Tag
论文提出 CogDrills-4D（CogDrift-4D），一种层级式 4D Transformer 多模态框架，通过 18 个月内七次非脚本化交互的行为数据建模时间性行为衰退，用于轻度认知障碍（MCI）早期检测，分类准确率达 0.733±0.13。
PCCG-2 presents an EOS gate study where a frozen Qwen3-4B answer computation is controlled by a learned 101-parameter gate that only alters the native termination score, alongside cross-vendor evidence of zero-byte 'Void' responses across 11 frontier models from OpenAI, Anthropic, Google, and Moonshot.
The paper introduces SAFE, a controlled benchmark to study whether AI models acquire safety-relevant evidence before acting, revealing distinct acquisition policies across frontier models like GPT-5.5, o3, and Claude variants.
This paper investigates the dissociation between attention-level proxies and behavioral performance in large language models under fine-tuning, revealing that attention sensitivity alone is unreliable for diagnosing in-context learning capabilities.
This paper investigates why large language models answer structurally unanswerable questions, concluding that models recognize impossibility but fail to route that recognition to refusal, highlighting a routing failure rather than an encoding failure in their decision-making.
TutorTrace introduces a dataset and taxonomy for classifying learner behavioral states in AI-assisted programming education, enabling adaptive tutoring systems through IDE telemetry and behavioral context.
This paper introduces Behavioral Lift to measure how reasoning behaviors in AI thinking models correlate with correct answers, revealing an amplification-lift gap where training amplifies behaviors not most predictive of success.
A behavioral audit of Nvidia's Nemotron 3.5 Lightning finds strong integrity, catching a subtle auth bug in reviewed code and passing safety probes, while noting a blind spot in action-based honesty.
This paper introduces Base Sequence Analysis, a framework that encodes LLM agent runtime behavior into compact sequences, revealing high-risk patterns like the 'P-X-P' trigram and a verification deficit. It presents Governor, a runtime intervention system that improves task success by 6.2% and reduces token consumption by 44%.
Researchers evaluate 28 LLMs on the St. Petersburg game to distinguish between outcome-level resemblance and mechanism-level alignment in risk decision-making, finding that LLMs often produce human-like bids without underlying human-consistent reasoning mechanisms. The study demonstrates that behavioral alignment can be superficial, urging high-stakes evaluations to go beyond outcome similarity.