arXiv subjects

Chengzhi Li

Publications and source records attributed to Chengzhi Li.

9 recordsLinked to original sources

arXiv2026

How Do LLMs and VLMs Understand Viewpoint Rotation Without Vision? An Interpretability Study

Over the past year, spatial intelligence has drawn increasing attention. Many prior works study it from the perspective of visual-spatial intelligence, where models have access to visuospatial information from visual inputs. However, in the absence of visual information, whether linguistic intelligence alone is sufficient to endow models with spatial intelligence, and how models perform relevant tasks with text-only inputs still remain unexplored. Therefore, in this paper, we focus on a fundamental and critical capability in spatial intelligence from a linguistic perspective: viewpoint rotation understanding (VRU). Specifically, LLMs and VLMs are asked to infer their final viewpoint and predict the corresponding observation in an environment given textual description of viewpoint rotation and observation over multiple steps. We find that both LLMs and VLMs perform poorly on our proposed dataset while human can easily achieve 100% accuracy, indicating a substantial gap between current model capabilities and the requirements of spatial intelligence. To uncover the underlying mechanisms, we conduct a layer-wise probing analysis and head-wise causal intervention. Our findings reveal that although models encode viewpoint information in the hidden states, they appear to struggle to bind the viewpoint position with corresponding observation, resulting in a hallucination in final layers. Finally, we selectively fine-tune the key attention heads identified by causal intervention to improve VRU performance. Experimental results demonstrate that such selective fine-tuning achieves improved VRU performance while avoiding catastrophic forgetting of generic abilities. Our dataset and code will be released at https://github.com/Young-Zhen/VRU_Interpret .

cs.AI↗

arXiv2026

Understanding Temporal Logic Consistency in Video-Language Models through Cross-Modal Attention Discriminability

Large language models (LLMs) often generate self-contradictory outputs, which severely impacts their reliability and hinders their adoption in practical applications. In video-language models (Video-LLMs), this phenomenon recently draws the attention of researchers. Specifically, these models fail to provide logically consistent responses to rephrased questions based on their grounding outputs. However, the underlying causes of this phenomenon remain underexplored. In this work, we adopt an interpretability-driven approach to analyze, statistically summarize, and intervention the potential factors of the phenomenon. We find that one of the primary reasons for the inconsistency in responses lies in the inability of cross-modal attention heads to effectively distinguish video tokens across different timestamps. To address this, we propose an attention enhancement method called Temporally Conditioned Attention Sharpening (TCAS), which constructs an enhancement objective based on attention distinctions to enhance the model's temporal resolution capability, thereby improving its temporal understanding logic consistency. Experimental results demonstrate that our method significantly enhances the temporal logic consistency of Video-LLMs. Further analyses reveal that our method indeed improves the temporal discriminability of attention heads, validating our conclusions. Additionally, our method even achieves performance improvements in general video temporal grounding tasks, suggesting that temporal logic consistency is an important factor in temporal understanding.

cs.CV↗

arXiv2026

Look, Listen and Segment: Towards Weakly Supervised Audio-visual Semantic Segmentation

Audio-Visual Semantic Segmentation (AVSS) aligns audio and video at the pixel level but requires costly per-frame annotations. We introduce Weakly Supervised Audio-Visual Semantic Segmentation (WSAVSS), which uses only video-level labels to generate per-frame semantic masks of sounding objects. We decompose WSAVSS into looking, listening, and segmentation, and propose Progressive Cross-modal Alignment for Semantics (PCAS) with two modules: *Looking-before-Listening* and *Listening-before-Segmentation*. PCAS builds a classification task to train the audio-visual encoder using video labels, injects visual semantic prompts to enhance frame-level audio understanding, and then applies progressive contrastive alignment to map audio categories to image regions without mask annotations. Experiments show PCAS achieves state-of-the-art performance among weakly supervised methods on AVS and remains competitive with fully supervised baselines on AVSS, validating its effectiveness.

cs.MM↗

arXiv2026

The Struggle Between Continuation and Refusal: A Mechanistic Analysis of the Continuation-Triggered Jailbreak in LLMs

With the rapid advancement of large language models (LLMs), the safety of LLMs has become a critical concern. Despite significant efforts in safety alignment, current LLMs remain vulnerable to jailbreaking attacks. However, the root causes of such vulnerabilities are still poorly understood, necessitating a rigorous investigation into jailbreak mechanisms across both academic and industrial communities. In this work, we focus on a continuation-triggered jailbreak phenomenon, whereby simply relocating a continuation-triggered instruction suffix can substantially increase jailbreak success rates. To uncover the intrinsic mechanisms of this phenomenon, we conduct a comprehensive mechanistic interpretability analysis at the level of attention heads. Through causal interventions and activation scaling, we show that this jailbreak behavior primarily arises from an inherent competition between the model's intrinsic continuation drive and the safety defenses acquired through alignment training. Furthermore, we perform a detailed behavioral analysis of the identified safety-critical attention heads, revealing notable differences in the behaviors of safety heads across different model architectures. Grounded in these mechanistic findings, we propose Head Competition Steering (HCS), a mechanistically grounded inference-time strategy that explicitly leverages the competition between safety heads and continuation heads to suppress harmful generation, and further distill its behavioral signal into a student model via knowledge distillation, achieving inference-time safety improvements without additional computational overhead.

cs.AI↗

arXiv2026

TAMMs: Change Understanding and Forecasting in Satellite Image Time Series with Temporal-Aware Multimodal Models

Temporal Change Description (TCD) and Future Satellite Image Forecasting (FSIF) are critical, yet historically disjointed tasks in Satellite Image Time Series (SITS) analysis. Both are fundamentally limited by the common challenge of modeling long-range temporal dynamics. To explore how to improve the performance of methods on both tasks simultaneously by enhancing long-range temporal understanding capabilities, we introduce **TAMMs**, the first unified framework designed to jointly perform TCD and FSIF within a single MLLM-diffusion architecture. TAMMs introduces two key innovations: Temporal Adaptation Modules (**TAM**) enhance frozen MLLM's ability to comprehend long-range dynamics, and Semantic-Fused Control Injection (**SFCI**) mechanism translates this change understanding into fine-grained generative control. This synergistic design makes the understanding from the TCD task to directly inform and improve the consistency of the FSIF task. Extensive experiments demonstrate TAMMs significantly outperforms state-of-the-art specialist baselines on both tasks. Our dataset can be found at https://huggingface.co/datasets/IceInPot/TAMMs .

cs.CV↗

arXiv2026

Can LLMs See Without Pixels? Benchmarking Spatial Intelligence from Textual Descriptions

Recent advancements in Spatial Intelligence (SI) have predominantly relied on Vision-Language Models (VLMs), yet a critical question remains: does spatial understanding originate from visual encoders or the fundamental reasoning backbone? Inspired by this question, we introduce SiT-Bench, a novel benchmark designed to evaluate the SI performance of Large Language Models (LLMs) without pixel-level input, comprises over 3,800 expert-annotated items across five primary categories and 17 subtasks, ranging from egocentric navigation and perspective transformation to fine-grained robotic manipulation. By converting single/multi-view scenes into high-fidelity, coordinate-aware textual descriptions, we challenge LLMs to perform symbolic textual reasoning rather than visual pattern matching. Evaluation results of state-of-the-art (SOTA) LLMs reveals that while models achieve proficiency in localized semantic tasks, a significant "spatial gap" remains in global consistency. Notably, we find that explicit spatial reasoning significantly boosts performance, suggesting that LLMs possess latent world-modeling potential. Our proposed dataset SiT-Bench serves as a foundational resource to foster the development of spatially-grounded LLM backbones for future VLMs and embodied agents. Our code and benchmark will be released at https://github.com/binisalegend/SiT-Bench .

cs.CV↗

arXiv2025

LISA-3D: Lifting Language-Image Segmentation to 3D via Multi-View Consistency

Text-driven 3D reconstruction requires masks that understand free-form instructions and remain stable under viewpoint changes. We present LISA-3D, a two-stage framework that adapts the instruction-following segmenter LISA with geometry-aware Low-Rank Adaptation (LoRA) layers while keeping the SAM-3D reconstructor frozen. During training, paired RGB-D frames and camera poses define a differentiable reprojection loss that enforces cross-view agreement without additional 3D-text annotations. At deployment, the adapted segmenter can produce an RGBA prompt for SAM-3D from one RGB image; when registered RGB-D views are available, optional logit fusion further improves the prompt. On ScanRefer and Nr3D, geometry-aware tuning improves both 2D masks and lifted 3D reconstructions while updating only 11.6M parameters. Our results separate geometry-aware training gains from optional multi-view inference gains, providing a modular route from language grounding to object-centric 3D reconstruction.

cs.CV↗

arXiv2025

Investigating disc-corona interaction in axisymmetric accretion disc models

The interaction between the accretion disc and its corona plays a critical role in the energy balance and emission mechanisms in astrophysical systems such as active galactic nuclei and X-ray binaries. However, the detailed physics of disc-corona interactions, including the mechanisms driving disc evaporation and the impact of accretion rate and viscosity, remain poorly understood. Our study aims to extend the well-known disc evaporation model to investigate the disc-corona interaction in a 2D axisymmetric, time-dependent hydrodynamic model, focusing on the effects of viscosity, accretion rate, and their influence on disc evaporation, luminosity, and corona formation. We develop a hydrodynamic model consisting of a thin accretion disc, a corona, and a vacuum region. Our model is implemented in Athena++, with the gas-vacuum interface tracking algorithm to handle the vacuum regions. We perform simulations incorporating turbulent viscosity, thermal conduction, Bremsstrahlung cooling, and artificial disc cooling, starting from an adiabatic state to explore the disc-corona interaction. We demonstrate the presence of acoustic shock heating. We find that viscosity dominates the intensity of disc evaporation, that the accretion rate primarily determines the disc truncation radius and the disc luminosity, and that there may be a positive correlation between the corona luminosity and the evaporation intensity. We find the warm gas required by the warm corona model. We also compare our results with observations and simulations, and estimate the $y$-parameters to explore the potential effects of Compton cooling as well as the potential effects of the warm corona.

astro-ph.HE↗

arXiv2019

On the cubic perturbations of the symmetric 8-loop Hamiltonian

We study arbitrary cubic perturbations of the symmetric 8-loop Hamiltonian, which are linear with respect to the small parameter. It is shown that when the first 4 coefficients in the expansion of the displacement functions corresponding to both period annuli inside the loop vanish, the system becomes integrable with the following three strata in the center manifold: Hamiltonian, reversible in $y$ and reversible in $x$. In the latter case, the first integral is of Darboux type and we calculate it explicitly. Next we prove that the cyclicity of each of period annuli inside the loop is five and the total cyclicity of both is at most nine. For this, we use Abelian integrals method together with careful study the geometry of the separatrix solutions of related Riccati equations in connection to the second-order Melnikov functions.

math.CA↗