mlsanityData to 14 Oct 2026

Perplexity Can Miss SAE Feature Damage Under Quantization

arXiv 2606.03002 ↗First posted 2 Jun 2026cs.LG cs.AIcited by 5

On the map: Mechanistic interpretability → SAE features and their applications

Abstract (arXiv)

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.

Key facts. Cited by 5 papers in the map (data through 14 Oct 2026). Placed in SAE features and their applications, one of 13 directions in Mechanistic interpretability. Builds on 23 of its 25 references that are in the map. First posted on arXiv 2 Jun 2026.

Builds on

23 of 25 references.

Closest ancestors · 6

Same direction, newest first

  1. 2026Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet28 May
  2. 2026How Pruning Reshapes Features: Sparse Autoencoder Analysis of Weight-Pruned Language Models26 Mar
  3. 2026Stable and Steerable Sparse Autoencoders with Weight Regularization4 Mar
  4. 2025Sparse Autoencoders Trained on the Same Data Learn Different Features28 Jan
  5. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Features4 Nov
  6. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Models15 Sep

Other references · 17

  1. 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,418
  2. 2015Learning both Weights and Connections for Efficient Neural Networkscited by 2,031
  3. 2022GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformerscited by 1,919
  4. 2023Pythia: A Suite for Analyzing Large Language Models Across Training and Scalingcited by 1,786
  5. 2022SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Modelscited by 1,386
  6. 2023AWQ: Activation-aware Weight Quantization for LLM Compression and Accelerationcited by 1,298
  7. 2023SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shotcited by 1,121
  8. 2022Toy Models of Superpositioncited by 971
  9. 2024Gemma: Open Models Based on Gemini Research and Technologycited by 916
  10. 2023A Simple and Effective Pruning Approach for Large Language Modelscited by 800
  11. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  12. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 95
5 more
  1. 2024Evaluating Quantized Large Language Modelscited by 77
  2. 2025Steering Large Language Model Activations in Sparse Spacescited by 58
  3. 2024Enhancing Neural Network Interpretability with Feature-Aligned Sparse Autoencoderscited by 25
  4. 2025Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoderscited by 14
  5. 2025Attribution-Guided Distillation of Matryoshka Sparse Autoencoderscited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 5.

5 papers.

  1. 2026Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Stalenesscited by 5
  2. 2026Perplexity Cost Understates What Activation Quantisation Breakscited by 0
  3. 2026Cosine Similarity Is Not Evidence: Measuring the Noise Floor of Interpretability Transfer Under Quantizationcited by 0
  4. 2026Linger and Lose: Knowledge Collapse in Low-Bit Language Modelscited by 0
  5. 2026Saying, Not Knowing: Aggressively GGUF-Quantized Small Language Models Still Write Rare Words They Can No Longer Definecited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Language-Conditional Dequantization: Recovering What Quantization Steals from Non-English Languages
  2. 2026The Structure of Quantization Damage in LLMs: Why the Next Bit Should Be Spent Globally
  3. 2026When Pruning Meets Interpretability: Preserving Sparse Autoencoder Robustness in LLMs
  4. 2026Are Sparse Autoencoder Benchmarks Reliable?
  5. 2024Investigating Sensitive Directions in GPT-2: An Improved Baseline and Comparative Analysis of SAEs
  6. 2026Are LLM Uncertainty and Correctness Encoded by the Same Features? A Functional Dissociation via Sparse Autoencoders
  7. 2026Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models
  8. 2025Measuring Sparse Autoencoder Feature Sensitivity