Despite the remarkable reasoning performance, eliciting the long chain-ofthought (CoT) ability in large language models (LLMs) typically requires costly reinforcement learning or supervised fine-tuning on high-quality distilled data. We investigate the internal mechanisms behind this capability and show that a small set of high-impact activations in the last few layers, greatly govern the long-form reasoning attributes, e.g., output length and self-reflection. Through simply amplifying these activations and adding "wait" tokens, the long CoT ability can be invoked without training, leading to significantly increased self-reflection rate and accuracy. In addition, we also find that the activation changes follow predictable trajectories, i.e., a sharp rise after special tokens and a subsequent exponential decay. Based on these insights, we introduce a general training-free activation control technique. It utilizes a few contrastive examples to identify the relevant activations, and then incorporates simple analytic functions to adjust their values at inference time to elicit long CoTs. Extensive experiments have verified the effectiveness of our methods in efficiently eliciting the long CoT ability of LLMs and improving the performance. Besides, we further propose a parameter-efficient fine-tuning method that trains only the last-layer activation amplification module and a few LoRA layers, outperforming LoRA on reasoning benchmarks with much fewer parameters.

Adversarial sample attacks perturb benign inputs to induce DNN misbehaviors. Recent research has demonstrated the widespread presence and the devastating consequences of such attacks.

Neural Information Processing Systems http://nips.cc/

Entropy neurons are characterized by an unusually high weight norm and influence the final layer normalization (LayerNorm) scale to effectively scale down the logits.

In this work, we first explore whether the parameters activated by the MoE layer remain highly sparse at inference. We perform a sparsification study on several representative MoE models. For each expert, we rank parameters by the magnitude of their activations from the gate projection and progressively prune the activated subset. Pruning up to 60% of parameters within that subset causes only negligible task-performance degradation; substantial drops occur only after more than 90% are removed. We further decompose experts into neuron-granular MoE and visualize their activation values, finding that most neuron activations are near zero. This observation motivates us to select only high-activation neuron experts during pretraining. Based on this insight, we propose Mixture of Neuron Experts (MoNE). MoNE achieves neuron-granular expert selection by only applying a simple top-k selection within each expert, incurs negligible latency, and requires no additional routing parameters or inter-expert communication. Extensive experiments demonstrate that MoNE matches traditional MoE performance while activating only 50% of the MoE-layer parameters, and it consistently outperforms traditional MoE when compared at equal numbers of activated parameters. These results suggest that MoNE is a practical approach to improving parameter utilization and inference efficiency in MoE-like models.

Deep neural networks (DNNs) are increasingly being deployed in high-stakes applications, from self-driving cars to biometric authentication. However, their unpredictable and unreliable behaviors in real-world settings require new approaches to characterize and ensure their reliability. This paper introduces DeepProv, a novel and customizable system designed to capture and characterize the runtime behavior of DNNs during inference by using their underlying graph structure. Inspired by system audit provenance graphs, DeepProv models the computational information flow of a DNN's inference process through Inference Provenance Graphs (IPGs). These graphs provide a detailed structural representation of the behavior of DNN, allowing both empirical and structural analysis. DeepProv uses these insights to systematically repair DNNs for specific objectives, such as improving robustness, privacy, or fairness. We instantiate DeepProv with adversarial robustness as the goal of model repair and conduct extensive case studies to evaluate its effectiveness. Our results demonstrate its effectiveness and scalability across diverse classification tasks, attack scenarios, and model complexities. DeepProv automatically identifies repair actions at the node and edge-level within IPGs, significantly enhancing the robustness of the model. In particular, applying DeepProv repair strategies to just a single layer of a DNN yields an average 55% improvement in adversarial accuracy. Moreover, DeepProv complements existing defenses, achieving substantial gains in adversarial robustness. Beyond robustness, we demonstrate the broader potential of DeepProv as an adaptable system to characterize DNN behavior in other critical areas, such as privacy auditing and fairness analysis.