Deploying and fine-tuning Large Language Models (LLMs) on resource-constrained edge devices requires navigating a strict trade-off between memory footprint and task performance. Existing quantization-aware fine-tuning methods typically decouple weight precision and adapter capacity, overlooking that a layer’s ability to adapt is constrained by the information preserved in its frozen weights. Layers that are highly sensitive to quantization—whether due to representational specialization or accumulated error propagation—can become bottlenecks that adapter rank alone cannot recover. To address this issue, we introduce QR-Adaptor, a unified framework that jointly optimizes per-layer quantization bit-width and LoRA rank. We formulate resource allocation as a multi-objective discrete search guided by empirical layer-wise sensitivity, and implement it with a three-stage pipeline comprising KL-based sensitivity profiling, evolutionary exploration, and Bayesian refinement. Extensive experiments across LLaMA and Qwen models, including modern instruction tuning on OpenOrca and comparisons with strong PEFT baselines such as QDoRA, show that QR-Adaptor establishes a strong Pareto frontier: under a strict 4-bit memory budget, it matches or approaches 16-bit baselines while using substantially less memory.

While the massive scale of modern LLMs enables remarkable performance, their static, input-agnostic computational graph incurs substantial resource wastage and high latency during inference. Existing dynamic schemes, such as early-exit and layer-drop reduce FLOPs but break batch processing or introduce KV-cache inconsistency. We propose Deputy, a dynamic low-rank substitution framework that employs a lightweight decision module at each layer to dynamically determine the execution branch for different tokens: Attention layers choose between full and low-rank computation to mitigate the KV cache issue, while FFN layers additionally support skipping to further reduce computation. We fine-tune the LLM with LoRA and then derive an additional low-rank matrix C via a least-squares fit BC ≈ W_pre, where B is the shared LoRA matrix, so that only one extra low-rank matrix is introduced, effectively reducing memory overhead. Moreover, a hybrid KV cache strategy stores KV values generated by the low-rank branch, achieving a 38% reduction in cache storage. Experiments on Llama models demonstrate that Deputy reduces computation by approximately 40% compared to the original dense model while outperforming existing baseline methods.

Despite strong performance on code generation tasks, it remains unclear whether large language models (LLMs) genuinely reason about code execution. Existing code reasoning benchmarks primarily evaluate final output correctness under a single canonical implementation, leaving two critical aspects underexplored: (1) whether LLMs predictions are consistent to functionally equivalent implementations, and (2) whether LLMs can accurately reason about intermediate execution states. We introduce CoRE, a Code Reasoning benchmark that evaluates code reasoning through implementation invariance and process transparency. Extensive evaluations on eight frontier LLMs reveal two fundamental limitations. First, models exhibit a substantial robustness gap, with performance varying significantly across equivalent implementations. Second, we observe superficial execution, where models arrive at correct final outputs without correctly reasoning about intermediate execution states. Together, these findings demonstrate that output-only evaluations are insufficient for assessing code reasoning and position CoRE as a necessary benchmark for evaluating robust and faithful code reasoning.

Process Reward Models (PRMs), which assign fine-grained scores to intermediate reasoning steps within a solution trajectory, have emerged as a promising approach to enhance the reasoning quality of Large Language Models (LLMs).However, most existing PRMs rely on a unidirectional left-to-right (L2R) evaluation scheme, which restricts their utilization of global context.In light of this challenge, we propose a novel bidirectional evaluation paradigm, named Bidirectional 𝐏rocess 𝐑eward 𝐌odel (BiPRM).BiPRM incorporates a parallel right-to-left (R2L) evaluation stream, implemented via prompt reversal, alongside the conventional L2R flow.Then a gating mechanism is introduced to adaptively fuse the reward scores from both streams to yield a holistic quality assessment.Remarkably, compared to the original PRM, BiPRM introduces only a 0.3% parameter increase for the gating module, and the parallel execution of two streams incurs merely 5% inference time latency. Our extensive empirical evaluations spanning diverse benchmarks, LLM backbones, PRM objectives and sampling policies demonstrate that BiPRM consistently surpasses unidirectional baselines, achieving an average relative gain of 10.6% over 54 solution-level configurations and 37.7% in 12 step-level error detection scenarios.Generally, our results highlight the effectiveness, robustness and general applicability of BiPRM, offering a promising new direction for process-based reward modeling.

In-context learning (ICL) enhances the reasoning abilities of Large Language Models (LLMs) by prepending a few demonstrations. It motivates researchers to introduce more examples to provide additional contextual information for the generation. However, existing methods show a significant limitation due to the problem of excessive growth in context length which causes a large hardware burden. Additionally, shallow-relevant examples selected by out-off-shelf tools hinder LLMs from capturing useful contextual information for generation. In this paper, to approach these limitations, we propose UniICL, a novel Unified ICL framework that unifies demonstration compression, demonstration selection, and final response generation. Furthermore, to avoid repeated compression of the same demonstration and boost inference efficiency, we design a tailored compression strategy that allows UniICL caching compression results into Demonstration Bank(DB). Extensive out-of-domain evaluations prove the advantages of UniICL in both effectiveness and efficiency.

Large Vision-Language Models (LVLMs) suffer from serious hallucination problems, where the model-generated responses are inconsistent with the visual inputs. Existing hallucination mitigation methods are mainly based on preference alignment and require external human annotations or auxiliary models for preference data collection, which increase costs and limit sustainable improvement. To tackle these challenges, we propose Autonomous Preference Alignment via Self-Injection (APASI), a novel and generalizable method that mitigates hallucinations without external dependencies. APASI leverages the target LVLM to self-inject hallucinations into a generated response, creating a pair of responses with varying preference levels. During the self-injection process, the dis-preferred response is generated based on three key observations of hallucinations, ensuring it simulates real hallucination patterns. This fidelity offers an accurate learning signal for hallucination mitigation. Moreover, APASI incorporates an iterative alignment training strategy combined with curriculum learning to periodically update the preference data with increasing challenge, enabling stable and continuous enhancement of the LVLM. Extensive experiments across six benchmarks show that APASI not only effectively mitigates hallucinations for three baseline models but also achieves comparable or even superior performance to alignment-based methods with external dependency, thereby demonstrating its effectiveness and generalization capability.

Enhancing the attribution in large language models (LLMs) is a crucial task. One feasible approach is to enable LLMs to cite external sources that support their generations. However, existing datasets and evaluation methods in this domain still exhibit notable limitations. In this work, we formulate the task of attributed query-focused summarization (AQFS) and present WebCiteS, a Chinese dataset featuring 7k human-annotated summaries with citations. WebCiteS derives from real-world user queries and web search results, offering a valuable resource for model training and evaluation. Prior works in attribution evaluation do not differentiate between groundedness errors and citation errors. They also fall short in automatically verifying sentences that draw partial support from multiple sources. We tackle these issues by developing detailed metrics and enabling the automatic evaluator to decompose the sentences into sub-claims for fine-grained verification. Our comprehensive evaluation of both open-source and proprietary models on WebCiteS highlights the challenge LLMs face in correctly citing sources, underscoring the necessity for further improvement. The dataset and code will be open-sourced to facilitate further research in this crucial field.

The de facto way of utilizing black-box large language models (LLMs) to perform various downstream tasks is prompting. However, obtaining suitable prompts for specific tasks is still a challenging problem. While existing LLM-based methods demonstrate promising performance in task-oriented dialogue (TOD) task, they often require manual adjustment in prompt selection, or focus solely on dialogue understanding or generation. To address these issues, we propose an adaptive prompt generation framework to fully unleash the potential of LLMs for the comprehensive TOD system. Firstly, we design a trainable slot generator (TSG) that can generate domain and slot information in the belief state, which serves as prior knowledge for subsequent prompt generation. Next, we propose an adaptive prompt generator (APG) that utilizes the prior knowledge to generate prompts for the LLM, deriving the belief state and system response of the dialogue for evaluation. Finally, we evaluate our framework on the MultiWOZ 2.0 dataset. Extensive experiments demonstrate that our method outperforms existing methods. Our code and data will be released.

Event skeleton generation, aiming to induce an event schema skeleton graph with abstracted event nodes and their temporal relations from a set of event instance graphs, is a critical step in the temporal complex event schema induction task. Existing methods effectively address this task from a graph generation perspective but suffer from noise-sensitive and error accumulation, e.g., the inability to correct errors while generating schema. We, therefore, propose a novel Diffusion Event Graph Model (DEGM) to address these issues. Our DEGM is the first workable diffusion model for event skeleton generation, where the embedding and rounding techniques with a custom edge-based loss are introduced to transform a discrete event graph into learnable latent representations. Furthermore, we propose a denoising training process to maintain the model’s robustness. Consequently, DEGM derives the final schema, where error correction is guaranteed by iteratively refining the latent representations during the schema generation process. Experimental results on three IED bombing datasets demonstrate that our DEGM achieves better results than other state-of-the-art baselines. Our code and data are available at https://github.com/zhufq00/EventSkeletonGeneration.

We present Twin Answer Sentences Attack (TASA), an adversarial attack method for question answering (QA) models that produces fluent and grammatical adversarial contexts while maintaining gold answers. Despite phenomenal progress on general adversarial attacks, few works have investigated the vulnerability and attack specifically for QA models. In this work, we first explore the biases in the existing models and discover that they mainly rely on keyword matching between the question and context, and ignore the relevant contextual relations for answer prediction.Based on two biases above, TASA attacks the target model in two folds: (1) lowering the model’s confidence on the gold answer with a perturbed answer sentence; (2) misguiding the model towards a wrong answer with a distracting answer sentence. Equipped with designed beam search and filtering methods, TASA can generate more effective attacks than existing textual attack methods while sustaining the quality of contexts, in extensive experiments on five QA datasets and human evaluations.

Event extraction (EE) is crucial to downstream tasks such as new aggregation and event knowledge graph construction. Most existing EE datasets manually define fixed event types and design specific schema for each of them, failing to cover diverse events emerging from the online text. Moreover, news titles, an important source of event mentions, have not gained enough attention in current EE research. In this paper, we present Title2Event, a large-scale sentence-level dataset benchmarking Open Event Extraction without restricting event types. Title2Event contains more than 42,000 news titles in 34 topics collected from Chinese web pages. To the best of our knowledge, it is currently the largest manually annotated Chinese dataset for open event extraction. We further conduct experiments on Title2Event with different models and show that the characteristics of titles make it challenging for event extraction, addressing the significance of advanced study on this problem. The dataset and baseline codes are available at https://open-event-hub.github.io/title2event.

We present Mask-then-Fill, a flexible and effective data augmentation framework for event extraction. Our approach allows for more flexible manipulation of text and thus can generate more diverse data while keeping the original event structure unchanged as much as possible. Specifically, it first randomly masks out an adjunct sentence fragment and then infills a variable-length text span with a fine-tuned infilling model. The main advantage lies in that it can replace a fragment of arbitrary length in the text with another fragment of variable length, compared to the existing methods which can only replace a single word or a fixed-length fragment. On trigger and argument extraction tasks, the proposed framework is more effective than baseline methods and it demonstrates particularly strong results in the low-resource setting. Our further analysis shows that it achieves a good balance between diversity and distributional similarity.

We present IBR, an Iterative Backward Reasoning model to solve the proof generation tasks on rule-based Question Answering (QA), where models are required to reason over a series of textual rules and facts to find out the related proof path and derive the final answer. We handle the limitations of existed works in two folds: 1) enhance the interpretability of reasoning procedures with detailed tracking, by predicting nodes and edges in the proof path iteratively backward from the question; 2) promote the efficiency and accuracy via reasoning on the elaborate representations of nodes and history paths, without any intermediate texts that may introduce external noise during proof generation. There are three main modules in IBR, QA and proof strategy prediction to obtain the answer and offer guidance for the following procedure; parent node prediction to determine a node in the existing proof that a new child node will link to; child node prediction to find out which new node will be added to the proof. Experiments on both synthetic and paraphrased datasets demonstrate that IBR has better in-domain performance as well as cross-domain transferability than several strong baselines. Our code and models are available at https://github.com/find-knowledge/IBR.

The memes serve as an important tool in online communication, whereas some hateful memes endanger cyberspace by attacking certain people or subjects. Recent studies address hateful memes detection while further understanding of relationships of entities in memes remains unexplored. This paper presents our work at the Constraint@ACL2022 Shared Task: Hero, Villain and Victim: Dissecting harmful memes for semantic role labelling of entities. In particular, we propose our approach utilizing transformer-based multimodal models through a VCR method with data augmentation, continual pretraining, loss re-weighting, and ensemble learning. We describe the models used, the ways of preprocessing and experiments implementation. As a result, our best model achieves the Macro F1-score of 54.707 on the test set of this shared task.

Representations of events described in text are important for various tasks. In this work, we present SWCC: a Simultaneous Weakly supervised Contrastive learning and Clustering framework for event representation learning. SWCC learns event representations by making better use of co-occurrence information of events. Specifically, we introduce a weakly supervised contrastive learning method that allows us to consider multiple positives and multiple negatives, and a prototype-based clustering method that avoids semantically related events being pulled apart. For model training, SWCC learns representations by simultaneously performing weakly supervised contrastive learning and prototype-based clustering. Experimental results show that SWCC outperforms other baselines on Hard Similarity and Transitive Sentence Similarity tasks. In addition, a thorough analysis of the prototype-based clustering method demonstrates that the learned prototype vectors are able to implicitly capture various relations between events.

Current approaches to empathetic response generation focus on learning a model to predict an emotion label and generate a response based on this label and have achieved promising results. However, the emotion cause, an essential factor for empathetic responding, is ignored. The emotion cause is a stimulus for human emotions. Recognizing the emotion cause is helpful to better understand human emotions so as to generate more empathetic responses. To this end, we propose a novel framework that improves empathetic response generation by recognizing emotion cause in conversations. Specifically, an emotion reasoner is designed to predict a context emotion label and a sequence of emotion cause-oriented labels, which indicate whether the word is related to the emotion cause. Then we devise both hard and soft gated attention mechanisms to incorporate the emotion cause into response generation. Experiments show that incorporating emotion cause information improves the performance of the model on both emotion recognition and response generation.

Sentence function is an important linguistic feature referring to a user’s purpose in uttering a specific sentence. The use of sentence function has shown promising results to improve the performance of conversation models. However, there is no large conversation dataset annotated with sentence functions. In this work, we collect a new Short-Text Conversation dataset with manually annotated SEntence FUNctions (STC-Sefun). Classification models are trained on this dataset to (i) recognize the sentence function of new data in a large corpus of short-text conversations; (ii) estimate a proper sentence function of the response given a test query. We later train conversation models conditioned on the sentence functions, including information retrieval-based and neural generative models. Experimental results demonstrate that the use of sentence functions can help improve the quality of the returned responses.

Neural conversation models such as encoder-decoder models are easy to generate bland and generic responses. Some researchers propose to use the conditional variational autoencoder (CVAE) which maximizes the lower bound on the conditional log-likelihood on a continuous latent variable. With different sampled latent variables, the model is expected to generate diverse responses. Although the CVAE-based models have shown tremendous potential, their improvement of generating high-quality responses is still unsatisfactory. In this paper, we introduce a discrete latent variable with an explicit semantic meaning to improve the CVAE on short-text conversation. A major advantage of our model is that we can exploit the semantic distance between the latent variables to maintain good diversity between the sampled latent variables. Accordingly, we propose a two-stage sampling approach to enable efficient diverse variable selection from a large latent space assumed in the short-text conversation task. Experimental results indicate that our model outperforms various kinds of generation models under both automatic and human evaluations and generates more diverse and informative responses.

Sequence-to-sequence neural generation models have achieved promising performance on short text conversation tasks. However, they tend to generate generic/dull responses, leading to unsatisfying dialogue experience. We observe that in the conversation tasks, each query could have multiple responses, which forms a 1-to-n or m-to-n relationship in the view of the total corpus. The objective function used in standard sequence-to-sequence models will be dominated by loss terms with generic patterns. Inspired by this observation, we introduce a statistical re-weighting method that assigns different weights for the multiple responses of the same query, and trains the common neural generation model with the weights. Experimental results on a large Chinese dialogue corpus show that our method improves the acceptance rate of generated responses compared with several baseline models and significantly reduces the number of generated generic responses.