arXiv:2603.07335v1 Announce Type: new Abstract: High-performing vision language models still produce incorrect answers, yet their failure modes are often difficult to explain. To make model internals more accessible and enable systematic debugging, we introduce VisualScratchpad, an interactive interface for visual concept analysis during inference. We apply sparse autoencoders to the vision encoder and link the resulting visual concepts to text tokens via text-to-image attention, allowing us to examine which visual concepts are both captured by the vision encoder and utilized by the language model. VisualScratchpad also provides a token-latent heatmap view that suggests a sufficient set of latents for effective concept ablation in causal analysis. Through case studies, we reveal three underexplored failure modes: limited cross-modal alignment, misleading visual concepts, and unused hidden cues. Project page: https://hyesulim.github.io/visual_scratchpad_projectpage/

arXiv:2506.05850v3 Announce Type: replace-cross Abstract: Reinforcement learning with verifiable reward (RLVR) has been instrumental in eliciting strong reasoning capabilities from large language models (LLMs) via long chains of thought (CoT). During RLVR training, we formalize and systemically study an empirical phenomenon whereby a multilingual model's CoT reverts to its dominant pre-training language (e.g., English) even when prompted in another language, which we term Cross-lingual Collapse. Because the long-CoT regime magnifies exposure to linguistic priors, the underlying trade-off between maximizing reasoning depth and preserving target-language fidelity has remained under-characterized. To examine this trade-off, we train LLMs with Group-Relative Policy Optimization (GRPO) on translated versions of math datasets widely used to elicit long-CoT reasoning. Throughout training, we track both task accuracy and the language consistency of reasoning chains. Our experiments yield three findings: (i) under RLVR, CoT in LLMs systematically drifts toward the pre-training dominant language as reasoning performance rises; (ii) English-centric priors, long-CoT GRPO optimization, task difficulty, and high-entropy decoding jointly amplify this drift, and the pattern persists beyond mathematics; and (iii) interventions that favor target-language traces--via a language-consistency reward, decoding-time controls, or more balanced backbones--mitigate collapse but reveal a persistent performance-fidelity trade-off.