Multi-turn jailbreak attacks pose a growing threat to LLMs by exploiting gradual escalation and cross-turn coordination. Many existing defenses either rely on costly retraining—often degrading utility—or analyze each turn independently, failing to capture risk accumulation across interactions. We observe that safety behavior in multi-turn interaction is trajectory-dependent: dialogue history reshapes the model’s conditioning context, making isolated turn-level evaluation insufficient. We present THRD, a training-free framework that explicitly models temporal risk accumulation for multi-turn jailbreak defense. THRD integrates four modules: a Turn-level Risk Assessor (TRA) for instantaneous risk estimation, a Historical Context Analyzer (HCA) for cross-turn intent escalation detection, a Response Evaluator (RE) for identifying facilitative outputs, and a Decision Module that combines these signals through a time-evolving scoring mechanism with attenuation-based modulation and trend-aware adjustment. Experiments against representative multi-turn attacks—including tree-search-based and multi-agent collaborative methods—across two target models show that THRD reduces ASR to 0.2–4.0% while preserving model utility within 1.5% degradation on MMLU and GSM8K. Ablation studies confirm non-redundant module contributions and consistent performance across both models. Over 70% of attacks require Turn 2 or later to detect, highlighting the importance of cross-turn information.

Test-Time Scaling has shown notable efficacy in addressing complex problems through scaling inference compute. However, within Large Audio-Language Models (LALMs), an unintuitive phenomenon exists: post-training models for structured reasoning trajectories results in marginal or even negative gains compared to training for direct answering. To investigate it, we introduce CAFE, an evaluation framework designed to quantify audio reasoning errors. Evaluation results reveal that LALMs encounter a critical bottleneck: Audio Perception Decay, which compromises reasoning performance. To address it, we propose MPAR², a strategy that encourages a step-by-step reasoning process with rich perceptual awareness. With reinforcement learning, MPAR² improves performance on CAFE from 31.74% to 63.51% and effectively mitigates audio perception decay, concurrently enhancing reasoning ability to achieve a strong 74.59% accuracy on the MMAU benchmark. Further analysis shows that MPAR² optimizes audio attention and enhances reasoning token budget efficiency.

Extended chain-of-thought (CoT) traces improve LLM reasoning but incur substantial computational and memory costs. While existing CoT compression methods mitigate this by condensing thought steps into compact representations via memory tokens and retaining only these representations at inference time, the loss of fine-grained information makes subsequent steps more error-prone. To alleviate this, we propose HybridThinker, where in addition to preserving these representations, thought steps are also temporarily retained to provide fine-grained details. However, we observe that naively keeping thought steps accessible to subsequent steps during training lets the model bypass memory tokens by retrieving information directly from these steps, leaving the model’s ability to compress and retrieve information through memory tokens insufficiently trained. We therefore introduce a hybrid training scheme, in which only some thought steps are directly accessible through attention to subsequent steps, while the other thought steps are masked, forcing the model to use memory tokens for compression and retrieval. Across four reasoning benchmarks, HybridThinker matches Vanilla on Qwen2.5-7B and approaches it on Llama3.1-8B, improving over LightThinker by 5.8 and 2.7 points, respectively, with similar inference time. Ablation studies confirm that both temporary thought-step retention and the hybrid training scheme contribute to these gains.

This paper addresses the efficiency challenge of Neural Architecture Search (NAS) by formulating the task as a ranking problem. Previous methods require numerous training examples to estimate the accurate performance of architectures, although the actual goal is to find the distinction between “good” and “bad” candidates. Here we do not resort to performance predictors. Instead, we propose a performance ranking method (RankNAS) via pairwise ranking. It enables efficient architecture search using much fewer training examples. Moreover, we develop an architecture selection method to prune the search space and concentrate on more promising candidates. Extensive experiments on machine translation and language modeling tasks show that RankNAS can design high-performance architectures while being orders of magnitude faster than state-of-the-art NAS systems.

Traditional neural machine translation is limited to the topmost encoder layer’s context representation and cannot directly perceive the lower encoder layers. Existing solutions usually rely on the adjustment of network architecture, making the calculation more complicated or introducing additional structural restrictions. In this work, we propose layer-wise multi-view learning to solve this problem, circumventing the necessity to change the model structure. We regard each encoder layer’s off-the-shelf output, a by-product in layer-by-layer encoding, as the redundant view for the input sentence. In this way, in addition to the topmost encoder layer (referred to as the primary view), we also incorporate an intermediate encoder layer as the auxiliary view. We feed the two views to a partially shared decoder to maintain independent prediction. Consistency regularization based on KL divergence is used to encourage the two views to learn from each other. Extensive experimental results on five translation tasks show that our approach yields stable improvements over multiple strong baselines. As another bonus, our method is agnostic to network architectures and can maintain the same inference speed as the original model.

Neural architecture search (NAS) has advanced significantly in recent years but most NAS systems restrict search to learning architectures of a recurrent or convolutional cell. In this paper, we extend the search space of NAS. In particular, we present a general approach to learn both intra-cell and inter-cell architectures (call it ESS). For a better search result, we design a joint learning method to perform intra-cell and inter-cell NAS simultaneously. We implement our model in a differentiable architecture search system. For recurrent neural language modeling, it outperforms a strong baseline significantly on the PTB and WikiText data, with a new state-of-the-art on PTB. Moreover, the learned architectures show good transferability to other systems. E.g., they improve state-of-the-art systems on the CoNLL and WNUT named entity recognition (NER) tasks and CoNLL chunking task, indicating a promising line of research on large-scale pre-learned architectures.

In encoder-decoder neural models, multiple encoders are in general used to represent the contextual information in addition to the individual sentence. In this paper, we investigate multi-encoder approaches in document-level neural machine translation (NMT). Surprisingly, we find that the context encoder does not only encode the surrounding sentences but also behaves as a noise generator. This makes us rethink the real benefits of multi-encoder in context-aware translation - some of the improvements come from robust training. We compare several methods that introduce noise and/or well-tuned dropout setup into the training of these encoders. Experimental results show that noisy training plays an important role in multi-encoder-based NMT, especially when the training data is small. Also, we establish a new state-of-the-art on IWSLT Fr-En task by careful use of noise generation and dropout methods.

This paper describes the Kingsoft AI Lab’s submission to the WMT2019 news translation shared task. We participated in two language directions: English-Chinese and Chinese-English. For both language directions, we trained several variants of Transformer models using the provided parallel data enlarged with a large quantity of back-translated monolingual data. The best translation result was obtained with ensemble and reranking techniques. According to automatic metrics (BLEU) our Chinese-English system reached the second highest score, and our English-Chinese system reached the second highest score for this subtask.

Chinese idioms (Cheng Yu) have seen five thousand years’ history and culture of China, meanwhile they contain large number of scientific achievement of ancient China. However, existing Chinese online idiom dictionaries have limited function for scientific exploration. In this paper, we first construct a Chinese idiom knowledge graph by extracting domains and dynasties and associating them with idioms, and based on the idiom knowledge graph, we propose a Science Toolkit for Ancient China (STAC) aiming to support scientific exploration. In the STAC toolkit, idiom navigator helps users explore overall scientific progress from idiom perspective with visualization tools, and idiom card and idiom QA shorten action path and avoid thinking being interrupted while users are reading and writing. The current STAC toolkit is deployed at http://120.92.208.22:7476/demo/#/stac.

Transformer is the state-of-the-art model in recent machine translation evaluations. Two strands of research are promising to improve models of this kind: the first uses wide networks (a.k.a. Transformer-Big) and has been the de facto standard for development of the Transformer system, and the other uses deeper language representation but faces the difficulty arising from learning deep networks. Here, we continue the line of research on the latter. We claim that a truly deep Transformer model can surpass the Transformer-Big counterpart by 1) proper use of layer normalization and 2) a novel way of passing the combination of previous layers to the next. On WMT’16 English-German and NIST OpenMT’12 Chinese-English tasks, our deep system (30/25-layer encoder) outperforms the shallow Transformer-Big/Base baseline (6-layer encoder) by 0.4-2.4 BLEU points. As another bonus, the deep model is 1.6X smaller in size and 3X faster in training than Transformer-Big.

Chinese Grammatical Error Diagnosis (CGED) is a natural language processing task for the NLPTEA2018 workshop held during ACL2018. The goal of this task is to diagnose Chinese sentences containing four kinds of grammatical errors through the model and find out the sentence errors. Chinese grammatical error diagnosis system is a very important tool, which can help Chinese learners automatically diagnose grammatical errors in many scenarios. However, due to the limitations of the Chinese language’s own characteristics and datasets, the traditional model faces the problem of extreme imbalances in the positive and negative samples and the disappearance of gradients. In this paper, we propose a sequence labeling method based on the Policy Gradient LSTM model and apply it to this task to solve the above problems. The results show that our model can achieve higher precision scores in the case of lower False positive rate (FPR) and it is convenient to optimize the model on-line.

Spoken Language Understanding (SLU), which typically involves intent determination and slot filling, is a core component of spoken dialogue systems. Joint learning has shown to be effective in SLU given that slot tags and intents are supposed to share knowledge with each other. However, most existing joint learning methods only consider joint learning by sharing parameters on surface level rather than semantic level. In this work, we propose a novel self-attentive model with gate mechanism to fully utilize the semantic correlation between slot and intent. Our model first obtains intent-augmented embeddings based on neural network with self-attention mechanism. And then the intent semantic representation is utilized as the gate for labelling slot tags. The objectives of both tasks are optimized simultaneously via joint learning in an end-to-end way. We conduct experiment on popular benchmark ATIS. The results show that our model achieves state-of-the-art and outperforms other popular methods by a large margin in terms of both intent detection error rate and slot filling F1-score. This paper gives a new perspective for research on SLU.

Multi-choice question answering in exams is a typical QA task. To accomplish this task, we present an answer localization method to locate answers shown in web pages, considering structural information and semantic information both. Using this method as basis, we analyze sentences and paragraphs appeared on web pages to get predictions. With this answer localization system, we get effective results on both validation dataset and test dataset.

Corpus is a valuable resource for information retrieval and data-driven natural language processing systems,especially for spoken dialogue research in specific domains. However,there is little non-English corpora, particular for ones in Chinese. Spoken by the nation with the largest population in the world, Chinese become increasingly prevalent and popular among millions of people worldwide. In this paper, we build a large-scale and high-quality Chinese corpus, called CSDC (Chinese Spoken Dialogue Corpus). It contains five domains and more than 140 thousand dialogues in all. Each sentence in this corpus is annotated with slot information additionally compared to other corpora. To our best knowledge, this is the largest Chinese spoken dialogue corpus, as well as the first one with slot information. With this corpus, we proposed a method and did a well-designed experiment. The indicative result is reported at last.