains interactions correlated with semantic congruence, expectedness, or discourse coherence. The noncausal comparison reuses clauses, yet changing their linkage also changes how appropriate the consequences are. It therefore does not cleanly distinguish an epistemic–outcome interaction from a broader congruence signal. The manuscript acknowledges the missing unrelated congruence comparison (Appendix E.4, p. 35). Moreover, the relation-control statistics come from discovery domains after default-condition layer selection, rather than providing another independent held-out test (Table 12, p. 31). This matters to significance and interpretation: the evidence supports a reproducible intervention more strongly than a specifically relational representation. The nearly orthogonal Gaussian direction further demonstrates that the branchwise sign criterion does not uniquely identify the factorial direction, although this result does not itself establish a semantic-congruence explanation. The factorial direction's stronger held-out response remains an informative positive finding. I regard this as a material limit on the explanatory claim, not a requirement that the authors prove directional uniqueness. The narrower existence result and the explicit analysis of its limits remain sufficient for weak acceptance.

Questions for authors

Q1. Focused validation — compare the controlled readout at the generation dose. Does Llama's novel-domain distortion-over-alignment ordering persist for teacher-forced contrasts at ±1.5? A comparison at the generation dose, with mean-token scoring as a length-weighting check, would clarify whether the behavioral inversion survives these direct alternatives. This would affect how strongly I interpret the controlled-versus-generated comparison in W1. Q2. Clarification — explain the disagreement about persistence. What explains the opposite human and automated attenuation directions on the identical subset in Table 9? Please specify the early/late checkpoint sets and weights, and explain whether the existing scoring protocol or disagreement cases account for the difference. This would affect my confidence in the persistence interpretation, beyond the better-supported average behavioral direction. Q3. Clarification — identify the domains used to select the behavioral dose. Appendix D.1 states that the final generation seeds differ from the pilot seeds, but does not specify the pilot domains. Which domains and scores were used to choose |α| = 1.5? If they include the final novel domains, please distinguish confirmation with new seeds from generalization to domains unused for dose selection. This would clarify the independence of the behavioral evidence, without changing the separate freezing of the axis and layer. Q4. Focused validation — distinguish relational structure from general congruence. Would one unrelated congruent/incongruent control direction reproduce the branch pattern under the existing evaluation? Selecting it on discovery data and freezing it for novel domains would address the alternative interpretation in W2. The result would affect the semantic specificity I attribute to the contribution; establishing uniqueness among all directions is unnecessary.

3 comments
AuthorsOpenReview ↗

Response to Reviewer roQX (1/2)

We thank the reviewer for distinguishing the well-supported fixed-direction intervention result from stronger interpretations of semantic specificity and behavioral persistence. The revised manuscript addresses Q1–Q4 in §4, §7 and Appendices D, H and I; Q3 also led to a correction (Appendices C.2, C.4 and D.1; see the General response).

Q1 — Matched-dose teacher-forced readout. We conducted a post-review evaluation on the same 24 held-out Llama domains with the frozen block-13 direction, the original candidate consequences, and the prompt-only RMS-scaled intervention. Before evaluating the generation-matched dose, we replayed the original five-dose grid and reproduced the submitted summary statistics exactly (absolute difference 0.0; identical medians, sign counts, and bootstrap intervals). At α = ±1.5, oriented endpoint secants under total candidate log-likelihood are T_A = 0.568 and T_D = 1.008, so G = T_A − T_D = −0.440 (95% domain-bootstrap CI [−0.829, −0.048]; two-sided exact sign-flip p = 0.042). The distortion-over-alignment ordering thus persists at the generation dose under this controlled score. Length-normalized mean-token scoring provides weaker evidence: on the original grid G = −0.027 (bootstrap CI [−0.052, −0.001], but two-sided exact p = 0.061); at ±1.5 G = −0.023 (CI [−0.048, 0.002], p = 0.094). The bootstrap interval and exact test differ at the original-grid boundary, and the generation-dose comparison is inconclusive under mean-token scoring (Table 19). The original prefill-only generation audit instead yields a larger rated accurate-side trajectory effect (Q_traj,A = 0.201, Q_traj,D = 0.050; B_traj = 0.151, CI [0.039, 0.279], one-sided p = 0.0094). Dose mismatch alone therefore does not account for the ordering under total likelihood, but score normalization, specified candidates versus sampled text, and judge calibration still differ. We do not claim a mechanism (§4; Appendix H).

Q2 — Human versus automated attenuation. Checkpoints are u ∈ {0.2, 0.4, 0.6, 0.8, 1.0}, with Q_early = [Δ(0.2) + Δ(0.4)]/2, Q_late = [Δ(0.8) + Δ(1.0)]/2, and C_atten = Q_early − Q_late. The middle checkpoint is excluded, weights are equal, and the late window includes the terminal checkpoint, so C_atten and Q_terminal are not independent. On the identical 480-item subset, distortion attenuation is −0.271 (human, 95% CI [−0.646, 0.125]), +0.042 (GPT-5.6 Sol, [−0.167, 0.250]) and +0.104 (Gemini-3.6 Flash, [−0.104, 0.313]); all three intervals include zero, so the subset cannot resolve the direction of attenuation. Neutral-score rates are 30.6%, 62.9% and 70.4%. Removing the terminal checkpoint from the late window moves the human estimate to −0.021, locating much of the difference in terminal judgments without explaining it. The human sample has one external rater and one matched seed per domain and branch, and the terminal checkpoint often reflects the 192-token cap rather than a natural ending. We now describe fine temporal attenuation as instrument-sensitive and secondary to the more robust trajectory-average direction (§4; Appendix D.7).

AuthorsOpenReview ↗

Response to Reviewer roQX (2/2)

Q3 — Behavioral-dose selection and independence. Thank you; this question led to a correction. Several passages of the submitted version (§4, §7, Appendices C.2, C.4 and D.1, Tables 5–7, and the reproducibility statement) incorrectly described a separate Llama dose-selection pilot. In fact, |α| = 1.5 came from an exploratory Qwen 10-seed pilot on the same 24 novel-domain contexts (IV01–IV24) (seeds 520000–520009), comparing 0.5, 1.0 and 1.5 with the weakest-branch criterion min(Q_A, Q_D). There was no Llama dose-selection pilot. The Llama manifest fixed 1.5 and the final seeds 520010–520039 before any Llama outcome was inspected (an author-recorded chronology; Appendix C.9). The Llama direction and layer were selected on discovery domains only and the behavioral seeds were new, but the dose was not selected on domain-disjoint evidence. To answer directly: the dose-selection domains were the final novel domains, evaluated on a different model. The behavioral evidence is therefore new-seed evaluation at a cross-model, pilot-selected dose, not new-domain validation of dose selection. All affected passages have been corrected.

Q4 — Unrelated congruence alternative. We agree that the interaction contrast cancels designed additive effects without excluding generic valence matching, expectedness, or discourse coherence. For a reviewer-requested control, we prepared 48 epistemically irrelevant positive/negative premises in the 24 discovery domains (same contexts, connectors and outcome clauses) and 48 more in 24 disjoint manipulation-check domains. Two non-target automated judges from different providers (GPT-5.6 Sol and Gemini-3.6 Flash) each rated 144 items, 96 control premises plus 48 original epistemic premises as references, for 288 single-item evaluations under criteria fixed before the ratings. Only 86/96 controls passed, against a required 96/96.

The reference-item ratings also revealed an affective contrast. Both judges rated every original "accurate" premise mildly positive (+1) and the "distorted" premises more negatively (GPT: −2 for all 24; Gemini: −2 for 11, −1 for 13), so the original epistemic contrast carries a substantial affective contrast that the factorial construction does not isolate. These ratings describe the stimuli, not what the direction encodes inside Llama. The proposed controls had weaker valence contrasts (GPT 2.21 vs. 3.00, paired difference −0.79 [−0.96, −0.63]; Gemini 1.88 vs. 2.46, −0.58 [−0.83, −0.33]) and much less natural narratives (about 1.6–2.0 vs. 3.3–3.5 on a five-point scale; descriptive contrasts in Table 21).

Our pre-fixed rule blocks a control direction when screening fails. We chose not to issue a revised set during the discussion period: even strongly worded outcome-irrelevant premises (e.g., hideous, repellent) were rated only mildly negative by both judges (Table 20 includes one such example), so a revision within this design would likely remain a weaker valence manipulation, and our prewritten interpretation rule treats a null result from a weaker-valence control as weak evidence. These are limitations of the proposed stimulus design, not evidence that all unrelated-valence controls would be ineffective. No control direction was constructed or evaluated, and no target-model inference used the control stimuli. Appendix I reports this as an unsuccessful feasibility audit, and §7 retains generic congruence as an unresolved alternative.

Accordingly, the central claim remains narrow: one frozen direction produces opposite signed changes in the same teacher-forced outcome-preference margin across premise conditions in held-out Llama domains. It does not identify a unique or specifically epistemic–causal internal direction, and the relation-control results do not establish causal-rule recomputation.

Thank you for the detailed responses. I have no further questions at this stage.

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

The paper examines how to steer activations/responses in a context-sensitive way, rather than in a fixed, uniform manner. Typical steering methods have a uniform impact across contexts; this paper develops a method to steer a network in a context-/content-specific way. The method developed in this paper mostly works, though it does not lead to all of the downstream impacts that one might hope or expect

Critical strengths
  • The method developed addresses an important real-world challenge (since the required “nudge” is often context-specific)
  • The paper provides a number of different analyses, including examination of downstream behavioral effects of the steering intervention
  • The conclusions are appropriately tempered, acknowledging ways in which the relational steering fails to have the same degree of influence as in non-relational cases
Critical weaknesses
  • The paper only considers two models (Llama & Qwen). Given the differences that were found between the models, it would have been helpful to have tested more models, or at least, identified potential explanations for why the two models had significant differences.
  • The problem is an important one, but the paper does not do a particularly good job of justifying its value. It would be helpful to have a continuing example (perhaps a toy one) in the paper. Similarly, the Related Work framing is very short.
Questions for authors

No questions that would materially affect assessment

1 comment
AuthorsOpenReview ↗

Response to Reviewer NvgV

We thank the reviewer for recognizing the importance of context-sensitive intervention and our downstream analyses. We agree that the motivation, running example, and relationship to prior work needed to be more accessible. The revised Introduction (§1) now begins with a concrete question: can one unchanged signed activation direction move the same outcome-preference readout in opposite directions under two premise conditions, without a gate or sign switch? A positive aggregate response is insufficient if either branch moves the wrong way.

A continuing example. We moved the existing maintenance-record item from Table 3 into the Introduction and reused its accurate/distorted records and coordinated/conflicting outcomes in the factorial explanation (§2). The same example clarifies why the controlled experiment scores two specified consequences, whereas in §4 the model freely generates continuations that are subsequently rated. This is an explanatory reuse of an existing discovery-domain item, not a claim that this item was part of the held-out behavioral evaluation. The interaction removes designed additive effects but does not guarantee a uniquely epistemic representation.

What the model differences establish. On 24 held-out domains under the shared factorial grammar, Llama passes the strict two-branch criterion (S_A = 0.587, S_D = 1.073), whereas Qwen has a positive aggregate (I = 1.517) but does not pass the distortion-branch test (S_D = 0.276, p ≈ 0.097). We cannot attribute this difference to training, architecture, or localization without controlled comparisons. Existing diagnostics give only partial, non-causal hints: Qwen's weaker distortion branch is not explained by loss of the injected perturbation downstream (Appendix F.2), and under explicit reversal Qwen tracks its unsteered preference more closely than Llama (Appendix E.4). In Llama, distortion dominates the teacher-forced total-likelihood readout, while accurate premises yield a larger rated trajectory-average effect under prefill-only generation. A post-review ±1.5 controlled comparison preserves distortion dominance under total likelihood (T_A − T_D = −0.440, CI [−0.829, −0.048]); under mean-token scoring the ordering is borderline (Appendix H). These distinct regimes remain a limitation, not a diagnosed mechanism.

Positioning and revisions. We position the fixed direction as a diagnostic probe of what one unchanged intervention can and cannot do, not as a deployable controller; adaptive methods remain the right tools for reliable control. Section 6 now situates our work relative to (i) contrastive steering, input sensitivity, and adaptive controllers, (ii) task/function and relational representations, (iii) identifiability and construct validity, (iv) controlled-readout versus generated-answer discrepancies, and (v) causal interpretation and content effects. Prior work already establishes that steering can be context-sensitive; our narrower addition is the factorial construction plus a held-out two-branch opposite-sign criterion on one common outcome margin, together with tests of its limits. Section 7 foregrounds two models, 48 synthetic domains, the shared grammar, prefill-only generation, and the unresolved generic-valence/discourse-coherence alternative. We do not assert a unique epistemic direction or causal-rule recomputation. Please also see the General response for a correction to the description of the behavioral dose selection.

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

The paper studies if it is possible to construct a single steering vector that can influence the model behaviour in the opposite ways dependent on the context. The studied setting is a model preference of a favourable outcome given the context with either accurate or distorted information: the model should prefer the favourable outcome given accurate information in the context, and prefer the unfavourable outcome given distorted information in the context. Experiments are done on a set of 48 synthetic domains and two models: Llama 3.1-8B and Qwen3-8B. Results suggest that some context-dependent steering effect exists, but results vary a lot across models and evaluation settings.

Critical strengths
  • The studied question of a possibility of constructing a single steering vector that can influence the model behaviour in the opposite ways dependent on the context is interesting and meaningful
  • Experimental settings include diverse evaluation settings: teacher-forced outcome preferences, open-ended generation behavior, removal of causal relationship, reversal of the relationship
Critical weaknesses
  • The paper is extremely hard to read. I tried my best to understand the methodology and experimental settings and results, but I am still not confident that I correctly understand all the aspects.
  • Overall, the presented results seem closer to a preliminary proof-of-concept, rather than a fully developed research paper. Experimental coverage is limited: 2 models, 48 synthetic domains and only one pair of (concept, context). Results are inconsistent across models and evaluation settings. Although the authors investigate some of these inconsistencies, the analysis does not establish any convincing explanation. Behavioural analysis relies on LLM judges with only 1 human assessor, whose assessment does not agree well with that of LLM judges. The human evaluation also does not reproduce some of the reported temporal effects.
  • The construction of a steering vector is logical, but it is unclear if it represents the intended meaning. It may instead happen to capture context-related lexical cues or level of disagreement between contextual cues. The authors evaluate the constructed steering vector against several controls, but it does not provide convincing support for the intended meaning.

Overall, the paper is extremely poorly presented and the experimental evidence is far from sufficient to support the presented conclusions. I do not think that it is possible for the authors to adequately address all the limitations during the rebuttal.

Questions for authors
  • Can you expand analysis to more models, concepts and context variations?
  • Can you provide convincing explanation of why inconsistencies across models and evaluation settings occur?
  • Can you provide thorough analysis of what the constructed steering direction represents?
4 comments
AuthorsOpenReview ↗

Response to Reviewer WJnP (1/2)

We thank the reviewer for the careful and critical reading. We agree that the original presentation obscured the research question, and that two 8B models, 48 synthetic domains, and one accuracy-by-outcome construct do not support broad deployment or semantic-generalization claims. We have revised the Introduction (§1), Figure 1, and the method exposition (§2) around a narrower operational test: can one fixed signed activation intervention, with no premise-dependent sign switch or gate, increase the same favorable-minus-unfavorable outcome margin under an accurate premise and decrease it under a distorted premise? A factorial interaction cancels the designed additive main effects; the intervention-based opposite-sign test provides the empirical evidence. The maintenance-record example (Table 3) now connects the 2×2 construction in §§1–2 with the distinction between candidate scoring and unrestricted generation in §4. Evaluating each branch separately is what exposes the paper's main findings: aggregate transfer can hide a failing branch (Qwen), and a controlled readout need not predict the ordering of behavioral effects (Llama).

Q1 — Generality across models, concepts, and contexts. We have not added further model families or concept pairs during discussion. The 24 held-out domains assess transfer to new semantic domains within the same factorial grammar, not to novel grammars or naturalistic prompts. The two model families are not equivalent successes. In Llama-3.1-8B, the held-out branch-oriented effects are S_A = 0.587 (p = 4.8×10⁻⁵) and S_D = 1.073 (p = 7.2×10⁻⁷), meeting the strict two-branch criterion; the aggregate I = 1.660 is positive in all 24 domains. Qwen3-8B-Base shows an aggregate I = 1.517, but its distortion branch (S_D = 0.276, p ≈ 0.097) fails the strict criterion. Sections 3 and 7 present this as a held-out demonstration of opposite-sign effects under one fixed intervention, bounded to the tested factorial grammar rather than broadly reliable control.

Q2 — Model, readout, and judge inconsistencies. We cannot establish the mechanistic cause of the Llama–Qwen difference, and the dominant teacher-forced branch itself changes between discovery and held-out domains. We can address a more focused alternative for Llama's controlled-versus-generated ordering. A post-review comparison on the same 24 held-out domains at the generation dose α = ±1.5, reusing the frozen direction and candidates, retains distortion dominance under total candidate log-likelihood (T_A − T_D = −0.440, 95% CI [−0.829, −0.048], two-sided p = 0.042). Under mean-token scoring the ordering is borderline at both doses (−0.027 on the original grid, CI [−0.052, −0.001]; −0.023 at ±1.5, CI [−0.048, 0.002]; Appendix H). By contrast, prefill-only generation has the larger judge-rated trajectory-average effect in the accurate branch (Q_traj,A = 0.201 vs. Q_traj,D = 0.050; difference 0.151, 95% CI [0.039, 0.279]). A dose difference alone therefore does not explain the ordering under total likelihood; different readouts, generated texts, length weighting, and judge calibration remain, and we do not identify the mechanism.

The human calibration involves one external rater and 480 matched items, not a multi-rater replication. On this subset, distortion attenuation is −0.271 for the human and +0.042/+0.104 for GPT/Gemini, with all three 95% CIs including zero; human–automated exact agreement is 41.7%/39.4%. We therefore treat fine-grained persistence as measurement-sensitive and do not claim human confirmation of automated attenuation (§4; Appendix D.7).

AuthorsOpenReview ↗

Response to Reviewer WJnP (2/2)

Q3 — What does the direction represent? The factorial interaction and within-domain clause reuse constrain additive effects and wording variation, but they do not exclude nonadditive lexical cues, general valence matching, expectedness, or discourse coherence. The outcome-main-effect vYv_Y is not an unrelated congruence control. A discovery-selected, nearly orthogonal Gaussian direction also passes the sign criterion, although the factorial direction has a larger held-out local weakest-branch response (0.565 vs. 0.216; post-hoc selection-aware difference CI [0.134, 0.676]). Removing the stated causal link weakens discovery-domain responses but also changes discourse coherence; explicit rule reversal fails to produce the jointly reversed pattern. A post-hoc baseline-conditioned diagnostic further suggests that Llama's reversal response is not well explained by simple amplification of its contemporaneous unsteered (α = 0) outcome preference, although the semantic source of this response remains unresolved (§5; Appendix E.4). None of these results identifies a pure epistemic–causal representation (§§3, 5). Section 6 now discusses identifiability, construct validity, and content effects alongside prior context-sensitive steering.

On the "level of disagreement between contextual cues" you raised: we prepared a post-review valence-matching control, and its pre-fixed stimulus screen failed (86/96 control premises passed). Its most informative result supports your concern: both automated judges rated every original "accurate" premise mildly positive (+1) and the "distorted" premises clearly negative (−1 to −2), so the original epistemic contrast also carries an affective contrast. The proposed control premises had weaker valence contrasts and much less natural causal narratives, and we did not issue a revised set because even strongly worded outcome-irrelevant premises were rated only mildly negative. No control vector was constructed or tested. This is an unsuccessful stimulus feasibility audit, not evidence for or against a congruence direction (Appendix I, examples in Table 20; contrast diagnostics in Table 21), and §7 leaves semantic specificity unresolved.

Correction. As described in the General response, the behavioral magnitude |α| = 1.5 was selected in an exploratory Qwen pilot using the same IV-domain contexts and fixed before any Llama outcome was inspected; there was no separate Llama dose-selection pilot (§4; Appendices C.2, C.4 and D.1).

Our supported contribution is therefore a held-out, fixed-intervention two-branch readout effect in Llama, aggregate transfer in Qwen, and a measurement-sensitive account of its behavioral limits, not a general-purpose or semantically unique controller.

Reviewer_WJnPOpenReview ↗

Response to the rebuttal

I thank the authors for their responses and revision of the paper. The authors clearly state that they cannot resolve some of my concerns, and one of them seems to have been supported (on the level of disagreement between contextual cues). Next, authors changed the position of the paper to a "a narrower operational test", which is a substantial change, in which case a papers needs to be re-written and resubmitted to a full new review cycle. Overall, I find the paper in its current form substantially weak, so I maintain my score.

AuthorsOpenReview ↗

Clarification regarding the scope of the revision

Dear Reviewer WJnP,

Thank you for your follow-up and for reconsidering the revised manuscript. We respect your assessment and agree that the semantic source of the steering effect remains unresolved. In particular, the affective-valence contrast identified in our post-review stimulus audit reinforces the relevance of your concern, although it does not establish that generic valence matching causes the observed intervention effects.

We would like to clarify one point regarding the characterization of the revision as a substantial change in the paper's research question.

The original submission already asked whether one fixed activation direction, applied with the same sign, could produce opposite effects across premise conditions without switching directions or gating the intervention (original Introduction, lines 45–50). Its first contribution also explicitly introduced the factorial direction and branchwise sign criterion for testing this property.

The submitted abstract already reported the main limits: the sign pattern is not unique to the factorial direction, strict branchwise transfer holds only in Llama, and explicit reversal does not produce the opposite joint pattern. It concluded that the results establish relational writability "while delimiting what it implies"; the submitted §5 likewise stated that a generic semantic-congruence account had not been exhaustively excluded. This question, these limits, the factorial construction, and the primary experimental results are unchanged in the revision. Our earlier wording, "revised ... around a narrower operational test," was imprecise and may have suggested a change in the research question. We intended to describe a clearer articulation of the evidential scope of the original test, not a replacement of that test.

The revision also makes three changes: a factual correction concerning behavioral dose selection, a post-review matched-dose analysis (Appendix H), and an unsuccessful stimulus-feasibility audit (Appendix I). The latter two are explicitly identified as retrospective. These changes do not alter the original representation-level hypotheses, the frozen intervention-direction and layer selection, or the primary held-out branchwise results.

We understand that the remaining limitations may still affect your assessment of the paper's significance. Our intention was to clarify what the original experiments establish, rather than introduce a new research objective.

Thank you again for your careful consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: summary of revisions and one correction

We thank all reviewers for their careful reading. We have uploaded a revised manuscript; the responses below refer to its sections and appendices.

Correction. While answering Reviewer roQX's Q3, we found that several passages of the submitted version incorrectly described a separate Llama dose-selection pilot. In fact, the behavioral magnitude |α| = 1.5 was selected in an exploratory Qwen pilot on the same IV01–IV24 domain contexts and was fixed, together with the 30 new Llama generation seeds, in the Llama manifest before any Llama outcome was inspected. No Llama output informed the dose. The Llama direction and layer remain discovery-only. The Llama behavioral evaluation is therefore new-seed evidence at a cross-model, pilot-selected dose, not evidence that is independent of dose selection on these domains. We have corrected the affected passages in the revised manuscript.

Scope of the claim. The central claim is narrow: in held-out Llama domains, one frozen signed direction produces opposite signed changes in the same teacher-forced outcome-preference margin across accurate and distorted premises. Qwen shows a positive held-out aggregate interaction, but its distortion branch does not pass the strict branchwise criterion. Section 2 now states that we call the factorial interaction "relational" only in this statistical sense; it cancels the designed additive effects but does not by itself rule out generic valence matching, expectedness, or discourse coherence.

Why this matters. Prior work has shown that steering effects depend on input context, and adaptive methods exploit this dependence by choosing whether, how strongly, or where to steer. We ask a more specific question: whether one fixed signed intervention can move the same outcome-preference margin in opposite directions across premise conditions, without context-dependent changes to direction, sign, or dose. This distinction matters because an aggregate-only success criterion can conceal a branch that fails the strict directional test (Qwen), while a controlled readout need not predict the ordering of effects in generated behavior (Llama). The held-out results establish this operational property in Llama, while the Gaussian, behavioral, and relation controls delineate its limits regarding directional uniqueness, generated behavior, and mechanism-sensitive causal-rule recomputation.

Main revisions.

  1. Readability: a rewritten Introduction and a running maintenance-record example reused in §1, §2 and §4; revised Figure 1.
  2. Related work (§6): new discussion of identifiability and construct validity of steering directions, controlled readouts versus generated answers, and content effects.
  3. Appendix H (Reviewer roQX Q1): a post-review controlled readout at the generation dose ±1.5. Under total log-likelihood, the distortion-dominant ordering persists; under mean-token scoring it is borderline at both doses.
  4. Appendix I (Reviewers roQX Q4, WJnP): A post-review stimulus-feasibility audit for a proposed valence-matching control; 86/96 control premises passed, below the required 96/96. Both automated judges rated all original accurate premises mildly positive and distorted premises more negatively, revealing an affective-valence contrast within the original epistemic manipulation. Table 20 provides illustrative stimuli and judge ratings; Table 21 summarizes affective-contrast diagnostics. No control direction was constructed or evaluated; semantic specificity remains unresolved.
  5. Behavioral persistence: fine-grained attenuation is now presented as instrument-sensitive (§4, Appendix D.7), and the trajectory-average direction as the more robust behavioral result.

We respond to each reviewer below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.