Perplexity Can Miss SAE Feature Damage Under Quantization

• URL: http://arxiv.org/abs/2606.03002v2
• Date: Thu, 04 Jun 2026 19:39:22 GMT
• Title: Perplexity Can Miss SAE Feature Damage Under Quantization
• Abstract summary: Quantized models are judged acceptable when perplexity or downstream accuracy remains close to the full-precision original.<n>We test this by using a frozen sparse-autoencoder (SAE) as a fixed measurement basis.
• Score: 0.16921396880325779
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Quantization is a standard path to deploying large language models, and quantized models are typically judged acceptable when perplexity or downstream accuracy remains close to the full-precision original. But behavioral parity need not imply feature fidelity: the sparse-autoencoder (SAE) features used to interpret a full-precision model may change after weight rounding. We test this directly by using a frozen SAE as a fixed measurement basis, encoding full-precision and round-to-nearest (RTN) quantized activations on identical tokens, and measuring per-feature survival by Pearson correlation across bit-widths from INT8 to INT4 on Pythia-70M and Gemma-2-2B. Our central finding is that perplexity can miss feature damage: on Gemma-2-2B, INT7 improves perplexity while degrading 18.7% of active SAE features, and under sliding-window evaluation INT6 also improves perplexity while only 51.3% of active features survive. Feature survival is graded rather than cliff-like, with 62.4% of active Pythia features and 51.3% of active Gemma features surviving at INT6; most non-surviving features are blurred rather than fully damaged. Survival is also predictable from full-precision feature statistics alone, with cross-validated AUC 0.92--0.97 and peak activation as the strongest marginal predictor. Finally, RTN quantization and matched-perplexity magnitude pruning damage strongly overlapping feature sets, with Jaccard overlap 0.79--0.86 and damage-score Spearman correlation 0.98. These results show that behavioral metrics alone are insufficient evidence that full-precision interpretability findings transfer to quantized models, motivating feature-level audits of compression.

Related papers

• Where Decoder Cosine Similarity Fails for SAE Feature Flow Discovery [0.20415910628419062]
  We study how state features and update features interact to produce downstream residual features.<n>Ultimately, our results suggest that feature flow atlases can serve as diagnostics of representation-update mechanisms.
  arXiv  Detail & Related papers  (2026-09-11T08:44:13Z)
• Steering Under Compression: Dose-Response, Capability Cost, and Failure Asymmetry in Quantized LLMs [0.0]
  Inference-time steering enables behavioral control of large language models without parameter modification.<n>Post-training quantization reduces memory and compute costs for deployment.<n>We systematically study activation steering under weight-only quantization.
  arXiv  Detail & Related papers  (2026-09-06T08:41:42Z)
• Distributional Energy-Based Models for Uncertainty-Aware Structured LLM Reasoning [40.342912574072024]
  Large Language Models produce structured outputs such as travel plans or code solutions.<n>Individual reasoning steps may appear correct while the output as a whole violates budgets, fails test cases, or contradicts earlier deductions.<n>We propose a learned quality scorer with deterministic analytical constraint penalties for verifying structured LLM outputs.
  arXiv  Detail & Related papers  (2026-05-15T17:08:27Z)
• Verbal Confidence Saturation in 3-9B Open-Weight Instruction-Tuned LLMs: A Pre-Registered Psychometric Validity Screen [0.0]
  We tested whether seven instruction-tuned open-weight models produce verbalised confidence that meets minimal validity criteria.<n>We found Categorical elicitation did not rescue validity.<n> Token-level logprobability did not usefully predict verbalised confidence under the observed variance regime.
  arXiv  Detail & Related papers  (2026-04-24T04:45:21Z)
• Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders [10.172598963520961]
  Large language models can be uncertain yet correct, or confident yet wrong, raising the question of whether their output-level uncertainty and their actual correctness are driven by the same internal mechanisms or by distinct feature populations.<n>We introduce a 2x2 framework that partitions model predictions along correctness and confidence axes, and uses sparse autoencoders to identify features associated with each dimension independently.
  arXiv  Detail & Related papers  (2026-04-21T20:34:14Z)
• Sparse Autoencoder Decomposition of Clinical Sequence Model Representations: Feature Complexity, Task Specialisation, and Mortality Prediction [1.7865154997539017]
  Sparse autoencoders (SAEs) have been applied to large language models and protein language models, but not systematically to electronic health record (EHR) foundation models.<n>We train TopK SAEs on FlatASCEND, a 14.5-million- parameter autoregressive clinical sequence model, at all 10 residual stream extraction points on INSPECT (outpatient) and MIMIC-IV (ICU)<n>SAE decomposition reveals progressive abstraction across transformer depth: layer-0 features are near-perfect token detectors (45.7% singleton), while layer-6 features span approximately 30 token types across multiple clinical categories (0.5% singleton)
  arXiv  Detail & Related papers  (2026-04-13T12:08:32Z)
• Understanding NPM Malicious Package Detection: A Benchmark-Driven Empirical Analysis [10.599261033874884]
  NPM ecosystem has become a primary target for software supply chain attacks.<n>We conduct a benchmark-driven empirical analysis of NPM malware detection.<n>We build a dataset of 6,420 malicious and 7,288 benign packages annotated with 11 behavior categories and 8 evasion techniques.
  arXiv  Detail & Related papers  (2026-03-29T07:04:31Z)
• Uncertainty-Aware Mapping from 3D Keypoints to Anatomical Landmarks for Markerless Biomechanics [39.155383773210836]
  Markerless biomechanics increasingly relies on 3D skeletal keypoints extracted from video.<n>In this work, we investigate predictive uncertainty as a measure of confidence for mapping 3D pose keypoints to 3D anatomical landmarks.
  arXiv  Detail & Related papers  (2026-03-27T09:42:23Z)
• How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models [0.0]
  We present the first systematic study of how unstructured pruning reshapes the feature geometry of language models.<n>We investigate five research questions spanning seed stability, feature survival, SAE transferability, feature fragility, and causal relevance.<n>Our most striking finding is that rare SAE features--those with low firing rates--survive pruning far better than frequent ones.
  arXiv  Detail & Related papers  (2026-03-26T11:12:42Z)
• Multi-Axis Trust Modeling for Interpretable Account Hijacking Detection [1.0152838128195467]
  This paper proposes a Hadith-inspired multi-axis trust modeling framework.<n>We translate five trust axes - long-term integrity (adalah), behavioral precision (dabt), contextual continuity (isnad), cumulative reputation, and anomaly evidence - into a compact set of 26 semantically meaningful behavioral features for user accounts.<n>We evaluate the framework on the CLUE-LDS cloud activity dataset with injected account hijacking scenarios.
  arXiv  Detail & Related papers  (2026-02-20T19:36:30Z)
• A Unified View of Attention and Residual Sinks: Outlier-Driven Rescaling is Essential for Transformer Training [86.64715217940274]
  Outliers function jointly with normalization.<n>Outliers serve more as rescale factors rather than contributors.<n>Outliers can be absorbed into learnable parameters or mitigated via explicit gated rescaling.
  arXiv  Detail & Related papers  (2026-01-30T13:29:45Z)
• Explainable Admission-Level Predictive Modeling for Prolonged Hospital Stay in Elderly Populations: Challenges in Low- and Middle-Income Countries [65.4286079244589]
  Prolonged length of stay (pLoS) is a significant factor associated with the risk of adverse in-hospital events.<n>We develop and explain a predictive model for pLos using admission-level patient and hospital administrative data.
  arXiv  Detail & Related papers  (2026-01-07T23:35:24Z)
• Transparent Early ICU Mortality Prediction with Clinical Transformer and Per-Case Modality Attribution [42.85462513661566]
  We present a lightweight, transparent multimodal ensemble that fuses physiological time-series measurements with unstructured clinical notes from the first 48 hours of an ICU stay.<n>A logistic regression model combines predictions from two modality-specific models: a bidirectional LSTM for vitals and a finetuned ClinicalModernBERT transformer for notes.<n>On the MIMIC-III benchmark, our late-fusion ensemble improves discrimination over the best single model while maintaining well-calibrated predictions.
  arXiv  Detail & Related papers  (2025-11-19T20:11:49Z)
• Segmentation variability and radiomics stability for predicting Triple-Negative Breast Cancer subtype using Magnetic Resonance Imaging [36.402807672379296]
  This study investigates the impact of segmentation variability on feature stability and predictive performance in radiomics-based prediction of Triple-Negative Breast Cancer subtype using Magnetic Resonance Imaging.<n>While incorporating peritumoral information may reduce feature stability, it does not diminish feature predictive capability.
  arXiv  Detail & Related papers  (2025-04-02T12:48:01Z)
• RULSurv: A probabilistic survival-based method for early censoring-aware prediction of remaining useful life in ball bearings [39.58317527488534]
  We introduce a novel and flexible method for early fault detection using Kullback-Leibler divergence and RUL estimation.<n>We demonstrate our approach in the XJTU-SY dataset using a 5-fold cross-validation strategy across three different operating conditions.<n>Our approach achieves a mean cumulative relative accuracy (CRA) of 0.7586 over 5 bearings under the highest load, which improves over several state-of-the-art baselines.
  arXiv  Detail & Related papers  (2024-05-02T16:17:29Z)
• Uncertainty-boosted Robust Video Activity Anticipation [72.14155465769201]
  Video activity anticipation aims to predict what will happen in the future, embracing a broad application prospect ranging from robot vision to autonomous driving. Despite the recent progress, the data uncertainty issue, reflected as the content evolution process and dynamic correlation in event labels, has been somehow ignored. We propose an uncertainty-boosted robust video activity anticipation framework, which generates uncertainty values to indicate the credibility of the anticipation results.
  arXiv  Detail & Related papers  (2024-04-29T12:31:38Z)
• Adaptive Feature Selection for No-Reference Image Quality Assessment by Mitigating Semantic Noise Sensitivity [55.399230250413986]
  We propose a Quality-Aware Feature Matching IQA Metric (QFM-IQM) to remove harmful semantic noise features from the upstream task.<n>Our approach achieves superior performance to the state-of-the-art NR-IQA methods on eight standard IQA datasets.
  arXiv  Detail & Related papers  (2023-12-11T06:50:27Z)
• Removing Spurious Features can Hurt Accuracy and Affect Groups Disproportionately [83.68135652247496]
  A natural remedy is to remove spurious features from the model. We show that removal of spurious features can decrease accuracy due to inductive biases. We also show that robust self-training can remove spurious features without affecting the overall accuracy.
  arXiv  Detail & Related papers  (2020-12-07T23:08:59Z)

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.