Research map / Label-efficient and robust learning / Recognition and long-tail
Beyond Interpretability: The Gains of Feature Monosemanticity on Model Robustness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that monosemantic features (via non-negative contrastive learning, sparse autoencoders, non-negative finetuning, and a sparse LoRA variant MonoLoRA) improve robustness to label/input noise, few-shot finetuning, and LLM alignment degradation, challenging the accuracy-interpretability tradeoff.
Problem
Monosemantic features are believed to sacrifice accuracy for interpretability; the paper asks whether monosemanticity brings robustness gains under label noise, input noise, distribution shift, few-shot finetuning, and LLM finetuning.
Approach
Treat feature monosemanticity as a source of learning robustness rather than a cost: use non-negative contrastive learning (NCL) and sparse autoencoders (SAE) to obtain monosemantic features, apply non-negative finetuning to keep them monosemantic during supervised training, and add non-negative sparsity to LoRA (MonoLoRA); support this with a superposition toy model and an LDA-style separability
Key results
- Under 90% label noise on ImageNet-100 linear probing, NCL improves top-1 accuracy by +13.7% over polysemantic CL features.
- Non-negative finetuning yields up to 11.9% improvement under large label noise rates on ImageNet-100, and better validation accuracy with 10% training data (+3.9% reported in intro).
- On Llama-2-7B-Chat, MonoLoRA achieves lower average ShieldGemma alignment score (3.53 vs 7.90 for LoRA) on SST-2 while raising task accuracy to 94.84 vs 92.78.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CL features with SAE monosemantic features | ResNet-18 linear probing · CIFAR-100, 90% label noise | accuracy: 35.7 -> 39.1 | Helped |
| Replace CL features with NCL monosemantic features | ResNet-18 linear probing · CIFAR-100, 90% label noise | accuracy: 35.7 -> 44.9 | Helped |
| Replace CL features with SAE monosemantic features | ResNet-18 linear probing · ImageNet-100, 90% label noise | accuracy: 34.4 -> 45.9 | Helped |
| Replace CL features with NCL monosemantic features | ResNet-18 linear probing · ImageNet-100, 90% label noise | accuracy: 34.4 -> 48.1 | Helped |
| Use monosemantic rather than standard polysemantic finetuning | ImageNet-100 finetuning · Few-shot training data | validation accuracy: +3.9% top-1 accuracy with 10% samples | Helped |
| Use non-negative finetuning rather than standard finetuning | ImageNet-100 finetuning · Noisy-label finetuning | accuracy: at most 11.9% improvement | Helped |
| Use monosemantic features instead of polysemantic features | ImageNet-100 classifiers · Input distribution shifts | classification accuracy | Helped |
| Add non-negative constraints to LoRA updates | Llama-2-7B-Chat · SST2 | accuracy: 92.78 -> 94.84 | Helped |
| Add non-negative constraints to LoRA updates | Llama-2-7B-Chat · SST2 | ShieldGemma alignment score (Avg): 7.90 -> 3.53 | Helped |
| Add non-negative constraints to LoRA updates | Llama-2-7B-Chat · Dolly | RougeL: 14.08 -> 14.48 | Helped |
| Add non-negative constraints to LoRA updates | Llama-2-7B-Chat · Dolly | ShieldGemma alignment score (Avg): 6.12 -> 6.04 | Helped |
| Use monosemantic rather than polysemantic features | Toy model classifier · Clean data | classification accuracy | Hurt |
Surprising findings
- Monosemantic features perform comparably on clean data yet outperform polysemantic features under noise, reversing the expected accuracy-interpretability tradeoff.
- Theory shows the monosemantic feature has better linear separability than the polysemantic one only when label noise rate is at least 0.25.
Limitations
- Theoretical analysis is limited to a toy binary model (n=2, m=1) and a specific antipodal-pair assumption for polysemantic features.
- Robustness gains are shown mainly on ImageNet-100 and CIFAR-100 with ResNet-18 and a small LLM finetuning evaluation.
- Few-shot and LLM results are preliminary, with LLM evaluation on a single small SST-2 and Dolly setup.
Experimental setup
Datasets: ImageNet-100 CIFAR-100 SST-2 Dolly Beavertails ImageNet-Sketch ImageNet-Stylized
Benchmarks: ImageNet-100 CIFAR-100 SST-2 Dolly Beavertails ImageNet-Sketch ImageNet-Stylized
Metrics: top-1 accuracy RougeL alignment score (ShieldGemma) Beavertails score LDA separability criterion J
Baselines: SimCLR polysemantic features (CL) Standard CE finetuning LoRA Base Llama-2-7B-Chat
Builds on: SimCLR Non-negative contrastive learning Sparse autoencoder LoRA Superposition hypothesis toy model
Closest papers by meaning
- Capturing Polysemanticity with PRISM: A Multi-Concept Feature Description Framework (NeurIPS 2025 · similarity 0.88)
- Measuring and Guiding Monosemanticity (NeurIPS 2025 · similarity 0.88)
- Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.85)
- Local vs. Global Interpretability: A Computational Complexity Perspective (ICML 2024 · similarity 0.85)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.85)
- Rethinking Evaluation of Sparse Autoencoders through the Representation of Polysemous Words (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.