Temporal knowledge drift is encoded as a geometrically orthogonal direction in LLM residual streams, independent of correctness and uncertainty.
Closing the confidence-faithfulness gap in large language models
6 Pith papers cite this work. Polarity classification is still indexing.
6
Pith papers citing it
citation-role summary
other 1
citation-polarity summary
years
2026 6roles
other 1polarities
unclear 1representative citing papers
LLMs exhibit Bayesian-like hypothesis updating with strong-sampling bias and an evaluation-generation gap but generalize poorly outside observed data.
RLCM trains LLMs with a margin-enhanced process reward that widens the gap between correct and incorrect reasoning steps, improving calibration on math, code, logic, and science tasks without hurting accuracy.
Seven 3-9B instruction-tuned LLMs produce verbal confidence that saturates at high values and fails psychometric validity criteria for Type-2 discrimination under minimal elicitation.
An NSM-based explication parser with fixed semantic rules produces emotion labels for events, achieving 0.33 accuracy on held-out crowd-sourced data while shifting empirical risk to an inspectable parser.
Fine-tuning Gemma 3 4B on unfiltered self-consistency targets produces a binary verbal correctness discriminator with AUROC 0.774 on TriviaQA, outperforming logit entropy after a modal-filtered pre-registration failed.
citing papers explorer
-
The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations
Temporal knowledge drift is encoded as a geometrically orthogonal direction in LLM residual streams, independent of correctness and uncertainty.
-
Hypothesis generation and updating in large language models
LLMs exhibit Bayesian-like hypothesis updating with strong-sampling bias and an evaluation-generation gap but generalize poorly outside observed data.
-
Process Supervision of Confidence Margin for Calibrated LLM Reasoning
RLCM trains LLMs with a margin-enhanced process reward that widens the gap between correct and incorrect reasoning steps, improving calibration on math, code, logic, and science tasks without hurting accuracy.
-
Verbal Confidence Saturation in 3-9B Open-Weight Instruction-Tuned LLMs: A Pre-Registered Psychometric Validity Screen
Seven 3-9B instruction-tuned LLMs produce verbal confidence that saturates at high values and fails psychometric validity criteria for Type-2 discrimination under minimal elicitation.
-
Faithful by Definition: Emotion Analysis via Natural Semantic Metalanguage Explications
An NSM-based explication parser with fixed semantic rules produces emotion labels for events, achieving 0.33 accuracy on held-out crowd-sourced data while shifting empirical risk to an inspectable parser.
-
Distilling Self-Consistency into Verbal Confidence: A Pre-Registered Negative Result and Post-Hoc Rescue on Gemma 3 4B
Fine-tuning Gemma 3 4B on unfiltered self-consistency targets produces a binary verbal correctness discriminator with AUROC 0.774 on TriviaQA, outperforming logit entropy after a modal-filtered pre-registration failed.