acceptodds
Under review as a conference paper at ICLR 2027

Read–Write Factorization: Preference Geometry for Personalized Decision Making

Abstract

Personalizing large language models requires both *reading* a preference from context and *writing* it into a decision, yet activation steering typically ties these operations. We show that this tying fails along three geometric degrees of freedom: *where* to read and write, *how many* coordinates are needed, and *which directions* support readout versus causal control. Across two 7B architectures, extraction-by-application intervention surfaces are approximately separable, with rank 1 explaining 81–91% of variance in homogeneous-preference conditions. Heterogeneous team preferences are low-rank but not scalar (effective rank 3.14), with the dominant coordinate tracking preference severity (ρ = 0.952). Most strikingly, high-decoding read directions are nearly orthogonal to a gradient-defined write direction (|cos| ≤ 0.026), and their span captures only 0.07% of its squared norm: decodability does not determine causal control. A sign-symmetric decomposition further separates directional control from sign-invariant perturbation. Together, these results establish *read–write factorization*: preference representation and causal control need not share a single steering geometry. Behaviorally, the dissociation transfers to PrefEval's original implicit-preference classification task: write but not read steering improves accuracy (34% → 51%). On a CL-Bench contextual-binding task, separately parameterizing representation and coupling raises binding accuracy from 46.7% to 90.5% ± 1.5%, with ablations attributing most of the gain to coupling the context-dependent representation rather than to the learned direction alone.

Primary area:
interpretability and explainable AI
Keywords:
preference personalization, activation steering, representation geometry, read-write factorization, LLM alignment

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_zuqLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that activation-steering approaches to LLM personalization implicitly "tie" three choices that should be treated separately: (i) the layer where a preference direction is extracted versus the layer where it is applied; (ii) a scalar (rank-1) preference axis versus a low-rank subspace; and (iii) the direction that best decodes a preference ("read") versus the direction that most changes a downstream decision ("write"). The authors call the separation read–write factorization and test each assumption with its own diagnostic. Most experiments use Qwen-2.5-7B-Instruct and Mistral-7B-Instruct-v0.3 on a synthetic binary team-decision task (escalate vs. review) and on PrefEval.

Critical strengths

1.Conceptual novelty and clarity: The paper clearly distinguishes between read and write directions and provides strong evidence that they are geometrically and causally distinct. This challenges a widely held assumption in activation steering and opens new research directions. The formalization of extraction vs. application depth, preference rank, and read-write separation is well-motivated and rigorously tested. 2.A simple, informative crossed-depth experiment and unusually candid scoping. Table 1 shows that late-extracted directions retain their effect at mid depth (2.91 vs. 2.87), while mid-extracted directions lose much of their effect when applied late (0.37). The practical insight and transparent reporting of limitations strengthen the paper.

Critical weaknesses
  1. The paper's presentation could be improved to enhance readability and clarity. Some aspects of the notation, mathematical exposition, and figure presentation make the methodology difficult to follow, and greater consistency would help readers understand the proposed framework and reproduce the experiments (see Question 1).

  2. The experimental validation is not yet sufficiently comprehensive to fully support some of the paper's central claims. More detailed ablations are needed to distinguish the contributions of individual components. Furthermore, some reported results appear inconsistent with the paper's broader conclusions, raising concerns about whether the evidence is interpreted comprehensively. Clarification of key implementation details is also needed to strengthen the validity and reproducibility of the experimental findings (see Questions 2–5).

Questions for authors

1.The paper's readability is compromised. For example, equation numbering is inconsistent (only Eqs. 1–5 are numbered; key definitions like Δodd are not) with no stated convention—please clarify and unify, and the formulas in Figure 3 are too small to read clearly—please enlarge the font.

2.In the CL-Bench experiment of TABLE 3, the random direction with inherited coupling achieves 82.5% BDA, while the learned direction reaches 90.5%. Can you provide a more detailed ablation that isolates the contribution of the learned direction versus the coupling interface and context representation? 3. In §4.3, the 32B replication shows a non-significant crossed-depth asymmetry. Given that 14B and 72B are significant, what might explain this? Could it be related to model scale, training data, or architecture-specific properties? Additional analysis would help clarify the robustness of the depth asymmetry claim.

4.In Eq. 4, how exactly is q(x) added to the decision-token logits under randomized A/B labels? Does any inference-time computation use information about which option is correct?

5.The paper's claim that read and write directions are "essentially none" overlapping and "consistent with random" across scales is contradicted by its own Table 11, which reports read-span coverage of 0.44–9.75% on other models—far above the random baseline (~0.11%)—making the conclusion a selective summary of Qwen-7B alone.

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

This paper investigates read–write factorization in LLM personalization: whether preference extraction and behavioral intervention should share layer locations, representational rank, and directions. Cross-layer intervention matrices, preference-spectrum analyses, and comparisons with gradient-defined write directions reveal distinct extraction/application depths, low-rank multidimensional preferences, and near-orthogonality between highly decodable read directions and local write directions. PrefEval and a CL-Bench-derived task explore behavioral consequences. Overall, the geometric analyses are valuable, while the behavioral experiments do not yet sufficiently isolate the contribution of read–write separation itself.

Critical strengths

1.Clear and systematic problem decomposition. 2.The central geometric experiments include useful controls. 3.The paper reports replication scope and negative results transparently.

Critical weaknesses

1.The CL-Bench result does not sufficiently establish improved preference binding or actual decisions. 2.Squared-norm coverage must be distinguished from recoverable first-order leverage. 3.The most direct PrefEval behavioral evidence lacks a complete dedicated protocol. The improvement from 34% to 51% on the original four-option task is meaningful, but calibration/validation/test separation, validation size, final write strength, and the exact sequence-probability gradient objective are not clearly specified.

Questions for authors

see weakness

Reviewer_yHz3OpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a study on the limitation of activation steering.

Critical strengths

N/A

Critical weaknesses

This paper is clear AI slop. I am unable to understand it.

Questions for authors

N/A

Comments on the paper

1 comment

Dear Reviewers,

Thank you for the review. The authors have responded to the points you raised, and the response period closes on 18 November (AoE). I would be grateful if you could post a brief reply on whether your concerns are resolved, partly resolved, or still open, so that the authors have a fair chance to address them. If anything still needs clarifying, please ask now while they can still respond.

Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.