acceptodds
Under review as a conference paper at ICLR 2027

QUERY-ADAPTIVE NEURON STEERING FOR LLMs

Abstract

Effectively eliciting the capabilities of large language models (LLMs) at inference time remains an important challenge. Existing studies address this challenge through context engineering, which shapes user queries, or fixed steering, which directly intervenes in model computation. However, predefined contexts and fixed intervention strategies may not adequately accommodate the diverse knowledge and reasoning requirements of individual queries. This raises a natural question: Can internal steering in LLMs be made query-adaptive? In this work, we investigate whether query-induced internal activations can guide adaptive steering in LLMs. Our pilot studies reveal subject- and layer-dependent neuron activation patterns and show that amplifying subject-relevant neurons can improve model performance. Based on these findings, we propose Q-NeuS, a query-adaptive neuron steering framework that dynamically determines which neurons to modulate and by how much for each input. Specifically, Q-NeuS constructs a compact subject-guided candidate neuron space and employs a lightweight Query-Adaptive Gate to generate neuron-wise modulation coefficients. To learn effective intervention strategies, we introduce Neu-GRPO, which adapts group-relative optimization to train the Query-Adaptive Gate using the relative performance of multiple neuron intervention configurations. The base LLM remains frozen throughout training and inference, and no task or subject labels are required at inference time. Extensive experiments across multiple LLMs and benchmarks demonstrate that Q-NeuS consistently improves model performance and exhibits broad applicability across model scales and strong cross-benchmark transferability.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM, neuron steering, activation intervention, adaptive gating

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NM4LOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Q-NeuS, a method that adjusts FFN neuron activations according to each query while keeping the base language model frozen. A shared gate uses intermediate activations and layer embeddings to generate neuron-wise modulation coefficients. During training, subject-specific activation statistics identify candidate neurons, and sampled intervention configurations are evaluated through correctness, answer-margin, and intervention-strength rewards. Neu-GRPO then uses relative rollout feedback to optimize the gate. At inference, query activations guide modulation without subject labels. Experiments cover seven models from two families and three in-domain benchmarks, with improvements across all 21 combinations. Cross-benchmark experiments assess transfer from MMLU-Pro, while comparisons with fixed steering and LoRA-GRPO examine the benefits of the learned intervention.

Critical strengths
  1. The method improves all 21 in-domain model–benchmark combinations, covering two model families and sizes from 3B to 32B.
  2. The preliminary neuron analyses motivate the design, and the reward and sampling ablations provide useful evidence about its components.
Critical weaknesses
  1. Section 4.2 defines rollout rewards but leaves the Neu-GRPO update largely implicit. The policy probabilities, advantage calculation, clipping, and gradient path require a complete specification.
  2. The fixed-intervention baseline changes optimization and intervention coverage alongside query conditioning. A learned query-independent gate would isolate the contribution of adaptivity.
  3. Transfer varies by task. LiveBench shows four losses across seven models, and several advantages over LoRA-GRPO are small. Repeated runs would help establish the reliability of these differences.
Questions for authors
  1. What exact objective does Neu-GRPO optimize, and how are probabilities and gradients computed for the sampled intervention configurations?
  2. How much improvement remains with a learned query-independent gate under the same training budget?
5 comments
AuthorsOpenReview ↗

Response to Reviewer NM4L

Thanks for your positive feedback and constructive comments. We would like to address your concerns in weaknesses and questions as follows.

W1&Q1 Complete Mathematical Formulation and Gradient Propagation Mechanism of Neu-GRPO

Thank you for pointing out that the original manuscript did not explicitly specify the Neu-GRPO update. For each query qq, we sample KK neuron masks M(k)\mathbf{M}^{(k)} from subject-guided candidate sets and hold them fixed during the parameter update. Under each mask, the GateNetwork generates neuron-wise modulation coefficients. We apply softmax to the valid single-token answer-option logits at the final prompt position to obtain πϕ(a∣q,M(k))\pi_\phi(a\mid q,\mathbf{M}^{(k)}). Thus, the objective uses answer probabilities conditioned on the sampled masks, rather than probabilities of the masks themselves. Each rollout samples an answer a(k)a^{(k)} from this distribution and receives a reward R(k)R^{(k)}. Rewards for the same query are normalized within the group:

A(k)=R(k)−μRσR+ϵa.A^{(k)}=\frac{R^{(k)}-\mu_R}{\sigma_R+\epsilon_a}.

Neu-GRPO then minimizes the following GateNetwork loss:

=-\frac{1}{K}\sum_{k=1}^{K} \operatorname{sg}(A^{(k)}) \log\pi_\phi(a^{(k)}\mid q,\mathbf{M}^{(k)}),$$ where $\mu_R$ and $\sigma_R$ are the mean and standard deviation of the $K$ rewards, $\epsilon_a=10^{-8}$ provides numerical stability, and $\operatorname{sg}$ denotes stop-gradient. During the update, the masks, sampled answers, and advantages are held fixed, while answer probabilities are recomputed under the same masks with gradient tracking. Gradients pass from these probabilities through the frozen LLM’s differentiable operations, the modulated FFN activations, and the continuous coefficients to the GateNetwork parameters; the backbone parameters remain unchanged. We accumulate gradients over the $K$ rollouts, clip their global norm, and take one Adam step. The revised Appendix B provides the full update procedure and gradient derivation.
AuthorsOpenReview ↗

Response to Reviewer NM4L

W2&Q2 How to Validate the Independent Contribution of Query Adaptivity?

Thank you for your comments. We agree that comparisons with fixed intervention methods alone are insufficient to isolate the contribution of query adaptivity. Following your suggestion, we introduce a Query-Independent Trainable Gate with the same training budget on Qwen3-8B, along with two additional baselines: Averaged Coefficients and IA³ .

Query-Independent Trainable Gate. We replace the activation input to the GateNetwork with a fixed vector while retaining the layer embeddings, allowing it to learn layer-specific and neuron-wise modulation coefficients that remain unchanged across queries. The training configurationand intervention scope are kept identical to those of Q-NeuS.

Averaged Coefficients. We average the modulation coefficients generated by the trained GateNetwork over a calibration set containing queries from different subjects. The resulting fixed coefficients are then applied uniformly to all queries during inference.

IA³. IA³ is a parameter-efficient fine-tuning method that learns input-independent activation scaling parameters while keeping the base model frozen. We include it as an additional trainable static scaling baseline to compare with the dynamic modulation of Q-NeuS.

The detailed results are presented below.

In-Domain: MMLU-Pro. Cross-Benchmark: all remaining benchmarks.

MethodMMLU-ProGPQA-MGPQA-DSuperGPQALiveBenchAGIEvalGSM8KARC-C
Base44.2036.1633.8428.4351.3359.6592.6590.96
Query-Independent Gate44.28 (+0.08)35.71 (-0.45)31.81 (-2.03)28.68 (+0.25)50.44 (-0.89)59.35 (-0.30)88.70 (-3.95)90.96 (0.00)
Averaged Coefficients46.30 (+2.10)33.04 (-3.12)34.85 (+1.01)28.52 (+0.09)25.81 (-25.52)52.81 (-6.84)46.32 (-46.33)90.78 (-0.18)
IA³46.47 (+2.27)35.94 (-0.22)33.84 (0.00)28.50 (+0.07)50.21 (-1.12)59.86 (+0.21)90.98 (-1.67)91.30 (+0.34)
Q-NeuS48.53 (+4.33)36.38 (+0.22)36.87 (+3.03)29.54 (+1.11)51.96 (+0.63)59.28 (-0.37)93.25 (+0.60)91.64 (+0.68)

The Query-Independent Trainable Gate yields only a marginal improvement on the in-domain benchmark and shows limited transfer performance. Averaged Coefficients improves accuracy on a few benchmarks but exhibits substantial degradation on others, while IA³ generally underperforms Q-NeuS. These results suggest that static modulation strategies, whether learned directly or derived from a trained GateNetwork, may have limitations in accommodating heterogeneous queries.

In comparison, Q-NeuS outperforms all three static baselines on nearly all evaluated benchmarks, suggesting that dynamic neuron modulation provides greater performance benefits on these tasks and offering further evidence for the value of query adaptivity.

AuthorsOpenReview ↗

Response to Reviewer NM4L

W3 Variability in Cross-Task Transfer Performance and Opportunities for Further Reliability Validation

Thank you for pointing out this concern. We agree that repeated runs can help improve the reliability of experimental evaluations. To address this, we repeated the experiments with another random seed while keeping the data splits and all other training settings unchanged. The complete results are presented below.

In-Domain: MMLU-Pro. Cross-Benchmark: all remaining benchmarks.

ModelMMLU-ProGPQA-MGPQA-DSuperGPQALiveBenchAGIEvalGSM8KARC-C
Qwen2.5-3B37.22 (+5.79)31.03 (+0.90)34.34 (+2.02)22.44 (+3.18)23.53 (-1.10)52.55 (+0.54)83.32 (-0.15)83.62 (+1.03)
Qwen3-4B46.80 (+1.94)37.72 (+1.11)40.40 (+3.03)27.74 (+1.32)55.06 (0.00)58.68 (-0.24)90.98 (+0.53)89.76 (-0.43)
Qwen3-8B48.86 (+4.66)37.95 (+1.79)38.89 (+5.05)29.61 (+1.18)52.08 (+0.75)60.55 (+0.90)93.48 (+0.83)91.30 (+0.34)
Qwen3-14B53.70 (+2.81)39.06 (+0.89)37.88 (-1.01)32.51 (+3.03)56.72 (-0.49)65.98 (+0.57)93.48 (-0.83)92.92 (+0.68)
Qwen3-32B59.31 (+4.75)44.87 (+1.34)45.45 (+2.02)32.96 (+3.64)59.09 (+0.40)69.37 (0.00)95.38 (+0.46)94.80 (+0.26)
Ministral3-8B48.86 (+4.04)39.06 (+1.78)39.90 (+5.05)28.17 (+3.24)39.38 (-0.74)54.67 (+0.15)91.21 (+0.84)90.44 (-1.03)
Ministral3-14B53.82 (+3.34)35.71 (-0.90)35.86 (+1.52)30.31 (+1.72)42.26 (+0.21)58.99 (+0.18)94.24 (+0.76)93.17 (+0.85)
Win / Tie / Loss7 / 0 / 06 / 0 / 16 / 0 / 17 / 0 / 03 / 1 / 35 / 1 / 15 / 0 / 25 / 0 / 2

These results show that the main performance gains are maintained under another training seed, providing additional evidence for the reliability of Q-NeuS.

Regarding the comparable performance with LoRA-GRPO, our primary goal is to explore query-adaptive neuron modulation while keeping the pretrained LLM frozen. Unlike LoRA-GRPO, which learns shared parameter updates, Q-NeuS dynamically adjusts neuron-wise modulation coefficients based on query-induced activations. Despite updating only the GateNetwork, Q-NeuS achieves performance comparable to LoRA-GRPO on in-domain benchmarks, demonstrating the potential of query-adaptive neuron steering.

Reviewer_jMk8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether LLM performance can be improved by dynamically modulating internal neuron activations according to the input query, without updating the backbone model. The authors first observe that highly activated neurons exhibit subject- and layer-dependent patterns, and that amplifying selected neurons can improve performance, although fixed interventions do not consistently benefit all subjects.

Based on these observations, the paper proposes Q-NeuS, which introduces a query-adaptive gating network to predict neuron-wise amplification coefficients from the model's internal activations. The authors further introduce Neu-GRPO to train the gate using relative rewards from multiple neuron intervention configurations, while keeping the backbone frozen.

Experiments are conducted on seven LLMs ranging from 3B to 32B parameters across MMLU-Pro, SuperGPQA, and WildSci. The method achieves consistent improvements on the evaluated in-domain benchmarks and shows positive transfer to several unseen benchmarks. The paper also includes ablations on reward components and neuron sampling ratios, along with case studies illustrating query-dependent intervention patterns.

Critical strengths
  1. The paper addresses an interesting problem and provides a reasonable motivation.

I appreciate that the authors start from empirical observations rather than directly introducing another trainable module. The subject- and layer-dependent neuron activation patterns, together with the inconsistent performance of fixed interventions, provide a reasonable motivation for exploring more adaptive steering strategies. The overall idea of improving a frozen LLM through input-dependent internal modulation is worth investigating.

  1. The empirical evaluation is reasonably extensive, with encouraging results.

The experiments cover seven models across two model families and multiple benchmarks, and the reported in-domain improvements are fairly consistent. I also appreciate the cross-benchmark evaluation, since generalization beyond the training distribution is particularly relevant for this type of method. Although the transfer results are somewhat mixed, the overall empirical evidence suggests that the proposed intervention mechanism can be effective in practice.

Critical weaknesses
  1. The results show that learned steering helps, but I am less sure how much query adaptivity contributes. The fixed baseline uses predefined neuron choices and amplification strengths, so this comparison changes both how the coefficients are obtained and whether they depend on the query. Given the overlap in highly activated neurons across subjects, a learned but query-independent intervention seems a plausible alternative. Using the trained gate's average coefficients would be a useful check on one representative setting; there is no need to repeat the full evaluation for this point.

The distinction from prior adaptive steering also needs to be clearer. SADI (ICLR 2025) already studies input-dependent activation interventions. Q-NeuS has a different parameterization and training procedure, but the paper should explain what these add beyond the shared idea of adaptive steering.

  1. I could not reconstruct the Neu-GRPO parameter update from Section 4.2. The paper defines sampled neuron subsets and rollout rewards, but does not give the loss used to update the gate. The gate outputs continuous coefficients, while the subsets appear to be sampled externally. It is unclear what distribution enters the GRPO objective and how the relative rewards produce gradients for the gate. This is the more immediate issue for assessing the method.

I also wanted a clearer reason for using GRPO here. The training examples have labeled answers, and the gate could in principle be trained through the frozen backbone with an answer-prediction loss. This does not make Neu-GRPO inappropriate, but the paper should explain what problem the group-relative objective solves. A clear account of the update and its motivation would help more than further reward-component ablations.

Questions for authors
  1. Could you give the complete Neu-GRPO loss and explain how gradients reach the gate? In particular, what is treated as the policy, and how do the sampled neuron masks enter the update?

  2. Would the gains remain if the trained gate's coefficients were averaged over calibration queries and then held fixed? A check on one model and benchmark would help distinguish learned steering from query-dependent steering.

  3. What motivated Neu-GRPO over supervised training of the same gate? I am asking for the rationale behind this choice, rather than a new comparison across all seven models.

5 comments
AuthorsOpenReview ↗

Response to Reviewer jMk8

Thanks for your positive feedback and constructive comments. We would like to address your concerns in weaknesses and questions as follows.

W1&Q2 How to Validate the Independent Contribution of Query Adaptivity?

Thank you for your valuable suggestion. We agree that comparisons with existing fixed intervention methods alone are insufficient to isolate the contribution of query adaptivity. Following your suggestion, we introduce an Averaged Coefficients baseline on Qwen3-8B, along with two additional baselines: Query-Independent Trainable Gate and IA³.

Averaged Coefficients. We average the modulation coefficients generated by the trained GateNetwork over a calibration set containing queries from different subjects. The resulting fixed coefficients are then applied uniformly to all queries during inference.

Query-Independent Trainable Gate. We replace the activation input to the GateNetwork with a fixed vector while retaining the layer embeddings, allowing it to learn layer-specific and neuron-wise modulation coefficients that remain unchanged across queries. The training configuration and intervention scope are kept identical to those of Q-NeuS.

IA³. IA³ is a parameter-efficient fine-tuning method that learns input-independent activation scaling parameters while keeping the base model frozen. We include it as an additional trainable static scaling baseline to compare with the dynamic modulation of Q-NeuS.

The comparison results are presented below.

In-Domain: MMLU-Pro. Cross-Benchmark: all remaining benchmarks.

MethodMMLU-ProGPQA-MGPQA-DSuperGPQALiveBenchAGIEvalGSM8KARC-C
Base44.2036.1633.8428.4351.3359.6592.6590.96
Averaged Coefficients46.30 (+2.10)33.04 (-3.12)34.85 (+1.01)28.52 (+0.09)25.81 (-25.52)52.81 (-6.84)46.32 (-46.33)90.78 (-0.18)
Query-Independent Gate44.28 (+0.08)35.71 (-0.45)31.81 (-2.03)28.68 (+0.25)50.44 (-0.89)59.35 (-0.30)88.70 (-3.95)90.96 (0.00)
IA³46.47 (+2.27)35.94 (-0.22)33.84 (0.00)28.50 (+0.07)50.21 (-1.12)59.86 (+0.21)90.98 (-1.67)91.30 (+0.34)
Q-NeuS48.53 (+4.33)36.38 (+0.22)36.87 (+3.03)29.54 (+1.11)51.96 (+0.63)59.28 (-0.37)93.25 (+0.60)91.64 (+0.68)

As shown above, IA³ achieves lower accuracy than Q-NeuS on most benchmarks, while Averaged Coefficients provides improvements on only a few tasks and exhibits substantial performance degradation on others. The Query-Independent Trainable Gate yields only a marginal improvement on the in-domain benchmark and shows limited transfer performance. These results suggest that static modulation strategies, whether learned directly or derived from a trained GateNetwork, may have limitations in accommodating heterogeneous queries.

In comparison, Q-NeuS outperforms all three static baselines on nearly all evaluated benchmarks, suggesting that dynamic neuron modulation provides greater performance benefits on these tasks and offering further evidence for the value of query adaptivity.

AuthorsOpenReview ↗

Response to Reviewer jMk8

W2 Insufficient Clarification of the Distinctions from Existing Adaptive Neuron Steering Methods

Thank you for your comments. We appreciate the reviewer for highlighting SADI and agree that the distinction between the two methods deserves further clarification.

The key distinction lies in the dynamic adjustment of neuron modulation. SADI constructs a predefined binary mask based on activation differences between positive and negative samples and applies a fixed scaling factor δ\delta to the masked activations. Although the resulting interventions vary with the input activations, the intervention mask and scaling factor remain fixed across queries. In contrast, Q-NeuS dynamically generates neuron-wise modulation coefficients based on the internal activations induced by each query, allowing different neurons across Transformer layers to receive different modulation strengths. Rather than applying fixed scaling rules to input-dependent activations, Q-NeuS learns to dynamically adjust the modulation of individual neurons according to each query.

We will clarify this distinction in the Related Work section to better highlight the dynamic neuron modulation mechanism of Q-NeuS.

W3&Q1 Complete Mathematical Formulation and Gradient Propagation Mechanism of Neu-GRPO.

Thank you for pointing out that the original manuscript did not explicitly specify the Neu-GRPO update. For each query qq, we sample KK neuron masks M(k)\mathbf{M}^{(k)} from subject-guided candidate sets and hold them fixed during the parameter update. Under each mask, the GateNetwork generates neuron-wise modulation coefficients. We apply softmax to the valid single-token answer-option logits at the final prompt position to obtain πϕ(a∣q,M(k))\pi_\phi(a\mid q,\mathbf{M}^{(k)}). Thus, the objective uses answer probabilities conditioned on the sampled masks, rather than probabilities of the masks themselves. Each rollout samples an answer a(k)a^{(k)} from this distribution and receives a reward R(k)R^{(k)}. Rewards for the same query are normalized within the group:

A(k)=R(k)−μRσR+ϵa.A^{(k)}=\frac{R^{(k)}-\mu_R}{\sigma_R+\epsilon_a}.

Neu-GRPO then minimizes the following GateNetwork loss:

=-\frac{1}{K}\sum_{k=1}^{K} \operatorname{sg}(A^{(k)}) \log\pi_\phi(a^{(k)}\mid q,\mathbf{M}^{(k)}),$$ where $\mu_R$ and $\sigma_R$ are the mean and standard deviation of the $K$ rewards, $\epsilon_a=10^{-8}$ provides numerical stability, and $\operatorname{sg}$ denotes stop-gradient. During the update, the masks, sampled answers, and advantages are held fixed, while answer probabilities are recomputed under the same masks with gradient tracking. Gradients pass from these probabilities through the frozen LLM’s differentiable operations, the modulated FFN activations, and the continuous coefficients to the GateNetwork parameters; the backbone parameters remain unchanged. We accumulate gradients over the $K$ rollouts, clip their global norm, and take one Adam step. The revised Appendix B provides the full update procedure and gradient derivation.
AuthorsOpenReview ↗

Response to Reviewer jMk8

Q3 Why Choose Neu-GRPO over Supervised Fine-Tuning (SFT)?

Thank you for your comments. We agree that the motivation for choosing Neu-GRPO over SFT deserves further clarification. When ground-truth answers are available, SFT is a viable training approach. However, the primary goal of Q-NeuS is to learn how to adaptively modulate neuron activations according to each query, while answer labels do not directly specify the desired modulation of individual neurons. Our motivation for adopting Neu-GRPO is therefore to leverage answer-based feedback to compare the relative effectiveness of different intervention configurations for the same query and optimize the GateNetwork accordingly.

When the sampled neuron subsets fail to yield clear benefits for a given query, SFT may still update the GateNetwork through cross-entropy to increase the likelihood of the correct answer. Such updates are not explicitly guided by differences in intervention effectiveness and may cause the GateNetwork to learn from relatively less effective interventions. In contrast, Neu-GRPO compares multiple intervention configurations for the same query and assigns positive or negative advantages based on their relative rewards. We believe that this comparison helps emphasize relatively effective interventions while avoiding equally reinforcing less effective configurations, thereby reducing the risk of GateNetwork updates that lack evidence of relative intervention effectiveness.

To complement this rationale with empirical evidence, we conducted a comparison using Qwen3-8B and the same Gate architecture. On each of our three primary benchmarks, we trained two supervised baselines: SFT-CE, using only cross-entropy loss, and SFT-Full, which additionally includes answer-margin and intervention-strength terms in the training objective. Test accuracies on the held-out splits of these benchmarks are reported below.

MethodsMMLU-ProSuperGPQAWildSci
Base44.2026.8342.91
SFT-CE45.89 (+1.69)28.21 (+1.38)44.50 (+1.59)
SFT-Full38.29 (-5.91)19.10 (-7.73)39.54 (-3.37)
Neu-GRPO48.53 (+4.33)30.95 (+4.12)45.78 (+2.87)

As shown above, SFT-CE consistently improves upon the base model, demonstrating that answer-level supervision can effectively train the GateNetwork. However, its improvements remain smaller than those achieved by Neu-GRPO across all three benchmarks. In contrast, SFT-Full performs below the base model, suggesting that simply incorporating margin and intervention-strength terms into the supervised objective does not reproduce the benefits of Neu-GRPO. These results support the effectiveness of group-relative optimization for learning neuron modulation strategies.

Cross-Benchmark Generalization. We further evaluate the GateNetworks trained on MMLU-Pro by directly transferring them to unseen benchmarks without additional training.

In-Domain: MMLU-Pro. Cross-Benchmark: all remaining benchmarks.

MethodMMLU-ProGPQA-MGPQA-DSuperGPQALiveBenchAGIEvalGSM8KARC-C
Base44.2036.1633.8428.4351.3359.6592.6590.96
SFT-CE45.89 (+1.69)33.93 (-2.23)33.33 (-0.51)28.84 (+0.41)50.14 (-1.19)59.68 (+0.03)92.49 (-0.16)87.80 (-3.16)
SFT-Full38.29 (-5.91)33.48 (-2.68)31.82 (-2.02)27.80 (-0.63)51.22 (-0.11)59.73 (+0.08)92.65 (0.00)89.68 (-1.28)
Neu-GRPO48.53 (+4.33)36.38 (+0.22)36.87 (+3.03)29.54 (+1.11)51.96 (+0.63)59.28 (-0.37)93.25 (+0.60)91.64 (+0.68)

The results show that although SFT-CE improves accuracy on the MMLU-Pro test set, its performance varies substantially across other benchmarks. In contrast, Neu-GRPO improves in-domain performance while achieving more consistent cross-benchmark gains, suggesting stronger generalization.

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

This paper proposes Q-NeuS, a query-adaptive neuron steering framework that improves LLM performance without modifying the base model parameters. The method uses a lightweight Query-Adaptive Gate to generate neuron-wise modulation coefficients based on internal activations, computed dynamically. It also introduces Neu-GRPO to optimize the Gate using group-relative rewards from different neuron-intervention configurations.

Critical strengths
  1. The paper is well-written and clearly organized, with intuitive and informative figures that make the proposed method easy to understand.

  2. The paper provides a relatively comprehensive experimental evaluation across multiple LLM families, model scales, and benchmarks, demonstrating consistent performance improvements over the original models.

Critical weaknesses
  1. The proposed method lacks substantial novelty, as its core components, including adaptive gating, neuron-wise activation scaling, and reward-based optimization, are general neural network techniques. The framework appears largely independent of LLM-specific characteristics and could be applied to conventional neural networks without substantial modifications. This raises concerns about its methodological contribution to LLM research.

  2. The proposed method closely resembles IA³ in its use of learnable activation scaling, yet the paper does not include IA³ or an equivalent trainable static scaling baseline. Without such comparisons, it remains unclear whether the performance improvements stem from the proposed Query-Adaptive mechanism or simply from introducing trainable activation scaling into a frozen LLM.

  3. I could not find a complete mathematical formulation of Neu-GRPO. In particular, the paper does not clearly define the optimization objective, the policy being optimized, or how the gradients are computed to update the Query-Adaptive Gate. These missing details make it difficult to understand and reproduce the proposed training algorithm.

  4. The method exhibits a training–inference inconsistency. During training, Neu-GRPO intervenes only on sampled subsets of subject-specific neurons, whereas inference applies the Gate to all FFN neurons. The paper does not adequately justify this discrepancy or evaluate its impact on performance.

  5. The use of Neu-GRPO appears unnecessary. Since ground-truth answers are available, the Query-Adaptive Gate could be directly trained using supervised learning, without designing additional reward functions and introducing reinforcement learning. The paper does not justify why this more complicated training approach is needed or demonstrate its advantages over standard supervised learning.

  6. The experimental improvements over GRPO are marginal on several benchmarks, with performance degradation on some tasks. However, the paper does not report results across multiple random seeds or statistical significance tests, making it difficult to assess the reliability of the reported improvements.

Questions for authors
  1. Could the authors clarify which components of the proposed method specifically exploit the unique characteristics of LLMs, rather than relying on general neural network techniques that could be applied to conventional neural networks?

  2. Why does Neu-GRPO use RL instead of supervised learning?

9 comments
AuthorsOpenReview ↗

Response to Reviewer gBhD

Thank you for your valuable feedback and thoughtful comments. We would like to address your concerns in weaknesses and questions as follows.

W1 Limited Novelty Due to the Use of Conventional Neural Network Components

Thank you for your comments. We agree that adaptive gating, multiplicative activation scaling, and reward-based optimization are established techniques. Our primary contribution lies in demonstrating how query-induced internal activations can guide adaptive neuron-level steering and in developing a learning framework for such interventions in frozen pretrained LLMs.

Motivated by the subject- and layer-dependent activation patterns and uneven effects of fixed amplification observed in Section 3, Q-NeuS employs a layer-aware gate that maps query-induced FFN activations to neuron-wise modulation coefficients. During training, subject-guided candidate sets constrain the exploration of neuron subsets, while Neu-GRPO optimizes the gate using the relative performance of multiple intervention configurations for the same query. This framework learns to modulate individual FFN neurons according to the knowledge and reasoning requirements of each query, while keeping the base LLM frozen and requiring no subject labels at inference.

We would like to emphasize that the key innovation of Q-NeuS lies not in the individual neural network components, but in learning how to identify and selectively amplify neurons that are most relevant to each query, thereby enabling fine-grained, query-adaptive internal steering rather than relying on predefined neuron selections and fixed activation strengths. We will revise the manuscript to better clarify this methodological contribution and distinguish it from existing approaches.

AuthorsOpenReview ↗

Response to Reviewer gBhD

W2 Insufficient Comparisons with Trainable Static Scaling Baselines to Validate the Contribution of Query Adaptivity

Thank you for pointing out this limitation. We agree that comparisons with fixed intervention methods alone are insufficient to distinguish the contributions of learnable scaling and query adaptivity. To address this concern, we introduce three additional baselines on Qwen3-8B: IA³ , Query-Independent Trainable Gate and Averaged Coefficients .

IA³. IA³ is a parameter-efficient fine-tuning method that learns input-independent activation scaling parameters while keeping the base model frozen. We include it as an additional trainable static scaling baseline to compare with the dynamic modulation of Q-NeuS.

Query-Independent Trainable Gate. We replace the activation input to the GateNetwork with a fixed vector while retaining the layer embeddings, allowing it to learn layer-specific and neuron-wise modulation coefficients that remain unchanged across queries. The training configuration, computational budget, and intervention scope are kept identical to those of Q-NeuS.

Averaged Coefficients. We average the modulation coefficients generated by the trained GateNetwork over a calibration set containing queries from different subjects. The resulting fixed coefficients are then applied uniformly to all queries during inference.

The comparison results are presented below.

In-Domain: MMLU-Pro. Cross-Benchmark: all remaining benchmarks.

MethodMMLU-ProGPQA-MGPQA-DSuperGPQALiveBenchAGIEvalGSM8KARC-C
Base44.2036.1633.8428.4351.3359.6592.6590.96
IA³46.47 (+2.27)35.94 (-0.22)33.84 (0.00)28.50 (+0.07)50.21 (-1.12)59.86 (+0.21)90.98 (-1.67)91.30 (+0.34)
Query-Independent Gate44.28 (+0.08)35.71 (-0.45)31.81 (-2.03)28.68 (+0.25)50.44 (-0.89)59.35 (-0.30)88.70 (-3.95)90.96 (0.00)
Averaged Coefficients46.30 (+2.10)33.04 (-3.12)34.85 (+1.01)28.52 (+0.09)25.81 (-25.52)52.81 (-6.84)46.32 (-46.33)90.78 (-0.18)
Q-NeuS48.53 (+4.33)36.38 (+0.22)36.87 (+3.03)29.54 (+1.11)51.96 (+0.63)59.28 (-0.37)93.25 (+0.60)91.64 (+0.68)

As shown above, IA³ achieves lower accuracy than Q-NeuS on most benchmarks, while Averaged Coefficients provides improvements on only a few tasks and exhibits substantial performance degradation on others. The Query-Independent Trainable Gate yields only a marginal improvement on the in-domain benchmark and shows limited transfer performance. These results suggest that static modulation strategies, whether learned directly or derived from a trained GateNetwork, may have limitations in accommodating heterogeneous queries.

In comparison, Q-NeuS outperforms all three static baselines on nearly all evaluated benchmarks, suggesting that dynamic neuron modulation provides greater performance benefits on these tasks and offering further evidence for the value of query adaptivity.

AuthorsOpenReview ↗

Response to Reviewer gBhD

W3 Incomplete Mathematical Formulation of Neu-GRPO

Thank you for pointing out that the original manuscript did not explicitly specify the Neu-GRPO update. For each query qq, we sample KK neuron masks M(k)\mathbf{M}^{(k)} from subject-guided candidate sets and hold them fixed during the parameter update. Under each mask, the GateNetwork generates neuron-wise modulation coefficients. We apply softmax to the valid single-token answer-option logits at the final prompt position to obtain πϕ(a∣q,M(k))\pi_\phi(a\mid q,\mathbf{M}^{(k)}). Thus, the objective uses answer probabilities conditioned on the sampled masks, rather than probabilities of the masks themselves. Each rollout samples an answer a(k)a^{(k)} from this distribution and receives a reward R(k)R^{(k)}. Rewards for the same query are normalized within the group:

A(k)=R(k)−μRσR+ϵa.A^{(k)}=\frac{R^{(k)}-\mu_R}{\sigma_R+\epsilon_a}.

Neu-GRPO then minimizes the following GateNetwork loss:

=-\frac{1}{K}\sum_{k=1}^{K} \operatorname{sg}(A^{(k)}) \log\pi_\phi(a^{(k)}\mid q,\mathbf{M}^{(k)}),$$ where $\mu_R$ and $\sigma_R$ are the mean and standard deviation of the $K$ rewards, $\epsilon_a=10^{-8}$ provides numerical stability, and $\operatorname{sg}$ denotes stop-gradient. During the update, the masks, sampled answers, and advantages are held fixed, while answer probabilities are recomputed under the same masks with gradient tracking. Gradients pass from these probabilities through the frozen LLM’s differentiable operations, the modulated FFN activations, and the continuous coefficients to the GateNetwork parameters; the backbone parameters remain unchanged. We accumulate gradients over the $K$ rollouts, clip their global norm, and take one Adam step. The revised Appendix B provides the full update procedure and gradient derivation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers and the Area Chair

We sincerely thank all reviewers for their careful evaluations and constructive suggestions. We have addressed the reviewers’ concerns through additional experiments and clarifications.

1 Independent Contribution of Query Adaptivity

Reviewers gBhD, jMk8, and NM4L ask about the contribution of query adaptivity; we test static baselines.

To isolate the contribution of query adaptivity, we compare Q-NeuS with three Qwen3-8B baselines: IA³, a Query-Independent Trainable Gate, and Averaged Coefficients. In comparison, Q-NeuS outperforms them on most benchmarks, supporting the benefit of query-adaptive neuron modulation.

2 Incomplete Mathematical Formulation of Neu-GRPO

Reviewers gBhD, jMk8, and NM4L ask for a clearer formulation of Neu-GRPO; we provide the update. Following the Query-Adaptive Neuron Gating mechanism introduced in Section 4.1, we provide a detailed mathematical formulation of Neu-GRPO, including neuron intervention rollouts, grouprelative advantage estimation, the GateNetwork update, and gradient propagation through the frozen LLM in Appendix B.

3 Rationale for Choosing Neu-GRPO over SFT

Reviewers gBhD and jMk8 ask about the motivation for choosing Neu-GRPO over SFT; we compare the two.

We agree that the motivation for choosing Neu-GRPO over SFT deserves further clarification. When ground-truth answers are available, SFT is a viable training approach. However, the primary goal of Q-NeuS is to learn how to adaptively modulate neuron activations according to each query, while answer labels do not directly specify the desired modulation of individual neurons. Our motivation for adopting Neu-GRPO is therefore to leverage answer-based feedback to compare the relative effectiveness of different intervention configurations for the same query and optimize the GateNetwork accordingly. To evaluate this distinction, we trained two SFT baselines using Qwen3-8B on our three primary benchmarks: SFT-CE, which uses only cross-entropy loss, and SFT-Full, which additionally incorporates margin and intervention-strength terms into the supervised objective. The results demonstrate that while answer-level supervision can effectively train the GateNetwork, Neu-GRPO consistently achieves superior performance. Specifically, Neu-GRPO delivers greater in-domain improvements and more consistent gains across benchmarks, suggesting better generalization capability.

4 Results with an Additional Training Seed and Comparison with LoRA-GRPO

Reviewers gBhD and NM4L ask about the reliability of the experimental results and the LoRA-GRPO comparison; we report another seed and clarify the comparison.

We repeated training with an additional random seed while keeping the data splits and settings unchanged. The results show that the main improvements reported in the paper remain stable under the additional seed.

Regarding the comparable performance with LoRA-GRPO, our primary goal is to explore query-adaptive neuron modulation while keeping the pretrained LLM frozen. Unlike LoRA-GRPO, which learns shared parameter updates, Q-NeuS dynamically adjusts neuron-wise modulation coefficients based on query-induced activations. Despite updating only the GateNetwork, Q-NeuS achieves comparable performance on in-domain benchmarks.

5 Novelty and LLM-Specific Design

Reviewer gBhD asks about the novelty and LLM-specific design of Q-NeuS; we explain our contribution.

Our primary contribution lies in demonstrating how query-induced internal activations can guide adaptive neuron-level steering and in developing a learning framework for such interventions in frozen pretrained LLMs. Q-NeuS learns to selectively amplify individual FFN neurons according to each query's knowledge and reasoning requirements, enabling fine-grained, query-adaptive internal steering without predefined neuron selections or fixed activation strengths. The base LLM remains frozen, and no subject labels are required at inference.

Each component of Q-NeuS is designed around LLM computation. Its intervention targets intermediate FFN activations; its GateNetwork outputs neuron-wise modulation coefficients applied to these activations; and Neu-GRPO optimizes the GateNetwork to improve the LLM's answers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.