- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper studies how prompt-induced personality forms inside LLMs and uses the findings to design a steering method. The analysis adds Gaussian noise to the embeddings of trait-description tokens in prompts of the form "You are a {trait} AI assistant. Question: {q}?" and measures how patching clean states back recovers the final-layer, final-position shift, projected onto each example's own shift direction (Eq. 2). The reported pattern is that the signal first appears at trait tokens, spreads to content tokens, is aggregated toward the final prompt position mainly through attention (attention patching recovers more than FFN patching), and stabilizes in middle-to-late layers. Based on this, the authors propose Stage-wise Personality Steering (SPS), a training-free method with three parts: Trait Seeding adds a prompt-side contrastive vector at selected content tokens in the first three layers (Eq. 7), Attention Routing adds a bias to attention logits toward trait-aligned source tokens in a 6-layer window (Eq. 10), and Expression Anchoring adds a response-side contrastive vector at one layer chosen from the recovery curve (Eq. 11). On PersonalityBench with Llama-3.1-8B-Instruct, SPS reports an Overall Mean of 9.54 (out of 10) versus 9.42 for PERSONA and 9.30 for NPTI, with a 0.27-point MMLU drop. SPS alone is also run on Qwen3-4B, Qwen2.5-7B, and Mistral-7B.
Critical strengths
- The analysis-to-method structure is a reasonable idea. Grounding intervention locations in a patching analysis, instead of picking a layer by downstream score alone, is a useful direction for steering work. The analysis is per-example (recovery along each example's own shift) and does not depend on a fixed PCA subspace or a trained probe, which avoids one common source of artifacts.
- The main comparison includes a statistical test, and the ablation is complete. Table 1 reports paired-bootstrap CIs and a permutation test against the strongest baseline, which many steering papers omit. Table 4 covers all seven stage combinations, and Fig. 3(b) shows where each stage's gains fall across layers.
Critical weaknesses
-
The representation-dynamics findings largely restate known Transformer information flow and are not shown to be specific to personality. Because only trait-token embeddings are corrupted, the difference necessarily appears first at those positions (Fig. 1a). Because attention is the only mechanism that moves information across positions, it is expected that attention patching recovers more of a cross-position shift than FFN patching (Fig. 1c). A final-position state that becomes recoverable in middle layers and plateaus later (Fig. 1d) is also the typical pattern in prior patching work on factual recall (e.g., Meng et al., 2022). The paper does not include a control attribute (e.g., topic, sentiment, or a non-personality instruction) to show which parts of this pattern are personality-specific. The analysis also uses a single template and 50 questions per trait. In addition, the method only loosely follows the analysis: the seeding layers are fixed to the first three, and the routing window is fixed to to , so most placement choices are still heuristic. Fig. 1(e) shows that trait-description tokens carry the most recoverable signal, yet Trait Seeding is applied to question-content tokens and BOS. The link between the analysis and the design is weaker than the paper's framing suggests.
-
The empirical evidence is too narrow and too close to ceiling to support the state-of-the-art and capability-preservation claims.
- Ceiling and margin. The score is the sum of two pole means with a maximum of 10. PERSONA, NPTI, and P² already reach 9.29–9.42, and the SPS gain over PERSONA is 0.12. The ablation shows Trait Seeding adds only 0.06 Mean on top of the AR+EA variant, which also has lower Std (0.81 vs. 0.83), so the full method is not the most stable configuration.
- Statistical reporting. The text says the Conscientiousness interval spans zero, but Table 1 reports [+0.008, +0.180] for that trait. The overall CI [+0.091, +0.187] also looks tight relative to a permutation p = 0.0217. The resampling unit (questions, poles, or traits) is not stated, so these numbers are hard to interpret.
- Baselines and other models. Baselines appear only on Llama. Table 2 reports SPS alone on the other three models, so it shows SPS works there but not that it beats existing methods. How baseline hyperparameters were set is not described in the main text.
- Capability. General capability is measured only by MMLU, only on Llama, and only for SPS. No baseline MMLU numbers are reported, so "limited degradation" cannot be compared against anything. Multiple-choice MMLU with short outputs also says little about the effect of decoding-time interventions (Attention Routing and Expression Anchoring act at every generated token) on longer reasoning or open-ended generation.
- Judge. Trait vectors are built from pairs filtered by an LLM judge, and evaluation uses GPT-4.1 as the judge. If the same or a similar judge is used for both, the method may be partly optimized toward the judge. There is no human evaluation and no separate fluency or coherence metric in the main text.
Questions for authors
Items 1–3 could change my assessment. Items 4–6 are clarifications.
- Can you run the same corruption-and-patching analysis on a non-personality instruction (e.g., a topic or formatting instruction)? Which parts of the stage-wise pattern differ from personality?
- Please report baselines (at least PERSONA and NPTI) on Qwen2.5-7B, Qwen3-4B, and Mistral-7B, and report MMLU (plus at least one generation-based benchmark such as GSM8K) for the baselines as well. How were baseline hyperparameters and SPS strengths (, , , ) selected, and was any tuning done on PersonalityBench itself?
- Please reconcile the Conscientiousness CI in Table 1 with the text, state the resampling unit for the bootstrap and permutation tests, and explain why the full method is preferred over AR+EA given its lower Std.
- Is the personality instruction present in the prompt at inference when SPS is applied, or does SPS steer from a neutral prompt? Is the same setting used for all baselines?
- Which judge filters the contrastive pairs, and does it overlap with the GPT-4.1 evaluator? How are pole scores mapped to the 0–10 range, and what exactly does Std measure (between-pole or across questions)?
- In Eq. 3–4, the prompt-side and response-side means use the same notation, and in Eq. 5 is not defined. If , then Eq. 8 reduces to . Is that the case, or do the filtered sets and differ?