The paper investigates whether sparse autoencoder (SAE) features that recur across different language settings in Gemma 2 and Gemma 3 actually have consistent causal effects on translation performance. By reproducing Wu et al.’s discovery method and extending it to multilingual prompts, the authors find over 20 frequently activating features, yet causal validation reveals that almost all have negligible or inconsistent impacts. Only one feature—Gemma 2’s (L10, 5717) and Gemma 3’s (L20, 2456)—consistently improves COMET scores when amplified and worsens them when ablated across 23 language settings, indicating a language‑agnostic translation‑initiation direction.
arXiv:2604. 03532v2 Announce Type: replace-cross Abstract: Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult.
arXiv:2608. 12334v1 Announce Type: cross Abstract: Despite the impressive multilingual capabilities of Large Language Models, the latent dynamics dictating language selection remain poorly understood.
The paper investigates how multilingual large language models can unintentionally switch languages during generation. It compares three techniques—ValSel, FreqSel, and AnnSel—for pinpointing latent variables that control language choice in cross‑layer transcoders. Using new multilingual benchmarks and targeted interventions on Gemma‑2‑2B and Qwen3‑4B, the study finds all methods can steer output language, with FreqSel performing best and AnnSel providing interpretable selections via explicit annotations.
arXiv:2609.08322v1 Announce Type: cross Abstract: Multilingual LLMs show stereotype-related behavior that varies across languages, but behavioral scores do not show where the relevant information is...
The paper proposes the interlingua hypothesis, suggesting that large language models translate by encoding a source sentence into a latent, task‑agnostic feature space and then decoding a target sentence from that space. Three lines of evidence support this: (1) BLEU variance across language pairs is largely explained by language‑specific competences without pair‑specific interactions; (2) many model components influence both monolingual and translation tasks; and (3) fine‑tuning on monolingual data recovers most translation gains seen with aligned documents. These findings converge to support the hypothesis and point toward new ways to understand and improve LLM translation.
The paper investigates how multilingual large language models (LLMs) encode and express stereotypes across different languages. By applying linear probing, attribution patching, sparse autoencoders (SAEs), and feature ablation to Llama‑3.1‑8B, Qwen3‑8B, and Gemma‑2‑9B, the authors find that probe performance peaks much earlier than attribution, indicating a separation of 36‑53% of model depth. They observe that only a small fraction (6‑18%) of residual‑stream features exhibit language‑agnostic effects, and none are category‑agnostic, highlighting the need to measure decodability, output influence, and cross‑lingual ablation effects separately.
arXiv:2609.07876v1 Announce Type: cross Abstract: Recent methods in language model interpretability employ techniques such as sparse autoencoders to decompose residual stream contributions into linea...
arXiv:2607. 20596v1 Announce Type: new Abstract: Sparse autoencoder (SAE) features are used to interpret and steer large language models, yet whether a feature's causal role is stable across SAE families remains untested.
The paper investigates weight‑space merging of independently fine‑tuned multilingual machine translation models. Experiments show that merging is more successful when models share a target language, yet it still cannot match the peak performance of language‑specific checkpoints. When target languages differ, performance drops sharply, and analysis reveals that overlapping neuron activation and incompatible upper‑layer geometries cause these failures.
The paper investigates why large language models sometimes hallucinate when asked about facts in a language different from the one in which the facts were learned. By training small Transformer models on synthetic multilingual datasets, the authors show that the degree of correlation between facts and their learning language (informativeness) and the ease of language identification (extractability) determine whether models develop unified or separate representations across languages. Unified representations enable cross‑lingual fact transfer, while separate representations do not. The study proposes a unifying perspective on cross‑lingual transfer and suggests training methods to promote representational unification.
arXiv:2606. 08365v1 Announce Type: cross Abstract: Sparse autoencoder (SAE) features are increasingly used to steer language models, but feature steering is rarely clean: the same intervention can behave inconsistently across contexts and perturb unrelated features.
arXiv:2608. 04021v1 Announce Type: cross Abstract: Cloze-style probes that vary how often a target token appears implicitly assume that more copies of a target affect prediction the same way regardless of where the readout slot sits.