MoRFI: Monotonic Sparse Autoencoder Feature Identification

• URL: http://arxiv.org/abs/2604.26866v1
• Date: Wed, 29 Apr 2026 16:32:57 GMT
• Title: MoRFI: Monotonic Sparse Autoencoder Feature Identification
• Abstract summary: Large language models (LLMs) acquire most of their factual knowledge during the pre-training stage, through next token prediction.<n> Subsequent stages of post-training often introduce new facts outwith the parametric knowledge, giving rise to hallucinations.<n>We conduct a controlled fine-tuning experiment, focusing on closed-book QA, and find latent directions that causally contribute to hallucinations.
• Score: 27.973959474864127
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models (LLMs) acquire most of their factual knowledge during the pre-training stage, through next token prediction. Subsequent stages of post-training often introduce new facts outwith the parametric knowledge, giving rise to hallucinations. While it has been demonstrated that supervised fine-tuning (SFT) on new knowledge may exacerbate the problem, the underlying mechanisms are still poorly understood. We conduct a controlled fine-tuning experiment, focusing on closed-book QA, and find latent directions that causally contribute to hallucinations. Specifically, we fine-tune Llama 3.1 8B, Gemma 2 9B and Mistral 7B v03 on seven distinct single QA datasets, controlling for the percentage of new knowledge and number of training epochs. By measuring performance on the test set, we validate that incrementally introducing new knowledge increases hallucinations, with the effect being more pronounced with prolonged training. We leverage pre-trained sparse autoencoders (SAEs) to analyze residual stream activations across various checkpoints for each model and propose Monotonic Relationship Feature Identification (MoRFI) for capturing causally relevant latents. MoRFI filters SAE features that respond monotonically to controlled fine-tuning data mixtures of a target property. Our findings show that exposure to unknown facts disrupts the model's ability to retrieve stored knowledge along a set of directions in the residual stream. Our pipeline reliably discovers them across distinct models, recovering knowledge through single-latent interventions.

Related papers

• Knowledgeless Language Models: Suppressing Parametric Recall for Evidence-Grounded Language Modeling [21.14539989233461]
  We introduce Knowledge--''Less'' Language Models (KLLMs)<n> KLLMs are pretrained on corpora in which named entities are anonymized.<n>We show that KLLMs rely less on parametric knowledge and more on external evidence, leading to improved reliability under realistic conditions.
  arXiv  Detail & Related papers  (2026-07-14T14:48:53Z)
• TRE: Training-Free Hallucination Detection for Diffusion Language Models [9.622962795441852]
  We propose TRE, a training-free hallucination detection metric for D-LLMs.<n> TRE estimates hallucination risk directly from the entropy signals of a single generation, without requiring any detector training or repeated sampling.<n>Extensive experiments on multiple D-LLMs and QA datasets demonstrate that TRE achieves competitive performance, while enjoying strong generalizability, efficiency, and robustness.
  arXiv  Detail & Related papers  (2026-06-29T00:47:48Z)
• Measuring the Impact of Lexical Training Data Coverage on Hallucination Detection in Large Language Models [26.89705770151822]
  Hallucination in large language models (LLMs) is a fundamental challenge, particularly in open-domain question answering.<n>Prior work attempts to detect hallucination with model-internal signals such as token-level entropy or generation consistency.<n>We investigate whether data coverage itself can serve as a detection signal.
  arXiv  Detail & Related papers  (2025-11-22T06:59:55Z)
• Detecting Token-Level Hallucinations Using Variance Signals: A Reference-Free Approach [0.0]
  Large Language Models (LLMs) have demonstrated impressive generative capabilities across diverse tasks but remain susceptible to hallucinations.<n>We introduce a reference-free, token-level hallucination detection framework that leverages the variance in token log-probabilities across multiple generations.<n>Our approach is model-agnostic, interpretable, and suited for real-time or post-hoc analysis.
  arXiv  Detail & Related papers  (2025-07-05T19:20:59Z)
• Are LLMs Really Not Knowledgable? Mining the Submerged Knowledge in LLMs' Memory [15.986679553468989]
  Large language models (LLMs) have shown promise as potential knowledge bases.<n>LLMs often struggle with question-answering tasks and are prone to hallucinations.<n>We develop SkipUnsure, a method to improve answer accuracy by leveraging detected but unexpressed knowledge.
  arXiv  Detail & Related papers  (2024-12-30T10:29:18Z)
• Hallucination Detox: Sensitivity Dropout (SenD) for Large Language Model Training [17.347067827911406]
  We investigate the relationship between the uncertainty in training dynamics and the emergence of hallucinations.<n>We propose Sensitivity Dropout (SenD), a novel training protocol designed to reduce hallucination variance during training.<n>SenD improves test-time reliability of Pythia and Meta's Llama models by up to 17% and enhances factual accuracy in Wikipedia, Medical, Legal, and Coding domains.
  arXiv  Detail & Related papers  (2024-10-20T18:18:23Z)
• Exploring Diffusion Time-steps for Unsupervised Representation Learning [72.43246871893936]
  We build a theoretical framework that connects the diffusion time-steps and the hidden attributes. On CelebA, FFHQ, and Bedroom datasets, the learned feature significantly improves classification.
  arXiv  Detail & Related papers  (2024-01-21T08:35:25Z)
• Mitigating Shortcut Learning with Diffusion Counterfactuals and Diverse Ensembles [104.60508550106618]
  We propose DiffDiv, an ensemble diversification framework exploiting Diffusion Probabilistic Models (DPMs)<n>We show that DPMs can generate images with novel feature combinations, even when trained on samples displaying correlated input features.<n>We show that DPM-guided diversification is sufficient to remove dependence on shortcut cues, without a need for additional supervised signals.
  arXiv  Detail & Related papers  (2023-11-23T15:47:33Z)
• R-Tuning: Instructing Large Language Models to Say `I Don't Know' [66.11375475253007]
  Large language models (LLMs) have revolutionized numerous domains with their impressive performance but still face their challenges. Previous instruction tuning methods force the model to complete a sentence no matter whether the model knows the knowledge or not. We present a new approach called Refusal-Aware Instruction Tuning (R-Tuning) Experimental results demonstrate R-Tuning effectively improves a model's ability to answer known questions and refrain from answering unknown questions.
  arXiv  Detail & Related papers  (2023-11-16T08:45:44Z)
• Can LMs Learn New Entities from Descriptions? Challenges in Propagating Injected Knowledge [72.63368052592004]
  We study LMs' abilities to make inferences based on injected facts (or propagate those facts) We find that existing methods for updating knowledge show little propagation of injected knowledge. Yet, prepending entity definitions in an LM's context improves performance across all settings.
  arXiv  Detail & Related papers  (2023-05-02T17:59:46Z)
• Understanding Self-supervised Learning with Dual Deep Networks [74.92916579635336]
  We propose a novel framework to understand contrastive self-supervised learning (SSL) methods that employ dual pairs of deep ReLU networks. We prove that in each SGD update of SimCLR with various loss functions, the weights at each layer are updated by a emphcovariance operator. To further study what role the covariance operator plays and which features are learned in such a process, we model data generation and augmentation processes through a emphhierarchical latent tree model (HLTM)
  arXiv  Detail & Related papers  (2020-10-01T17:51:49Z)
• Automatic Recall Machines: Internal Replay, Continual Learning and the Brain [104.38824285741248]
  Replay in neural networks involves training on sequential data with memorized samples, which counteracts forgetting of previous behavior caused by non-stationarity. We present a method where these auxiliary samples are generated on the fly, given only the model that is being trained for the assessed objective. Instead the implicit memory of learned samples within the assessed model itself is exploited.
  arXiv  Detail & Related papers  (2020-06-22T15:07:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.