acceptodds
Under review as a conference paper at ICLR 2027

Exploring Role-Playing Mechanisms in Large Language Models

Abstract

While large language models exhibit strong role-playing capabilities, the internal mechanisms that support role-conditioned generation remain insufficiently understood. To address this gap, we introduce Contrastive Decomposed Forward Pass for Complete Sequences (CoDePass), which detects role-related computational components by comparing their contributions across matched prompts and complete responses. Our experiments reveal that role playing is organized through a sparse subset of attention heads: recurrent heads retrieve profile and question information, track identity cues, and maintain positional structure, while multi-layer perceptrons (MLPs) progressively transform these signals from explicit role instructions into identity-conditioned response representations. Further behavioral analyses show that while context retrieval, response timing, and baseline semantic construction rely on universally reusable mechanisms across models and characters, identity selection and persona rendering remain highly specific to the assigned role. Finally, targeted fine-tuning of as little as 2% of the model parameters improves role-playing quality over capacity-matched random and full-parameter tuning on held-out characters while preserving general capabilities. These results reveal a division of labor: sparse attention heads control which profile and identity information reaches the response pathway, while MLPs already provide the organized transformation that converts this information into role-related responses. Our code is anonymously available at https://anonymous.4open.science/r/CoDePass-0E45.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Role-Playing, Dialogue Generation, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vTspOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CoDePass to identify role-related components by contrasting their contributions to complete reference responses under matched prompts. It analyzes attention-head and MLP patterns and reports gains on held-out RoleBench characters from fine-tuning selected attention components. The results support the usefulness of the ranking, but the proposed division of labor and improvements in character fidelity are not yet convincingly established.

Critical strengths

The selective fine-tuning results have meaningful controls. Ranking and training use 75 characters, while evaluation uses six disjoint characters. Random SFT matches layerwise head counts, shared GQA K/V groups, and parameter count, and the advantage persists across multiple training seeds and random selections. These comparisons provide useful evidence that the ranking identifies parameters whose adaptation transfers across characters.

Critical weaknesses
  1. The proposed cue-to-identity-to-answer mechanism is stronger than the evidence. Figure 6a measures the pre-MLP residual, which combines attention and earlier MLP contributions, so this plot alone does not establish that MLPs produce the observed identity alignment. Figure 6b isolates MLP updates but evaluates their alignment with the response direction. Knockout and functional-class fine-tuning support component importance and tuning utility, but do not establish the claimed roles in identity selection, profile retrieval, and response timing.

  2. The role-playing evaluation does not establish improved character fidelity. Selected cases in Figures 40–42 show Targeted SFT receiving higher ROUGE-L scores despite incorrect character motives or facts. The length controls and qualitative analysis are useful, but do not determine whether character knowledge and belief consistency improve across the held-out evaluation. This limits the use of the fine-tuning results as validation of the proposed role-playing mechanism.

Questions for authors
  1. What evidence attributes the mid-layer increase in identity alignment to MLPs, given that Figure 6a probes the pre-MLP residual? Can the respective contributions of attention and MLP updates be separated, and can a targeted intervention establish a selective effect on role identity?

  2. In Figure 41, Targeted SFT replaces Ripley’s fraudulent income with inheritance and investments yet obtains 33.33 ROUGE-L; Random Set 2 describes manipulation but scores 21.28. On the existing held-out outputs, does an assessment of character facts and beliefs, independent of reference overlap, still favor Targeted SFT?

  3. Section 4.3 attributes Base-to-Instruct changes to instruction tuning, while Appendix C.5.2 also changes the completion interface to a chat interface. Do the conclusions persist under a shared template? What evidence establishes MLP stability beyond the nonsignificant test in Table 1?

If the authors can substantiate the proposed role-specific functions and confirm gains in character fidelity under an assessment independent of reference overlap, this could strengthen my confidence in the paper’s conclusions.

6 comments
AuthorsOpenReview ↗

Response to vTsp (1/3): W1; Q1

Thank you for recognizing the clarity of our presentation and the rigorous controls in our selective fine-tuning experiments—specifically that our 75-to-6 disjoint character split and layerwise/GQA-matched Random SFT confirm cross-character parameter transfer.

W1 & Q1. The proposed cue-to-identity-to-answer mechanism is stronger than the evidence.

W1-1 & Q1: Isolating direct MLP updates (ΔMLP\Delta_{\text{MLP}}) from attention updates (ΔAttn\Delta_{\text{Attn}}) in identity alignment.

By decomposing the residual stream into isolated sublayer writes, we directly separate the contributions of attention mechanisms and MLP updates. This directly resolves the concern: middle-layer MLPs—not attention heads—are the primary drivers of the mid-layer surge in identity alignment and identity consolidation.

While Figure 6a plotted the pre-MLP residual state entering each block, we now separately compute the direct sublayer write projections cos⁡(ΔAttnl,uT)\cos(\Delta_{\text{Attn}}^l, u_T) and cos⁡(ΔMLPl,uT)\cos(\Delta_{\text{MLP}}^l, u_T) across depth (120 matched prompts across six characters), where ΔAttnl=Attnoutl−Attninl\Delta_{\text{Attn}}^l = \text{Attn}_{\text{out}}^l - \text{Attn}_{\text{in}}^l and ΔMLPl=MLPoutl−MLPinl\Delta_{\text{MLP}}^l = \text{MLP}_{\text{out}}^l - \text{MLP}_{\text{in}}^l. Table A shows that attention updates dominate early indicator/context routing (L0–L10), whereas direct MLP writes ΔMLPl\Delta_{\text{MLP}}^l drive the significant mid-layer surge in identity alignment (L11–L24, p<.001p<.001) and late-layer response generation (L25–L35):

LayersAttention / INDMLP / INDAttention / IDMLP / IDMLP / Response
Early (L0–L10)+0.0084+0.0031+0.0018+0.0024-0.0016
Middle (L11–L24)+0.0042+0.0019+0.0021+0.0148+0.0038
Late (L25–L35)+0.0015+0.0012+0.0034+0.0089+0.0196

Table A: Cosine alignment of isolated sublayer writes (ΔAttnl\Delta_{\text{Attn}}^l vs. ΔMLPl\Delta_{\text{MLP}}^l) across depth.

W1-2 & Q1: Targeted causal intervention on mid-layer identity and routing functions.

Targeted sublayer interventions establish a clean functional double dissociation across profile retrieval, identity synthesis, and response-boundary timing.

Motivated by recent multi-component causal tracing research showing that functional circuits operate as coordinated component groups [3], we perform targeted interventions on each functional group across 200 matched questions and compare against layer-matched random head/direction controls. Each intervention selectively impairs its corresponding target function without disrupting non-target capabilities (Table B): zeroing the 3 role-context heads selectively lowers profile content recall; zeroing the 2 role-indicator heads or removing the projection of mid-layer MLP writes (L12–L22) onto uIDu_{\text{ID}} selectively reduces the Identity Probe score and LLM Judge score (while keeping content recall and normal boundary completion intact); and zeroing the positional heads selectively disrupts normal response-boundary termination while preserving persona quality:

Intervention (200 Questions)Identity Probe ΔProfile Content Recall (%)LLM Judge (1–5)Boundary Completion (%)
Unmodified+0.000023.923.91100.0
Zero Context Heads (3)-0.000618.143.88100.0
Zero Indicator Heads (2)-0.014223.803.42100.0
Zero Positional Heads (2)-0.000423.753.8986.5
Remove Mid-MLP uIDu_{\text{ID}} Write-0.018623.843.36100.0
Matched Random Controls-0.000523.853.89100.0

Table B: Dissociation effects of functional sublayer interventions vs. layer-matched random controls.

Therefore, targeted interventions establish a selective causal effect on role identity, and together with isolated sublayer writes, directly substantiate the claimed division of labor across profile retrieval, identity selection/synthesis, and response timing.

Refs. [1] Cai: ThinkPersona (ACL'26); [2] Lai: RoleCDE (ACL'26); [3] Yan: Multi-component Tracing (ACL'26); [4] Yi: Too Good to be Bad (ACL'26); [5] Liao: MOA (ACL'26).

AuthorsOpenReview ↗

Response to vTsp (2/3): W2; Q2

W2 & Q2. The role-playing evaluation does not establish improved character fidelity.

W2-1 & Q2: Character knowledge and belief consistency improve across the held-out evaluation.

An evaluation of character facts, motives, and beliefs on the existing held-out set—completely independent of reference overlap and combined with blinded human preference judgments—confirms that Targeted SFT significantly outperforms Random SFT and Full SFT.

First, reference-based lexical overlap (ROUGE-L) on RoleBench is a mainstream evaluation paradigm in recent role-playing literature, where factual content recovery and surface rhetorical style represent two distinct objectives. Recent studies—including ThinkPersona [1] and RoleCDE [2]—evaluate RoleBench using ROUGE-L.

Second, Figures 40–43 in Appendix E.6 were curated specifically from failure cases to transparently illustrate the shared conflict faced by safety-aligned LLMs when handling morally transgressive personas [2,4]. In fact, in Figure 41, Full SFT (25.64 R-L) and Random Sets 1 & 3 (30.30 R-L) commit the exact same error as Targeted SFT (replacing Ripley's fraudulent income with "inheritance and investments").

Finally, to test whether character knowledge and belief consistency improve without lexical-overlap bias, we re-evaluate the existing held-out outputs using an independent LLM judge (deepseek-flash, given only the character profile and question, without the reference answer) and blinded human evaluators (300 items, 3 annotators, κ=.64\kappa=.64). Targeted SFT achieves the highest reference-free factual accuracy, belief/motive consistency, Overall score, and blinded human win rate among all models, and leads across all fine-tuned baselines:

MethodRef-Free Fact Score (1–5) ↑Belief / Motive (1–5) ↑Voice (1–5) ↑Overall ↑Targeted Human W/T/L
Qwen3-8B-Instruct3.10 ± 0.003.24 ± 0.003.42 ± 0.003.25 ± 0.00172 / 77 / 51
Full SFT3.26 ± 0.043.41 ± 0.033.12 ± 0.033.26 ± 0.03127 / 107 / 66
Random SFT3.21 ± 0.053.36 ± 0.053.15 ± 0.043.24 ± 0.04184 / 56 / 60
Targeted SFT3.54 ± 0.033.68 ± 0.023.29 ± 0.033.50 ± 0.02-

Table C: Reference-free independent judge evaluation (1–5 scale) and blinded human preference.

W2-2: This limits the use of the fine-tuning results as validation of the proposed role-playing mechanism.

Our mechanistic conclusions rest on a three-part interpretability framework rather than fine-tuning alone, and the reference-free fidelity gains above further confirm that Targeted SFT validates the localized mechanism.

First, as recognized by Reviewers VRNr and ueGy, our primary contribution is the first systematic component-level mechanistic study of role playing in decoder-only LLMs—integrating CoDePass counterfactual attribution, progressive knockouts, and sublayer behavioral diagnostics.

Second, the proposed role-playing mechanism is established primarily through RQ1 (counterfactual localization and knockout necessity) and RQ2 (value-weighted interaction maps, isolated ΔAttn/ΔMLP\Delta_{\text{Attn}}/\Delta_{\text{MLP}} writes, and functional double dissociation in Tables A–B), rather than relying solely on RQ3 fine-tuning. Selective fine-tuning in RQ3 serves as a downstream prescriptive validation: combined with the reference-free fact and belief improvements in Table C, it indicates that the sparse attention components identified by CoDePass indeed govern character adaptation more effectively and efficiently (1.52% parameters, 35.9% faster training) than random or full-model updates.

Refs. [1] Cai: ThinkPersona (ACL'26); [2] Lai: RoleCDE (ACL'26); [3] Yan: Multi-component Tracing (ACL'26); [4] Yi: Too Good to be Bad (ACL'26); [5] Liao: MOA (ACL'26).

AuthorsOpenReview ↗

Response to vTsp (3/3): Q3

Q3. Section 4.3 attributes Base-to-Instruct changes to instruction tuning, while Appendix C.5.2 also changes the completion interface to a chat interface. Do the conclusions persist under a shared template? What evidence establishes MLP stability beyond the nonsignificant test in Table 1?

Q3-1: Shared prompt template between Base and Instruct.

The Base-to-Instruct reorganization conclusion holds fully when both checkpoints are evaluated under the exact same plain-text template (Shared Plain) or chat template (Shared Chat).

To rule out interface formatting as a confounder, we re-run the Jack Sparrow Base-vs.-Instruct attribution comparison (corresponding to Appendix Table 19, 100 identical questions per template) using identical input and target token sequences on both checkpoints. Table D confirms that under either shared template, attention heads undergo a highly significant distributional shift and top-set turnover, while MLP blocks maintain high overlap:

Prompt InterfaceAttention K-S D (p)Attention Top-32 OverlapMLP K-S D (p)MLP Top-8 Overlap
Native (Table 19).584 (2.0e-182)17 / 32 (53.1%).556 (1.9e-5)6 / 8 (75.0%)
Shared Plain.214 (<1e-20)16 / 32 (50.0%).222 (.340)6 / 8 (75.0%)
Shared Chat.228 (<1e-22)15 / 32 (46.9%).250 (.212)6 / 8 (75.0%)

Table D: Base-vs.-Instruct attribution comparisons under identical prompt templates.

Q3-2: Evidence for MLP stability beyond the two-sample K-S p-value.

Four sample-size-invariant rank correlation, cosine similarity, distributional divergence, and functional trajectory metrics consistently confirm that the MLP backbone is substantially more conserved and stable than attention heads.

First, as noted in Appendix C.5.1, because the two component families have different cardinalities (1,152 attention heads vs. 36 MLP blocks), we do not rely on the K-S test pp-value alone.

Second, Table E reports four cardinality-independent stability metrics under the shared plain-text interface—layerwise Spearman rank correlation (ρ\rho), attribution vector cosine similarity, Jensen-Shannon divergence (JSD), and direct sublayer-write identity trajectory correlation (ΔMLP\Delta_{\text{MLP}} vs. ΔAttn\Delta_{\text{Attn}})—which mutually reinforce our RQ3 fine-tuning finding (that updating a small fraction of attention parameters suffices, whereas updating all MLPs degrades performance, Table 24):

Component FamilySpearman ρ ↑Attribution Cosine ↑JSD ↓Identity Trajectory r ↑
Attention Heads0.38840.72670.10920.8332
MLP Blocks0.78420.89140.04180.8578

Table E: Cardinality-invariant Base/Instruct stability metrics under the shared plain-text interface.

Refs. [1] Cai: ThinkPersona (ACL'26); [2] Lai: RoleCDE (ACL'26); [3] Yan: Multi-component Tracing (ACL'26); [4] Yi: Too Good to be Bad (ACL'26); [5] Liao: MOA (ACL'26).

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

To explore the internal mechanisms that support role-conditioned generation, the paper proposes Contrastive Decomposed Forward Pass for Complete Sequences(CoDePass), which detects role-related computational components by comparing their contributions across matched prompts and complete responses. Authors find that sparse attention heads retrieve and route identity cues into the response pathway, while MLPs progressively transform these signals into role-conditioned responses. The paper also designs a targeted fine-tuning strategy of as little as 2% of the model parameters, improving role-playing quality while preserving general capabilities.

Critical strengths

The paper introduces CoDePass, and provides the first component-level analysis of role playing in decoder-only LLMs, systematically combining component localization, progressive knockout, complementary diagnostics, and component-guided fine-tuning.

The diagnostics reveals a clear division of labor: Specialized attention heads route context and identity cues. Meanwhile, MLPs progressively transform these explicit role instructions into identity-conditioned answers.

Targeted fine-tuning over a minimal parameter subset has practical value. It enhances role-playing performance while updating only 1.52% of the parameters and reducing training time by 35.9% relative to Full SFT.

Critical weaknesses

The core knockout experiments and most mechanistic analyses appear to center on Qwen3-8B-Instruct. And different tuned parameter budgets are used across different architectures. Evaluating other models under a unified parameter budget would provide stronger evidence for the generality of the proposed mechanism.

The paper lacks the total GPU time and peak memory usage during the attribution stage, the overhead relative to a standard forward pass, and how computational cost scales with sequence length and the number of components.

The paper shows that targeted fine-tuning might weaken a character’s unique persona, which authors can explore the potential reasons.

Questions for authors

Please refer to the weakness

5 comments
AuthorsOpenReview ↗

Response to VRNr (1/2): W1

Thank you for recognizing our first component-level analysis of role playing in decoder-only LLMs, the clear division of labor between attention routing and MLP transformations, and the practical value of fine-tuning a small parameter subset to substantially reduce training time.

W1. Core knockout experiments and mechanistic analyses center on Qwen3-8B-Instruct, and different tuned parameter budgets are used across architectures.

W1-1: Cross-architecture knockout and mechanistic consistency on Llama-3.1-8B-Instruct.

Both our manuscript cross-architecture diagnostics and new progressive knockout experiments on Llama-3.1-8B-Instruct confirm that the discovered role-playing mechanism and component sensitivity generalize across architectures.

First, our mechanistic diagnostics already cover Llama-3.1-8B-Instruct in manuscript Figure 1b and Appendices C.1, C.3, D.2.2, and D.4. Specifically, Table 21 indicates that role-context and positional head functions transfer stably on Llama, corroborating recent findings on sparse personalization heads [1] and confirming that this division of labor is a cross-architecture regularity rather than a Qwen-specific phenomenon.

Second, we add the progressive knockout evaluation on Llama-3.1-8B-Instruct under the identical protocol (200 role-specific questions, removing 20 heads or 1 MLP block per step). Consistent with Qwen's trajectory in Figure 2, Table A shows that Llama's attention routing exhibits initial redundancy over Steps 1–3 (20–60 heads) before degrading steeply across Steps 5–10 (100–200 heads), whereas ablating even the first top-ranked MLP block disrupts role generation immediately:

Heads / MLPsRandom Heads Ret. (%)CoDePass Heads Ret. (%)Random MLP Ret. (%)CoDePass MLP Ret. (%)
20 / 198.56104.69100.017.08
60 / 3102.2499.8964.203.94
100 / 5100.2969.7428.400.57
160 / 899.1255.4011.200.25
200 / 1098.5144.924.150.10

Table A: Llama-3.1-8B-Instruct progressive knockout, normalized to its unmodified baseline.

W1-2: Unified parameter budgets across architectures.

Under strictly unified parameter budgets, Targeted SFT consistently outperforms capacity-matched Random SFT across all four Qwen and Llama models.

First, in Appendix Table 22, we initially set a higher 5.00% parameter budget for Llama because its CoDePass attention-head attribution is slightly more dispersed across early-to-middle layers than Qwen's (comparing manuscript Figure 1b with 1a).

Second, to provide a strictly unified comparison—while following the layerwise head-count and GQA-group matched controls recognized by Reviewer vTsp—we evaluate both ~1.5% and 5.0% unified budgets across all four models in Table B. Across every model and budget, Targeted SFT achieves superior role-playing performance while preserving general benchmark capabilities. Moreover, we observe that expanding the trainable parameter budget is not always beneficial across models and can slightly degrade both role-playing and general performance:

ModelBudgetRandom (Role / General R-L)Targeted (Role / General R-L)Δ Role / GeneralTargeted MMLU / ARC
Qwen3-8B-Instruct1.52% (124.8M)21.31 ± 0.10 / 34.07 ± 0.7423.39 ± 0.11 / 39.08 ± 0.12+2.08 / +5.0174.21 ± 0.12 / 65.93 ± 0.32
Qwen3-8B-Instruct5.01% (410.0M)22.18 ± 0.15 / 35.12 ± 0.5223.51 ± 0.09 / 39.24 ± 0.21+1.33 / +4.1274.08 ± 0.18 / 65.85 ± 0.28
Qwen3-8B-Base1.49% (121.6M)23.84 ± 0.32 / 36.16 ± 0.2325.14 ± 0.27 / 38.33 ± 0.64+1.30 / +2.1776.76 ± 0.11 / 68.43 ± 0.39
Qwen3-8B-Base4.99% (408.9M)23.89 ± 0.10 / 36.41 ± 0.1824.55 ± 0.10 / 38.29 ± 0.21+0.66 / +1.8874.34 ± 0.77 / 66.21 ± 0.82
Llama-3.1-8B-Instruct1.53% (122.7M)23.62 ± 0.10 / 32.17 ± 0.3524.48 ± 0.08 / 36.04 ± 0.41+0.86 / +3.8767.42 ± 0.35 / 62.25 ± 0.29
Llama-3.1-8B-Instruct5.00% (401.6M)24.04 ± 0.04 / 32.68 ± 0.2924.24 ± 0.20 / 36.10 ± 0.42+0.20 / +3.4267.39 ± 0.40 / 62.17 ± 0.52
Llama-3.1-8B-Base1.51% (121.6M)24.65 ± 0.09 / 35.12 ± 0.4225.18 ± 0.08 / 36.42 ± 0.48+0.53 / +1.3064.75 ± 0.38 / 61.42 ± 0.31
Llama-3.1-8B-Base5.00% (401.6M)24.92 ± 0.10 / 35.31 ± 0.4125.33 ± 0.06 / 36.89 ± 0.56+0.41 / +1.5864.82 ± 0.34 / 61.49 ± 0.25

Table B: Unified parameter budget comparisons across Qwen and Llama.

Refs. [1] Zhang: Preference Heads (ACL'26); [2] Liao: MOA (ACL'26); [3] Yi: Too Good to be Bad (ACL'26); [4] Lai: RoleCDE (ACL'26).

AuthorsOpenReview ↗

Response to VRNr (2/2): W2-W3

W2. Computational cost.

CoDePass is a gradient-free, one-time offline attribution pass whose computational cost scales linearly with sequence length and attributed layers, adding minimal memory overhead and requiring very little time per 32-example discovery cell.

First, because CoDePass runs in torch.no_grad() inference mode without storing backward activation graphs, its peak memory overhead is minimal (+4.87 GB at 256 tokens).

Second, by vectorizing component readout inner products ⟨zc,t,WU[yt]⟩\langle z_{c,t}, W_U[y_t]\rangle along the sequence dimension, attributing all 1,152 attention heads and 36 MLP blocks on a single H100 GPU takes only 21.8–78.4 seconds per 32-example discovery cell. Manuscript Table 20 shows that 32 examples per cell recover 92.64% Top-64 overlap (r=0.9883r=0.9883), and this one-time ranking transfers zero-shot to unseen characters while saving 35.9% training time per subsequent epoch:

TokensStandard Fwd (ms / GB)Vectorized CoDePass (s / GB)1 / 4 / 12 / All Layers (s)32-Item Cell Time
25620.1 / 16.480.68 / 21.350.04 / 0.14 / 0.38 / 0.6821.8 s
51222.6 / 16.481.38 / 26.310.08 / 0.28 / 0.76 / 1.3844.2 s
102437.1 / 16.522.45 / 36.270.15 / 0.56 / 1.48 / 2.4578.4 s

Table C: H100 (80GB) timing and memory profiling across sequence lengths and attributed layers.

W3. Exploring potential reasons why targeted fine-tuning might weaken a character's unique persona in certain cases.

The occasional persona smoothing in Appendix E.6 stems from two mechanisms documented in recent role-playing literature—multi-objective style-content tension under concise references and safety alignment conflicts on antagonistic characters—and affects Targeted SFT significantly less than Full SFT. We will incorporate these discussions into the revised manuscript.

First, multi-objective tension under concise reference supervision: MOA [2] points out that factual role-knowledge recovery and surface rhetorical style are often weakly correlated or competing objectives. Because RoleBench references are much shorter than unadapted Qwen outputs, supervised fine-tuning substantially boosts content precision (39.49 for Targeted SFT, Table 6) while occasionally trimming rhetorical catchphrases (such as Keating's "Carpe Diem" in Figure 38).

Second, safety alignment conflicts on antagonistic/negative personas: recent studies (Too Good to be Bad [3] and RoleCDE [4]) indicate that safety-aligned LLMs systematically rewrite deceitful, manipulative, or arrogant villain motives (such as Tom Ripley or Coriolanus in Figures 39–41) into socially acceptable explanations—which explains why Full SFT and Random SFT commit the exact same errors in those figures.

Finally, quantitative evaluation on the held-out set using a reference-free independent LLM judge (deepseek-flash, given only the character profile and question without the reference answer across Knowledge, Belief, and Voice on a 1–5 scale) confirms that Targeted SFT preserves character voice and antagonistic beliefs better than both Full SFT and Random SFT (Table D):

MethodVoice (1–5) ↑Antagonistic Belief Score (1–5) ↑
Full SFT3.12 ± 0.033.32 ± 0.04
Random SFT3.15 ± 0.043.35 ± 0.06
Targeted SFT3.29 ± 0.033.62 ± 0.03

Table D: Reference-free evaluation of persona voice and antagonistic belief retention (Tom Ripley & Coriolanus).

Refs. [1] Zhang: Preference Heads (ACL'26); [2] Liao: MOA (ACL'26); [3] Yi: Too Good to be Bad (ACL'26); [4] Lai: RoleCDE (ACL'26).

AuthorsOpenReview ↗

Follow-up on the response

Dear Reviewer VRNr,

We hope this message finds you well. We are writing to briefly follow up on the analyses and experiments added in response to your review.

For cross-architecture generality and unified parameter budgets (W1), we have added progressive knockout experiments on Llama-3.1-8B-Instruct under the same protocol as Qwen, together with approximately 1.5% and 5.0% parameter-budget comparisons across all four Qwen and Llama variants using matched random controls. For attribution cost (W2), we report H100 timing and peak-memory measurements relative to a standard forward pass, including scaling across sequence lengths and attributed layers; the vectorized implementation takes 21.8–78.4 seconds per 32-example discovery cell. For persona smoothing (W3), we analyze the effects of concise-reference supervision and safety alignment, and add reference-free evaluations indicating that Targeted SFT better preserves character voice and antagonistic beliefs than Full SFT and Random SFT.

These additions are also consistent with strengths identified elsewhere in the reviews. Reviewer ueGy (#1) writes that “The targeted fine-tuning experiments provide practical evidence that the identified components are useful for model adaptation.” Reviewer vTsp (#3) recognizes the matched layerwise head counts, shared GQA groups, and parameter counts, noting that “the advantage persists across multiple training seeds and random selections.” The new Llama interventions, unified-budget comparisons, attribution profiling, and persona evaluations directly extend these findings to the generality, computational cost, and character-preservation questions raised in your review.

If you have any further thoughts—especially on the cross-architecture evidence, attribution profiling, or persona-preservation analysis—we would greatly appreciate your guidance and will respond promptly during the remaining discussion period.

Thank you again for your careful review.

With sincere appreciation,
All Authors

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

This paper investigates the internal mechanisms underlying role-playing behavior in large language models. The authors propose Contrastive Decomposed Forward Pass for Complete Sequences (CoDePass), an attribution method that compares component contributions under matched role-conditioned and counterfactual prompts across complete reference responses.

The paper addresses three research questions: identifying components important for role-playing, characterizing their functional patterns, and determining whether these components can guide efficient fine-tuning.

The analysis suggests that role-playing relies on a sparse subset of attention heads for context retrieval, identity selection, and positional information, while MLPs contribute to distributed transformations of role-related representations.

Based on these findings, the authors introduce component-guided fine-tuning that selectively updates important attention projections. Experiments on Qwen3-8B and Llama-3.1-8B variants demonstrate improvements over capacity-matched random fine-tuning and, in several settings, full-parameter fine-tuning.

Critical strengths

S1. Systematic investigation of role-playing mechanisms.

The paper addresses an interesting and relatively underexplored research question: how role-conditioned behavior is implemented within the internal computations of LLMs.

The proposed CoDePass framework provides a structured approach to examining attention heads and MLPs using matched counterfactual prompts. The analysis covers different characters, prompt conditions, model architectures, and training stages. The combination of component attribution, intervention experiments, and behavioral diagnostics provides useful insights into the possible division of labor between attention routing and MLP transformations.

S2. Encouraging results from component-guided fine-tuning.

The targeted fine-tuning experiments provide practical evidence that the identified components are useful for model adaptation.

On Qwen3-8B-Instruct, Targeted SFT updates only 1.52% of model parameters and improves role-specific ROUGE-L from 21.31 to 23.39 compared with capacity-matched Random SFT, while outperforming Full SFT at 22.42. The method also improves general-scene ROUGE-L from 34.07 to 39.08 relative to Random SFT.

The authors report results across three training seeds, multiple random component selections, and additional model architectures. These experiments strengthen the evidence that the component rankings can guide parameter-efficient adaptation.

Critical weaknesses

W1. The evaluation metrics do not sufficiently establish improvements in role-playing fidelity.

The main experimental evaluation relies heavily on ROUGE-L and SacreBLEU against reference responses. While these metrics measure lexical overlap, they are insufficient to establish improvements in character identity, personality consistency, conversational naturalness, and behavioral fidelity.

This concern is particularly relevant because Appendix Table 26 shows that Targeted SFT substantially reduces average response length from 100.1 words in the original model to 28.9 words. Such a large change in generation behavior may contribute considerably to the observed improvements in reference-based metrics.

More importantly, the qualitative error analysis in Appendix E.6 reveals cases where Targeted SFT loses distinctive character voices, changes character beliefs, or produces incorrect character motivations and relationships. For example, the model occasionally replaces character-specific ideological positions with generic responses.

These findings suggest that higher lexical overlap does not necessarily correspond to better role-playing fidelity. Additional evaluations using character-consistency benchmarks, human preference judgments, or independent semantic evaluators would substantially strengthen the paper.

W2. The causal evidence is insufficient to establish that the identified components are specifically responsible for role-playing.

The paper interprets CoDePass-ranked attention heads and MLPs as components supporting role-conditioned generation. However, the current intervention experiments do not adequately distinguish role-specific mechanisms from general components important for language generation.

In Section 4.2, progressive knockout experiments show that removing highly ranked components substantially degrades ROUGE-L compared with random removal. Yet, a decrease in generation quality does not necessarily demonstrate that the removed components are specifically responsible for role-playing rather than general linguistic processing or response generation.

Appendix Table 16 further shows that large interventions frequently cause generation failures or maximum-token-limit hits. For instance, removing 300 no-role-ranked attention heads results in 188 limit hits out of 200 examples, compared with 15 for random head removal.

This raises the possibility that the observed degradation partly reflects general generation disruption rather than selective impairment of role-playing capabilities.

A more convincing causal analysis would examine whether ablating the identified components disproportionately affects role-conditioned behavior compared with matched non-role tasks, ideally using controlled interventions that preserve general generation quality.

Questions for authors

Q1. Role-playing evaluation. Could the authors provide additional evaluations using character-fidelity metrics, human judgments, or established role-playing benchmarks beyond ROUGE-L and SacreBLEU? This is particularly important because the qualitative analysis identifies cases where Targeted SFT changes character beliefs or loses distinctive personality traits despite improved lexical-overlap scores.

Q2. Response-length confounding. Appendix Table 26 reports that Targeted SFT reduces average response length from 100.1 to 28.9 words. Could the authors provide a length-controlled evaluation to determine how much of the ROUGE-L improvement is attributable to improved content selection rather than shorter, more reference-aligned responses?

Q3. Role-specific causal validation. Could the authors evaluate the effect of ablating CoDePass-ranked components on general language generation or non-role tasks? This would help determine whether the identified attention heads specifically implement role-playing behavior or are broadly important for language generation.

Q4. Knockout interpretation. Appendix Table 16 shows substantial increases in generation-limit hits after removing highly ranked components. Could the authors report role-playing quality conditional on successful generation and provide more localized interventions? This would clarify whether the observed performance degradation reflects role-specific impairment or general generation failures.

Q5. Alternative component-selection baselines. Could the authors compare CoDePass-guided fine-tuning with other importance-based selection methods, such as gradient magnitude, Fisher information, or activation-based component ranking? The current capacity-matched random baseline demonstrates that component selection matters, but it does not fully establish the advantage of CoDePass over simpler importance-estimation approaches.

Q6. Generalization and transfer. Could the authors clarify how well the identified components transfer to broader role-playing settings, particularly multi-turn conversations, longer character profiles, and characters outside the RoleBench distribution? These settings would help establish whether the reported mechanisms generalize beyond the evaluated question-answering format.

4 comments
AuthorsOpenReview ↗

Response to ueGy (2/3): W2; Q3-Q4

W2. The causal evidence is insufficient to establish that the identified components are specifically responsible for role-playing.

W2-1 & Q3: Distinguishing role-specific mechanisms from general language generation.

CoDePass theoretically cancels out generic language generation contributions via matched counterfactual differences, and empirically impairs role-playing capabilities selectively while preserving non-role general capabilities even better than random head knockout.

First, following best practices in mechanistic interpretability [6], CoDePass contrasts matched counterfactual prompts (x+x^+ vs. x−x^-) over the exact same target sequence yy, naturally canceling out generic syntactic and lexical generation contributions in the difference term ∣sc(x+,y)−sc(x−,y)∣|s_c(x^+,y)-s_c(x^-,y)|.

Second, attention routing exhibits cooperative redundancy across the first 20–60 heads and enters a steep degradation regime across 80–160 heads—within this active regime, ablating CoDePass-ranked heads selectively degrades role-conditioned generation while leaving non-role general capabilities intact. Table C compares retention on RoleBench (200 items) against non-role tasks (Non-Role QA, MMLU, and ARC-C) when removing 80, 120, and 160 heads. Because CoDePass precisely targets role-routing heads while avoiding generic reasoning heads, its non-role retention is even slightly higher than random head removal, whereas role-playing performance drops steeply:

Knockout SetRole R-L Ret. (%)Role Judge Ret. (%)Non-Role QA Ret. (%)MMLU Ret. (%)ARC Ret. (%)
Unmodified (0)100.00100.00100.00100.00100.00
Random Heads (120)99.9598.4099.2098.8598.47
CoDePass Heads (80)84.2085.6099.6899.6099.55
CoDePass Heads (120)61.5671.2099.4599.3299.38
CoDePass Heads (160)54.1258.3099.2499.1299.18

Table C: RoleBench vs. non-role benchmark retention under attention-head knockout.

Therefore, both theoretical formulation and new empirical analyses indicate that CoDePass effectively distinguishes role-specific mechanisms from general components essential for language generation.

W2-2 & Q4: Knockout interpretation and generation-limit hits.

Within the active intervention regime, generation-limit hits are near zero while role-playing quality already drops sharply; moreover, even when strictly restricted to normally completed generations, role-playing capability remains selectively impaired while general continuation perplexity stays stable.

First, the 188/200 limit hits in Appendix Table 16 occur only at the extreme endpoint of removing 300 attention heads. Within the active degradation regime of Figure 2 (removing 80–160 heads), limit hits are near zero, yet role-playing performance already drops by nearly half.

Second, incorporating the soft-intervention perspective from multi-component causal tracing [7], Table D reports performance strictly on the subset of normally completed generations (No-Limit-Hit) as well as under soft scaling (α=0.3\alpha=0.3), indicating selective persona degradation while neutral continuation perplexity (HellaSwag PPL) remains stable:

InterventionLimit Hits / 200Completed R-L Ret. (%)LLM JudgeHellaSwag PPL ↓
Unmodified0100.003.8718.40
CoDePass Top-80 (Zero)084.203.3118.41
CoDePass Top-120 (Zero)161.862.7618.44
CoDePass Top-160 (Zero)254.652.2618.48
CoDePass Top-160 (Soft, α=0.3)074.302.9518.42
Random-160298.603.8018.89

Table D: Comparison on the normally completed generation subset and neutral continuation perplexity.

Therefore, the observed performance degradation is primarily driven by the selective impairment of role-playing capability rather than general generation disruption.

Refs. [1] Cai: ThinkPersona (ACL'26); [2] Lai: RoleCDE (ACL'26); [3] Liao: MOA (ACL'26); [4] Wang: CoSER (ICML'25); [5] Yi: Too Good to be Bad (ACL'26); [6] Zhang & Nanda: Activation Patching (ICLR'24); [7] Yan: Multi-component Tracing (ACL'26).

AuthorsOpenReview ↗

Response to ueGy (3/3):Q5-Q6

Q5. Alternative component-selection baselines.

CoDePass is specifically tailored for tasks where target behavioral conditioning is tightly entangled with general language generation. Building on rigorous structural controls, CoDePass consistently outperforms activation-norm, gradient/Taylor, and Fisher-information rankings under the exact same parameter budget.

First, unlike generic pruning or importance metrics that measure unconditional token traffic on a single prompt, CoDePass is purpose-built to disentangle role-specific computation from general linguistic processing by contrasting matched counterfactual prompts over complete sequences.

Second, as Reviewer vTsp recognized, our capacity-matched Random SFT already enforces rigorous structural controls—strictly matching layerwise head counts, shared GQA key/value groups, and parameter budgets across disjoint training and evaluation characters.

Finally, under the identical parameter budget on Qwen3-8B-Instruct, CoDePass consistently surpasses traditional importance selectors across both role-playing and general benchmarks. Generic importance metrics prioritize high-traffic language generation components regardless of whether they carry role-conditioning signals [6], whereas CoDePass precisely isolates the counterfactual role differential (Table E):

Selection / Reference MethodRole-Specific R-LGeneral R-LLLM Judge ↑MMLUARCHellaSwag
Qwen3-8B-Instruct18.0911.843.2571.9463.3155.87
Full SFT22.4235.913.2673.8565.9073.85
Random SFT21.3134.073.2473.3264.6173.46
Activation Norm22.1835.163.3072.0558.5072.03
Gradient / Taylor22.5735.823.2873.5062.0073.50
Fisher Information23.0335.913.3172.1660.5073.18
Targeted SFT (CoDePass)23.3939.083.5074.2165.9375.16

Table E: Unified evaluation of sparse selectors under the identical parameter budget.

Q6. Generalization and transfer to multi-turn conversations, longer profiles, and out-of-distribution characters.

The localized components not only generalize across unseen characters, but also transfer zero-shot to multi-turn dialogues, long character profiles, and external OOD benchmarks without re-running attribution.

First, manuscript Table 5 already indicates significant transfer gains from the 75 attribution/training characters to 6 disjoint test characters, which Reviewer vTsp explicitly noted provides clear evidence that the adaptation of identified parameters transfers across characters.

Second, without re-running attribution or fine-tuning, we directly transfer our RoleBench-trained models to long character profiles (800–1,500 words), multi-turn interaction histories (3 turns), and external OOD continuations on CoSER [4]. Table F confirms that both the routing selectivity of functional heads and the performance advantage of Targeted SFT persist stably across all complex settings:

Setting (60 Questions)Head Routing RetentionFull (R-L / Judge)Random (R-L / Judge)Targeted (R-L / Judge)
Short / Single Turn97.38%25.35 / 3.2723.43 / 3.2226.45 / 3.51
Three Turns91.40%24.82 / 3.2424.35 / 3.2126.46 / 3.48
Long Profile (>1,000 words)93.20%24.64 / 3.2123.33 / 3.1826.25 / 3.46
Long Profile / Three Turns89.50%25.04 / 3.2224.81 / 3.1925.83 / 3.44
CoSER Fixed History (OOD)94.52%8.56 / 2.918.43 / 2.749.36 / 3.28

Table F: Zero-shot transfer to long profiles, multi-turn dialogues, and the external CoSER benchmark.

Refs. [1] Cai: ThinkPersona (ACL'26); [2] Lai: RoleCDE (ACL'26); [3] Liao: MOA (ACL'26); [4] Wang: CoSER (ICML'25); [5] Yi: Too Good to be Bad (ACL'26); [6] Zhang & Nanda: Activation Patching (ICLR'24); [7] Yan: Multi-component Tracing (ACL'26).

AuthorsOpenReview ↗

Follow-up on the response

Dear Reviewer ueGy,

We hope this message finds you well. We are writing to briefly follow up on the analyses and experiments added in response to your review.

For role-playing fidelity and response-length confounding (W1/Q1–Q2), we have added reference-free independent judge evaluations of character knowledge, beliefs, and voice, together with a blinded human preference study on 300 items and both length-matched and length-adjusted comparisons. For role-specific causality and knockout interpretation (W2/Q3–Q4), we compare role-conditioned and non-role performance under matched interventions, report results conditional on normally completed generations, and include soft interventions with neutral continuation perplexity controls. We further compare activation-norm, gradient/Taylor, and Fisher-information selectors under the same parameter budget (Q5), and evaluate transfer to longer profiles, multi-turn conversations, and CoSER without additional attribution or fine-tuning (Q6).

These additions are also consistent with strengths identified elsewhere in the reviews. Reviewer VRNr (#2) highlights “a clear division of labor” between specialized attention heads and MLP transformations. Reviewer vTsp (#3) writes that “The selective fine-tuning results have meaningful controls,” and notes that these comparisons provide useful evidence that the ranking identifies parameters whose adaptation transfers across characters. The new reference-free evaluations, length controls, and localized interventions directly examine the character-fidelity and role-specific causal questions raised in your review.

If you have any further thoughts, we would greatly appreciate your guidance and will respond promptly during the remaining discussion period.

Thank you again for your careful review.

With sincere appreciation,
All Authors

AuthorsOpenReview ↗

Second Follow-up

Dear Reviewer ueGy,

We hope you are doing well. As the author–reviewer discussion deadline is imminent, we are writing a brief follow-up concerning several statements in the review.

The review characterizes the evaluation as relying heavily on lexical overlap, raises response length as a potential confounder, links large-knockout degradation to possible general generation disruption, and questions whether the identified components support role-specific functions and broader transfer.

Our rebuttal clarifies that these concerns either relate to extreme knockout settings or are directly examined by additional controlled evaluations: it provides reference-free judging, blinded human evaluation, length-controlled comparisons, non-role task controls, normally completed-generation results, and broader transfer evaluations. It also includes isolated attention/MLP writes and targeted functional interventions in our response to vTsp (#3), together with Llama knockouts and unified-budget comparisons in our response to VRNr (#2). These additions directly test the objections rather than merely restating our claims.

Your Strengths section recognizes “Systematic investigation of role-playing mechanisms” and “Encouraging results from component-guided fine-tuning,” together with the multi-seed and cross-architecture controls. Our added material focuses directly on the accompanying concerns about character fidelity, role-specific mechanisms, and generalization.

We would greatly appreciate knowing whether any specific questions remain after considering the rebuttal. Even brief feedback would help us use the remaining discussion period to provide any further clarification and ensure that the added analyses are considered before the deadline.

Thank you for your time and consideration.

With sincere appreciation,
All Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.