Zirui Ding, Xinxin Fan, Hao Yuan, Shaoye Luo, Men Niu, Quanliang Jing, Yunfeng Lu
Large language models (LLMs) are usually evaluated by final-output correctness or safety, yet an endpoint label does not reveal how a failure develops throughout the reasoning process. We therefore ask whether ordinary reasoning failures and adversarially induced failures can be characterized using the same internal coordinates. Towards this end, we in this paper study a white-box diagnostic framework CauTrace to disclose the root-cause of why LLM fails from three fundamental inspections: Functional States, Propagation of Representation, Temporal-State Coordination and Reasoning Trajectory. First, seven functional states are creatively defined to represent the support and constraint processing in LLMs. Second, the propagation of representation change measures how the changes among the seven states persist across subsequent Transformer blocks. Third, temporal-state coordination and trajectory topology characterize when the commitment and verification occur relative to the support, and how the dominant state sequence evolves over the full inference process. Finally, we evaluate CauTrace on both clean reasoning failures and paired adversarial perturbations. Across Qwen and LLaMA-family models, the failed traces show weaker Attention-MLP agreement on support-state updates and less verification activity relative to commitment, while the exact representation-level signatures and timing shifts vary by architecture. Held-out repair and steering tests demonstrate consistent operational signals that were in line with our claims for certain models and specific settings, with aligned steering producing the predicted directional changes in process measures. Overall, CauTrace provides a unified diagnostic framework for analyzing failure dynamics through changes in support, commitment, and verification, which would shed light on the in-depth research on LLM security.
Yang Li, Yike Wang, Di Wan, Chujun He, Zongsi Xu, Yitao Yu, Shuai Zhong, Minghao Pan 等 23 位作者
Simulating economic responses at society scale requires decision rules that respond to household circumstances and accounting rules that track their consequences. We introduce S$^2$-MAS, a multi-agent simulation framework with a resource of 500 million synthetic household personas. It combines trainable candidate-scoring policies, reusable financial scenarios and explicit state transitions. Trained with SFT + RLCD on 2.93 million examples, YN-mini-MLX achieves 74.55% accuracy and 38.37% question-macro F1 on held-out consumer financial-choice tasks, improving on the training-question majority baseline by 2.78 and 6.21 percentage points. A completed application evaluates 12 million distinct households in two purchase-funding conditions: 24 million one-shot choices feed 288 million monthly ledger states in 3 h 59 min on 14 GPUs. Jointly increasing purchase goals and approved limits produces heterogeneous consumption, cash and repayment paths, with much of the consumption variation arising from transitions between deferral and purchase. In a separate 1,060-household retrospective consumption experiment, standalone YN-master-9B has 3.42% lower annual MAE than the deterministic fully rational benchmark, alongside a 19.7% lower predicted population mean than observed. Information-timing tests show how year labels and later outcome information change decisions. Together, these experiments establish an executable alternative for studying heterogeneous household behavior with learned policies and explicit financial accounting.
Speech interaction systems have gradually evolved from traditional cascaded systems that use text as intermediate information to end-to-end speech language models that directly process speech representations. In some applications, the speech frontend runs on the user's device and sends speech representations to a remote server for inference, while the original audio remains local. However, these representations may still contain speaker identity information. If an attacker obtains them, it may use this information to imitate the user's voice, threatening their voice privacy. To investigate this risk, we propose \stia, a Speech Token Inversion Attack that generates speech with attacker-specified text in a victim's voice using a single short leaked representation, without access to the original audio or its transcript. Specifically, \stia converts discrete, continuous, or mixed speech representations into conditioning sequences and trains a conditional flow-matching speech generation model using pairs of different utterances from the same speaker in public speech datasets. The model combines speaker characteristics contained in the leaked representation with the specified text to generate speech in the victim's voice. Experiments show that, with a Moshi representation leak corresponding to 3 seconds of speech, \stia achieves an average CAM++ speaker cosine similarity of 0.78 and a target-text word error rate of 3.80\%. These results show that speech frontend representations reveal reusable speaker identity information and highlight the need to consider privacy risks when designing speech frontend representation methods.