Yixin Zhu
Publications
Visual Distraction Undermines Moral Reasoning in Vision-Language Models
Moral reasoning is fundamental to safe Artificial Intelligence (AI), yet ensuring its consistency across modalities becomes critical as AI systems evolve from text-based assistants to embodied agents. Current safety techniques demonstrate success in textual contexts, but concerns remain about generalization to visual inputs. Existing moral evaluation benchmarks rely on textonly formats and lack systematic control over variables that influence moral decision-making. Here we show that visual inputs fundamentally alter moral decision-making in state-of-the-art (SOTA) Vision-Language Models (VLMs), bypassing text-based safety mechanisms. We introduce Moral Dilemma Simulation (MDS), a multimodal benchmark grounded in Moral Foundation Theory (MFT) that enables mechanistic analysis through orthogonal manipulation of visual and contextual variables. The evaluation reveals that the vision modality activates intuition-like pathways that override the more deliberate and safer reasoning patterns observed in text-only contexts. These findings expose critical fragilities where language-tuned safety filters fail to constrain visual processing, demonstrating the urgent need for multimodal safety alignment.
Learning Physics-Grounded 4D Dynamics with Neural Gaussian Force Fields
Predicting physical dynamics from raw visual data remains a major challenge in AI. While recent video generation models have achieved impressive visual quality, they still cannot consistently generate physically plausible videos due to a lack of modeling of physical laws. Recent approaches combining 3D Gaussian splatting and physics engines can produce physically plausible videos, but are hindered by high computational costs in both reconstruction and simulation, and often lack robustness in complex real-world scenarios. To address these issues, we introduce Neural Gaussian Force Field (NGFF), an end-to-end neural framework that integrates 3D Gaussian perception with physics-based dynamic modeling to generate interactive, physically realistic 4D videos from multi-view RGB inputs, achieving two orders of magnitude faster than prior Gaussian simulators. To support training, we also present GSCollision, a 4D Gaussian dataset featuring diverse materials, multi-object interactions, and complex scenes, totaling over 640k rendered physical videos (~4 TB). Evaluations on synthetic and real 3D scenarios show NGFF's strong generalization and robustness in physical reasoning, advancing video prediction towards physics-grounded world models.
Code over Words: Overcoming Semantic Inertia via Code-Grounded Reasoning
LLMs struggle with Semantic Inertia: the inability to inhibit pre-trained priors (e.g., "Lava is Dangerous") when dynamic, in-context rules contradict them. We probe this phenomenon using Baba Is You, where physical laws are mutable text rules, enabling precise evaluation of models' ability to override learned priors when rules change. We quantatively observe that larger models can exhibit inverse scaling: they perform worse than smaller models when natural language reasoning requires suppressing pre-trained associations (e.g., accepting "Lava is Safe"). Our analysis attributes this to natural language encoding, which entangles descriptive semantics and logical rules, leading to persistent hallucinations of familiar physics despite explicit contradictory rules. Here we show that representing dynamics as executable code, rather than descriptive text, reverses this trend and enables effective prior inhibition. We introduce Code-Grounded Vistas (LCV), which fine-tunes models on counterfactual pairs and identifies states with contradictory rules, thereby forcing attention to logical constraints rather than visual semantics. This training-time approach outperforms expensive inference-time search methods in both efficiency and accuracy. Our results demonstrate that representation fundamentally determines whether scaling improves or impairs contextual reasoning. This challenges the assumption that larger models are universally better, with implications for domains that require dynamic overriding of learned priors.
BotzoneBench: Scalable LLM Evaluation via Graded AI Anchors
Large Language Models (LLMs) are increasingly deployed in interactive environments requiring strategic decision-making, yet systematic evaluation of these capabilities remains challenging. Existing benchmarks for LLMs primarily assess static reasoning through isolated tasks and fail to capture dynamic strategic abilities. Recent game-based evaluations employ LLM-vs-LLM tournaments that produce relative rankings dependent on transient model pools, incurring quadratic computational costs and lacking stable performance anchors for longitudinal tracking. The central challenge is establishing a scalable evaluation framework that measures LLM strategic reasoning against consistent, interpretable standards rather than volatile peer models. Here we show that anchoring LLM evaluation to fixed hierarchies of skill-calibrated game Artificial Intelligence (AI) enables linear-time absolute skill measurement with stable cross-temporal interpretability. Built on the Botzone platform's established competitive infrastructure, our BotzoneBench evaluates LLMs across eight diverse games spanning deterministic perfect-information board games to stochastic imperfect-information card games. Through systematic assessment of 177,047 state-action pairs from five flagship models, we reveal significant performance disparities and identify distinct strategic behaviors, with top-performing models achieving proficiency comparable to mid-to-high-tier specialized game AI in multiple domains. This anchored evaluation paradigm generalizes beyond games to any domain with well-defined skill hierarchies, establishing a scalable and reusable framework for assessing interactive AI capabilities.