- 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 along a bipolar merging path (piecewise-linear or quadratic) through , , and . FPA-Merge keeps the top- coordinates by attribution (default ) 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 .
Critical strengths
- 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 () gives 69.0, while gives a stable support ( vs. ) and around 90 TRAIT. So the path integration matters, and the method is not a disguised magnitude heuristic.
- 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
-
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 at (App. C.2). Eq. 9 and Algorithm 1 define 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, , 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.
-
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 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.
- 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 .
- What happens if you remove only the coordinates with negative attribution and keep all positive ones? And does applying only the bottom- coordinates move TRAIT toward the opposite pole? These results would directly support or refute the cancellation account over a generic sparsity effect.
- 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 at or a Fisher-weighted mask? This would show how much of the gain comes from path integration versus having target-specific gradient signal.
- 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?
- 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?
- 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
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.
- 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.
- 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.
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 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 be the fine-tuned endpoint for the target pole, be the displacement, and be the path, where . Single-point saliency uses the gradient at 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 .
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.
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.
For standard DaRE, given the full fine-tuning displacement Δ, each coordinate is retained independently with probability q, so mi∼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.