Summary: Inspired by Finding 8 of *Look Before You Leap*, this work introduces a lightweight, black-box paradigm for Code LLM risk assessment that shifts uncertainty estimation from surface-level textual similarity to execution semantics. By demonstrating that syntactic consensus can mask substantial functional divergence, it suggests that grounding a model’s confidence in the runtime behavior of its generated programs is a promising direction for building trustworthy code automation.

Summary: We propose characterizing LLM behavior unsupervisedly by jointly analyzing internal neuron-level representations and external uncertainty signals. By capturing complementary behavioral cues, this approach enables more effective and sample-efficient testing.

Summary: The internal hidden states of a code model contain critical signals regarding the trustworthiness of its outputs. However, effectively leveraging these signals from modern, massive LLMs requires an approach that matches their sheer complexity. Because a model's generative power emerges from massive scale, we argue that its risk assessment framework should evolve to match this by applying scalable pre-training approaches to these highly informative internal layers.

Summary: This work represents an early exploratory work that leverages external uncertainty analysis to evaluate model trustworthiness, highlighting that the open-ended, generative mechanisms of LLMs require fundamentally different uncertainty measurements than the fixed-class probability outputs of classical DNNs.

Summary: This work demonstrates that for DNN models (more specifically, Transformer), systematic and rigorous exhaustive testing can move beyond empirical evaluations to provide verifiable, certified results regarding a model's robustness against patch attacks.

Summary: This work performed a large-scale measurement study of the EOSIO ecosystem through graph analysis. It reveals EOSIO's superficial prosperity and pervasive malicious activities.