Linear representations of high-level concepts in LLMs are formalized via counterfactuals in input and output spaces, unified under a causal inner product that enables consistent probing and steering.
hub
Language models represent space and time
27 Pith papers cite this work. Polarity classification is still indexing.
27
Pith papers citing it
hub tools
citation-role summary
background 4
citation-polarity summary
roles
background 4representative citing papers
Turn-averaged SAEs reconstruct average activations over conversation turns to represent high-level turn characteristics with a fixed number of features, simplifying long-context interpretability compared to per-token SAEs.
Evaluation of two latent reasoning models against controls shows observable latent patterns appear without the proposed mechanisms, have graded causal effects on behavior, and concentrate in structured low-rank directions, arguing that patterns are insufficient evidence for reasoning.
Bridge augments a graph neural network backbone with time-aware retrieval from a memory of region-time windows to improve cold-start and cross-city urban delivery demand forecasting.
Transformer activations show spectral anti-concentration for concepts in the tail while syntax prefers high-variance directions, forming a dual geometry.
Large language models encode relational bindings via a cell-based representation: a low-dimensional linear subspace in which each cell corresponds to an entity-relation index pair and attributes are retrieved from the matching cell.
Subword tokenization impairs phonological knowledge encoding in LMs, but an IPA-based fine-tuning method restores it with minimal impact on other capabilities.
Attribute retrieval in LLMs follows non-contiguous, redundant layer paths identified via iterative patching, implying highly distributed knowledge storage.
LLM representations encode essay quality in a linearly decodable form that emerges across layers and includes identifiable scoring neurons whose distribution shifts with essay length.
SLM adds a dedicated spatial modality and training dataset to LLMs, enabling geometric spatial reasoning and outperforming prompt-based symbolic methods on the new SpatialEval benchmark.
Supervised fine-tuning lets LLMs linearly encode action validity and state predicates, with broader state-space coverage during training improving world-model recovery.
Experiments reveal that topological cues robustly support LLM navigation planning while incorrect semantic cues derail it, with linguistic format effects varying by model size and compression.
Sparse autoencoders scaled to 34 million features on Claude 3 Sonnet yield interpretable, steerable representations of concrete and abstract concepts that generalize across languages and modalities.
Introduces 9 synthetic annotation tasks and benchmarks for behavioral cloning, finding hierarchical skill learning, scaling benefits, effective multi-task pretraining, and shared internal representations of task phases and mistakes.
PerturbedVAE disentangles perturbation-specific signals from invariant gene expression structure to recover causal representations and improve out-of-distribution prediction in single-cell perturbation modeling.
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
Fluent AI users adopt an active, iterative collaboration mode that produces more visible failures but better recovery and success on hard tasks, whereas novices experience more invisible failures from passive use.
The Master Key Hypothesis states that capabilities are low-dimensional directions transferable across models through linear subspace alignment, with UNLOCK demonstrating gains such as 12.1% accuracy improvement on MATH when transferring CoT from 14B to 7B models.
Reasoning in LLMs produces a transient geometric pulse in which concept manifolds untangle into linearly separable subspaces immediately before computation and compress afterward.
Muon learns more robust and transferable features than Adam and SGD, shown via corruption robustness tests, transfer experiments, layer-wise probes, effective rank measurements, and a theoretical proof on margins in a multi-component classification problem.
DCO is an inference-time intervention that decomposes attention head outputs orthogonally to a dynamic context anchor and suppresses outlier components via Z-score to improve contextual faithfulness in Llama models.
A survey proposing a three-pillar framework to evaluate LLMs as tools for measuring latent psychological constructs and reviewing applications in personality and mental health.
H-probes locate low-dimensional subspaces encoding hierarchy in LLM activations for synthetic tree tasks, show causal importance and generalization, and detect weaker signals in mathematical reasoning traces.
Analysis estimates 18.7% of Common Crawl documents contain geospatial information like coordinates and addresses, with little difference by language.
citing papers explorer
-
Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution
Turn-averaged SAEs reconstruct average activations over conversation turns to represent high-level turn characteristics with a fixed number of features, simplifying long-context interpretability compared to per-token SAEs.
-
Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
Evaluation of two latent reasoning models against controls shows observable latent patterns appear without the proposed mechanisms, have graded causal effects on behavior, and concentrate in structured low-rank directions, arguing that patterns are insufficient evidence for reasoning.
-
Bridge: Retrieval-Augmented Spatiotemporal Modeling for Urban Delivery Demand
Bridge augments a graph neural network backbone with time-aware retrieval from a memory of region-time windows to improve cold-start and cross-city urban delivery demand forecasting.
-
Concepts Whisper While Syntax Shouts: Spectral Anti-Concentration and the Dual Geometry of Transformer Representations
Transformer activations show spectral anti-concentration for concepts in the tail while syntax prefers high-variance directions, forming a dual geometry.
-
Cell-Based Representation of Relational Binding in Language Models
Large language models encode relational bindings via a cell-based representation: a low-dimensional linear subspace in which each cell corresponds to an entity-relation index pair and attributes are retrieved from the matching cell.
-
How Tokenization Limits Phonological Knowledge Representation in Language Models and How to Improve Them
Subword tokenization impairs phonological knowledge encoding in LMs, but an IPA-based fine-tuning method restores it with minimal impact on other capabilities.
-
Factual Retrieval in LLMs Is a Redundant, Distributed and Non-Contiguous Process
Attribute retrieval in LLMs follows non-contiguous, redundant layer paths identified via iterative patching, implying highly distributed knowledge storage.
-
From Texts to Scores: Tracing the Emergence of Essay Quality Representations in Large Language Models
LLM representations encode essay quality in a linearly decodable form that emerges across layers and includes identifiable scoring neurons whose distribution shifts with essay length.
-
From Symbolic to Geometric: Enabling Spatial Reasoning in Large Language Models
SLM adds a dedicated spatial modality and training dataset to LLMs, enabling geometric spatial reasoning and outperforming prompt-based symbolic methods on the new SpatialEval benchmark.
-
A Close Look At World Model Recovery In Supervised Fine-Tuned LLM Planners
Supervised fine-tuning lets LLMs linearly encode action validity and state predicates, with broader state-space coverage during training improving world-model recovery.
-
The Sword, Shield, and Achilles' Heel: Characterizing the Linguistic Inductive Bias of Large Language Models for Spatial Reasoning in Navigation Planning
Experiments reveal that topological cues robustly support LLM navigation planning while incorrect semantic cues derail it, with linguistic format effects varying by model size and compression.
-
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
Sparse autoencoders scaled to 34 million features on Claude 3 Sonnet yield interpretable, steerable representations of concrete and abstract concepts that generalize across languages and modalities.
-
A Systematic Study of Behavioral Cloning for Scientific Data Annotation
Introduces 9 synthetic annotation tasks and benchmarks for behavioral cloning, finding hierarchical skill learning, scaling benefits, effective multi-task pretraining, and shared internal representations of task phases and mistakes.
-
What Makes a Representation Good for Single-Cell Perturbation Prediction?
PerturbedVAE disentangles perturbation-specific signals from invariant gene expression structure to recover causal representations and improve out-of-distribution prediction in single-cell perturbation modeling.
-
Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
-
A paradox of AI fluency
Fluent AI users adopt an active, iterative collaboration mode that produces more visible failures but better recovery and success on hard tasks, whereas novices experience more invisible failures from passive use.
-
The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment
The Master Key Hypothesis states that capabilities are low-dimensional directions transferable across models through linear subspace alignment, with UNLOCK demonstrating gains such as 12.1% accuracy improvement on MATH when transferring CoT from 14B to 7B models.
-
Emergent Manifold Separability during Reasoning in Large Language Models
Reasoning in LLMs produces a transient geometric pulse in which concept manifolds untangle into linearly separable subspaces immediately before computation and compress afterward.
-
Muon Learns More Robust and Transferable Features than Adam
Muon learns more robust and transferable features than Adam and SGD, shown via corruption robustness tests, transfer experiments, layer-wise probes, effective rank measurements, and a theoretical proof on margins in a multi-component classification problem.
-
Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization
DCO is an inference-time intervention that decomposes attention head outputs orthogonally to a dynamic context anchor and suppresses outlier components via Z-score to improve contextual faithfulness in Llama models.
-
A Survey of Large Language Models for Perception and Measurement of Human Psychology
A survey proposing a three-pillar framework to evaluate LLMs as tools for measuring latent psychological constructs and reviewing applications in personality and mental health.
-
H-Probes: Extracting Hierarchical Structures From Latent Representations of Language Models
H-probes locate low-dimensional subspaces encoding hierarchy in LLM activations for synthetic tree tasks, show causal importance and generalization, and detect weaker signals in mathematical reasoning traces.
-
Quantifying Geospatial in the Common Crawl Corpus
Analysis estimates 18.7% of Common Crawl documents contain geospatial information like coordinates and addresses, with little difference by language.
-
LaGO: Latent Action Guidance for Online Reinforcement Learning
LaGO improves online RL success rates over vanilla PPO by using pretrained LLMs as latent action priors, raising rates from 15.1% to 27.2% on CLEVR-Robot and 2.7% to 15.2% on Meta-World.
-
Integrating Large Language Model Agents with Digital Twins for Industrial Autonomous Systems
A TPSR-based framework with four LLM roles integrates language model reasoning into industrial automation via digital twins, achieving high task executability in case studies.
-
Geometry of Human Perceptual Domains Emerges Transiently in LLM Representations
Perceptual geometry for color, pitch, emotion and taste emerges transiently in intermediate layers of transformer LLMs despite purely textual training.