Adaptive scheduling of interventions in discrete diffusion language models, timed to attribute-specific commitment schedules discovered with sparse autoencoders, delivers precise multi-attribute steering up to 93% strength while preserving generation quality.
hub Mixed citations
Probing classifiers: Promises, shortcomings, and advances.Computational Linguistics, 48(1):207–219
Mixed citation behavior. Most common role is background (62%).
28
Pith papers citing it
130
external citations · Crossref
Background
62% of classified citations
abstract
Probing classifiers have emerged as one of the prominent methodologies for interpreting and analyzing deep neural network models of natural language processing. The basic idea is simple -- a classifier is trained to predict some linguistic property from a model's representations -- and has been used to examine a wide variety of models and properties. However, recent studies have demonstrated various methodological limitations of this approach. This article critically reviews the probing classifiers framework, highlighting their promises, shortcomings, and advances.
hub tools
citation-role summary
background 8
method 3
baseline 1
dataset 1
citation-polarity summary
representative citing papers
Factual associations in autoregressive transformers are localized to mid-layer feed-forward modules and can be edited via rank-one model editing while preserving both specificity and generalization on counterfactual tests.
QAOD projects away question-aligned directions from answer representations to isolate domain-agnostic factuality signals, enabling efficient hallucination detection with top in-domain AUROC and up to 21% better OOD transfer.
KamonBench is a grammar-based dataset of 20,000 synthetic Japanese crests with multi-format annotations that enables direct evaluation of factor recovery beyond caption accuracy in vision-language models.
Symmetry under affine reparameterizations of hidden coordinates selects a unique hierarchy of shallow coordinate-stable probes and a probe-visible quotient for cross-model transfer.
EEG foundation models encode 68.6% of a 63-feature clinical lexicon in a representation-causal way, with frequency-domain features dominant; these recover 79.3% of the models' advantage over random baselines on average.
Behavioral directions from one LLM family transfer to others via projection into a shared anchor coordinate space, yielding 0.83 ten-way detection accuracy and steering effects up to 0.46% on held-out models.
Tabular foundation models show substantial depthwise redundancy, so a looped single-layer version achieves comparable results with 20% of the original parameters.
Finite-answer projections of continuation probabilities stabilize before the answer is parseable, showing 17-31 token mean lead in delayed-verdict tasks with Qwen3-4B-Instruct.
Latent space probing on CogVideoX achieves 97.29% F1 for adult content detection on a new 11k-clip dataset with 4-6ms overhead.
Transformer represents but does not causally transmit staged algorithmic intermediates for base-digit extraction, diverging from probe predictions.
Sparse autoencoders on EEG transformers extract clinical features, identify three steering regimes, expose age-pathology entanglements and wrecking-ball failures, and map interventions to frequency spectra.
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 tabular foundation model with LLM-as-Observer features predicts AI agent decisions in controlled games, outperforming baselines by 4 AUC points and 14% lower error at K=16 interactions.
Instructions trigger a production-centered mechanism in language models, with task-specific information stable in input tokens but varying strongly in output tokens and correlating with behavior.
Chemically meaningful steering for properties like cLogP and TPSA emerges in entangled Transformer-VAE latent spaces only after controlling for SELFIES representation confounds through residualization and decoded traversals.
Conceptors as soft projection matrices from bipolar activations offer a multidimensional, compositional, and geometrically principled method for semantic steering in LLMs that outperforms single-vector baselines in multi-dimensional subspaces.
An encoding probe reconstructs transformer representations from acoustic, phonetic, syntactic, lexical and speaker features, showing independent syntactic/lexical contributions and training-dependent speaker effects.
LLMs encode accurate but brittle internal beliefs about latent game states and convert them poorly into actions, creating systematic gaps that explain strategic failures.
Architecture and training determine whether transformers retain a readable internal signal that lets activation monitors catch errors missed by output confidence.
Prophecy infers formal properties of feed-forward neural networks by extracting rules from neuron activation patterns that imply desirable output behaviors.
At sufficient scale, LLMs linearly represent the truth value of factual statements, as shown by visualizations, cross-dataset generalization, and causal interventions that flip truth judgments.
BLOOM is a 176B-parameter open-access multilingual language model trained on the ROOTS corpus that achieves competitive performance on benchmarks, with improved results after multitask prompted finetuning.
Overthinking in medical QA is linearly decodable at 71.6% accuracy yet fixed residual-stream steering yields no correction across 29 configurations, while enabling selective abstention with AUROC 0.610.
citing papers explorer
-
Why Do LLMs Struggle in Strategic Play? Broken Links Between Observations, Beliefs, and Actions
LLMs encode accurate but brittle internal beliefs about latent game states and convert them poorly into actions, creating systematic gaps that explain strategic failures.
-
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model
BLOOM is a 176B-parameter open-access multilingual language model trained on the ROOTS corpus that achieves competitive performance on benchmarks, with improved results after multitask prompted finetuning.
-
Decodable but Not Corrected by Fixed Residual-Stream Linear Steering: Evidence from Medical LLM Failure Regimes
Overthinking in medical QA is linearly decodable at 71.6% accuracy yet fixed residual-stream steering yields no correction across 29 configurations, while enabling selective abstention with AUROC 0.610.