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.

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.

While LLMs demonstrate impressive reasoning capabilities, their internal decision dynamics remain opaque. To render these process interpretable and intervenable, we propose Dynamic Entropy Tracing, a mechanism-aware framework that interprets the evolving “choice state” of attention heads during CoT generation through stepwise head-wise option-logit and entropy tracing. Our analysis reveals distinct functional behaviors at attention heads: Steadfast Heads, characterized by consistently low entropy and producing a sharp, option-selective logit pattern with a stable top choice, and Wavering Heads, characterized by consistently high entropy and producing flat or oscillatory option logits without a persistent winner. Leveraging these traces, we identify a set of intervention targets and perform Selective Head Fine-Tuning, updating solely these selected heads against a frozen backbone. Experiments across the LLaMA and Qwen families reveal a striking plasticity hierarchy: fine-tuning just 30 Wavering Heads recovers over 98% of the performance achieved by full-parameter tuning, and in some settings modestly exceeds it. In contrast, intervening on Steadfast Heads yields much less gains. Our findings translate process-level mechanistic observables into a principled criterion for selective fine-tuning, offering a fundamental insight: the most effective tuning knobs are not the components that signal the final decision, but those that retain uncertainty, and thus plasticity, during its formation.

Despite the remarkable performance across numerous tasks, Large Language Models (LLMs) still exhibit notable deficiencies in temporal reasoning, even in simple event ordering tasks. For instance, a slight alteration in the temporal phrasing of the question (e.g., changing "Is event A before B?” to “Is event A after B?”) can lead LLMs to hallucinate and produce inconsistent answers, reflecting a lack of robust temporal reasoning. Although many prior studies have focused on benchmarking and improving the temporal reasoning ability of LLMs, little is known about the intrinsic mechanisms within LLMs when performing temporal reasoning. In this work, we investigate the mechanistic interpretability of temporal ordering within event temporal reasoning through a structured "Identify-Interpret-Verify” pipeline. We first employ path patching to identify a sparse subset of attention heads that are causally responsible for reasoning outcomes. Detailed pattern analysis reveals that these key heads specialize in attending to either temporal keywords (semantic cues) or structural delimiters (syntactic cues). Furthermore, we rigorously validate the observed mechanism through comprehensive intervention-based experiments, ranging from head ablation to targeted attention modulation. We demonstrate that dynamically modulating the attention of these specific heads can robustly enhance model performance, which serves as strong empirical evidence that our identified mechanism faithfully captures the internal logic of temporal ordering in LLMs.

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.

With the widespread applications of large language models (LLMs), aligning LLMs with human values has emerged as a critical challenge. For alignment, we always expect LLMs to be honest, positive, harmless, etc. And LLMs appear to be capable of generating the desired outputs after the alignment tuning process, such as the preference tuning via reinforcement learning from human feedback (RLHF). However, it also raises a question about after alignment, do LLMs genuinely obtain a value distinction between positives and negatives, beyond the generation of positive outputs? In this work, we start by investigating this question from the token distribution perspective. Our findings reveal that compared to the unaligned versions, LLMs after alignment exhibit a larger logits gap between positive and negative tokens at each generation step, which suggests that LLMs do obtain a value distinction of positives and negatives after alignment. Meanwhile, it also motivates us to achieve alignment by directly constructing such value distinction, thus alleviating the excessive reliance on computational resources required by training-time alignment. Specifically, we propose a representation editing method that intervenes the last hidden representation by amplifying the logits difference between positive and negative tokens (defined as anchor words). Experimental results demonstrate that the proposed method not only achieves effective alignment, but also requires fewer computational resources compared to training-time alignment methods