acceptodds
Under review as a conference paper at ICLR 2027

Not All Updates Help: Functional Path Attribution for Personality Vector Merging

Abstract

Personality vector merging enables training-free personality control by adding parameter displacements obtained from personality fine-tuning. However, existing methods typically treat the entire displacement as a coherent personality vector. We show that this assumption overlooks substantial functional cancellation: along parameter-space merging paths, many parameter displacements contribute negatively to the target personality objective, and including them can weaken the resulting personality behavior. We propose Functional Path Attribution, which integrates parameter-wise contributions along personality merging paths and selectively retains the most functionally constructive updates to construct sparse personality vectors. Across Big Five personality control experiments, these vectors achieve stronger target-trait expression than their source fully fine-tuned models while substantially better preserving general capabilities, outperforming random, magnitude-based, and other parameter-selection baselines. These results show that personality vectors are functionally heterogeneous and that path-aware selective merging provides a more effective approach to parameter-space personality control.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Large Language Models, Model Merging, Sparse Parameter Space, Personality LLM

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

The paper studies personality vector merging, where a Big Five trait is represented as the parameter displacement between a base LLM and a personality-fine-tuned checkpoint. The authors argue that this displacement is functionally heterogeneous: some coordinate updates push toward the target pole and others counteract it, which they call Functional Cancellation. They propose Functional Path Attribution (FPA), an Integrated-Gradients-style decomposition in parameter space of a contrastive NLL personality score StpS_t^p along a bipolar merging path (piecewise-linear or quadratic) through θt−\theta_t^-, θ0\theta_0, and θt+\theta_t^+. FPA-Merge keeps the top-ρ\rho coordinates by attribution (default ρ=5%\rho = 5\%) and applies a shared mask to both pole displacements. Experiments on Qwen3-4B-Instruct (plus Extraversion on Qwen3.5-9B) use BIG5-CHAT for adaptation, TRAIT for personality, and GPQA, AIME25, and MATH-500 for capability. On Extraversion-High, FPA-Merge reaches TRAIT 90.1 versus 68.1 for Full SFT while keeping GPQA near the base model (57.17 vs. 60.56). It exceeds the Full SFT endpoint in 7 of 10 trait-pole settings. The paper also reports multi-trait composition, transfer to a MedMCQA expert (60.24 vs. 39.56 for dense merging), and continuous control over α\alpha.

Critical strengths
  1. The data-aware selection signal works well, and the paper isolates this from sparsity alone. Without rescaling, random 5% retention (DARE) gives TRAIT 38.6, close to the base model's 37.6 (Table 9). FPA's 5% support gives 90.1. The quadrature ablation (Tables 10–11) is also informative: the endpoint gradient (n=1n=1) gives 69.0, while n≥4n \ge 4 gives a stable support (J≥0.94J \ge 0.94 vs. n=64n=64) and around 90 TRAIT. So the path integration matters, and the method is not a disguised magnitude heuristic.
  2. The practical result is useful. A 5% sparse vector that transfers to a separately fine-tuned expert with a small capability loss (MedMCQA 63.33 → 60.24, vs. 39.56 dense) is a concrete step for parameter-space behavior control. The paper is also clearly written and well documented in the appendix (splits, configs, prompts, decoding, quadrature nodes).
Critical weaknesses
  1. The baseline comparison does not appear to be matched on effective update strength, and this affects the headline trade-off. Sec. 5 states that sparsification baselines are rescaled, which for DARE means multiplying retained updates by 1/(1−p)=20×1/(1-p) = 20\times at ρ=5%\rho = 5\% (App. C.2). Eq. 9 and Algorithm 1 define vtp=Mt⊙Δtpv_t^p = M_t \odot \Delta_t^p with no rescaling, so FPA-Merge seems to be applied unscaled. If so, the baselines in Table 1 apply a 20×-amplified sparse update, while FPA applies the original-scale coordinates. That could explain much of the baselines' capability loss (e.g., DARE GPQA 43.18 vs. FPA 57.17). Table 9 reports unrescaled baselines only for TRAIT, not for capability. The fair comparison is a Pareto view: for each method, sweep the scale (and for dense merging, α>1\alpha > 1, which Fig. 4 does not explore) and compare capability at matched TRAIT. Without this, the main claim "stronger trait control and better capability preservation" is not established against the baselines. Reporting both quantities on a common curve would likely settle it.

  2. The evidence for Functional Cancellation as a mechanism is weaker than presented. Fig. 6 shows cumulative attribution peaking at about 15× the full-vector total when coordinates are added in descending FPA order. This shape is guaranteed for any sorted list containing negative values, so it does not test the hypothesis. It is also a sum of first-order attributions computed along the dense path, not the score change of the masked model, whose path differs. The behavioral result (top 20% → TRAIT 97.3 vs. 68.1 dense) is real but has competing explanations. Sparse masking may mainly remove updates that disturb general behavior or shift TRAIT's forced-choice calibration, rather than removing updates that oppose the trait. Direct tests would be: (a) remove only the negatively attributed coordinates and keep all positive ones; (b) apply only the most negative coordinates and check that they move TRAIT toward the opposite pole; (c) measure the actual StpS_t^p of masked models and compare it with the summed attributions. In addition, the full baseline comparison is shown only for Extraversion-High (Table 1). Other traits are compared only against Full SFT (Fig. 2), all results are single-seed, and personality is measured only by a logit-based multiple-choice benchmark. Together, this makes the generality of the mechanism claim hard to judge.

Questions for authors

Items 1–3 could change my assessment. Items 4–6 are clarifications.

  1. Are FPA-Merge's retained updates rescaled? If not, please report GPQA/AIME25 for the unrescaled baselines (Table 9 shows only TRAIT). Please also report a TRAIT-vs-capability curve for each method under a scalar sweep, including dense merging with α>1\alpha > 1.
  2. What happens if you remove only the coordinates with negative attribution and keep all positive ones? And does applying only the bottom-kk coordinates move TRAIT toward the opposite pole? These results would directly support or refute the cancellation account over a generic sparsity effect.
  3. FPA uses 200 labeled paired samples and gradients, while the parameter-selection baselines use no data. Could you compare against a data-aware saliency baseline at the same data budget, such as ∣Δi⋅∇iS∣|\Delta_i \cdot \nabla_i S| at θ0\theta_0 or a Fisher-weighted mask? This would show how much of the gain comes from path integration versus having target-specific gradient signal.
  4. Can you provide the Table 1 comparison (at least Task Vector, DARE, and Update Magnitude) for the other nine trait-pole settings, and variance across seeds for the main numbers?
  5. Have you evaluated personality in open-ended generation (e.g., LLM-judge or questionnaire-based scoring of free responses), given that TRAIT is scored from option logits?
  6. What is the wall-clock and memory cost of attribution (16 full-parameter gradient evaluations over 200 samples per trait for the PL path), compared with one SFT run?
6 comments
AuthorsOpenReview ↗

Response to Reviewer yzCc-Q1

Thank you for your careful review of our paper and for recognizing the value of data-aware parameter selection and transfer to expert models. The revised manuscript compares general capabilities at similar levels of personality steering and directly intervenes on negatively attributed updates to test Functional Cancellation. We also added saliency baselines with the same data budget, five-seed experiments, and generative personality evaluations.

Q1: Rescaling and comparison at similar steering strengths

Rescaling changes both the baselines' steering strength and their general capabilities. The original FPA-Merge merges the selected updates at their original magnitude. To see what role rescaling plays in Table 1, Table 12 (Appendix C.2) compares TRAIT, GPQA, and AIME25 for DaRE, Weight Magnitude, Update Magnitude, and MLP Merge with and without rescaling at the same outer coefficient α=+1\alpha=+1.

For standard DaRE, given the full fine-tuning displacement Δ\Delta, each coordinate is retained independently with probability qq, so mi∼Bernoulli(q)m_i\sim\mathrm{Bernoulli}(q). This gives

\qquad \mathbb E\left[\frac{m_i\Delta_i}{q}\right]=\Delta_i.$$ This is the basis for the 20-fold compensation in standard DaRE at a 5% retention ratio. The two magnitude baselines and MLP Merge use deterministic selection and apply $1/q$ as an empirical scaling rule, with factors of 20 for the magnitude baselines and approximately 1.50 for MLP Merge. Table 12 shows that rescaling increases TRAIT for all four baselines and lowers their GPQA and AIME25 scores. For example, disabling rescaling for DaRE yields higher capability scores, but TRAIT returns to nearly the base model's level. This shows that some of the differences in capability scores for the baselines in Table 1 are indeed related to the update magnitude. It also explains why we used the rescaled variants to show their stronger steering at a fixed coefficient. FPA selects coordinates based on their signed functional contributions to the target personality. For Extraversion-High, the selected 5% of coordinates account for 58.69% of the total positive attribution, and the updates achieve a TRAIT score of 90.1 at their original magnitude. Because these updates already achieve strong steering at their original scale, we use that scale by default. **Comparing capabilities at similar TRAIT scores.** The same outer coefficient can yield different steering strengths, so in Figure 8 we extend the bipolar coefficient sweep for Extraversion to $\alpha\in[-2,2]$ and compare GPQA and AIME25 at points where the methods reach similar TRAIT scores. Disabling rescaling lets some baselines retain more of their general capabilities, but their personality steering is also weaker. Dense merging and standard DaRE can achieve stronger steering with larger coefficients. Even when dense merging can use $\alpha>1$, FPA-Merge (PL) retains more capability than dense merging at points with similar TRAIT scores for the high pole. The approximately matched points for the low pole also show that FPA-Merge (PL) has higher GPQA and AIME25 scores than both dense merging and standard DaRE. This shows that the trade-off advantage in Table 1 remains visible when the baselines can use a wider range of coefficients and steering strength is matched. We also tested rescaling by the reciprocal of the fraction of positive attribution retained, using the same FPA mask. Scaling the updates by $1/0.5869\approx1.704$ further strengthens steering toward the high pole of Extraversion, but both capability scores decline (Table 1, Appendix C.8). This ablation further supports retaining the original update magnitude in the default configuration.
AuthorsOpenReview ↗

Response to Reviewer yzCc-Q2

Q2: Direct interventions on negatively attributed updates and Functional Cancellation

The cumulative attribution curve in Figure 6 gives a prediction that can be tested directly. Negatively attributed updates should weaken the expression of the target personality. When the full displacement for Extraversion-High is restored, both the measured TRAIT score and the contrastive score on the validation set fall. Section 6.4 of the revised manuscript further applies two interventions based on attribution sign across all ten Big Five trait–pole settings to test which updates cause this change.

  1. Remove only the negatively attributed updates. After removing the updates on these coordinates, which account for about 40% of the total, and retaining the original SFT updates on the remaining coordinates, TRAIT is closer to the target pole than with Full SFT in all ten settings (Table 4). This shows that these updates weaken steering toward the target when applied together with the rest of the displacement.
  2. Apply only the 5% of updates with the lowest attributions. TRAIT moves from the base model toward the opposite pole in all ten settings (Table 4). The base score for Extraversion-High is 37.6. At a retention ratio of about 5% and the original update scale, selecting the coordinates with the highest attributions gives 90.1, random selection gives 38.6, and selecting the coordinates with the lowest attributions gives 21.3 (Tables 4, 9, and 12). Similar levels of sparsity produce opposite steering directions, indicating that the selected updates determine this difference in direction.

We also directly computed the target contrastive score from the log-likelihoods of paired responses on the models after applying the masks. After removing negatively attributed updates, the scores are higher than those of Full SFT in all ten settings. Applying only the updates with the lowest attributions yields scores below the base model in all ten settings. Because the mask changes the actual merging path, we use the measured scores to quantify the magnitude of the change and test whether the sum of attributions along the original path predicts its direction. In all 20 interventions, the measured score change relative to the base model has the same sign as the sum of the attributions of the retained coordinates (Table 16). This result also corroborates the direction of the interventions beyond TRAIT's scoring of answer choices.

When attributions are computed using only responses for the target pole, restoring the full displacement also weakens steering toward the high pole of Extraversion (Appendix C.4, Table 15 and Figure 9). Together, these results show that negatively attributed updates counteract the updates in the full SFT displacement that promote the target personality. By using this difference in direction, FPA-Merge can select sparse vectors that steer more strongly than the full displacement.

AuthorsOpenReview ↗

Response to Reviewer yzCc-Q3&Q4

Q3: Saliency baselines with the same data budget and path integration

The target gradient itself provides an effective selection signal, and integrating along the path further improves steering toward both poles. We added four saliency baselines in Section 6.5, Table 5, using the same 200 labeled pairs and 5% retention ratio as FPA. As suggested by the reviewer, ranking with ∣Δip∇iSp(θ0)∣|\Delta_i^p\nabla_i S^p(\theta_0)| at the base model yields Extraversion TRAIT scores of 57.20 and 28.10 for the high and low poles, respectively. Both scores shift toward the target pole relative to the base model. Retaining the sign of saliency further raises the high-pole score to 76.70 and lowers the low-pole score to 14.90.

Holding the signed ranking rule fixed, we compare gradients at the base model, the midpoint, and the SFT endpoint. The best single-point location varies with the target pole. The midpoint is best for the high pole (80.10), and the base model is best for the low pole (14.90). FPA-Merge (PL) achieves 90.10 and 13.90, respectively, moving 10.0 and 1.0 percentage points further toward the target than the best single-point baseline for each pole. FPA-Merge (PL) also achieves the highest AIME25 scores in the table for both poles, and for each pole its GPQA score differs by less than 2 percentage points from the highest single-point baseline score.

The role of path integration can be understood through the total derivative of the contrastive score. Consider the segment of the PL path from the base model to the target pole. Let θp\theta_p be the fine-tuned endpoint for the target pole, Δp=θp−θ0\Delta^p=\theta_p-\theta_0 be the displacement, and γp(t)=θ0+tΔp\gamma_p(t)=\theta_0+t\Delta^p be the path, where t∈[0,1]t\in[0,1]. Single-point saliency A^ip(t0)=Δip∇iSp(γp(t0))\widehat A_i^p(t_0)=\Delta_i^p\nabla_i S^p(\gamma_p(t_0)) uses the gradient at t0t_0 to approximate the average gradient along the full segment. Path integration accumulates the gradients along the path to obtain the signed net contribution of coordinate ii.

Aip=∫01Δip ∇iSp(γp(t)) dt.A_i^p=\int_0^1 \Delta_i^p\,\nabla_i S^p(\gamma_p(t))\,dt.

Summing over all coordinates and applying the chain rule gives

=\int_0^1 \sum_i \Delta_i^p\,\nabla_i S^p(\gamma_p(t))\,dt =\int_0^1 \frac{d}{dt}S^p(\gamma_p(t))\,dt =S^p(\theta_p)-S^p(\theta_0).$$ Thus, path integration decomposes the change in the target contrastive score due to the full update into contributions from individual coordinates. A coordinate's contribution can be large at one location, but it can weaken or even change sign as other parameters change along with it. Accumulating signed contributions along the path incorporates these changes into the ranking. ## Q4: Effectiveness across traits and stability across random seeds **FPA-Merge's steering effectiveness extends to all of the Big Five traits.** Table 18 in Appendix C.6 extends the comparison with Task Vector, standard DaRE, and Update Magnitude to the nine settings outside Table 1 at the same magnitude of the merging coefficient. In all ten trait–pole settings, both FPA-Merge variants move TRAIT from the base model toward the target pole, and at least one variant achieves the strongest target steering in seven settings. The advantage in capability retention also appears for other traits. For example, in seven of the eight non-Extraversion settings, PL scores higher on AIME25 than both Task Vector and standard DaRE. These results show that selecting updates based on functional contribution can balance personality steering and general capability retention beyond Extraversion. **With fixed SFT endpoints, the advantages of FPA-Merge (PL) persist across the five-seed results.** For both poles of Extraversion, we resampled the 200 attribution pairs for FPA, resampled the DaRE masks, and varied the generation seeds for the four methods (Appendix C.6, Table 19). The TRAIT scores of FPA-Merge (PL) are $90.74\pm1.34$ and $13.10\pm0.86$ for the high and low poles, respectively (mean ± sample standard deviation). TRAIT varies little at either pole, and average steering is stronger than for each baseline at both poles. PL also leads in mean GPQA and AIME25 at both poles. Thus, under the tested attribution sampling, random masking, and generation settings, FPA-Merge's advantages in personality steering and capability retention remain visible after averaging across seeds.
Reviewer_rXThOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates functional cancellation in personality vector merging, where parameter updates induced by personality fine-tuning can make opposing contributions to the target personality objective. To address this issue, the authors propose Functional Path Attribution (FPA), which uses path-integrated gradients to estimate the signed contributions of individual parameter updates along personality merging trajectories. Based on these attributions, FPA-Merge constructs sparse personality vectors by retaining the most functionally constructive parameter coordinates. Experiments across the Big Five personality traits show that FPA-Merge achieves stronger target-trait expression while better preserving general model capabilities than dense merging and parameter-selection baselines. The authors further demonstrate that the resulting sparse vectors support continuous personality control, multi-trait composition, and transfer to task-specialized models.

Critical strengths
  1. Interesting empirical finding. The paper identifies functional cancellation within personality fine-tuning updates, showing that some parameter changes can counteract the desired personality behavior rather than contribute positively. This finding provides a useful perspective on the functional heterogeneity of personality vectors and motivates more selective parameter merging.

  2. Strong empirical performance and practical applicability. FPA-Merge achieves substantial improvements in personality control while better preserving general model capabilities compared to dense merging and other baselines. Notably, retaining only 5% of parameter coordinates outperforms the corresponding Full SFT models in 7 out of 10 trait-pole settings. The paper further demonstrates continuous personality control, multi-trait composition, and transfer to a task-specialized expert model, supporting the effectiveness and flexibility of the proposed approach.

Critical weaknesses
  1. Limited methodological novelty. The core attribution mechanism of FPA is largely based on existing parameter-space integrated gradients. Although the paper introduces task-specific designs for personality vector merging, the methodological contribution appears relatively incremental, as the main approach builds upon an established attribution technique rather than introducing a fundamentally new one.

  2. Limited experimental validation. Although the method is evaluated across all Big Five traits, most detailed ablations and functional cancellation analyses focus on Extraversion, and the three trait-pole settings where FPA-Merge does not outperform Full SFT are not thoroughly investigated. Furthermore, the main experiments rely on Qwen3-4B-Instruct, while the additional evaluation on Qwen3.5-9B is restricted to Extraversion. Both models also belong to the Qwen family, leaving the generalizability of the findings across personality traits and model families insufficiently established.

Questions for authors
  1. Analysis across personality traits. Most of the functional cancellation analyses and ablation studies focus on Extraversion. Could the authors provide additional evidence of functional cancellation across the other Big Five traits? In particular, could the authors explain why FPA-Merge does not outperform Full SFT in the three trait-pole settings identified in Section 6.1?

  2. Generalization across models. The main experiments are conducted on Qwen3-4B-Instruct-2507, while the evaluation on Qwen3.5-9B is limited to Extraversion. Could the authors provide more comprehensive results or analyses on Qwen3.5-9B, particularly across other personality traits? Additionally, is there any evidence that FPA-Merge generalizes to models outside the Qwen family?

3 comments
AuthorsOpenReview ↗

Response to Reviewer rXTh-Q1&Q2

We thank the reviewer for recognizing the empirical finding of Functional Cancellation and for suggesting further tests of its applicability across traits and models. The revised manuscript adds attribution-based interventions across the Big Five, complete results on Qwen3.5-9B, and validation across model families on Llama.

Q1: Functional Cancellation across traits and the three 5% sparsity settings

Across all ten Big Five trait–pole settings, negatively attributed updates collectively weaken target-trait expression. Section 6.4 of the revised manuscript adds two complementary interventions. After removing the negatively attributed coordinates and keeping the remaining SFT updates, TRAIT shows stronger target-pole steering than the corresponding Full SFT in all ten settings. Applying only the 5% of updates with the lowest signed attributions shifts TRAIT from the base model toward the opposite pole in all ten settings (Table 4). The measured contrastive scores are higher than those of Full SFT after removing the negatively attributed updates and lower than those of the base model after applying only the bottom 5% of updates. The sign of the measured score change agrees with that of the sum of attributions over the retained coordinates in all 20 interventions (Table 16).

For the three settings in Figure 2 of Section 6.1 where the results at a 5% retention ratio did not surpass Full SFT, we retain all nonnegatively attributed coordinates, which account for approximately 59%–60%. TRAIT increases from the Full SFT scores of 93.7 and 91.3 to 95.8 and 92.5 for Conscientiousness-High and Agreeableness-High, respectively, and decreases from 16.0 to 10.8 for Neuroticism-Low. All three move closer to the target pole (Table 4). These results show that beyond the fixed 5% budget, the SFT displacement still contains updates that can strengthen target expression. Increasing the parameter retention ratio can further improve FPA-Merge's steering to the point of surpassing Full SFT.

The methodological contribution of this work lies in identifying Functional Cancellation in personality merging and using it to construct sparse personality vectors. Updates from the same personality fine-tuning run can strengthen or weaken the target trait, and these two types of updates partially cancel each other in dense merging. Building on existing parameter-space integrated gradients, FPA-Merge combines a contrastive personality objective, a bipolar merging path, and sparse update construction. We estimate each coordinate's signed net contribution along the path connecting the two fine-tuned endpoints and the base model, then use these contributions to select updates that promote the intended personality change (Sections 4.1–4.2).

Signed attribution also provides a basis for adjusting update directions. In Appendix C.8, at a fixed 5% retention ratio, selecting coordinates by absolute attribution and reversing the selected negatively attributed updates brings TRAIT closer to the target pole than the original PL variant in seven of the ten settings.

Q2: Generalization across model scales and families

The Qwen3.5-9B results now cover all ten Big Five trait–pole settings. Table 27 in Appendix G.1 reports TRAIT, GPQA, and AIME25 and includes other baseline methods for comparison. Both FPA-Merge paths shift TRAIT from the base model toward the target pole in all ten settings, and at least one path achieves the strongest target-pole steering in eight settings. FPA-Merge (PL) scores higher on GPQA than Full SFT and all four merging baselines in all ten settings. For example, on Openness-Low, FPA-Merge (PL) reduces TRAIT from 21.2 with Full SFT to 19.4, while increasing GPQA from 35.10% to 60.10%.

Beyond the Qwen family, we add bipolar experiments on Llama-3.1-8B-Instruct. In Table 28 of Appendix G.2, we use Llama's own high- and low-pole SFT checkpoints to construct Extraversion vectors. Both FPA-Merge variants shift TRAIT from the base model toward the corresponding target pole. For example, on Extraversion-Low, FPA-Merge (PL) reduces TRAIT from 29.60 for the base model to 6.40, below the Full SFT score of 13.60, and achieves a GPQA score of 22.98%, above the Full SFT score of 5.30%. This exploratory result provides evidence for the generalization of FPA-Merge across model families.

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns, particularly through the expanded functional cancellation analyses across all Big Five traits and the additional evaluations on Qwen3.5-9B and Llama-3.1-8B-Instruct.

The new results provide stronger evidence that functional cancellation occurs across different personality traits and help explain the three settings where FPA-Merge did not outperform Full SFT at 5% sparsity. The expanded Qwen3.5-9B experiments and the preliminary Llama results also strengthen the empirical support for the method's generalizability.

However, I still have some reservations regarding the methodological novelty. While the identification of functional cancellation is an interesting empirical finding, the proposed attribution mechanism remains largely based on existing parameter-space integrated gradients. The additional explanations help clarify the contributions but do not fundamentally change my assessment of the methodological advancement. Furthermore, the cross-family evaluation on Llama remains limited to Extraversion.

Overall, I acknowledge that the rebuttal has addressed several of my experimental concerns and strengthened the paper. However, given the remaining concerns about methodological novelty and the limited cross-family validation, I will maintain my original rating of 3 (Weak Acceptance).

AuthorsOpenReview ↗

Clarification on Novelty and Experiments on Llama3.1-8B

Thank you for your follow-up. We have now extended the Llama-3.1-8B-Instruct evaluation to both poles of all five Big Five traits, addressing the previous limitation to Extraversion. The full results are provided below.

Regarding novelty, FPA-Merge is a model-merging method built on established parameter-space attribution. We agree that path integration itself is not new. Our contribution is identifying functional cancellation and developing a behavior-aware merging rule: a personality-specific objective, bipolar trajectories, and signed attributions guide which updates are retained to construct sparse personality vectors. The contribution concerns how personality updates are selected and merged, supported by the demonstrated steering–capability benefits.

With the expanded cross-family evidence and this clarification, we would appreciate your reconsideration of the overall assessment and rating.

Llama-3.1-8B-Instruct results (%)

For TRAIT, higher is better for high targets and lower is better for low targets. Bold indicates the best TRAIT score for each trait–target pair.

(a) Personality control (TRAIT)

TraitTargetBaseFull SFTTask Vec.DaREWeight Mag.Update Mag.FPA-PLFPA-Q
OpennessHigh46.8092.8047.4093.2077.0078.2091.7093.80
OpennessLow46.8010.9068.3011.4036.3052.202.702.80
ConscientiousnessHigh81.2097.6052.5097.3092.4070.8090.5089.20
ConscientiousnessLow81.2047.6038.2047.8076.6056.8017.5017.00
ExtraversionHigh29.6076.1044.5076.8049.9044.6082.0087.90
ExtraversionLow29.6013.6057.3014.2024.3040.406.407.10
AgreeablenessHigh68.3087.2055.3086.5081.7050.8090.5092.90
AgreeablenessLow68.3043.0038.5042.2067.1036.908.107.80
NeuroticismHigh22.2040.5049.6038.0024.9067.9066.9074.50
NeuroticismLow22.2010.0056.209.3015.0068.904.805.60

(b) GPQA-Diamond (avg@10)

TraitTargetBaseFull SFTTask Vec.DaREWeight Mag.Update Mag.FPA-PLFPA-Q
OpennessHigh26.265.811.526.3117.930.0031.5730.81
OpennessLow26.263.030.003.2821.720.0019.4421.97
ConscientiousnessHigh26.263.281.524.2916.670.0029.2928.79
ConscientiousnessLow26.266.820.008.0823.990.0022.4725.00
ExtraversionHigh26.262.780.252.2717.930.0030.5629.80
ExtraversionLow26.265.300.004.2921.210.0022.9823.99
AgreeablenessHigh26.264.291.775.0522.220.0028.5428.28
AgreeablenessLow26.2618.690.0019.4422.220.0027.2728.54
NeuroticismHigh26.2611.360.0010.3519.700.0027.0223.48
NeuroticismLow26.266.311.265.0518.940.0025.0027.53

(c) AIME25 (mean@64)

TraitTargetBaseFull SFTTask Vec.DaREWeight Mag.Update Mag.FPA-PLFPA-Q
OpennessHigh2.080.000.000.420.420.002.080.83
OpennessLow2.080.421.250.420.830.000.830.42
ConscientiousnessHigh2.080.001.670.830.420.000.830.00
ConscientiousnessLow2.080.420.001.671.670.000.830.00
ExtraversionHigh2.080.420.420.421.250.000.831.67
ExtraversionLow2.081.250.000.420.000.001.831.83
AgreeablenessHigh2.080.420.420.000.420.002.080.00
AgreeablenessLow2.082.080.000.000.000.001.250.42
NeuroticismHigh2.080.000.831.251.250.002.500.00
NeuroticismLow2.081.251.671.670.420.002.500.42
Reviewer_9qLJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates Functional Cancellation in personality vector merging, where parameter updates from personality fine-tuning may either reinforce or oppose the target personality. The authors propose Functional Path Attribution (FPA), which computes coordinate-wise integrated-gradient attributions along a bipolar parameter-space path using a contrastive personality objective. FPA-Merge retains the highest-attributed coordinates to construct sparse personality vectors. Experiments on Big Five personality control demonstrate improved target-trait expression and general-capability preservation, alongside continuous control, multi-trait composition, and expert-model transfer.

Critical strengths
  1. Substantial personality-control gains with strong general-capability preservation.

Table 1 shows that FPA-Merge (PL) achieves a TRAIT score of 90.1 on Extraversion-High, outperforming Full SFT (68.1), Task Vector (75.0), and Prompt (71.2). Meanwhile, it achieves 57.17/47.56/93.4 on GPQA/AIME25/MATH-500, compared with 38.64/39.74/83.0 for Full SFT. These results demonstrate a substantially improved trade-off between personality control and general capabilities across diverse baseline methods.

  1. Effective continuous control, multi-trait composition, and expert-model transfer.

Figure 2 shows that FPA-Merge improves target-trait expression across all ten trait-pole settings and surpasses Full SFT in seven. Figure 4 demonstrates continuous personality control with reduced capability degradation. Notably, Table 2 shows that FPA-Merge preserves 60.24 MedMCQA accuracy on an expert model, compared with 39.56 after dense merging and 63.33 before merging, supporting its compositionality and transferability.

  1. Informative attribution analyses and ablation studies.

Figures 5–6 reveal opposing parameter contributions and declining personality strength as lower-ranked updates are restored, supporting Functional Cancellation. Appendix C.4 shows that contrastive attribution substantially outperforms target-only attribution (90.1 vs. 68.5 at 5% retention). Appendix C.3 further shows that path-based attribution improves over endpoint-gradient selection (89.7 with two quadrature points and 90.1 with eight, versus 69.0), supporting the importance of path information.

Critical weaknesses
  1. The incremental contribution over existing parameter-space attribution methods remains insufficiently isolated.

Parameter-space integrated gradients have been studied in FAIG and IGU-LoRA. FPA's distinctive components are its contrastive personality objective, bipolar path, and sparse-vector construction. While Appendix C.4 supports the contrastive objective, Appendix C.3 only compares path integration against an endpoint-gradient baseline. Since two quadrature points already achieve 89.7 versus 90.1 with eight, comparisons with base-model and midpoint gradients under identical retention and merging settings would better distinguish the benefits of path integration from gradient sampling location.

  1. Comprehensive strong-baseline comparisons are concentrated on Extraversion-High.

Table 1 compares numerous methods, but Figure 2 primarily contrasts FPA-Merge with Base and Full SFT across the remaining trait-pole settings. In particular, Conscientiousness-High, Agreeableness-High, and Neuroticism-Low are the three settings where FPA-Merge does not surpass Full SFT. Extending Task Vector and DaRE comparisons to these settings would clarify the consistency of the gains. Performance curves across merging coefficients would also better characterize personality-control and capability-preservation trade-offs.

  1. Multi-trait composition lacks direct comparisons with established interference-aware merging methods.

Section 6.2 primarily compares FPA-Merge with dense merging. Methods such as TIES-Merging explicitly address interference among parameter updates and provide relevant baselines for multi-trait composition. Appendix D reports substantial cross-trait overlap among selected parameters, further motivating these comparisons. Evaluating such methods under matched multi-trait configurations and parameter budgets would clarify the contribution of functional attribution beyond sparsification alone.

  1. Offline attribution costs and cross-model applicability are insufficiently characterized.

FPA-Merge requires pretrained personality-specific checkpoints and full-parameter gradient computations at multiple path positions. Although attribution samples and quadrature settings are reported, GPU time, peak memory, and storage overhead remain unclear. Moreover, the Qwen3.5-9B experiments cover only Extraversion, while expert transfer is evaluated on one medical model derived from the same base checkpoint. More complete cost reporting and a clearer characterization of the supported transfer settings would strengthen the practical assessment.

Questions for authors

see weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9qLJ-Q1

We thank the reviewer for recognizing FPA-Merge's performance in personality control, capability preservation, and expert-model transfer, and for identifying comparisons that needed strengthening. The revision adds gradient baselines at different points along the path, baseline comparisons across all ten Big Five trait–pole settings, a multi-trait comparison with TIES-Merging, and experiments across models. We also measured the cost of offline attribution.

Q1: The contribution of path integration relative to single-point gradients

Building on existing integrated attribution methods in parameter space, FPA-Merge combines a contrastive personality objective, a bipolar path, and sparse update construction. To distinguish the effect of gradient sampling location from that of path integration, we compare signed saliency at the base model, midpoint, and fine-tuned endpoint in Table 5 of Section 6.5. The single-point baselines and FPA-Merge use the same 200 labeled sample pairs and a 5% retention ratio. The ranking rule and merging settings are held constant across the three single-point baselines.

Consider the segment of the PL path from the base model to the target pole. Let Δp=θp−θ0\Delta^p=\theta_p-\theta_0 and γp(u)=θ0+uΔp\gamma_p(u)=\theta_0+u\Delta^p, where u∈[0,1]u\in[0,1]. A single-point method ranks coordinates by Δip∇iSp(γp(u0))\Delta_i^p\nabla_i S^p(\gamma_p(u_0)), using the gradient at u0u_0 to approximate the average gradient along the entire segment. Path integration accumulates the signed contribution of coordinate ii over the entire update

\qquad \sum_i A_i^p=\int_0^1\frac{d}{du}S^p(\gamma_p(u))\,du =S^p(\theta_p)-S^p(\theta_0).$$ By the chain rule, the sum of the attributions across all coordinates equals the change in the contrastive score induced by the full update. As parameters change together along the path, a coordinate's positive contribution at one point may weaken or even change sign later. Our goal is to retain the coordinates with the largest net contributions over the entire update path within the 5% budget, so we rank coordinates by their accumulated signed contributions. Table 5 shows that among the three single-point locations tested, the midpoint gives the best result for Extraversion-High (80.10), and the base model gives the best result for Extraversion-Low (14.90). FPA-Merge (PL) achieves 90.10 and 13.90 at the two poles, respectively, outperforming the best single-point result at each pole. **This is why we chose path attribution.** Accumulating contributions along the path lets us use the same selection rule for both target poles without having to identify the best single-point location for each pole separately. Two-point quadrature achieves 89.70 on Extraversion-High, 9.60 percentage points above the strongest single-point result in Table 5, showing that aggregating gradients from a few points along the path can yield gains (Appendix C.3, Table 13). We use eight-point quadrature by default so that the set of coordinates to be merged agrees more closely with the result from finer quadrature. Using 64-point quadrature as the reference, the Jaccard similarities of the sets selected by two-point and eight-point quadrature are 0.832 and 0.960, respectively (Table 14).
AuthorsOpenReview ↗

Response to Reviewer 9qLJ-Q2&Q3

Q2: Baseline comparisons across ten trait–pole settings and continuous control

In Table 18 of Appendix C.6, we extend the comparison with Task Vector and standard DaRE to all ten Big Five trait–pole settings. At a fixed merging coefficient magnitude of ∣α∣=1|\alpha|=1 , FPA-Merge (PL) outperforms both baselines on target-pole TRAIT in seven settings and on the reported AIME25 scores in nine settings. In the three settings the reviewer highlighted, Conscientiousness-High, Agreeableness-High, and Neuroticism-Low, standard DaRE's TRAIT scores are closer to the target poles. FPA-Merge (PL) achieves AIME25 scores of 44.58%, 41.67%, and 41.67%, respectively, exceeding the higher score of the two baselines in each setting by 2.92%–7.50%.

Table 4 of Section 6.4 helps explain why these three settings did not outperform Full SFT at a 5% retention ratio. When we retain all coordinates with nonnegative attribution (about 60% of the coordinates), TRAIT rises from 93.7 with Full SFT to 95.8 for Conscientiousness-High and from 91.3 to 92.5 for Agreeableness-High, and falls from 16.0 to 10.8 for Neuroticism-Low. In all three settings, the scores are closer to the target poles than those of Full SFT. This shows that steering in these settings can be strengthened by retaining a broader set of updates that contribute positively to the target. Table 18 shows the capability preservation advantage at a fixed 5% budget.

The additional ablation in Appendix C.8 shows that selecting by absolute attribution and reversing the updates with negative attribution, while still retaining 5% of the coordinates and fixing ∣α∣=1|\alpha|=1, reduces TRAIT for Neuroticism-Low from 18.1 with the original FPA-Merge(PL) variant to 14.7, below the Full SFT score of 16.0. This shows that, in this setting, changing coordinate selection and update direction can also strengthen target-pole steering within the original sparsity budget.

Figure 8 of Appendix C.2 extends the bipolar sweep for Extraversion to α∈[−2,2]\alpha\in[-2,2] and adds Task Vector and standard DaRE. At α=−1.5\alpha=-1.5, the TRAIT scores of FPA-Merge (PL), Task Vector, and DaRE fall within the narrow range of 13.1–14.2. At comparable levels of low-pole steering, their AIME25 scores are 43.75%, 38.33%, and 22.50%, respectively. We also add a five-seed evaluation for both Extraversion poles in Table 19 of Appendix C.6. With fixed SFT endpoints, FPA-Merge (PL) achieves the strongest average TRAIT steering and the highest average GPQA and AIME25 scores in both directions.

Q3: Comparison with TIES-Merging at the same per-vector retention ratio

Table 2 in Section 6.2 adds a comparison with TIES-Merging. We use the five-trait mixed-polarity configuration from Figure 3, [O,C,E,A,N]=[+,−,+,+,−][O,C,E,A,N]=[+,-,+,+,-]. Both methods retain 5% of the coordinates in each input vector, and we apply the merged updates to the base model and the MedMCQA expert initialized from the same base. TIES-Merging selects coordinates by displacement magnitude and resolves direction conflicts during merging. FPA-Merge selects coordinates by the signed functional contribution of each input update to the corresponding personality target.

On both recipient models, FPA-Merge brings all five dimensions closer to their target poles than TIES-Merging does. On the expert model, for example, the Extraversion-High score rises from 52.1 with TIES-Merging to 94.2 with FPA-Merge, while the Conscientiousness-Low score falls from 78.2 to 57.3. This comparison shows that, at the same per-vector retention ratio, selecting updates by their functional contributions to the target personality can achieve stronger multi-trait steering after merging multiple personality vectors.

AuthorsOpenReview ↗

Response to Reviewer 9qLJ-Q4

Q4: Offline attribution cost, validation across models, and expert-model transfer

Theoretical computational cost. Let KK be the quadrature points per PL branch, NaN_a the attribution pairs, bab_a the microbatch size, and CC the separately computed contrastive channels. With MaM_a GPUs, r=min⁡(2,Ma)r=\min(2,M_a) branches run in parallel. Assuming similar branch runtimes,

TPL≈⌈2r⌉KC⌈Naba⌉(tforwarda+tbackwarda)+Textraa,T_{\mathrm{PL}} \approx \left\lceil\frac{2}{r}\right\rceil KC\left\lceil\frac{N_a}{b_a}\right\rceil (t_{\mathrm{forward}}^a+t_{\mathrm{backward}}^a)+T_{\mathrm{extra}}^a,

where tforwardat_{\mathrm{forward}}^a and tbackwardat_{\mathrm{backward}}^a are microbatch times, and TextraaT_{\mathrm{extra}}^a covers loading, interpolation, accumulation, and saving. For data-parallel SFT with MsM_s GPUs, microbatch size bsb_s, AA accumulation steps, and SS optimizer steps (effective batch size MsbsAM_sb_sA),

TSFT≈S[A(tforwards+tbackwards)+tsync+topt]+Textras,T_{\mathrm{SFT}} \approx S\left[A(t_{\mathrm{forward}}^s+t_{\mathrm{backward}}^s)+t_{\mathrm{sync}}+t_{\mathrm{opt}}\right]+T_{\mathrm{extra}}^s,

where tsynct_{\mathrm{sync}}, toptt_{\mathrm{opt}}, and TextrasT_{\mathrm{extra}}^s denote synchronization, optimizer, and other overhead, respectively.

Experimental configuration and GPU memory estimate. Qwen3.5-9B uses 200 attribution pairs, a microbatch size of 8, and a maximum sequence length of 1024. The two PL branches each have 8 quadrature points, with two contrastive channels computed separately at each point and 25 microbatches processed per channel. Thus, bipolar attribution for each trait requires 2×8×2×25=8002\times8\times2\times25=800 microbatch forward and backward passes. Attribution uses FP32 parameters and gradients. For P=8,953,803,264P=8{,}953{,}803{,}264 parameters, the full weights and full gradients each occupy 4P/230≈33.364P/2^{30}\approx33.36 GiB. If both reside in full on one GPU at the same time, they occupy 8P/230≈66.718P/2^{30}\approx66.71 GiB in total. The quadrature points are processed sequentially and reuse GPU memory, so increasing KK primarily increases the number of forward and backward passes. Attribution does not need to maintain SFT optimizer state.

Measured cost. We measured bipolar PL attribution across four traits on Qwen3.5-9B using two H100 80GB GPUs. Memory reports the highest single-GPU allocation.

Model and stageGPUsWall-clock time per traitGPU-hours per traitPeak allocated memory on one GPU
Qwen3.5-9B, FPA-Merge (PL)2 × H100 80GB42.29–59.21 min1.41–1.9768.32 GiB

For fixed SFT endpoints and attribution settings, attribution is computed offline once per trait. Its cache supports different retention ratios, while the resulting vectors support coefficient sweeps and steering of same-base experts without recomputation.

Across model scales and families. Table 27 of Appendix G.1 extends the Qwen3.5-9B evaluation to all ten Big Five trait–pole settings. Both FPA-Merge path variants move TRAIT toward the target pole in all ten settings. FPA-Merge(PL) outperforms Full SFT on target-pole TRAIT in seven settings, and at least one FPA-Merge variant achieves the best target-pole TRAIT in eight settings. FPA-Merge(PL) outperforms Full SFT and all four merging baselines on GPQA in ten settings and on AIME25 in nine. For example, on Openness-Low, FPA-Merge(PL) retains only 5% of the coordinates and reduces TRAIT from 21.2 with Full SFT to 19.4, while increasing GPQA from 35.10% to 60.10%.

Using Llama-3.1-8B-Instruct's own SFT endpoints (Appendix G.2, Table 28), we also obtain Extraversion steering at both poles. At the low pole, FPA-Merge(PL) achieves TRAIT 6.40 versus 29.60 (Base) and 13.60 (Full SFT), with GPQA 22.98% versus Full SFT's 5.30%. This provides preliminary evidence of cross-family applicability.

Reusing vectors on expert models derived from the same base. The updates in Eq.10 of Section 4.3 are defined with respect to the source base model's parameter coordinates. An expert initialized from the same base with corresponding parameter coordinates can receive these vectors directly. Figure 3 and Table 3 in Section 6.2 illustrate this use. When the same set of five-trait updates is applied to the MedMCQA expert, all five traits move toward their target poles. MedMCQA accuracy goes from 63.33% before merging to 60.24%, while dense merging at the same coefficients yields 39.56%. This transfer reuses the constructed vectors, allowing controllable personality steering to be added to experts derived from the same base without retraining personality endpoints or computing attribution again.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.