acceptodds
Under review as a conference paper at ICLR 2027

Mind Your Own Role: Reducing Role Overstepping for Effective and Efficient Multi-Agent Collaboration via Dynamic Role Activation Steering

Abstract

Multi-Agent Systems (MAS) have shown strong capability in solving complex tasks by assigning different agents specialized roles and corresponding subtasks. However, agents may overstep their assigned responsibilities and perform subtasks belonging to other roles, which can disrupt workflow, reduce system efficiency, and degrade task performance. To address this issue, we investigate role-related patterns in the model's internal activation space and uncover two key observations: (i) role instructions induce consistent directional shifts in activation space, and (ii) role-adherent and role-overstepping trajectories exhibit distinct activation patterns. Based on these observations, we propose Dynamic Role Activation Steering (DRAS), a lightweight and effective inference-time activation intervention approach to reduce role overstepping in multi-agent collaboration. Specifically, DRAS constructs a role steering vector and a lightweight role-overstepping detector for each role. During LLM generation, the detector estimates the probability of role overstepping from the current activation state, and DRAS dynamically adjusts the steering strength accordingly to guide the agent toward its assigned responsibilities. Experimental results show that DRAS effectively reduces the number of role-overstepping trajectories by up to 86.6% compared with Vanilla MAS, thereby improving task quality score by 10.0%–11.1%, achieving a 1.50–1.61 task-completion speedup, and reducing token consumption by 26.8%–38.7%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Multi-Agent collaboration, Activation Engineering, Role Overstepping

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_v6SEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Dynamic Role Activation Steering (DRAS) to reduce cases where an agent performs another role's tasks. It extracts a role vector by comparing activations with and without role instructions, and trains a linear detector on compliant and overstepping trajectories. During generation, the detector controls how strongly the role vector is applied. Experiments with ChatDev and MetaGPT using two backbone models show fewer role violations, higher software Quality scores, and lower runtime and token use.

Critical strengths
  • DRAS modifies activations at one model layer without updating the backbone weights or adding a response-regeneration step. Table 3 reports substantially lower offline memory and computation costs than fine-tuning a separate LoRA model for each role.

  • Quality improves and both runtime and token use fall relative to Vanilla MAS in all four system–model combinations. The authors use separate datasets for offline preparation and final evaluation, and check a sample of the LLM judge's decisions against human annotations.

Critical weaknesses
  • DRAS requires clearly defined role responsibilities and does not cover collaboration without a clear division of work. All experiments and baseline comparisons are also limited to software development. The method has not been evaluated in other domains.

  • In Table 4, static steering produces fewer role violations, runs faster, and uses fewer tokens, while DRAS achieves higher Quality scores. Both approaches have advantages, but the paper does not explain how the fixed strength was chosen or how changing it affects quality and efficiency. Without this comparison, it is hard to tell when to prefer one approach over the other.

  • The detector is trained to judge whether a completed trajectory contains overstepping, but it is used to assess risk while generation is still in progress. The paper does not explain why this training should work for real-time detection. More evaluation is needed beyond the single example in Figure 5.

Questions for authors
  • How was the static steering strength selected and tuned in Table 4? How do Quality, runtime, and token use change with different fixed strengths?

  • What is the source of the 2.5B Qwen2.5 model shown in Figure 1b (page 2) and mentioned in the accompanying text (lines 69–70)? The official Qwen2.5 model cards do not list this variant.

6 comments
AuthorsOpenReview ↗

Author's Response (1)

Thank you for providing your valuable and constructive feedback! Please find below the responses to each comment on weaknesses and questions.

[W1] DRAS requires clearly defined role responsibilities and does not cover collaboration without a clear division of work. All experiments and baseline comparisons are also limited to software development. The method has not been evaluated in other domains.

Response: Thank you for this comment. This work focuses on role-based multi-agent collaboration. We evaluate DRAS on software development tasks because software development is a widely used and well-established application setting for such multi-agent systems [1-3]. However, we would like to clarify that the design of DRAS is based on role definitions and internal model activations and is not specific to software development. Therefore, the same procedure can be applied to other collaborative tasks with clearly defined roles for generation-time role control. We will add this discussion to the revised manuscript.

[1] Yue Hu et al. Self-Evolving Multi-Agent Collaboration Networks for Software Development. ICLR, 2025.

[2] Zhengren Wang et al. MaintainCoder: Maintainable Code Generation Under Dynamic Requirements. NeurIPS, 2025.

[3] Iris Xu et al. BOAD: Discovering Hierarchical Software Engineering Agents via Bandit Optimization. ICLR, 2026.

[W2 & Q1] How was the static steering strength selected and tuned in Table 4? How do Quality, runtime, and token use change with different fixed strengths?

Response: Thank you for the question. (i) To isolate the effects of different steering strategies, Static Steering and Binary Gating use the same base steering strengths as DRAS. (ii) For Static Steering, weak steering may be insufficient to correct overstepping, while sufficiently strong steering may disrupt role-adherent generation and produce garbled outputs, leading to premature termination and task failure. Therefore, Static Steering’s lower token consumption in the reported results is partly due to the premature termination of some failed tasks, rather than more efficient completion of successful tasks.

AuthorsOpenReview ↗

Author's Response (2)

[W3] The detector is trained to judge whether a completed trajectory contains overstepping, but it is used to assess risk while generation is still in progress. The paper does not explain why this training should work for real-time detection. More evaluation is needed beyond the single example in Figure 5.

Response: Thank you for the question. The approach of training probes on activations from complete responses and using them for real-time detection and dynamic control of steering strength during generation has also been adopted in ACT [1] and AAC [2]. Other studies on internal representations [3,4] also identified predictive signals in internal model states before the corresponding errors or violations appear in the output. These studies provide some support for the feasibility of this approach.

To further evaluate real-time role-overstepping detection beyond the single example in Figure 5, we kept the detector unchanged and conducted an additional prefix evaluation for the CEO role in ChatDev using Qwen2.5-7B. We selected 100 trajectories with role overstepping and 100 trajectories without role overstepping, all generated without steering. For trajectories with overstepping, we annotated the onset of the first role-overstepping behavior and recorded the probe’s role-overstepping predicted probability (p) at 20, 10, 5, and 1 generation steps before this onset. For trajectories without overstepping, we used the final generated token as the reference position and recorded the probabilities at the same offsets. The table reports the mean and standard deviation of predicted probability at each position for each group.

20 steps before10 steps before5 steps before1 step before
With overstepping: onset of the first violation0.36 ± 0.1050.54 ± 0.0920.79 ± 0.0950.83± 0.075
Without overstepping: final generated token0.29± 0.1000.35± 0.1240.42 ± 0.1010.31± 0.103

The results show that, for trajectories that subsequently overstep, the probe’s predicted probability rises noticeably before the first violation occurs. DRAS uses precisely this signal to increase steering strength before overstepping occurs, thereby reinforcing role adherence. For trajectories without overstepping, the predicted probability remains low at the evaluated positions. These results support the detector's ability to detect early warning signals of role overstepping before it occurs.

[1] Tianlong Wang et al. Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categories. WWW, 2025.

[2] Eric Yocam et al. Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models. 2026.

[3] Ben Snyder, Marius Moisescu, and Muhammad Bilal Zafar. On Early Detection of Hallucinations in Factual Question Answering. KDD, 2024.

[4] Dhananjay Ashok and Jonathan May. Language Models Can Predict Their Own Behavior. NeurIPS, 2025.

[Q2] What is the source of the 2.5B Qwen2.5 model shown in Figure 1b (page 2) and mentioned in the accompanying text (lines 69–70)? The official Qwen2.5 model cards do not list this variant.

Response: Thank you for pointing this out. We checked our experimental code and confirmed that the model used was Qwen2.5-3B. The "2.5B" label in Figure 1b was a labeling error introduced when preparing the plot, where "2.5" in the model family name was mistakenly used as the model size. The correct model sizes are 14B, 7B, and 3B, and we will correct this error in both Figure 1b and the corresponding text.

Thank you for the responses and experiments. However, my concern about applicability remains, as the effectiveness of DRAS beyond software development has not been evaluated.

Besides, for the comparison with static steering, does DRAS still offer a better trade-off between task quality and efficiency when the static steering strength is tuned independently?

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

This paper studies role overstepping in LLM-based multi-agent collaboration and proposes Dynamic Role Activation Steering (DRAS). The method constructs role steering vectors from activation differences between inputs with and without role instructions, trains linear probes on role-adherent and role-overstepping trajectories, and adjusts the steering strength during generation according to the probe output. The end-to-end evaluation covers four combinations of two multi-agent frameworks and two backbone LLMs. All four configurations show improvements in task quality and runtime efficiency over Vanilla MAS.

Critical strengths
  1. A practical problem addressed with a simple method. The paper addresses the difficulty of maintaining responsibility boundaries in role-based multi-agent systems through lightweight activation interventions during generation. DRAS does not update the backbone LLM parameters, and its online intervention does not require an additional LLM for post-hoc repair. This offers a practical approach to controlling role behavior during collaboration.

  2. Consistent end-to-end results with useful validation. Table 2 reports improvements in task quality and efficiency over Vanilla MAS across all four configurations. The authors use SWE-Dev for offline preparation and SRDD for final evaluation, and manually validate the GPT-4o role-overstepping judgments for the three selected target roles in ChatDev–Qwen. The paper also includes ablations of the dynamic steering mechanism and analyses of the intervention layer and base steering strength. These results and design choices provide support for the method’s practical value.

Critical weaknesses
  1. Early detection of role overstepping remains insufficiently validated. Section 4.2 trains the probe on last-token residual-stream activations from collected role trajectories, whereas Section 4.3 applies the probe to the activations at the last token of the current sequence during generation. The paper does not clearly report how this difference between training and online inputs affects detection performance. Figure 4b does not separately report detection performance on generation prefixes, and Figure 5 presents only one case study. The available evidence therefore does not establish that the detector consistently provides useful signals before role overstepping occurs. Moreover, the probe dataset contains equal numbers of role-adherent and role-overstepping trajectories, while the probe output directly scales the steering strength. Its probability calibration under the actual generation distribution also warrants examination.

  2. The technical contribution beyond existing adaptive steering methods needs further validation. The Static Steering and Binary Gating ablations in Table 4 provide some support for dynamic control. However, these experiments do not sufficiently disentangle the contribution of the role-specific design from that of existing probe-driven adaptive steering mechanisms. Clarifying the differences from related methods such as ACT (Wang et al., 2025), together with a comparison adapted to role overstepping or component-level comparisons, would help establish the technical contribution.

Questions for authors
  1. Probe training and evaluation on generation prefixes. Has the probe been trained or evaluated on generation prefixes? Section 4.2 describes last-token activations from collected role trajectories. What inputs are used to compute the accuracy in Figure 4b? Beyond the single case in Figure 5, are there prefix-level results on independent tasks that quantify detection before the first instance of role overstepping, as well as false-positive rates on trajectories that ultimately remain role-adherent and on their prefixes? Please also clarify the prefix-level labels and evaluation criteria.

  2. Hyperparameter selection for Static Steering and Binary Gating. Do Static Steering and Binary Gating both use the base steering strengths in Table 7, or are their strengths selected separately on the validation set? How was the 0.5 threshold for Binary Gating determined? Please clarify whether the search ranges, selection criteria, and tuning budgets are consistent across the strategies, so that the influence of parameter selection on the quality–efficiency trade-off in Table 4 can be assessed.

  3. Stability of the Quality gains in Table 2. DRAS achieves absolute Quality gains of 0.0071–0.0158 over the highest-scoring comparison method in each configuration. However, the paper does not clearly report the number of independent runs, making it difficult to assess the stability of these gains.

  4. Sources and partitioning of the probe training trajectories. Do the probes for Qwen and Llama use trajectories generated by their respective backbone LLMs, or do they share the same trajectory texts? In the 8:2 split into an extraction set and a validation set, are trajectories associated with the same source task kept within the same set?

5.The experiments focus on software-development tasks and provide support for the method’s effectiveness in that setting. These results do not establish whether the method generalizes to other task domain.

6 comments
AuthorsOpenReview ↗

Author's Response (1)

Thank you for providing your valuable and constructive feedback! Please find below the responses to each comment on weaknesses and questions.

[W1&Q1] Early detection of role overstepping remains insufficiently validated. Probe training and evaluation on generation prefixes.

Response: Thank you for the question. The approach of training probes on activations from complete responses and using them for dynamic control of steering strength during generation has also been adopted in ACT [1] and AAC [2]. Other studies on internal representations [3,4] also identified predictive signals in internal model states before the corresponding errors or violations appear in the output. These studies provide some support for the feasibility of this approach.

To further examine this in the agent role control, we kept the detector unchanged and conducted an additional prefix evaluation for the CEO role in ChatDev using Qwen2.5-7B. We selected 100 trajectories with role overstepping and 100 trajectories without role overstepping, all generated without steering. For trajectories with overstepping, we annotated the onset of the first role-overstepping behavior and recorded the probe’s role-overstepping predicted probability (p) at 20, 10, 5, and 1 generation steps before this onset. For trajectories without overstepping, we used the final generated token as the reference position and recorded the probabilities at the same offsets. The table reports the mean and standard deviation of predicted probability at each position for each group.

20 steps before10 steps before5 steps before1 step before
With overstepping: onset of the first violation0.36 ± 0.1050.54 ± 0.0920.79 ± 0.0950.83± 0.075
Without overstepping: final generated token0.29± 0.1000.35± 0.1240.42 ± 0.1010.31± 0.103

The results show that, for trajectories that subsequently overstep, the probe’s predicted probability rises noticeably before the first violation occurs. DRAS uses precisely this signal to increase steering strength before overstepping occurs, thereby reinforcing role adherence. For trajectories without overstepping, the predicted probability remains low at the evaluated positions. These results support the detector's ability to detect early warning signals of role overstepping before it occurs.

[1] Tianlong Wang et al. Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categories. WWW, 2025.

[2] Eric Yocam et al. Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models. 2026.

[3] Ben Snyder, Marius Moisescu, and Muhammad Bilal Zafar. On Early Detection of Hallucinations in Factual Question Answering. KDD, 2024.

[4] Dhananjay Ashok and Jonathan May. Language Models Can Predict Their Own Behavior. NeurIPS, 2025.

[W2] The technical contribution beyond existing adaptive steering methods needs further validation.

Response: Thank you for this suggestion. Activation engineering methods, such as those used in ACT, have become fundamental tools for controlling model behavior. In recent years, more studies have focused on applying these established techniques to problems in different application scenarios. For example, Zhang et al. [1] extract user-specific writing style directions for personalized text generation; Sui et al. [2] use activation steering to mitigate overthinking and overacting in coding agents; Panfilov et al. [3] use linear probes and activation steering to study strategic dishonesty in LLMs; and Hua et al. [4] use activation steering to suppress evaluation-awareness in LLMs. These studies also build on established activation steering principles rather than introducing fundamentally new activation steering techniques.

Similarly, our main contribution is to identify role-related activation patterns in multi-agent collaboration and use them to develop a lightweight approach to role control during generation. Our work focuses on addressing role overstepping rather than introducing a new general-purpose activation steering mechanism. As described in the paper, DRAS builds on established activation steering and linear probing techniques, similar to ACT.

[1] Jinghao Zhang et al. Personalized Text Generation with Contrastive Activation Steering. ACL, 2025.

[2] Yuan Sui et al. TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering. NeurIPS, 2026.

[3] Alexander Panfilov et al. Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs. ICLR, 2026.

[4] Tim Hua et al. Steering Evaluation-Aware Language Models To Act Like They Are Deployed. ICLR, 2026.

AuthorsOpenReview ↗

Author's Response (2)

[Q2] Hyperparameter selection for Static Steering and Binary Gating.

Response: Thank you for the question. (i) To isolate the effects of different steering strategies, Static Steering and Binary Gating use the same base steering strengths as DRAS. (ii) Probe outputs typically fall within 0.2–0.4 for role-adherent trajectories and 0.6–0.9 for role-overstepping trajectories. We therefore chose 0.5, between these ranges, as the threshold for activating steering. (iii) Table 4 shows that DRAS outperforms Binary Gating in role-overstepping mitigation, task quality, and efficiency. For Static Steering, weak steering may be insufficient to correct overstepping, while sufficiently strong steering may disrupt role-adherent generation and produce garbled outputs, leading to premature termination and task failure. Therefore, Static Steering’s lower token consumption in the reported results is partly due to the premature termination of some failed tasks, rather than more efficient completion of successful tasks.

[Q3] Stability of the Quality gains in Table 2.

Response: Thank you for the question. we conducted additional evaluations of DRAS and the highest-scoring comparison method LoRA FT under the ChatDev–Qwen setting in Table 2 using three different random seeds. The table below reports the mean and standard deviation of task quality and token consumption.

Task Quality ScoreTokens
LoRA FT0.8304 ± 0.003226.97K ± 1.68K
DRAS(Ours)0.8381 ± 0.001724.23K ± 0.50K

The results show that DRAS maintains its advantages in task quality and token efficiency across multiple runs. In addition, DRAS incurs lower offline preparation costs than LoRA FT (Table 3 in the main paper).

[Q4] Sources and partitioning of the probe training trajectories.

Response: Thank you for the question. (i) For Qwen and Llama, we separately generated role trajectories using their respective backbone models, extracted activations, and trained role-overstepping detectors independently. (ii) For the 8:2 split, we partitioned the data by source task, ensuring that all trajectories from the same task were assigned to either the extraction set or the validation set, preventing overlap in task contexts between the two sets.

[Q5] The experiments focus on software-development tasks and provide support for the method’s effectiveness in that setting.

Response: Thank you for this comment. We evaluate DRAS on software development tasks because software development is a widely used and well-established application setting for such multi-agent systems [1-3]. However, we would like to clarify that the design of DRAS is based on role definitions and internal model activations and is not specific to software development. Therefore, the same procedure can be applied to other collaborative tasks with clearly defined roles for generation-time role control. We will add this discussion to the revised manuscript.

[1] Yue Hu et al. Self-Evolving Multi-Agent Collaboration Networks for Software Development. ICLR, 2025.

[2] Zhengren Wang et al. MaintainCoder: Maintainable Code Generation Under Dynamic Requirements. NeurIPS, 2025.

[3] Iris Xu et al. BOAD: Discovering Hierarchical Software Engineering Agents via Bandit Optimization. ICLR, 2026.

Reviewer_sRuUOpenReview ↗

Response

Thank you for your response. My main concerns remain: it is still unclear what additional benefit DRAS’s role-specific design offers over existing adaptive steering methods, and DRAS has not been evaluated beyond software development.

Reviewer_M3M7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dynamic Role Activation Steering (DRAS) to mitigate role overstepping in LLM-based multi-agent systems. The method constructs role-specific steering vectors from contrastive activations and trains lightweight linear probes to detect role-overstepping risks. During inference, DRAS dynamically adjusts the steering strength based on the predicted overstepping probability. Experiments on ChatDev and MetaGPT with Qwen2.5-7B and Llama-3.1-8B demonstrate improvements in role adherence, task quality, and computational efficiency.

Critical strengths

Important and practical problem. The paper addresses role overstepping in multi-agent collaboration, which can negatively affect task quality and execution efficiency. This is a meaningful problem for practical MAS deployment.

Lightweight and effective approach. DRAS dynamically steers model activations without modifying backbone parameters or requiring response regeneration. This design offers advantages in deployment flexibility and computational efficiency.

Promising experimental results. Experiments across two MAS frameworks and two LLM backbones demonstrate consistent improvements in role adherence, task quality, and efficiency. The ablation studies further support the benefits of dynamic steering.

Critical weaknesses

Limited methodological novelty. The core mechanism of combining linear probes with adaptive activation steering closely resembles existing Conditional Activation Steering methods, particularly Adaptive Activation Steering (Wang et al., 2025). The paper lacks direct comparisons with these closely related approaches, making its methodological contribution difficult to assess.

Insufficient evidence for consistent role-specific directions. The paper mainly relies on PCA visualizations to support the claim that role instructions induce consistent activation shifts. However, PCA separation does not necessarily establish directional consistency across tasks or demonstrate that the extracted directions specifically encode role-related information.

Potential mismatch between detector training and inference. The role-overstepping detector is trained using activations from complete trajectories but is applied to partially generated sequences during inference. The paper does not sufficiently evaluate whether the detector can reliably identify overstepping risks before violations occur.

Questions for authors

Comparison with prior methods: How does DRAS differ fundamentally from existing probe-based adaptive steering approaches, such as Adaptive Activation Steering (Wang et al., 2025)? Could the authors provide direct comparisons with ACT and CAST adapted to the role-overstepping setting?

Validation of role-specific directions: Can the authors quantitatively validate the directional consistency of role-induced activation shifts across different tasks, for example, through cosine similarity analysis? Would random, generic instruction, or cross-role steering vectors achieve similar results?

Online detection reliability: Given that the detector is trained on complete trajectories, how well does it generalize to intermediate generation states? Could the authors report token-level detection performance, false-positive rates, and the ability to predict overstepping before it occurs?

Statistical reliability and generalization: Are the improvements over strong baselines statistically significant across multiple runs? How well does DRAS generalize to smaller LLMs or multi-agent tasks beyond software development?

6 comments
AuthorsOpenReview ↗

Author's Response (1)

Thank you for providing your valuable and constructive feedback! Please find below the responses to each comment on weaknesses and questions.

[W1&Q1] Comparison with prior methods: How does DRAS differ fundamentally from existing probe-based adaptive steering approaches, such as Adaptive Activation Steering (Wang et al., 2025)? Could the authors provide direct comparisons with ACT and CAST adapted to the role-overstepping setting?

Response: Activation engineering methods, such as those used in CAST and ACT, have become fundamental tools for controlling model behavior. In recent years, more studies have focused on applying these established techniques to problems in different application scenarios. For example, Zhang et al. [1] extract user-specific writing style directions for personalized text generation; Sui et al. [2] use activation steering to mitigate overthinking and overacting in coding agents; Panfilov et al. [3] use linear probes and activation steering to study strategic dishonesty in LLMs; and Hua et al. [4] use activation steering to suppress evaluation-awareness in LLMs. These studies also build on established activation steering principles rather than introducing fundamentally new activation steering techniques.

Similarly, our main contribution is to identify role-related activation patterns in multi-agent collaboration and use them to develop a lightweight approach to role control during generation. Our work focuses on addressing role overstepping rather than introducing a new general-purpose activation steering mechanism. As described in the paper, DRAS builds on established activation steering and linear probing techniques, similar to ACT.

[1] Jinghao Zhang et al. Personalized Text Generation with Contrastive Activation Steering. ACL, 2025.

[2] Yuan Sui et al. TACT: Mitigating Overthinking and Overacting in Coding Agents via Activation Steering. NeurIPS, 2026.

[3] Alexander Panfilov et al. Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs. ICLR, 2026.

[4] Tim Hua et al. Steering Evaluation-Aware Language Models To Act Like They Are Deployed. ICLR, 2026.

[W2&Q2] Validation of role-specific directions: Can the authors quantitatively validate the directional consistency of role-induced activation shifts across different tasks, for example, through cosine similarity analysis? Would random, generic instruction, or cross-role steering vectors achieve similar results?

Response: Thank you for the suggestion. We conducted additional analyses to examine directional consistency across tasks and the role relevance of the extracted directions. First, the mean pairwise cosine similarity between the activation difference vectors for the tasks used in Figure 2(a) is 0.774 ± 0.159. This provides quantitative evidence of directional consistency across tasks. Second, we replaced the target role vector with a random vector, a vector extracted from a generic instruction, or a vector for a different role. We evaluated the CEO role in ChatDev on the SRDD dataset using Qwen2.5-7B as the backbone model. For a fair comparison, we used static steering with the same intervention layer and steering strength (α = 1.2) across all vector-based methods. The results are reported below.

Steering vectorRole overstepping rate ↓(%)Nonsensical or garbled outputs↓(%)
No steering39.8%--
Random vector---100 %
Generic instruction55.7%23.3%
Other role vector (CPO)61.9%30.7%
Target role vector (CEO)3.9%10.2 %

The results show that steering with the target role vector achieved the largest reduction in role overstepping, while the alternative vectors failed to reduce role overstepping and substantially increased nonsensical or garbled outputs. These results demonstrate the effectiveness and importance of using a steering direction aligned with the agent's assigned role for role control.

AuthorsOpenReview ↗

Author's Response (2)

[W3&Q3] Online detection reliability: Given that the detector is trained on complete trajectories, how well does it generalize to intermediate generation states? Could the authors report token-level detection performance, false-positive rates, and the ability to predict overstepping before it occurs?

Response: Thank you for the question. The approach of training probes on activations from complete responses and using them for dynamic control of steering strength during generation has also been adopted in ACT [1] and AAC [2]. Other studies on internal representations [3,4] also identified predictive signals in internal model states before the corresponding errors or violations appear in the output. These studies provide some support for the feasibility of this approach.

To further examine this in the agent role control, we kept the detector unchanged and conducted an additional prefix evaluation for the CEO role in ChatDev using Qwen2.5-7B. We selected 100 trajectories with role overstepping and 100 trajectories without role overstepping, all generated without steering. For trajectories with overstepping, we annotated the onset of the first role-overstepping behavior and recorded the probe’s role-overstepping predicted probability (p) at 20, 10, 5, and 1 generation steps before this onset. For trajectories without overstepping, we used the final generated token as the reference position and recorded the probabilities at the same offsets. The table reports the mean and standard deviation of predicted probability at each position for each group.

20 steps before10 steps before5 steps before1 step before
With overstepping: onset of the first violation0.36 ± 0.1050.54 ± 0.0920.79 ± 0.0950.83± 0.075
Without overstepping: final generated token0.29± 0.1000.35± 0.1240.42 ± 0.1010.31± 0.103

The results show that, for trajectories that subsequently overstep, the probe’s predicted probability rises noticeably before the first violation occurs. DRAS uses precisely this signal to increase steering strength before overstepping occurs, thereby reinforcing role adherence. For trajectories without overstepping, the predicted probability remains low at the evaluated positions. These results support the detector's ability to detect early warning signals of role overstepping before it occurs.

[1] Tianlong Wang et al. Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categories. WWW, 2025.

[2] Eric Yocam et al. Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models. 2026.

[3] Ben Snyder, Marius Moisescu, and Muhammad Bilal Zafar. On Early Detection of Hallucinations in Factual Question Answering. KDD, 2024.

[4] Dhananjay Ashok and Jonathan May. Language Models Can Predict Their Own Behavior. NeurIPS, 2025.

AuthorsOpenReview ↗

Author's Response (3)

[Q4] Statistical reliability and generalization: Are the improvements over strong baselines statistically significant across multiple runs? How well does DRAS generalize to smaller LLMs or multi-agent tasks beyond software development?

Response: Thank you for the question. (i) Regarding statistical reliability, we conducted additional evaluations of DRAS and the strong baseline LoRA FT under the ChatDev–Qwen setting in Table 2 using three different random seeds. The table below reports the mean and standard deviation of task quality and token consumption.

Task Quality ScoreTokens
LoRA FT0.8304 ± 0.003226.97K ± 1.68K
DRAS(Ours)0.8381 ± 0.001724.23K ± 0.50K

The results show that DRAS maintains its advantages in task quality and token efficiency across multiple runs. In addition, DRAS incurs lower offline preparation costs than LoRA FT (Table 3 in the main paper).

(ii) Regarding smaller models, we additionally compared role-overstepping rates with and without DRAS for Llama models of different sizes, using the CEO role in ChatDev. The table below reports the role-overstepping rates ↓ (%).

Vanilla MASDRAS (Ours)
Llama-3.1-8B-Instruct47.7%6.1%
Llama-3.2-3B-Instruct56.6%5.7%
Llama-3.2-1B-Instruct68.5%19.6%

The results show that DRAS consistently reduces role-overstepping rates for the tested 1B, 3B, and 8B models, supporting its effectiveness in improving role adherence in smaller models.

(iii) Regarding application scenarios, this work focuses on role-based multi-agent collaboration. We evaluate DRAS on software development tasks because software development is a widely used and well-established application setting for such multi-agent systems [1–3]. However, we would like to clarify that the design of DRAS is based on role definitions and internal model activations and is not specific to software development. Therefore, the same procedure can be applied to other collaborative tasks with clearly defined roles for generation-time role control. We will add this discussion to the revised paper.

[1] Yue Hu et al. Self-Evolving Multi-Agent Collaboration Networks for Software Development. ICLR, 2025.

[2] Zhengren Wang et al. MaintainCoder: Maintainable Code Generation Under Dynamic Requirements. NeurIPS, 2025.

[3] Iris Xu et al. BOAD: Discovering Hierarchical Software Engineering Agents via Bandit Optimization. ICLR, 2026.

Comments on the paper

1 comment

Dear Reviewers,

As the AC–reviewer discussion phase is about to begin, please highlight any disagreements or inconsistencies across the reviews, as well as any concerns that remain unresolved after the author rebuttal.

Please share your perspectives on these points and engage with one another to help us reach a well-supported assessment of the paper.

Thank you for your time and contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.