Source-linked AI summary
NeuReasoner: Towards Explainable, Controllable, and Unified Reasoning via Mixture-of-Neurons
Haonan Dong, Kehan Jiang, Haoran Ye, Wenhao Zhu, Zhaolu Kang, Guojie Song
TL;DR
LRMs exhibit intra-step, inter-step, and instance-level failures, while prior approaches address these levels separately and remain limited in explainability and unified control. NeuReasoner identifies failure-associated Mixture of Neurons and fluctuation patterns, then uses lightweight monitoring MLPs and SFT-learned special-token triggers for self-correction. Across six benchmarks and six backbone models, it improves performance by up to 27.0% while reducing token consumption by 19.6% ∼63.3%.
Problem
LRMs suffer distinct reasoning failures across intra-step, inter-step, and instance levels, while existing approaches target these failures separately and leave explainability and unified control limited.
Method
NeuReasoner performs white-box MoN analysis and uses lightweight MLP monitoring with SFT-learned special tokens to trigger controllable self-correction.
Results
Across six benchmarks and six backbone models, NeuReasoner achieves performance gains up to 27.0% while reducing token consumption by 19.6% ∼63.3%.
Takeaways & Limitations
NeuReasoner provides an explainable, controllable, and unified reasoning framework with superior performance and token efficiency across multiple domains.
Takeaways & Limitations
Monitoring MLPs introduce marginal inference latency, and NeuReasoner is not yet a fully automated end-to-end pipeline.
Abstract
from arXiv · showhide
Large Reasoning Models (LRMs) have recently achieved remarkable success in complex reasoning tasks. However, closer scrutiny reveals persistent failure modes compromising performance and cost: I) Intra-step level, marked by calculation or derivation errors; II) Inter-step level, involving oscillation and stagnation; and III) Instance level, causing maladaptive over-thinking. Existing endeavors target isolated levels without unification, while their black-box nature and reliance on RL hinder explainability and controllability. To bridge these gaps, we conduct an in-depth white-box analysis, identifying key neurons (Mixture of Neurons, MoN) and their fluctuation patterns associated with distinct failures. Building upon these insights, we propose NeuReasoner, an explainable, controllable, and unified reasoning framework driven by MoN. Technically, NeuReasoner integrates lightweight MLPs for failure detection with a special token-triggered self-correction mechanism learned via SFT. During inference, special tokens are inserted upon failure detection to actuate controllable remedial behaviors. Extensive evaluations across six benchmarks, six backbone models (8B~70B) against nine competitive baselines, demonstrate that NeuReasoner achieves performance gains of up to 27.0% while reducing token consumption by 19.6% ~ 63.3%.
1 Introduction
LRMs exhibit distinct intra-step, inter-step, and instance-level reasoning failures that affect correctness and token cost. NeuReasoner addresses these gaps through white-box MoN analysis and token-triggered self-correction, improving performance while reducing token consumption.
- Failure Attribution: LRM failures occur within steps through incorrect deductions, across steps through stagnation or oscillation, and across instances through maladaptive over-thinking.These failure modes respectively compromise final answers, waste tokens before collapse, and waste tokens on simpler queries.
- Research Dilemma: Prior methods separately target step supervision, broader exploration, or adaptive reasoning length, leaving explainability and unified control unresolved.The cited research gap follows approaches that address the three failure levels individually.
- Present Framework: NeuReasoner identifies failure-associated Mixture of Neurons and their fluctuation patterns through fine-grained white-box analysis.The analysis is motivated by the hypothesis that distinct neuron clusters contribute to different failure levels.
- Present Framework: NeuReasoner combines lightweight MLP failure detection with SFT-learned special-token triggers that activate controllable self-correction during inference.Special tokens are dynamically inserted when failures are detected.
- Experimental Validation: 27.0% maximum performance gains accompany 19.6% ∼63.3% lower token consumption across six benchmarks and six backbone models.The evaluation compares NeuReasoner against nine baselines and includes case studies of reasoning details.
2 Dive into Neurons
The paper uses neuron-level attribution to identify MoN experts associated with different failure levels, then analyzes their temporal and frequency-domain activation patterns. These experts differ in location and fluctuation behavior across intra-step, inter-step, and instance-level failures.
- MoN Identification: The analysis focuses on FFN neurons and attention heads in the middle layer, using attribution to identify failure-associated neuron clusters.DePass computes each neuron's independent contribution to target-token generation.
- MoN Identification: MoN is formed by intersecting the top-attributed neurons across the time steps associated with each failure level.The resulting neuron sets function as distinct experts.
- MoN Identification: Intra-step and instance-level experts predominantly cluster in FFNs, whereas inter-step experts concentrate in attention heads.This provides a layer-component distinction among the identified failure-associated experts.
- Fluctuation Analysis: Positive and negative sample pairs are constructed for failure analysis, with Fourier transforms used to characterize MoN fluctuations in time and frequency domains.The procedure uses KV-cache cloning and repeated sampling for intra-step and inter-step failures, while instance-level pairs contrast questions of different difficulty.
- Fluctuation Analysis: Intra-step experts show sharp spikes, inter-step experts show periodic fluctuations, and instance-level experts sustain activation for complex instances but rapidly collapse for simpler ones.These distinct patterns link MoN dynamics to the three failure levels.
3 Method
NeuReasoner predicts neuron fluctuation patterns, reconstructs training data for trigger-based self-correction, and monitors reasoning online to insert corrective controls when failures emerge.
- 3.1 Pattern Prediction: Lightweight MLPs predict MoN fluctuation patterns from Fourier-based features derived from neuron activation sequences.The method uses sliding windows for intra-step and inter-step reasoning, while instance-level prediction uses the initial K steps.
- 3.1 Pattern Prediction: Intra-step prediction targets high-frequency energy, spectral entropy, and total variation energy associated with intense fluctuations.The high-frequency set is defined over the latter half of the spectrum.
- 3.1 Pattern Prediction: Inter-step prediction targets dominant-frequency energy and spectral entropy, capturing periodic oscillations in reasoning trajectories.The dominant-frequency energy ratio measures the largest nonzero-frequency energy share.
- 3.1 Pattern Prediction: Instance-level prediction uses low-frequency concentration and lower spectral entropy to characterize easy queries that rapidly stabilize after initial fluctuations.The instance-level features distinguish this pattern from sustained activation on complex queries.
- 3.2 Trigger Training: Dataset reconstruction and SFT teach special tokens to trigger diagnose-then-correct behaviors for intra-step and inter-step failures.The reconstruction segments outputs into steps, perturbs critical steps with level-specific rewriting, and inserts corresponding trigger tokens and correction templates.
- 3.3 Online Monitoring: During inference, parallel MLP monitoring forces a detected failure token into the decoding prefix, activating learned correction or a no-thinking transition.Intra-step and inter-step triggers invoke SFT-trained correction, whereas instance-level detection inserts a prompt that transitions to NoThinking mode.
4 Experiments
NeuReasoner is evaluated across diverse backbones, benchmarks, baselines, and scalability settings, with results covering performance, token cost, explainability, controllability, and component contributions.
- Experimental Setup: Evaluations use six benchmarks, six backbone models spanning 7B–70B, and nine competitive baselines.The benchmarks cover mathematical, scientific, and code reasoning domains.
- Main Results: Performance gains range from 0.3–7.8 points, or 3.2%–27.0%, while token consumption falls by 19.6%–63.3% versus the vanilla model.On DeepSeek-R1-Distill-Qwen-7B + MATH500, accuracy improves by 5.0 points and token usage decreases by 48.1%.
- Main Results: NeuReasoner generalizes across tasks, achieving a 1.4-point gain and 34.2% lower cost on LiveCodeBench while attaining state-of-the-art results on AIME25, MATH500, and GPQA-Diamond.The comparison is made against DAPO on DeepSeek-R1-Distill-Qwen-32B.
- Scalability Analysis: NeuReasoner scales across model sizes and test-time sampling, with k = 8 matching vanilla performance at k = 64 and k = 64 improving vanilla by 3.5%–10.0%.On the DeepSeek-R1-Distill-Llama series + AIME25, gains are 7.8 points at 8B and 4.4 points at 70B.
- Framework Analysis: Case studies and ablations support explainability and controllability, while removing MLPintra causes the largest performance drop and removing MLPinst sharply increases token consumption.The ablations assess MLPintra, MLPinter, MLPinst, and all MLP detectors.
5 Conclusion
The paper categorizes reasoning failures, identifies associated Mixture of Neurons and fluctuation patterns through fine-grained neuronal analysis, and proposes NeuReasoner as a unified framework.
- Conclusion: NeuReasoner uses fine-grained neuronal analysis to identify Mixture of Neurons and their fluctuation patterns across failure levels.The framework is designed to be explainable, controllable, and unified across multiple reasoning domains.
- Conclusion: Experiments show superior performance and cost-efficiency across multiple domains.The conclusion presents these results as evidence for the framework’s overall effectiveness.
Limitations
NeuReasoner has practical limitations from monitoring overhead and incomplete automation, while its mechanistic interpretability addresses gaps in understanding why reasoning trajectories fail.
- Limitations: Monitoring MLPs introduce persistent marginal inference latency despite approximate O(1) feature-update optimizations.The paper directs readers to a detailed runtime comparison in Appendix F.2.
- Limitations: NeuReasoner is not yet a fully automated, end-to-end framework.The authors identify further optimization of this aspect as future work.
- Interpretability gap: Existing interpretability methods leave a lack of systematic, in-depth mechanistic analysis of why LRMs fail during reasoning trajectories.NeuReasoner addresses this gap by examining internal reasoning mechanisms.
D Dataset
The paper evaluates reasoning across mathematical, scientific, and coding benchmarks, alongside implementation materials and datasets used for reasoning supervision.
- Evaluation benchmarks: The evaluation suite covers mathematical, scientific, and code reasoning benchmarks, including MATH, GSM8K, AIME24, AIME25, GPQA-Diamond, and LiveCodeBench.These benchmarks span competition mathematics, arithmetic, expert scientific questions, and contamination-resistant coding evaluation.
- Mathematical benchmarks: AIME24 and AIME25 each contain 30 problems with integer answers in [0,999], supporting exact-match evaluation of advanced mathematical reasoning.AIME25 additionally provides a fresh annual test of generalization.
- Scientific benchmark: GPQA-Diamond contains 198 expert-authored questions across biology, chemistry, and physics with four-option multiple-choice answers.The split is designed to probe scientific reasoning under expert-level knowledge demands.
- Training data: Bespoke-Stratos-17k provides 17K reasoning examples paired with traces and final answers for supervised fine-tuning across math and coding domains.The dataset was created by replicating and improving the Berkeley Sky-T1 distillation pipeline.
E.1 Data Preprocessing
The preprocessing pipeline uses BS-17k to train failure-detection MLPs and the SFT component while excluding evaluation-overlapping samples to reduce leakage risk.
- Data preprocessing: BS-17k is used as the training dataset for both the failure-detection MLPs and SFT.The dataset supports both monitoring and trigger-behavior learning.
- Data preprocessing: Samples overlapping with MATH500, GSM8K, AIME24, AIME25, GPQA-Diamond, and LiveCodeBench are excluded through text matching.The stated purpose is to prevent potential data leakage.
- Backbone dependence: The framework is not fully automated end-to-end because backbone models vary in layer depth and architecture.This requires independent MoN extraction and MLP training for each backbone, although the processes remain lightweight.
E.3 MLPs Training
MLP training collects varied reasoning trajectories, labels failure segments, and fits a shared lightweight architecture with standard supervised optimization and evaluation settings.
- Data collection: Reasoning trajectories are sampled from DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B to diversify erroneous paths.The trajectories are paired with neuron activations before failure instances are filtered.
- Data labeling: An LLM-as-a-Judge annotates token segments corresponding to detected failures.These annotations provide supervision for the failure-detection models.
- MLP architecture: Each failure-detection MLP uses a three-layer fully connected architecture with GELU, dropout, binary cross-entropy, and AdamW.Input dimensions vary according to the respective feature representations.
- Evaluation: The experiments report accuracy and recall on the MLP test set.The supplied passage identifies these as the reported evaluation metrics.
- Experimental settings: Baseline comparisons standardize temperature at 0.6, top-p at 0.95, and the maximum token limit at 16,000, reporting average Pass@1 over five runs.RL-based baselines use A100 GPUs with bf16 mixed precision and DeepSpeed ZeRO-3.
F.1 AIME24 Results
The section points to additional AIME24 results and describes runtime comparison on an A100 GPU, where NeuReasoner adds only marginal overhead within an acceptable range.
- Table 5 presents additional experimental results on the AIME24 dataset.
- Inference runtime is compared between NeuReasoner and the Vanilla model on an A100 GPU.
- NeuReasoner introduces marginal inference overhead that remains within an acceptable range.
EXAMPLE OF INTRA-STEP LEVEL DATA
The example demonstrates an intra-step counting error caused by ignoring a parity constraint, then repairs the derivation by restricting factor pairs and obtains 12 valid positive integers.
- The initial reasoning incorrectly treats every factor pair of 2016 as producing a valid positive integer n.The overlooked condition is that n+1=(v−u)/2 must be an integer.
- The correction requires u and v to have the same parity, which forces both factors to be even because uv=2016 is even.
- Dividing both factors by 2 transforms uv=2016 into u'v'=504 while preserving the ordering condition.
- 504 has 24 divisors and is not a perfect square, yielding 12 unordered factor pairs and therefore 12 valid positive integers n.
EXAMPLE OF INTER-STEP LEVEL DATA
The example illustrates inter-step stagnation: repeated discriminant, modular, and symmetry checks fail to shrink the search, after which diagonalization provides a consistent route to 24 solutions.
- The problem asks for the number of ordered integer pairs satisfying x^2+xy+y^2=91.
- The loop is explicitly identified as inter-step stagnation because equivalent checks remain permissive or restate the same discriminant constraint.
- Diagonalizing the quadratic form converts the problem into u^2+3v^2=364 with a parity condition.
- Enumerating the admissible square cases gives 24 ordered integer pairs.