PRISM is a new activation-conditioned model that recovers full sets of simultaneous instructions from LLM hidden states via judge-guided GRPO training and outperforms prior activation-to-language methods on security-relevant tasks.
arXiv preprint arXiv:2401.06102 , year=
12 Pith papers cite this work, alongside 2 external citations. Polarity classification is still indexing.
12
Pith papers citing it
2
external citations · external index
citation-role summary
dataset 1
citation-polarity summary
roles
dataset 1polarities
use dataset 1representative citing papers
A new framework shows concept subspaces are not unique, estimator choice affects containment and disentanglement, LEACE works well but generalizes poorly, and HuBERT encodes phone info as contained and disentangled from speaker info while speaker info resists compact containment.
Model organism interpretability depends strongly on training methodology, with integrated training yielding less interpretable MOs than post-hoc SFT or DPO.
Attribute retrieval in LLMs follows non-contiguous, redundant layer paths identified via iterative patching, implying highly distributed knowledge storage.
Deep transformer layers can replace context-dependent value vectors with per-token lookup tables (Bank of Values), improving validation loss and the 21-benchmark average at 135M–780M while cutting FLOPs and the value cache.
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.
LLMs solve compositional factual recall either by computing intermediates or directly, with mechanism choice correlated to translation geometry in embedding spaces.
Introduces effective dimension d_ρ from spectral analysis of reasoning trajectories to distinguish task hardness (0.93 AUC on MATH500) and uses kinematic features for early correctness prediction from partial generations.
A five-stage causal feature analysis methodology is proposed and tested on GPT-2 for IOI, showing partial causality of SAE features, robustness differences under shifts, and deployment cost benefits.
Activation verbalization methods for LLMs largely reflect the verbalizer model's parametric knowledge rather than privileged information from the target model's activations.
A training-free region-aware attention recalibration strategy reduces object hallucinations in LVLMs on CHAIR, POPE, and MME benchmarks while preserving fluency.
ReasoningLens is a framework for hierarchical visualization, agentic auditing, and systemic profiling of long reasoning chains in large reasoning models.
citing papers explorer
-
PRISM: Recovering Instruction Sets from Language Model Activations
PRISM is a new activation-conditioned model that recovers full sets of simultaneous instructions from LLM hidden states via judge-guided GRPO training and outperforms prior activation-to-language methods on security-relevant tasks.
-
A framework for analyzing concept representations in neural models
A new framework shows concept subspaces are not unique, estimator choice affects containment and disentanglement, LEACE works well but generalizes poorly, and HuBERT encodes phone info as contained and disentangled from speaker info while speaker info resists compact containment.
-
The Model Organism Lottery: Model Organism Interpretability Strongly Depends on Training Methodology
Model organism interpretability depends strongly on training methodology, with integrated training yielding less interpretable MOs than post-hoc SFT or DPO.
-
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.
-
Do Value Vectors in Deep Layers Need Context from the Residual Stream?
Deep transformer layers can replace context-dependent value vectors with per-token lookup tables (Bank of Values), improving validation loss and the 21-benchmark average at 135M–780M while cutting FLOPs and the value cache.
-
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.
-
How Do Language Models Compose Functions?
LLMs solve compositional factual recall either by computing intermediates or directly, with mechanism choice correlated to translation geometry in embedding spaces.
-
Geometric Signatures of Reasoning: A Spectral Perspective on Task Hardness
Introduces effective dimension d_ρ from spectral analysis of reasoning trajectories to distinguish task hardness (0.93 AUC on MATH500) and uses kinematic features for early correctness prediction from partial generations.
-
From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models
A five-stage causal feature analysis methodology is proposed and tested on GPT-2 for IOI, showing partial causality of SAE features, robustness differences under shifts, and deployment cost benefits.
-
Do Activation Verbalization Methods Convey Privileged Information?
Activation verbalization methods for LLMs largely reflect the verbalizer model's parametric knowledge rather than privileged information from the target model's activations.
-
Mitigating Object Hallucinations in Vision-Language Models through Region-Aware Attention Recalibration
A training-free region-aware attention recalibration strategy reduces object hallucinations in LVLMs on CHAIR, POPE, and MME benchmarks while preserving fluency.
-
ReasoningLens: Hierarchical Visualization and Diagnostic Auditing for Large Reasoning Models
ReasoningLens is a framework for hierarchical visualization, agentic auditing, and systemic profiling of long reasoning chains in large reasoning models.