While Chain-of-Thought (CoT) prompting has become a standard paradigm for eliciting complex reasoning capabilities in Large Language Models, it inadvertently exposes a new attack surface for backdoor attacks. Existing CoT backdoor attacks typically manipulate the intermediate reasoning steps to steer the model toward incorrect answers. However, these corrupted reasoning traces are readily detected by prevalent process-monitoring defenses. To address this limitation, we introduce MirageBackdoor(MirageBD), the first backdoor attack to achieve Think Well but Answer Wrong. By unlocking the model’s post-output space alongside a tailored training procedure, MirageBD enables the triggered model to preserve clean CoTs while selectively steering the final answer toward a specific target, significantly enhancing the stealthiness of the attack. Experiments show that MirageBD generally achieves over 90% attack success rate across four datasets and five models with a poison ratio of only 5%. Moreover, even under rigorous evaluations such as trigger perturbations and CoT-based detection, MirageBD maintains robust performance and stealthiness, posing a critical challenge to existing safety guardrails.

Computational narrative analysis aims to capture rhythm, tension, and emotional dynamics in literary texts. Existing large language models can generate long stories but overly focus on causal coherence, neglecting the complex story arcs and orchestration inherent in human narratives. This suggests a structural misalignment between model- and human-generated narratives.We therefore position narrative analysis as a diagnostic proxy for generation and propose VISTA Space, a high-dimensional framework for narrative orchestration that unifies human and model perspectives while jointly characterizing narrative function and structure in a common space.We further introduce LitVISTA, a structurally annotated benchmark grounded in literary texts, which operationalizes VISTA Space for systematic evaluation of models’ narrative orchestration capabilities. Under an oracle setting with gold event anchors, we evaluate frontier LLMs including GPT, Claude, Grok, and Gemini. Results reveal systematic deficiencies, as current models struggle to jointly capture narrative function and structure and fail to form an integrated global view of literary narrative orchestration. End-to-end analysis further shows that failures are dominated by anchor identification and localization errors. Even advanced thinking modes yield mixed and often limited gains for literary narrative understanding.

While Retrieval-Augmented Generation (RAG) systems are designed to enhance factual fidelity by grounding LLMs in provided sources, the application of current watermarking techniques creates a paradoxical failure mode. These methods, being inherently fact-agnostic, force the model to deviate from the very source documents it is supposed to follow. This leads to “faithfulness hallucinations"—a critical flaw where the generated output contradicts its own grounding context. Consequently, these watermarks undermine the core value of RAG, rendering even the most secure schemes untrustworthy for high-stakes applications. To resolve this RAG-specific conflict, we introduce the Dual Factual Shield (DFS) framework, a novel architecture designed to enforce knowledge loyalty. The DFS framework employs a defense-in-depth strategy through two synergistic layers: a source-anchored algorithmic safeguard that shields critical terms from the retrieved context, and prompt-based semantic guidance that protects against factual corruption. To demonstrate its effectiveness, we enhance a state-of-the-art, spoofing-aware contrastive watermarking baseline with our framework. Experiments show that our framework drastically reduces the Knowledge Corruption Rate (KCR)—a new metric we introduce—while preserving its original high security and robustness. This work establishes a new paradigm for watermarking, evolving it from merely secure to truly trustworthy. We demonstrate that traceability and truth can, and must, coexist, paving the way for the responsible deployment of traceable AI in knowledge-critical domains.

Reasoning-enhanced large language models rely on intermediate reasoning signals to solve complex, multi-step tasks, making reasoning behavior a valuable form of intellectual property. Meanwhile, knowledge distillation enables an adversary to replicate this behavior in a realistic black-box setting by repeatedly querying a deployed model on a target domain and training a local student to imitate its outputs, including reasoning traces. Existing LLM watermarks primarily operate on surface text and decoding-time token biases, and thus fail to provide reliable attribution of reasoning behavior once it is transferred through knowledge distillation. ReasMark entangles the watermark with the target-domain input distribution by selecting watermark tokens from high-frequency prompts, so distillation queries naturally activate it. It then embeds the watermark by score-conditioned losses that create a detectable reasoning-length gap for black-box verification. Comprehensive experiments across multiple LLMs, datasets, and distillation settings demonstrate that ReasMark consistently outperforms existing baselines while preserving task utility.

Large Reason Models (LRMs) extend long reasoning process to solve complex tasks. However, due to the lack of fine-grained control, they often suffer from overthinking and erroneous reasoning problems, risking accuracy loss. To address this issue, we introduce Reasoning Direction Steering (RDS) to enable fine-grained control over LRMs’ reasoning behaviors by aligning reasoning trajectories with specific cognitive patterns. We develop a simple yet effective paradigm, Thinking Intervention, which explores two key dimensions - intervention positions and intervention styles - to achieve integration intervention throughout model reasoning processes. To validate the effectiveness of our approach, we conduct comprehensive experiments on multi-hop question answering tasks using state-of-the-art LRMs, including Qwen3-Series and R1-Series models. Experimental results demonstrate the efficacy of Thinking Intervention with 9.4% average improvement on R1-Series models and 1.9% improvement on Qwen3-Series models.

Large Language Models (LLMs) have achieved impressive results in Machine Translation by simply following instructions, even without training on parallel data. However, LLMs still face challenges on low-resource languages due to the lack of pre-training data. In real-world situations, humans can become proficient in their native languages through abundant and meaningful social interactions and can also learn foreign languages effectively using well-organized textbooks. Drawing inspiration from human learning patterns, we introduce the Translate After LEarNing Textbook (TALENT) approach, which aims to enhance LLMs’ ability to translate low-resource languages by learning from a textbook. TALENT follows a step-by-step process: (1) Creating a Textbook for low-resource languages. (2) Guiding LLMs to absorb the Textbook’s content for Syntax Patterns. (3) Enhancing translation by utilizing the Textbook and Syntax Patterns. We thoroughly assess TALENT’s performance using 112 low-resource languages from FLORES-200 with two LLMs: ChatGPT and BLOOMZ. Evaluation across three different metrics reveals that TALENT consistently enhances translation performance by 14.8% compared to zero-shot baselines. Further analysis demonstrates that TALENT not only improves LLMs’ comprehension of low-resource languages but also equips them with the knowledge needed to generate accurate and fluent sentences in these languages.

Machine Translation task has made great progress with the help of auto-regressive decoding paradigm and Transformer architecture. In this paradigm, though the encoder can obtain global source representations, the decoder can only use translation history to determine the current word. Previous promising works attempted to address this issue by applying a draft or a fixed-length semantic embedding as target-side global information. However, these methods either degrade model efficiency or show limitations in expressing semantics. Motivated by Functional Equivalence Theory, we extract several semantic kernels from a source sentence, each of which can express one semantic segment of the original sentence. Together, these semantic kernels can capture global semantic information, and we project them into target embedding space to guide target sentence generation. We further force our model to use semantic kernels at each decoding step through an adaptive mask algorithm. Empirical studies on various machine translation benchmarks show that our approach gains approximately an improvement of 1 BLEU score on most benchmarks over the Transformer baseline and about 1.7 times faster than previous works on average at inference time.

Controllable story generation is a challenging task in the field of NLP, which has attracted increasing research interest in recent years. However, most existing works generate a whole story conditioned on the appointed keywords or emotions, ignoring the psychological changes of the protagonist. Inspired by psychology theories, we introduce global psychological state chains, which include the needs and emotions of the protagonists, to help a story generation system create more controllable and well-planned stories. In this paper, we propose a Psychology-guided Controllable Story Generation System (PICS) to generate stories that adhere to the given leading context and desired psychological state chains for the protagonist. Specifically, psychological state trackers are employed to memorize the protagonist’s local psychological states to capture their inner temporal relationships. In addition, psychological state planners are adopted to gain the protagonist’s global psychological states for story planning. Eventually, a psychology controller is designed to integrate the local and global psychological states into the story context representation for composing psychology-guided stories. Automatic and manual evaluations demonstrate that PICS outperforms baselines, and each part of PICS shows effectiveness for writing stories with more consistent psychological changes.

In this paper we introduce the systems IIE submitted for the WMT20 shared task on German-French news translation. Our systems are based on the Transformer architecture with some effective improvements. Multiscale collaborative deep architecture, data selection, back translation, knowledge distillation, domain adaptation, model ensemble and re-ranking are employed and proven effective in our experiments. Our German-to-French system achieved 35.0 BLEU and ranked the second among all anonymous submissions, and our French-to-German system achieved 36.6 BLEU and ranked the fourth in all anonymous submissions.