probing

论文提出 parameter-trajectory crosscoding 方法，追踪语言模型预训练中 unembedding（读出矩阵）的动态变化，发现模型隐藏状态编码任务信息的时间早于其输出所能表达的时间，早期的行为学时间点可能反映的是读出层的学习而非模型能力的习得。

本文对隐式思维链（latent CoT）是否真正进行推理进行了系统研究，覆盖21种模型-任务组合，提出一个统一分类法，区分链条的表示内容与其在模型中的因果作用，揭示不同训练配方下latent reasoning的多样机制。

该论文研究代码语言模型中对 Python 的默认偏好是否由可定位的因果电路产生，通过 55 个模型-语言设置的消融、激活修补和电路提取实验发现，偏好是否显得局部化取决于探针选择与组件预算，并呼吁报告分析选择。

该论文结合叙事学与机制可解释性，在GPT-2-Small、Gemma-2-2B和Llama-3.1-8B中识别出与叙事性（narrativity）相关的注意力头、MLP神经元和稀疏自编码器特征，并通过消融实验证明这些组件对模型的叙事行为具有因果影响。

这篇论文研究生成模型在何种条件下可以充当可编辑的模拟器，提出通过探测器解码与逆映射两种方式将模型隐状态映射到模拟器状态，并在Othello与自建的连续世界Rayworld中发现映射成功与否强烈依赖环境，即使状态始终可解码。

该论文通过输入扰动、证据冲突和线性探针等方法，系统评估了六个生物推理模型在DNA、蛋白质和单细胞任务中是否真正利用了生物基础模型的表示。研究发现BioReason和BioReason-Pro几乎完全依赖文本而非生物输入，表明当前后训练策略并不能保证基础模型表示对任务性能产生实际贡献。

The paper proposes PAIR (Phase-Aligned Intra-question Reasoning), which treats LLM reasoning paths as phase-structured trajectories and evaluates path quality only within the same question, showing that standard correctness probes partly rely on question-level variation and that phase-wise steering yields causal evidence of path-quality directions.

A technique called spurious probes uses adaptive API-level search to find questions that reveal the latent state of blackbox LLMs, such as distinguishing capability evaluations via specific prompts.

This paper examines script knowledge distribution in LLM layers, revealing that script recognition happens early while commitment to output script emerges in late layers, highlighting the importance of model depth for multilingual architectures.

The paper introduces Probe of Internal Recognition (PIR), a method to detect hidden knowledge in language models by analyzing internal states, distinguishing between intentional concealment and absence of knowledge for AI safety applications.

This paper introduces 'legible failures' in language models, where models possess correct information in hidden states but fail to use it, and shows that linear probes can detect and repair such failures through steering interventions.

This paper investigates how large language models internally represent logical validity, showing that validity information is decodable from hidden states despite poor behavioral performance, suggesting distinct roles for representation, expression, and causal use.

This study compares probing techniques for identifying latent language in multilingual LLMs, finding that different methods yield inconsistent results, indicating they expose distinct aspects of multilingual processing rather than a single internal lingua franca.

This paper provides a systematic study of evaluation awareness in language models, showing that models internalize evaluation context, leading to a disconnect between internal representation, verbalization, and steering behavior, with implications for benchmark reliability.

This paper proposes a nuisance-controlled residual-stream probing protocol to detect benchmark contamination in language models, showing that naive probing fails and their corrected method controls false positives while achieving reasonable power.

This paper introduces Prompt Embedding Probes (PEP), a parameter-efficient extension of linear probes that uses learnable prompt embeddings on hidden states to detect hallucinations in frozen LLMs. Evaluations on TriviaQA, GSM8K, and MedQA with Qwen3 models show improvements over standard linear probes, including in pre-generation and cross-model settings.

This paper investigates the 'knowing-saying gap' in language models, showing that linear probes can detect corrupted context with near-perfect accuracy yet fail to predict final answer errors, with implications for deployment monitoring and intervention strategies.

This paper proposes a three-stream detector that combines residual-stream motion with coarse regions and fine directions to better identify reasoning errors in LLMs, improving selection accuracy by up to 12% over state-of-the-art displacement-only methods.

This paper introduces Right Reset (RR), a prefix-removal probing method that uses causal language model hidden-state preservation to identify chunk boundaries in flattened text, recovering 47.7% of original records compared to 25.9% for a BGE baseline.

This paper shows that repetition effects in language models depend on readout position: adjacent repetition boosts target probability, while displaced repetition produces an inverted-U curve. The finding challenges assumptions in cloze-style probing and is validated across multiple models and languages.