Cheng Yang

Siyue Chen, Yifu Guo, Yuquan Lu et al.

Standard accuracy metrics cannot explain why LLMs handle variable tracking but fail on semantically equivalent loops. We study an internal lifecycle of code reasoning in which models first brew the answer, making it linearly recoverable many layers before it becomes self-decodable, and then diverge into one of four resolution outcomes: Resolved, Overprocessed, Misresolved, or Unresolved. Understanding this lifecycle matters because similar task accuracies can mask fundamentally different failure modes that surface-level evaluation cannot detect. We introduce a dual diagnostic framework pairing layer-wise linear probing with Context-Stripped Decoding (CSD) and apply it to six code-reasoning task families across 16 models spanning Qwen, Llama, and DeepSeek architectures. All four outcomes carry substantial mass in every task family: overall Resolved is only 41.5%, with multiple tasks below 30%. Controlled sweeps over structure, depth, and operators expose task-specific failure bottlenecks: Function Call Resolved plunges from 61.1% to 2.5% as call depth increases from one to three. Across architectures and scales, the brewing scaffold remains stable, with normalized brewing duration 24-42% across all 16 models, while resolution success varies with capability. This indicates that the scaffold is a stable empirical regularity across the tested decoder-only Transformer families, whereas resolution success covaries with capability, scale, and training. Code: https://github.com/euyis1019/llm-brewing

Changwei Tu, Cheng Yang, Xianzhao Feng et al.

Blind recognition of polar-coded transmissions is an important task in non-cooperative wireless forensics and security-oriented signal analysis. When the code length is known or has been estimated, recovering the frozen/information bit-position pattern is a key step in identifying the underlying polar-code structure and enabling subsequent information recovery from intercepted observations. In this paper, blind recognition of polar codes is investigated from a hypothesis-testing perspective under the successive cancellation (SC)-based synthetic bit-channel representation. First, under an ideal SC-consistent condition, we formulate position-wise recognition as a binary hypothesis test between frozen-position and information-position models, which provides a theoretical benchmark for analyzing their intrinsic distinguishability. Second, we show that the adopted soft recognition metric admits an exact shifted log-likelihood-ratio interpretation. This justifies ln 2 as the neutral threshold under equal priors and costs, while unequal priors or costs lead to the corresponding Bayesian threshold shift. Third, under the ideal SC-consistent model and this neutral setting, we derive upper and lower bounds on the position-wise and sequence-level recognition error probabilities with multiple independent observations. The resulting overlap coefficient is further related to the classical Bhattacharyya parameter, establishing an interpretable link between blind-recognition difficulty and polar synthetic-channel reliability. Simulation results show that the derived bounds characterize the recognition performance under the ideal SC-consistent model and capture the effects of code length, the number of intercepted observations, and SNR. Further paired comparisons in the tested settings indicate that the SC-consistent recursion provides a good sequence-level match to the realistic SC-recursive procedure.