How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings

Technical Summary: How Far Do Auto-Interpretation Labels Generalize?

Problem Statement
Sparse autoencoders (SAEs) have become a standard tool for interpreting language models (LMs) by decomposing dense activations into sparse, human-interpretable features. Because models contain far more features than can be inspected manually, practitioners rely on auto-interpretation labels: natural-language descriptions generated by an LLM based on a feature's top-activating examples.

The central problem addressed is the generalizability of these labels. A label describing a concept (e.g., "deception" or "violent content") implies that the feature tracks that concept universally. However, it remains unclear whether a feature labeled for a concept in English actually detects that same concept across different languages, scripts, and rewordings. If a feature's behavior is heavily dependent on the specific surface form of the input (e.g., English text) rather than the underlying semantic concept, the label may be misleading for safety auditing or model steering in multilingual contexts.

Methodology
The authors construct a controlled factorial evaluation paradigm using Serbian digraphia as a testbed. Serbian is natively written in both Latin and Cyrillic scripts, related by deterministic, lossless transliteration. This allows the researchers to vary the script while holding language, wording, and meaning constant.

The study utilizes:

• Dataset: 300 sentences from the FLORES+ devtest split, covering news, instructional, and travel registers.
• Variants: Each sentence appears in four forms: English (Latin), Serbian (Latin), Serbian (Cyrillic), and Russian (Cyrillic).
• Conditions: For each variant, three conditions are tested:
  1. Original: The professional translation or its transliteration.
  2. Paraphrase: A meaning-preserving rewrite with varied surface form.
  3. Random: An unrelated sentence of similar length in the same language/script.
• Models & SAEs: Primary analysis uses Gemma-3-27B with Gemma Scope 2 SAEs (16k dictionary width, JumpReLU activation). Results are validated across model scales (1B, 12B) and architectures (Llama-3.1-8B with Llama Scope SAEs).
• Evaluation Metrics:
  • Feature Overlap: Measured via Jaccard similarity of active feature sets across layers to determine if features encode abstract meaning.
  • Auto-Interpretation Falsification: Features are selected if they fire on both English originals and Russian paraphrases (confirming cross-lingual semantic tracking). Their auto-generated labels (from Neuronpedia, using Claude Sonnet and Gemini) are then tested against Serbian Latin and Cyrillic versions. A "miss" occurs if a content-labeled feature fails to activate on the same meaning in a different script/language.

Key Findings

1. SAE Features Encode Abstract Meaning:
   The study confirms that SAE features genuinely encode semantic content that survives changes in script, language, and wording.

   • Script Invariance: Transliterating between Serbian Latin and Cyrillic preserves a high degree of feature overlap (mean Jaccard 0.73).
   • Cross-Lingual Semantic Features: Features activated by the same content in English and Russian (differing in script, language, and wording) show substantial overlap (peak Jaccard 0.57), significantly higher than random baselines (0.13). This proves that cross-lingual semantic features exist and are worth labeling.

2. Auto-Interpretation Labels Fail Systematically on Less-Represented Forms:
   Despite the existence of semantic features, the auto-generated labels do not accurately predict feature behavior across all forms.

   • Miss Rates: Content-labeled features miss the same meaning in Serbian up to 4× more often than within English. For example, at layer 53, the miss rate for English paraphrases is 7.3%, while the average miss rate for Serbian is 30.1%.
   • Script Asymmetry: Even between the two Serbian scripts (which are deterministic transliterations of each other), features miss Serbian Cyrillic more often than Serbian Latin. The Cyrillic-to-Latin miss ratio grows with network depth, reaching 1.35× at the final layer.
   • Training Distribution Correlation: These failures align with estimated training data coverage. English is dominant in pretraining corpora, Russian is high-resource but less dominant, and Serbian (particularly Cyrillic) is rare. The model's feature reliability degrades as the representation of the input form in training data decreases.

3. Depth-Dependent Divergence:
   The gap in reliability widens as the network depth increases. While within-language reliability (English) improves or stays stable in deeper layers, cross-language reliability (Serbian) degrades. This suggests that deeper layers may tie representations more tightly to specific input forms rather than abstract concepts.

4. Invisibility of Failure:
   Crucially, the auto-interpretation labels themselves provide no indication of these failures. A label like "Olympic athletes" appears accurate based on the feature's top activations (likely English examples) but fails to predict the feature's silence on the same concept in Serbian Cyrillic.

Significance and Claims
The paper claims that auto-interpretation labels should be understood not as guarantees about a concept in general, but as claims about a feature's behavior on its best-represented inputs.

• Limitation of Current Tools: Practitioners relying on these labels for safety auditing or model steering may be misled. A feature labeled for a safety-critical concept might detect it in English but fail to trigger in Serbian Cyrillic, allowing harmful content to pass unnoticed.
• Graded Failure: The failure is not binary (high-resource vs. low-resource) but graded, correlating with the frequency of the specific script/language in training data.
• Methodological Contribution: The study provides a controlled, factorial evaluation suite (released with the paper) that isolates script, language, wording, and meaning, offering a robust method for future work to test the generalizability of interpretability tools.

The authors conclude that while SAEs successfully extract semantic features, the current auto-interpretation pipelines do not keep pace with the complexity of multilingual generalization, creating a "blind spot" where labels are locally accurate but globally unreliable.