Tag
This paper introduces an unsupervised, training-free approach to discover prompt-conditional stylistic axes in LLM hidden activations using sampling and PCA, validated through human studies.
This paper explores methods to interpret and steer large language model agents in social simulations, comparing prompt-based, SAE-based, and probe-based techniques, and finds that SAE and probe methods often outperform basic prompting for control and interpretability.
This research investigates whether multilingual large language models develop shared internal representations for mathematical reasoning across languages, introducing a novel Geometry-Invariant Sparse Autoencoder (GI-SAE) method. It finds that cross-language feature sharing is model-dependent and that geometric similarity does not consistently imply functional interchangeability.
This paper investigates how large language models implicitly personalize outputs based on demographic cues, locating an internal activation signal that tracks these shifts and showing that removing this signal can suppress the behavior.
This paper introduces Feature Nonlocality (FNL), an entropy-based metric for measuring the semantic abstractness of SAE features in LLMs, and demonstrates applications in auditing jailbreak mitigation features and improving MATH-500 accuracy via steering.
This paper asks whether lesion parameters in LLaVA-Vicuna 13B can be recovered from aphasic picture-naming error profiles. The authors find that perturbation intensity is recoverable while layer index is only approximate, with 81.4% counterfactual fidelity and syndrome-discriminative generalization to stroke survivors, suggesting functional redundancy across transformer layers.
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 applies graph signal processing to analyze how LLMs internally represent numerical sequences during in-context learning, finding that attention-induced token graphs and hidden-state signals show systematic, context-dependent signatures related to input complexity.
This paper identifies an 'inverted detection-control' phenomenon where some discriminative steering vectors, despite aligning with positive concept representations, consistently promote the opposite behavior. The authors propose a method to detect such inverted steering vectors without generation, enabling sign flips that improve a detection-based steering pipeline across multiple LLMs and concepts.
This paper investigates internal representational differences between RL and SFT fine-tuned models on mathematical reasoning, finding that RL models exhibit more linearly separable hidden states and hierarchical layer importance. Token allocation variability under repeated sampling suggests training pipeline dependence rather than RL vs SFT alone.
This paper studies behavioral detection of unfaithful chain-of-thought reasoning in LLMs, finding that answer correctness structures detection performance: on incorrect answers, where most unfaithfulness occurs, behavioral signals are at chance, while on correct answers they offer modest separation.
This article presents a Tarski-inspired diagonal argument showing that no linear probe on an LLM's embedding space can reliably detect truth, drawing parallels to Gödel's incompleteness and Turing's halting problem. It critiques the linear representation hypothesis for truth in language models.
This paper introduces a framework using Jungian cognitive functions for activation steering in LLMs, demonstrating effective monotonic control over eight functions and revealing structured geometric relationships in activation space.
This paper investigates how instruction-tuned Transformer models (LLaMA and Mistral) encode the discourse relations of causation and antithesis using interpretability techniques on next-token prediction.
This paper investigates how alignment tuning introduces cue-induced biases such as sycophancy in LLMs, finding that biases are installed by alignment rather than pretraining and can be decoded and steered via hidden state directions.
This paper investigates whether neural network probes that predict correctness of language model outputs actually capture objective correctness or the model's own self-judgement, using conflict cases where the two disagree. The authors find that transferable directions predominantly preserve self-judgement polarity, challenging the interpretation of correctness readouts.
This paper investigates how language models separate a character's belief from reality, finding that they use a shared value slot for attributed values and a router at the query position to select the frame (belief or reality) to read out. It identifies two routes for asserted and derived beliefs, and shows that the slot itself carries no belief-reality tag; the separation lies in dissociated routing subspaces.
Anthropic discovered a hidden internal space (J-space) in LLMs like Claude that contains words influencing reasoning, advancing understanding of AI model internals.
A digest covering Anthropic's discovery of a hidden 'J-space' inside Claude's LLM, OpenAI's launch of the ChatGPT Work super app with GPT 5.6 models, humanoid robot surgery on live pigs, SK Hynix's record US listing, and Tencent's deal to acquire Manus.
This paper introduces Mining via Activation Geometry (MAG), an unsupervised framework that extracts reasoning features from LLM activations using natural-language instructions, enabling activation steering and effective training data selection for classifier probes.