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.
By Giang Son Nguyen, Nhi Ngoc-Yen Nguyen, Wray Buntine, Dung D. Le
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.
By Han-yu Wang
arXiv:2602. 06941v2 Announce Type: replace-cross Abstract: Large language models can recover mid-generation from task-misaligned activation steering, producing explicit verbal restarts (e.
By Alex McKenzie, Keenan Pepper, Stijn Servaes, Martin Leitgab, Murat Cubuktepe, Mike Vaiana, Diogo de Lucena, Judd Rosenblatt, Michael S. A. Graziano
arXiv:2602. 00945v2 Announce Type: replace-cross Abstract: LLMs are multilingual by training, yet their lingua franca is often English, reflecting English language dominance in pretraining.
By Anusa Saha, Tanmay Joshi, Vinija Jain, Aman Chadha, Amitava Das
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.
By Ryo Mitsuhashi, Sabri Boughorbel, Majd Hawasly
arXiv:2609.14151v1 Announce Type: cross
Abstract: Steering language models using a set of contrastive representations has been a canonical and computationally efficient method for controlling model b...
By Prajjwal Bhattarai, Tuka Alhanai
arXiv:2607.20995v2 Announce Type: replace
Abstract: Distinguishing animate from inanimate concepts in written language requires more than shallow text processing, as it involves recognizing complex s...
By Samuele Punzo, Giovanni Cin\`a, Sandro Pezzelle
The paper investigates whether different latent language probes—GMM-based representation probes and decoding-based probes—measure the same phenomenon in multilingual language models. Across various model families, training regimes, domains, tasks, checkpoints, and up to 27 languages, the authors find systematic disagreement: representation probes indicate earlier cross‑lingual mixing, while decoding probes reveal sharper, English‑biased language signals. These differences correlate with model multilinguality and training progression but remain relatively stable across domains, suggesting that current probes capture distinct aspects of multilingual processing rather than a single internal lingua franca.
By Deniz Bayazit, Badr AlKhamissi, Antoine Bosselut
Fine‑tuning reshapes internal representations of large language models, affecting attention patterns and layer‑wise activations. The study shows that components identified by EAP as important for task performance cluster in specific layers, yet these layers do not align with those undergoing the largest representational changes. Additionally, overlapping EAP components across different tasks do not guarantee cross‑task transfer and can even degrade performance when tasks differ in nature.
By Lingfang Li, Procheta Sen, Shubham Das, Danushka Bollegala
arXiv:2606. 08682v1 Announce Type: cross Abstract: Activation steering has emerged as a popular inference-time technique for modulating the behavior of large language models (LLMs).
By Qi Cao, Jian Lou, Meiting Liu, Wenjie Feng, Dan Li, See-Kiong Ng, Anh Tuan Luu
arXiv:2605. 07984v2 Announce Type: replace-cross Abstract: We study planning site formation in language models -- where internal representations of structurally-constrained future tokens form during the forward pass, and whether they causally drive generation.
By Nicole Ma, Nick Rui
The paper introduces Language Identity Head Ablation (LIHA), a causal method that zeroes individual attention heads in transformer models to measure language switch rates across multilingual prompts. Applying LIHA to GPT‑2 reveals a small set of first‑token broadcaster heads—most notably L6H1—that persistently attend to the initial prompt token and propagate language signals throughout generation, with compensatory head recruitment occurring hierarchically in higher layers. A controlled comparison between Qwen2.5‑1.5B‑Base and Qwen2.5‑1.5B‑Instruct shows that instruction tuning concentrates language‑identity influence in early layers, while experiments with Chinese and Russian confirm script‑specific first‑token broadcasting at layer 0.
By Arjun Pillai, Christian Hoang, Anjelo Jann Laroza