- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates why a fixed activation-steering coefficient works on one checkpoint but can produce little change in another one. They studied 15 instruction-tuned models and give personality attributes, and within each model they find the scale is stable across attributes.
Critical strengths
The problem is interesting and well motivated for steering literature.
RMSnorm derivative gives insights about why scale is informative.
Exhaustive calibration splits are provided.
Critical weaknesses
Feedback-based selection seems to add little over trivial (zero-feedback) baselines. In Tab 2 it seems like choosing the largest expanded amplitude beats the six-rating selector on 4 of 5 models, and the selector's small mean advantage seems to be coming entirely from Qwen3.
The last two authors’ surnames in one reference seems to be fabricated. The reference in paper: Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Ma, and Yixin Liang. Evaluating and inducing personality in pre-trained language models. Advances in Neural Information Processing Systems, 36, 2024. The correct author names for last two authors: Chi Zhang (instead of Chi Ma) and Yixin Liang (instead of Yixin Zhu): https://proceedings.neurips.cc/paper_files/paper/2023/file/21f7b745f73ce0d1f9bcea7f40b1388e-Paper-Conference.pdf
Questions for authors
Primary judge is Gemma-2-2B-it, does the main results hold under a different judge?
2 comments
Thanks for your rebuttal, I increased my score to Weak Accept.
Response to Reviewer UJfJ
Thank you for highlighting our problem motivation, RMSNorm derivation, and exhaustive split design.
1. Reference Correction (
Jiang et al., 2023)Thank you for catching the surname error in
Jiang et al.(NeurIPS 2023, Vol. 36, pp. 10622–10643) and for linking the official NeurIPS PDF. Per the linked paper (noting that your review text also inadvertently swapped the order in "Yixin Liang (instead of Yixin Zhu)"), we have restored the exact official authorship (Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu, 2023) on Page 11 (Lines 582–587).2. Weakness 1: Why Sparse Behavioral Selection Matters Over the Zero-Feedback
LargestBaselineWhy use behavioral feedback if
Largestgives a slightly higher 5-trait average on 4 of 5 models? Our 25-curve audit (App. G, Tab. 10, Page 19), zero-self-judging reanalysis (App. F, Tab. 9), and 7B-judge replication (App. I, Tab. 21) clarify how the two stages work together:Qwen3-0.6B(App. G, Tab. 10): Across all 25 dose curves Gt(ai;S), 8 of 25 peak strictly at an interior amplitude (a1 or a2) across 3 of the 5 models, and 3 reverse sign (G(a3)<0) at a3:App. G, Tab. 10)Qwen3-0.6BAgreeableness / NeuroticismQwen3-0.6BOpenness / ConscientiousnessQwen2.5-1.5BAgreeableness / NeuroticismQwen2.5-7BExtraversion / NeuroticismTab. 6). Expanding the grid liftsLargestto 1.172, showing the value of Stage 1 (scale-motivated range expansion).Qwen3-0.6B, and negative reversal on 3 traits).Largestgrows from +0.095 (1.266 vs. 1.172) to +0.241 (1.413 vs. 1.172) across all 5 models.Gemma-2-2B(self-judging) is excluded (App. F, Tab. 9), the 6-rating selector's mean advantage overLargestmore than doubles from +0.095 to +0.193 (1.140 vs. 0.947; reaching +0.334 [1.281 vs. 0.947] at 30 ratings), and also exceedsUniform(1.021).3. Question 1: Do the Main Results Hold Under a Different Judge?
Yes. Beyond our 1,260-response check (
Spearmanρ=0.951 on model rankings), we usedQwen2.5-7B-Instructto score all 13,650 stored responses across both complete 5-model grids (6,825Orig +6,825Exp) (Sec. 5.5, Page 10; App. I, Tabs. 21–22):Tab. 22): Across all 13,650 responses,Qwen2.5-7B-InstructandGemma-2-2B-itagree within ±1 point on 93.2% of ratings (92.6%Exp,93.8%Orig; Pearson r=0.775, Spearman ρ=0.781; 0.797 and 0.803 on Exp).Tab. 21): UnderQwen2.5-7B-Instruct, the 6-rating expanded selector improves all 25/25 model–trait pairs over the original grid (2.210 vs. 0.472, gain +1.738, 95% CI [1.172,2.440]; 2.408 at 30 ratings) and beats bothLargest(2.200; 1.953 vs. 1.853 excluding Gemma; 2.102 vs. 2.029 excluding Qwen2.5-7B) andUniform(2.069; 1.953 vs. 1.771 excluding Gemma; 2.102 vs. 1.950 excluding Qwen2.5-7B). OnQwen3-0.6B, 6 ratings again prevent largest-dose collapse (2.620 vs. 1.332, +1.288).