Tracing Stereotypes from Representation to Output in Multilingual LLMs
Ariun-Erdene Tumurchuluun, Yusser Al Ghussin, Pinzhen Chen, Josef van Genabith, Koel Dutta Chowdhury
cs.CL · cs.AI
Quick verdict
Shows decodability, output influence and cross-lingual transfer of bias features dissociate; careful study, but cross-SAE comparison is sparsity-confounded.
Summary
This mechanistic-interpretability study asks whether the layer where stereotype information is most decodable in multilingual LLMs coincides with where it most influences output. Working in Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B on MBBQ (English, Spanish, Dutch, Turkish; six social categories), the authors compare layer-wise linear probes on mean-pooled hidden states with attribution patching on the residual stream, then characterize sparse-autoencoder (SAE) features from four suites via lexical anchoring and glosses, and finally ablate individual features while measuring bias change on a separate benchmark, SHADES. Probe macro-F1 peaks in mid-layers whereas attribution peaks at the final layer, separated by 36-53% of model depth; controls include permuted-label probes, regularization sweeps, and an English-centric OLMo-7B reference. Of 139 glossed Llama-Scope features, 123 (88.5%) match their selection category and form semantic families. Under a specificity criterion, 6-18% of evaluated residual-stream features are language-agnostic and none are category-agnostic; the larger ablation effects of language-agnostic features appear only in Llama-Scope. Key limitations: the four SAE suites differ drastically in sparsity (average active features from ~50 to ~26,893), width, and training data, confounding cross-suite comparisons; thousands of feature-level t-tests lack multiple-comparison correction; mean pooling includes the appended answer token; and SHADES omits Turkish.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Peak linear decodability of stereotype completions precedes peak attribution-based influence by 36-53% of model depth in all three models (S4-S5).
- 88.5% (123/139) of glossed retained Llama-Scope features match the social category they were selected on and form recurring semantic families (S6.2, Table 3).
- Only 6-18% of evaluated residual-stream features are language-agnostic under the LSS<0.5 criterion, and none are category-agnostic, across all four SAE suites (S7.2, Table 4).
- Language-agnostic features reduce measured bias more than language-dependent ones (46-79% vs 32-55%) only in Llama-Scope; this pattern fails to replicate in the other SAE suites (Fig. 3).
- Decodability, estimated output influence, and cross-lingual ablation transfer are three distinct properties that must be measured separately.
Strengths
Discovery-evaluation separation avoids circularity
Features are selected on MBBQ contrastive pairs (top-3 per category-layer-stream) but their causal effects are measured on the separate SHADES benchmark (S3, S7.1), so intervention results are not evaluated on the data that produced the features.
Probing controls go beyond raw accuracy
Permuted-label control probes stay at chance (macro-F1 0.494-0.498), C=0.1 regularization lowers curves uniformly by 0.04-0.07 without moving peaks, weight norms show no spike at peaks (App. C, Figs. 6-7), and English-centric OLMo-7B peaks far lower (0.606 vs 0.845-0.899; English 0.757 vs 0.556 elsewhere).
Honest bidirectional intervention results
Negative results reported as findings
Scale and release
Weaknesses
Final-layer attribution peak is partly definitional
Cross-SAE comparison confounded by operating point
No multiple-comparison correction
Probe may decode the appended answer token
No random-ablation null; fixed transfer threshold
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.08322) asks whether three properties of multilingual stereotype behavior - linear decodability, estimated causal influence, and feature-level transfer - coincide or dissociate in Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B, with OLMo-7B as an English-centric control. Discovery uses MBBQ (English/Spanish/Dutch/Turkish; six social categories); interventions are scored on the separate benchmark SHADES (English/Spanish/Dutch). Four headline claims: (1) probe macro-F1 peaks in mid-layers while attribution peaks at the final layer, separated by 36-53% of model depth (Llama 53%, Gemma 48%, Qwen 36%, S5.2); (2) 123/139 (88.5%) glossed Llama-Scope features match their selection category and form semantic families (Table 3); (3) 6-18% of evaluated residual-stream features are language-agnostic and none are category-agnostic (Table 4); (4) language-agnostic features show larger ablation effects only in Llama-Scope (46-79% vs 32-55% bias reduction, Fig. 3), a pattern that fails to replicate in the other three SAE suites.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of this paper?
In three multilingual LLMs (Llama-3.1-8B, Qwen3-8B, Gemma-2-9B), the layer where stereotype information is most linearly decodable sits in mid-layers, while attribution-based influence peaks at the final layer - a gap of 36-53% of model depth. Sparse bias features are almost all both language- and category-specific, and the paper argues decodability, causal influence, and cross-lingual transfer are separate properties that must each be measured directly.
Which models, languages, and datasets were used?
Models: Llama-3.1-8B, Qwen3-8B, Gemma-2-9B, plus English-centric OLMo-7B as a control. Discovery data: MBBQ in English, Spanish, Dutch, and Turkish across six social categories. Evaluation data: SHADES (English, Spanish, Dutch - no Turkish) for measuring the effect of ablating individual sparse-autoencoder features. Four SAE suites were compared: Llama-Scope, Gemma-Scope, Llama-Multi, and Qwen-Multi.
Can you remove stereotypes from an LLM by deleting one feature?
No. The paper finds individual feature ablations are heterogeneous and bidirectional: only 1.8-9.6% of feature-bias pairs (depending on SAE suite) significantly reduced bias, some ablations increased measured bias, and mean effects were small. The authors conclude localization is a starting point for intervention analysis, not a guarantee an intervention will work.
Is code and data available for this paper?
Yes. The paper states code and data are available at github.com/ariunerdenetum/stereotype-tracing-mllm, and all underlying assets (model weights, SAE suites, MBBQ, SHADES, Neuronpedia glosses) are public. Reproduction requires GPU inference on 8-9B parameter models.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure-internal data (curve shapes and heatmap cell values in Figs. 2-5 and 8-13) appear in the extraction only as axis-label fragments; numeric claims in this review were verified against prose and Tables 1-7. The body text is complete through references, appendices, and Table 7, with no truncation observed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.