acceptodds
Under review as a conference paper at ICLR 2027

CGUS: Conditional Gate and Up-Branch Steering for Personality Composition

Abstract

Personality steering helps tailor large language models to different users and roles. However, directions extracted in isolation may not match the activation states encountered during composition, limiting fine-grained control. We propose Conditional Gate and Up-Branch Steering (CGUS), a framework with two complementary modules. The gate module identifies and edits sparse activations associated with broad personality domains and their facets. The conditional vector module extracts steering directions under specified gate interventions and adds them to the parallel up-projection branch, matching vector construction to the intervened state. Together, the modules support combining facet components on a shared personality controller, with high and low targets constructed separately against neutral references. We use local IPIP-300 questionnaires to assess broad-personality and facet steering, and external language-model ratings to assess personality expression and fluency on PersonalityBench scenarios. Preliminary results suggest improved broad-personality expression.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, personality steering, activation steering, conditional steering vectors, neuron editing, compositional control, personality facets

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_2hTBOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CGUS, a training-free steering method for SwiGLU feed-forward blocks that sits on top of the NPTI personality gate. It has two parts. The first is a gate operator (Eq. 3) that enhances or suppresses selected post-activation gate coordinates. The second is an up-branch steering vector (Eq. 5) added to the up projection before the elementwise product, extracted by replaying contrastive responses under an active gate state ("state-conditioned extraction"). Domain and facet controllers are meant to compose as P, P+A, P+B, and P+A+B (Eqs. 6–8). The paper also defines a post-parent actionability criterion (Eq. 4) that flags facet suppressions made redundant by a parent gate that has already zeroed the same coordinates. On the 459-question NPTI benchmark with Llama-3-8B-Instruct, the domain up-branch gives +0.062 over NPTI at ρ = 0.5 (the only dose whose corrected CI excludes zero) and collapses at ρ = 2. It gives no reliable gain on Gemma-2-9B-it. Adding one facet's gate and up-branch gives +0.040 to +0.051 personality with a 0.16 to 0.23 drop in fluency, relevance, and completion. About 44% of that gain comes from one facet-pole whose outputs are degenerate. The authors describe the results as development-stage findings.

Critical strengths
  1. The reporting is candid and partly preregistered. The paper separates preregistered doses from post-hoc extensions, uses energy-matched random directions and a second mask seed, applies Bonferroni correction, and reports negative results and degenerate cases (Table 14) instead of filtering them out. This is better practice than many steering papers.
  2. The actionability observation is a useful diagnostic. Showing that 51 of 52 suppression coordinates in one gate bank duplicate parent suppressions, and that the result is token-identical output (Case 2), is a concrete warning about stacking gate-style edits.
Critical weaknesses
  1. The paper's own results do not support the proposed method, and the central composition claim is never tested.

    • Domain up-branch. None of the three preregistered doses separates from NPTI or from random directions (Table 1). The only significant cell, ρ = 0.5, was added after seeing those results. Its gain is +0.062 on a 1–5 scale, and it lowers the quality pass rate from 61.65% (NPTI) to 56.05%. Higher doses fail, with 0% quality pass at ρ = 2. On Gemma no dose separates, and the one promising cell reverses sign on a second mask seed (+0.2889 to −0.1111, Table 11).
    • Facet branch. The personality gain (+0.040 to +0.051, Table 2) is smaller than the quality cost (fluency and relevance −0.18 to −0.23). The up-branch adds nothing measurable over the facet gate alone: every "vs. facet-gate-only" interval in Table 13 contains zero. So the up-branch, which is half of the method's name and the target of state-conditioned extraction, has no demonstrated effect in the facet study.
    • Composition. The title and Section 3.3 frame the method around composing facets (P+A, P+B, P+A+B), and Eq. 9 defines joint attainment J and an interaction score I. Neither J nor I is reported for any experiment. The only two-facet evidence is a short qualitative description of "earlier exploratory" runs in Section 7, which reports mutual attenuation. The parent-priority gate-merge study was generated but not scored (Appendix D gives only a cost estimate). Appendix F lists the confirmatory protocols as future work.
  2. The experiments do not match the method described, and the evaluation has validity problems.

    • Method mismatch. Sections 3.1–3.2 claim independent high/low endpoints and up-branch vectors read under the active gate state. But the domain-up study in Section 5.1 extracts its direction "with gate and up interventions off" and uses a single high-minus-low direction, with "its negation" for the low pole (Appendix D). That is the unconditioned, sign-reversed setup the paper argues against. The facet study uses raw, non-normalized vectors while the domain study uses unit vectors, so even the scaling convention differs between the two main experiments.
    • Judge validity. In Table 14 the low Artistic-Interests output degenerates into strings of single letters and receives personality = 5. A metric that gives the top score to incoherent text cannot separate trait expression from model failure, and this slot supplies about 44% of the facet macro gain. The judge is named only as "openai/gpt-6-luna:batch", with no description or reference, so it cannot be identified or reproduced. There is no human evaluation.
    • Baselines. The two main studies compare only against NPTI and random directions. PERSONA, DPN-LE, CAA, and prompting are not run on the official benchmark, even though they are discussed as the relevant prior work. DPN-LE appears only in the earlier IPIP study (Table 7).
    • Effect sizes in the earlier studies. The component studies that "motivated these designs" report effects such as +0.00125 to +0.003 on a 1–5 scale (Table 8) and a pilot on six scenarios (Table 6). These are too small or too underpowered to justify design choices.
    • Internal inconsistencies. The text in Section 5.1 says the quality pass rate "falls from 65.15% at ρ = 0.1", but Table 1 shows 65.15% at ρ = 0.01 and 63.27% at ρ = 0.1. The contributions claim "comprehensive" evaluations, while the abstract and conclusion state the results are not a validated recipe.
    • Technical framing. The actionability criterion follows directly from Eq. 3: once max(g, 0) = 0, further suppression is zero by construction. It is a useful sanity check, but presenting it as a main contribution overstates it. The stochastic mask z in Eq. 3 is not defined.

Taken together, the paper reads as a progress report. The authors' own conclusion, that the results "do not yet establish a dose, facet selection, or cross-model recipe that is simultaneously effective, quality-preserving, and reproducible", is accurate and is the main reason for my recommendation.

Questions for authors

Items 1–3 would need to be resolved for the paper to be reconsidered. Items 4–6 are clarifications.

  1. Can you report J and I (Eq. 9) for the P / P+A / P+B / P+A+B configurations on the official benchmark, with the parent-priority gate-merge study scored?
  2. Why does the domain-up study use gate-off extraction and a sign-reversed low pole, when the method claims state-conditioned extraction and independent endpoints? Is there a version of Section 5.1 that follows Section 3?
  3. Please compare against PERSONA, DPN-LE, CAA, and prompting on the same 459-question benchmark, and add a human evaluation or a second judge. What exactly is "openai/gpt-6-luna:batch"?
  4. How many questions does each domain-pole/facet slot contain? The per-facet increments in Table 12 look like they are computed on roughly 90 questions per slot.
  5. The low Artistic-Interests gate has 27,124 coordinates. Across how many layers is this support spread, and was any coordinate cap or threshold ablation run?
  6. Please define the stochastic mask z in Eq. 3 and explain how λ, e, and s are set for each study.
Reviewer_sbFeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CGUS, a training-free framework for compositional personality steering through gate and up-branch interventions in LLMs. The method aims to support hierarchical control over broad personality domains and fine-grained facets. Experiments investigate steering effectiveness, compositional behavior, and generation quality.

Critical strengths
  1. Meaningful problem with a reasonable technical motivation

    The paper addresses a meaningful problem, which is similar to attribute entanglement in multi-attribute controllable text generation. Existing activation-steering methods may suffer from interference when multiple attributes are controlled simultaneously. Investigating state-dependent interventions and compositional conflicts is therefore worthwhile.

Critical weaknesses
  1. Insufficient validation of multi-attribute composition

    The central claim of compositional personality control is insufficiently validated. The main experiments primarily evaluate a broad personality domain combined with a single facet, while confirmatory evaluations of simultaneous multi-facet control (P+A+BP+A+B) are missing. Standard multi-attribute controllable generation requires demonstrating joint target satisfaction rather than improvements on individual attributes.

  2. Insufficient evaluation of attribute disentanglement and interference

    The evaluation does not sufficiently demonstrate facet-specific control or attribute independence. Improvements in aggregate personality scores do not necessarily indicate successful control of the intended facet. More comprehensive evaluations of target-attribute success, non-target attribute preservation, and interference under joint control are needed.

  3. Limited effectiveness and substantial generation-quality degradation

    The practical effectiveness of CGUS remains questionable. Facet steering produces only small personality-score improvements (approximately 0.04–0.05 points), accompanied by noticeable degradation in fluency and relevance. Moreover, approximately 44% of the aggregate improvement comes from a facet condition exhibiting severe generation degeneration. These results do not establish a satisfactory control–quality trade-off.

  4. Unclear contribution of dual-branch steering

    The necessity of the proposed dual-branch design is not convincingly established. The additional up-branch intervention does not yield a statistically distinguishable improvement over gate-only steering. Moreover, the method shows no reliable gain on Gemma-2-9B-it. Stronger component analyses and cross-model validation are needed to substantiate the claimed methodological advantages.

Questions for authors
  1. Can the authors provide confirmatory experiments on simultaneous multi-facet control (P+A+BP+A+B), reporting joint target attainment, non-target attribute preservation, and generation quality against strong compositional baselines?
  2. Given the limited improvement over gate-only steering and the observed quality degradation, what evidence supports the necessity and practical effectiveness of the additional up-branch intervention?
Reviewer_CvfVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates hierarchical personality control in LLMs, focusing on composing fine-grained personality facets on top of broader personality domains. The authors propose Conditional Gate and Up-Branch Steering (CGUS), a training-free framework that combines gate interventions with state-conditioned up-branch steering vectors. The paper also introduces an actionability criterion to identify redundant gate operations under an active parent controller. Experiments evaluate domain-level steering on Llama-3-8B-Instruct and Gemma-2-9B-it, and single-facet composition on Llama. The results show limited personality improvements, often accompanied by generation-quality degradation, while the additional benefits of up-branch steering remain inconclusive, and the domain-up approach shows no reliable improvement on the second model family.

Critical strengths
  1. Meaningful research problem. The paper investigates hierarchical personality composition in LLMs, aiming to control fine-grained personality facets while preserving broader personality domains and generation quality. This is a meaningful research direction for achieving more flexible and precise personality control.

  2. Useful empirical findings and diagnostic analysis. The paper provides detailed analyses of intervention strength, generation-quality degradation, and redundant gate operations. In particular, the proposed actionability criterion helps identify ineffective facet-level interventions under an active parent controller, offering useful insights into the challenges of hierarchical personality composition.

Critical weaknesses
  1. Insufficient validation of the core compositional framework. Although CGUS is proposed for hierarchical personality composition, the main completed experiments focus on domain-level steering and single-facet interventions. The full two-facet composition setting (P + A + B) lacks a completed confirmatory evaluation, and the parent-priority gate-merge study has not been scored. Consequently, the experimental evidence is insufficient to establish whether CGUS achieves its primary goal of reliable multi-facet personality composition.

  2. Limited effectiveness and substantial generation-quality degradation. The reported personality improvements are generally small and often accompanied by noticeable reductions in fluency, relevance, and completion. Moreover, the additional contribution of the up branch over facet-gate-only steering is not statistically distinguishable from zero, and approximately 44% of the aggregate personality gain comes from a single facet-pole setting with severely degraded outputs. These results raise concerns about the incremental effectiveness and practical value of the proposed dual-branch design.

Questions for authors
  1. Multi-facet composition. Could the authors provide a direct comparison of the P, P + A, P + B, and P + A + B configurations? In particular, does jointly steering two facets introduce interference compared to controlling each facet individually, and how does it affect parent-domain stability and generation quality?

  2. Effectiveness of the up branch. The experiments show that the additional contribution of up-branch steering over facet-gate-only interventions is not statistically distinguishable from zero. Could the authors clarify the specific benefits of introducing the up branch? Are there any experimental conditions where it consistently improves personality control beyond gate-only steering?

  3. Generation-quality degradation. The reported personality improvements are often accompanied by substantial degradation in fluency, relevance, and completion. Moreover, approximately 44% of the aggregate personality gain comes from a single facet-pole setting with severely degraded outputs. Could the authors report the aggregate personality results excluding this degenerate setting and provide further analysis of the trade-off between personality improvement and generation quality?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.