Organizations: Northeastern University · Washington University in St. Louis · Nanjing University of Aeronautics and Astronautics · Tsinghua University · Security Capability Center · Tencent · Beijing University of Posts and Telecommunications
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.
Figures & tables
Figure 1: Comparison of paradigms for eliciting latent capabilities in LLMs. Prompt-based methods modify model inputs, predefined steering applies fixed internal interventions, while Q-NeuS dynamically adapts neuron-level interventions to each query.
Figure 2
Figure 3: Overview of Q-NeuS. Q-NeuS generates query-adaptive neuron-wise modulation coefficients from query-induced activations. During training, subject-guided neurons form intervention rollouts evaluated by correctness, margin, and intervention-strength rewards. Neu-GRPO optimizes the Query-Adaptive Gate with group-relative feedback while keeping the base LLM frozen.
Model
MMLU-Pro
SuperGPQA
WildSci
Base
Q-NeuS
Δ
Base
Q-NeuS
Δ
Base
Q-NeuS
Δ
Qwen2.5-3B
31.43
38.37
+6.94
19.58
24.37
+4.79
33.95
39.51
+5.56
Qwen3-4B
44.86
45.93
+1.07
26.29
29.79
+3.50
43.20
45.52
+2.32
Qwen3-8B
44.20
48.53
+4.33
26.83
30.95
+4.12
42.91
45.78
+2.87
Qwen3-14B
50.89
54.03
+3.14
30.35
35.46
+5.11
46.91
49.58
+2.67
Qwen3-32B
54.56
59.11
+4.55
33.97
38.73
+4.76
49.64
52.56
+2.92
Table 2: Performance of Q-NeuS across different LLMs on MMLU-Pro, SuperGPQA, and WildSci.
Model
In-Domain
Cross-Benchmark Generalization
MMLU-Pro
GPQA-M
GPQA-D
SuperGPQA
LiveBench
AGIEval
GSM8K
ARC-C
Qwen2.5-3B
38.37 ( +6.94 )
33.04 ( +2.91 )
36.87 ( +4.55 )
19.39 ( +0.13 )
24.15 ( -0.48 )
52.32 ( +0.31 )
81.20 ( -2.27 )
82.76 ( +0.17 )
Qwen3-4B
45.93 ( +1.07 )
36.61 (0.00)
39.90 ( +2.53 )
27.72 ( +1.30 )
55.06 ( -0.26 )
58.29 ( -0.63 )
90.75 ( +0.30 )
89.59 ( -0.60 )
Qwen3-8B
48.53 ( +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 )
Qwen3-14B
54.03 ( +3.14 )
40.62 ( +2.45 )
42.93 ( +4.04 )
32.71 ( +3.23 )
57.57 ( +0.36 )
65.55 ( +0.14 )
94.54 ( +0.23 )
92.66 ( +0.42 )
Qwen3-32B
59.11 ( +4.55 )
42.41 ( -1.12 )
43.43 (0.00)
33.46 ( +4.14 )
58.08 ( -0.61 )
69.41 ( +0.04 )
95.45 ( +0.53 )
94.80 ( +0.26 )
Table 3: In-domain performance and cross-benchmark generalization of Q-NeuS trained on MMLU-Pro. Each cell reports accuracy (%), with the change over the base model shown in parentheses.
Method
In-Domain
Cross-Benchmark Generalization
MMLU-Pro
GPQA-M
GPQA-D
SuperGPQA
LiveBench
AGIEval
GSM8K
ARC-C
Base
44.20
36.16
33.84
28.43
51.33
59.65
92.65
90.96
GRPO
48.33
35.71
33.84
30.77
51.36
60.07
93.18
91.21
Q-NeuS
48.53
36.38
36.87
29.54
51.96
59.28
93.25
91.64
Table 4: Cross-benchmark transfer comparison between Q-NeuS and GRPO on Qwen3-8B.
Method
Bio.
Bus.
Chem.
CS
Econ.
Eng.
Health
Hist.
Law
Math
Other
Phil.
Phys.
Psych.
Avg.
Base
70.34
33.96
25.11
50.00
60.59
43.59
53.94
58.44
32.58
36.53
42.47
38.61
41.38
60.25
44.20
Fix
-1.38
-1.26
-0.44
+1.22
+1.76
+3.59
+1.21
-1.30
-0.90
+2.58
+0.54
+0.00
-0.77
+0.00
+0.45
GRPO
+2.07
+13.84
+11.45
+3.66
+4.12
+1.54
+6.67
-1.30
-0.45
-0.74
+5.91
+2.97
+3.07
+4.35
+4.13
Ours
+3.45
+5.03
+7.49
+1.22
+5.88
+3.59
+13.94
+1.30
+4.07
-0.74
+4.30
+2.97
+2.30
+5.59
+4.33
Table 5: Subject-wise accuracy changes on MMLU-Pro with Qwen3-8B. The baseline row reports original accuracy, other rows show changes in percentage points (pp).
Variant
λc
λm
λs
Acc. (%)
Δ (pp)
w/o Rmargin
1.0
0
0.1
48.04
+3.84
w/o Rstrength
1.0
1.0
0
47.83
+3.63
w/o Rmargin,Rstrength
1.0
0
0
47.50
+3.30
Full Neu-GRPO
1.0
1.0
0.1
48.53
+4.33
Table 6: Neu-GRPO reward ablation on MMLU-Pro with Qwen3-8B. Baseline: 44.20%.
Sampling Ratio
Baseline (0%)
10%
20%
30%
40%
50%
60%
70%
80%
90%
Acc. (%)
44.20
46.76
47.13
47.38
48.24
48.16
48.53
48.29
48.29
46.80
Δ (pp)
+0.00
+2.56
+2.93
+3.18
+4.04
+3.96
+4.33
+4.09
+4.09
+2.60
Table 7: Effect of the neuron sampling ratio on MMLU-Pro with Qwen3-8B.
Figure 4: Case study of Fix and Q-NeuS across queries with different semantic relationships.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Layer 5
Figure 10: Mean absolute FFN neuron activations across Transformer layers for different MMLU-Pro subjects.
Model
Gate Params
Gate / Backbone
BF16 Weight Size
Qwen2.5-3B
45.17M
1.46%
90.34MB
Qwen3-4B
39.92M
0.99%
79.85MB
Qwen3-8B
50.41M
0.62%
100.83MB
Qwen3-14B
71.39M
0.48%
142.78MB
Qwen3-32B
104.95M
0.32%
209.91MB
Ministral3-8B
58.80M
0.69%
117.61MB
Appendix
Table 8: Parameter overhead of the shared Query-Adaptive Gate across different backbone models. Weight storage is reported in BF16.
Inference-time steering is crucial for deploying Large Language Models (LLMs) in real-world applications, where they must flexibly adjust behavior preferences. For instance, a model should dynamically adjust its caution level based on the risk of the scenario. This requirement introduces a challenge: modulating the intensity of LLM preferences. Prompt engineering can require task-specific prompt design, while fine-tuning requires additional optimization and data. These requirements can make repeated behavioral adjustments costly. To address this, we introduce NeuroSteer, a plug-and-play framework that enables flexible, quantitative adjustment of LLM preferences at inference time. NeuroSteer employs a Sparse Autoencoder (SAE) as its controller. Comparing neuron activation frequencies between positive and negative examples isolates "feature neurons" that correspond to target preferences such as politeness and positive sentiment. During inference, the activations of selected SAE features are amplified in the latent representations to steer the output distribution. Users can adjust the strength of this steering on a per-scenario or per-user basis, granting the LLM fine-grained, quantifiable trait modulation. Our experiments show that NeuroSteer can alter sentiment, toxicity, stance, and politeness at inference time. It achieves higher target-direction steering scores than the compared baselines in several settings, with task- and strength-dependent trade-offs in generation quality.
Keqi Chen, Zekai Sun, Yuhua Wen +5
Beijing University of Posts and Telecommunications, Beijing, China
Large language models can be steered at inference time through prompting or activation interventions, but activation steering methods often underperform compared to prompt-based approaches. We propose a framework that formulates prompt steering as a form of activation steering and investigates whether distilling successful prompt steering behavior into simpler, interpretable models can close this gap. Our analysis reveals that popular activation steering methods are not faithful to the mechanics of prompt steering, which applies strong interventions on some tokens while barely affecting others. Based on these insights, we introduce Prompt Steering Replacement (PSR) models that estimate token-specific steering coefficients from the activations themselves and are trained to imitate prompt-based interventions. Experiments on three steering benchmarks across multiple language models show that PSR models outperform existing activation steering methods, especially when controlling for high-coherence completions, and also compare favorably to prompting on AxBench and persona steering.
Large language models (LLMs) often exhibit undesirable behaviors, such as safety violations and hallucinations. Although inference-time steering offers a cost-effective way to adjust model behavior without updating its parameters, existing methods often fail to be simultaneously effective, utility-preserving, and training-efficient due to their rigid, one-size-fits-all designs and limited adaptability. In this work, we present FineSteer, a novel steering framework that decomposes inference-time steering into two complementary stages: conditional steering and fine-grained vector synthesis, allowing fine-grained control over when and how to steer internal representations. In the first stage, we introduce a Subspace-guided Conditional Steering (SCS) mechanism that preserves model utility by avoiding unnecessary steering. In the second stage, we propose a Mixture-of-Steering-Experts (MoSE) mechanism that captures the multimodal nature of desired steering behaviors and generates query-specific steering vectors for improved effectiveness. Through tailored designs in both SCS and MoSE, FineSteer maintains robust performance on general queries while adaptively optimizing steering vectors for targeted inputs in a training-efficient manner. Extensive experiments on safety and truthfulness benchmarks show that FineSteer outperforms state-of-the-art methods in overall performance, achieving stronger steering performance with minimal utility loss. Code is available at https://github.com/YukinoAsuna/FineSteer