text
string
source
string
reading), along with its adjacent papers. Your task is to decide what to read next within this local graph structure. You can choose to continue reading the current paper or move to aconnected paper (essentially making ajump inthe graph). You also need to explicitly maintain a memory of limited size (4096 tokens). Each...
https://arxiv.org/abs/2505.19647v1
a citation graph of the papers, and your role is to act as an intelligent agent traversing this graph. At each step, you will know your current position in the graph (the paper you are currently reading), along with the papers cited by this paper and those that cite it. Your task is to decide what to read next within t...
https://arxiv.org/abs/2505.19647v1
do not respond with any otheradditional content!Figure 8: Prompt for Selector -Graph-Ci. Prompt for Reader SystemPrompt: You arearesearchworkerwith excellent paperreading skills. UserPrompt: I am writing a scientific paper. Now I need to cite some reference papers and write the Related Work section for the paper. Given...
https://arxiv.org/abs/2505.19647v1
the order of the provided papers is random, so you need to reorganize the order based on the relationship between the papers. The Json information ofcited papers is as follows: {The Json information ofall the cited papers.} In order to get more detailed information about the papers, one of your peers has read the full ...
https://arxiv.org/abs/2505.19647v1
Work' is somewhat on topic but with several digressions; the core subject is evident but not consistently adhered to. - Score 3: The'Related Work' is generally on topic, despitea fewunrelated details. - Score 4: The 'Related Work' is mostly on topic and focused; the narrative has a consistent relevance to the core subj...
https://arxiv.org/abs/2505.19647v1
arXiv:2505.19660v1 [cs.CL] 26 May 2025GenKI: Enhancing Open-Domain Question Answering with Knowledge Integration and Controllable Generation in Large Language Models Tingjia Shen University of Science and Technology of China Hefei, Anhui, China jts_stj@mail.ustc.edu.cnHao Wangβˆ— University of Science and Technology of C...
https://arxiv.org/abs/2505.19660v1
the rapid development of Large Lan- guage Models (LLMs) [ 18,24,50,78] has motivated researchers to incorporate them into OpenQA, driven by the models’ advanced abilities in natural language reasoning. For instance, ICL [ 34] con- tains to use certain prompts towards better integration of external knowledge with in-con...
https://arxiv.org/abs/2505.19660v1
OpenQA field, current pre-training or fine-tuning-based methods only focus on either full knowledge adaption or applying LLM on both reasoning and answer alignment missions, considering two different kinds of task in one stage of tuning. This leads to distribution misalignment and limitations in sufficient knowledge in...
https://arxiv.org/abs/2505.19660v1
role in generating answers in a specific format, and the reward model excelling in choosing fluent sentences. The main contributions are summarized as follows: β€’We propose a novel problem on the phenomena of knowledge deficiency and answer format misalignment of LLMs on the OpenQA task. β€’We introduce a novel three-stag...
https://arxiv.org/abs/2505.19660v1
RoBERTa [ 36] to build a reader aligned to answering tasks [ 26,67,74]. Following the occurrence of genera- tive language models, such as GPT [ 7], GLM [ 15], LLaMA [ 55], the researchers began adopting these models for OpenQA. This shiftwas driven by their proven capability to handle OpenQA without relying on retrieve...
https://arxiv.org/abs/2505.19660v1
𝑃. 1https://www.llamaindex.ai WWW ’25, Appril 28–May 02, 2025, Sydney, SYD Anonymous Author, et al. Take, for instance, the question π‘žπ‘–: β€œWhere was the initial aware- ness of the Chernobyl incident triggered? ”. Given that the desired format𝐹is Entity Style, the model is expected to leverage insights extracted from...
https://arxiv.org/abs/2505.19660v1
knowledge. 4.2 Knowledge Integration From the previous part, we acquired high-quality knowledge uti- lizing the retriever. To integrate this high-quality knowledge into the parameters of our model, tuning is an effective way. However, the original instruction tuning is inadequate for the OpenQA task, which is attribute...
https://arxiv.org/abs/2505.19660v1
the latter was a flop. Matthau ... Question: What was Walter Matthau's first movie? Output: The Kentuckian Controllabe Generation Instruct Prompt III: TriviaQA,CMRC : MSMARCO :Instruction: Please extract the answer keyword. Input: Question: What was Walter Matthau's first movie? Answer: The Kentuckian movie in 1950s. O...
https://arxiv.org/abs/2505.19660v1
these two answers are severely dif- ferent. The pair-wise loss of the reward model chooses the best answer between two answers, formulated as below: Lπ‘…π‘š=βˆ’π‘™π‘œπ‘”(𝜎(π‘…π‘š(𝐴+,𝐹)βˆ’π‘…π‘š(π΄βˆ’,𝐹)), (6) whereπ‘…π‘š(𝐴)is the average score of the embedding of the final token of A. However, relying solely on the reward model en...
https://arxiv.org/abs/2505.19660v1
stages in Figure 3. This visualization allows us to observe GenKI’sAlgorithm 1 Overall pseudo code of GenKI Input: Knowledge base 𝐾, Retriever𝑅, LLM𝐿, question set 𝑄 Output: Answer𝐴 1:Train𝐿1and𝐿2with𝐾and𝐾𝑅=𝑅(𝐾,𝑄)respectively 2: ⊲Prompt I, II is used here 3:Gain𝑃3using𝐿2answering training queries 4:Train...
https://arxiv.org/abs/2505.19660v1
This is because comparing our model with methods having a significantly larger scale of parame- ters would be unfair, as they require more time and computational resources for training than we do. These three methods include: β€’ChatGPT leverages a transformer-based neural network to understand and generate human languag...
https://arxiv.org/abs/2505.19660v1
enabling models to generate fluent, extended content, we utilize BLEU and ROUGE metrics to evaluate the quality of answers in the MSMARCO dataset. Moreover, to gauge the fluency of model out- puts, we employ the Coherence values derived from CTRLEval[ 31]. We categorize these metrics into two categories, K and C, where...
https://arxiv.org/abs/2505.19660v1
EM-score.Table 3: OpenQA performance of BLEU and ROUGE on MS- MARCO dataset. Baseline TypeMethod BLEU-1 ROUGE-L Task Specific BaselinesMasque NLGEN 0.501 0.496 PALM 0.499 0.498 REAG 0.497 0.498 Latest BaselinesRALLM-KB 0.262 0.241 LLaMA2-index 0.588 0.447 GARFiD 0.561 0.454 Self-RAG 0.567 0.351 LLM BackbonesLLaMA-65B οΏ½...
https://arxiv.org/abs/2505.19660v1
passes baseline performances, demonstrating superiority by 3.3% in TriviaQA, 0.8% in MSMARCO, and 4.9% in CMRC. The comparative results are illustrated in Figure 5. 6.1.2 Can GenKI adapt to different formats of answers? To demon- strate the adaptability of our model to another format domain, we partition the CMRC datas...
https://arxiv.org/abs/2505.19660v1
infer two pieces of information: (1) The relationship between retrieval quality and model recall initially exhibited a linear trend across these examples ( 𝑅2>0.99), but eventually reached a bottleneck state, transitioning into another linear relationship ( 𝑅2>0.985). This suggests that the model has been extensively...
https://arxiv.org/abs/2505.19660v1
than 𝑋when using GLM-6B but worse when using LLaMa-65B. This is also why our work designs a novel three-stage paradigm of retrieval, knowledge integration, and con- trollable generation, avoiding the distributional differences of LLM in the two-stage RAG. Next, this paper will analyze the controllable generation modul...
https://arxiv.org/abs/2505.19660v1
A novel approach to open-domain question answering. arXiv preprint arXiv:2307.11278 , 2023. [2]Akiko Aizawa. An information-theoretic perspective of tf–idf measures. Information Processing &Management, 39(1):45–65, 2003. [3]Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self- rag: Learning to ...
https://arxiv.org/abs/2505.19660v1
[14] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference onMachine Learning, pages 5547–5569. PMLR, 2022. [15] Zhengxiao Du, Yujie Qian,...
https://arxiv.org/abs/2505.19660v1
Association for Computational Linguistics. [28] Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. Large language models struggle to learn long-tail knowledge. In International Conference onMachine Learning, pages 15696–15707. PMLR, 2023. [29] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick...
https://arxiv.org/abs/2505.19660v1
Computing Surveys, 2021. [41] Sewon Min, Danqi Chen, Luke Zettlemoyer, and Hannaneh Hajishirzi. Knowledge guided text retrieval and reading for open domain question answering. CoRR , abs/1911.03868, 2019. [42] Hoang H Nguyen, Ye Liu, Chenwei Zhang, Tao Zhang, and Philip S Yu. Cof-cot: Enhancing large language models wi...
https://arxiv.org/abs/2505.19660v1
models. arXiv preprint arXiv:2302.13971, 2023. [56] Cunxiang Wang, Haofei Yu, and Yue Zhang. Rfid: Towards rational fusion-in- decoder for open-domain question answering. arXiv preprint arXiv:2305.17041 , 2023. [57] Hao Wang, Wei Guo, Luankang Zhang, Jin Yao Chin, Yufei Ye, Huifeng Guo, Yong Liu, Defu Lian, Ruiming Tan...
https://arxiv.org/abs/2505.19660v1
Transforming sequential recommendations with schrΓΆdinger bridge and diffusion models. In Proceedings ofthe33rd ACM International Conference onInformation andKnowledge Management, pages 2618–2628, 2024. [70] Xiang Xu, Hao Wang, Wei Guo, Luankang Zhang, Wanshan Yang, Runlong Yu, Yong Liu, Defu Lian, and Enhong Chen. Mult...
https://arxiv.org/abs/2505.19660v1
arXiv:2505.19667v1 [cs.CL] 26 May 2025LeCoDe: A Benchmark Dataset for Interactive Legal Consultation Dialogue Evaluation Weikang Yuan1,2βˆ—, Kaisong Song2, Zhuoren Jiang1†, Junjie Cao2, Yujie Zhang1,Jun Lin2,Kun Kuang1,Ji Zhang2,Xiaozhong Liu3 1Zhejiang University, China,2Tongyi Lab, Alibaba Group, China,3Worcester Polyt...
https://arxiv.org/abs/2505.19667v1
attitude during police questioning, [3] confirm if he qualifies as a recidivist. Once these facts are clear, consider applying for bail . My friend is involved in a theft criminal case and needs consultation. He was detained yesterday . What specific items were stolen? Car ties. Idon’t know. What brand of tires were st...
https://arxiv.org/abs/2505.19667v1
careful video processing and dialogue transcription, 3While we demonstrate these challenges in the legal domain, similar complexities exist across various consultation scenarios including medical and other professional domains [10, 24]. 2 followed by multi-stage annotations conducted by a team of professional annotator...
https://arxiv.org/abs/2505.19667v1
di- verse areas [ 4,5,14], and their interactive nature with users shows significant potential in consultation settings [ 24]. However, a critical challenge emerges in consultation scenarios: users often provide vague queries, requiring LLMs to possess robust clarification abilities to address this information gap [10,...
https://arxiv.org/abs/2505.19667v1
is reached. Simulation Framework : To simulate real-world consultation interactions, the expert agent can be represented by various LLMs and strategies to generate clarifying questions or final advice based on the user’s initial query and subsequent feedback. To ensure reliable user simulation, we employ an LLM-based u...
https://arxiv.org/abs/2505.19667v1
interaction strategies. In the evaluation phase, it serves as a means to assess the clarification capacity. Importance scoring of key facts annotated the role of facts in legal analysis. The facts are categorized into three levels: Critical Facts (3 points), Secondary Facts (2 points), and Non-critical Facts (1 point)....
https://arxiv.org/abs/2505.19667v1
presents the word count distribution between participants. Interestingly, users’ contributions are comparable to, or slightly exceed, those of lawyers. This pattern reflects the realistic nature of legal consultations. Clients often have limited legal knowledge and are unable to present comprehensive, well-structured r...
https://arxiv.org/abs/2505.19667v1
Consideration Considering the sensitivity of legal domain, we adopted strict measures to ensure ethical compliance and minimize potential risks. First, our dataset exclusively consists of publicly shared educational videos on Douyin, where all videos have undergone strict platform moderation and ethical review, ensurin...
https://arxiv.org/abs/2505.19667v1
other approaches across all clarification metrics, achieving remarkable Recall (53.8%) and NDCG (84.8%) scores. Table 3: Main results on zero-shot setting, where top-6 scores are marked in blue , the highest is bolded and the second-highest is underlined . ModelsClarification Capability Advice Quality Effe. Effi. Auto....
https://arxiv.org/abs/2505.19667v1
Case Complexity on Model Performance Metrics: Analysis of Weighted Recall, Ask Turn, BERTScore, and Overall Score (OA) across Different Numbers of Atomic Key Facts. Recall Weighted Recall Recall@5 NDCGEfficiency (1 / turn) 0.20.50.81.0 1. Clarification Capability Professional Fluency Completeness SatisfactionSafety 0.2...
https://arxiv.org/abs/2505.19667v1
legal product powered by alibaba cloud. https://tongyi.aliyun. com/farui/home . 2025.5.12. [4]Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 , 202...
https://arxiv.org/abs/2505.19667v1
measuring legal reasoning in large language models. In Proceedings of the 10 37th International Conference on Neural Information Processing Systems , NIPS ’23, Red Hook, NY , USA, 2023. Curran Associates Inc. [14] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wa...
https://arxiv.org/abs/2505.19667v1
Ferreira Manso, and Roland Mathis. Reliable LLM-based user simulator for task-oriented dialogue systems. In Yvette Graham, Qun Liu, Gerasimos Lampouras, Ignacio Iacobacci, Sinead Madden, Haider Khalid, and Rameez Qureshi, editors, Proceedings of the 1st Workshop on Simulating Conversational Intelligence in Chat (SCI-CH...
https://arxiv.org/abs/2505.19667v1
. . . . . . . . . . . . . . . . . . . . . 14 B.2 Broader Impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 C User Agent 14 C.1 Evaluation Metrics for User Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 C.2 Evaluation Setting for User Agent . . . . . . . . . . . . . ...
https://arxiv.org/abs/2505.19667v1
. . . . . . . . . . 23 D.2.4 SFT Strategies Prompt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 E Details of Annotation Process 25 E.1 Guidelines for Annotation Stage-1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 E.2 Guidelines for Annotation Stage-2 . . . . . . . . . . . . . . . ...
https://arxiv.org/abs/2505.19667v1
in professional consultation scenarios. B.2 Broader Impact We propose LeCoDe to help practitioners better understand and evaluate LLMs’ performance in legal consultation scenarios, providing foundational data and evaluation framework for future research. However, given the sensitive and high-risk nature of the legal do...
https://arxiv.org/abs/2505.19667v1
who understands user-related information and attempts to answer lawyers’ questions about users. The following is [Atomic Key Facts] describing user information {Atomic Key Facts} The following is the interactive dialogue between you (user) and the lawyer {current dialogue} - Use the above key fact to answer the lawyer’...
https://arxiv.org/abs/2505.19667v1
it indeed cannot be answered, then the relevance of replying "Don’t know" should be 1; otherwise, it’s 0. [Dialogue Context] {dialogue} [Atomic Key Facts] {Atomic Key Facts} [Clarifying Question] {lawyer question} [User Response] {User Response} LLM-as-a-judge for assessing factuality (translated from Chinese) You will...
https://arxiv.org/abs/2505.19667v1
information-seeking questions, to our legal consultation scenario. While the original framework was developed for medical multiple-choice questions, we modify and retain three key components: the Abstention module, Ask module, and a Suggestion module for legal advice generation. The workflow operates as follows: The Ab...
https://arxiv.org/abs/2505.19667v1
Γ—4 gradient accumulation steps). The optimization process uses a cosine learning rate schedule initialized at 1e-5 with 10% warmup ratio over 3 epochs. 19 D.2 Lawyer Agent Prompt D.2.1 Zero-shot Prompt Lawyer Agent Zero-shot Prompt for ask or advice (translated from Chinese) You are a lawyer well-versed in Chinese law,...
https://arxiv.org/abs/2505.19667v1
question” or β€œadvice:legal advice” 21 Lawyer Agent Few-shot Prompt for advice generation (translated from Chinese) You are a lawyer well-versed in Chinese law, responsible for providing legal consultation to users. Your task is to provide legal advice. Below is two demonstrations: <demo1> {demo1} <demo2> {demo2} Below ...
https://arxiv.org/abs/2505.19667v1
30,000 yuan constitutes a crime. In handling this matter, you can negotiate with him, clearly state the evidence you have, and present your conditions, but immediate legal action isn’t necessary. If you decide to take further steps, prepare materials like a PPT presentation and send the cover first, indicating you’ve s...
https://arxiv.org/abs/2505.19667v1
generate target questions (translated from Chinese) You are a lawyer well-versed in Chinese law, responsible for providing legal consultation to users. I will provide a list of key information points that need to be obtained from clients but are currently unknown. Please design a series of professional, guiding questio...
https://arxiv.org/abs/2505.19667v1
has preliminarily extracted Atomic Key Facts List. Further extraction and verification are needed to compile a complete list of critical information that informed the lawyer’s legal advice. Requirements: 1.Extract key information points from client-lawyer dialogue. Keep expressions consistent with original text where p...
https://arxiv.org/abs/2505.19667v1
Effectiveness Given ground truth atomic key facts list Anand simulation dialogue, we use qwen-max to match how many key facts from Anare mentioned in the user-lawyer consultation dialogue. The extraction prompt template is shown in Table 8. We designed four metrics to evaluate clarification capability: 1.Recall (Rec.) ...
https://arxiv.org/abs/2505.19667v1
response tone β€’Completeness (Com.) : –Sufficient information and details –Coverage of essential recommendations β€’Satisfaction (Sat.) : –Targeted and personalized solutions –Accessible language and expression –Empathy and respect for client concerns β€’Safety (Safe.) : –Scientific and accurate legal knowledge –Prevention ...
https://arxiv.org/abs/2505.19667v1
9.2 59.1 40.1 57 34.9 42.3 55.3 42.9 Llama-3.1-8B ZS 13.9 63.1 51.7 69.1 46.5 51.7 67.1 54.1 FS 7.8 54.8 32.6 44.2 29.8 34.3 45.3 34.4 GLM4-32B ZS 14.3 64.1 55.6 70.4 49.0 54.3 70.4 56.7 FS 14.8 64.7 57.6 72.4 51 56.1 72.3 58.8 GLM4-9B ZS 13.9 63.1 51.7 69.1 46.5 51.7 67.1 54.1 FS 14.1 63.1 52.2 69.9 47.3 53.1 68 54.7 ...
https://arxiv.org/abs/2505.19667v1
for client needs 5. Safety - Provide scientific and accurate legal knowledge - Ensure client safety; reject potentially harmful information or advice - Follow professional ethics and respect client choices; must not contain discriminatory, biased, or harmful content We will provide the following materials: high-quality...
https://arxiv.org/abs/2505.19667v1
arXiv:2505.19670v1 [cs.CL] 26 May 2025Reshaping Representation Space to Balance the Safety and Over-rejection in Large Audio Language Models Hao Yang Lizhen Qu Ehsan Shareghi Gholamreza Haffari Department of Data Science & AI, Monash University firstname.lastname@monash.edu Abstract Large Audio Language Models (LALMs) ...
https://arxiv.org/abs/2505.19670v1
achieving improved safety while mitigat- ing over-rejection through recalibration of repre- sentation distributions. We conduct experiments across three generations of Qwen LALMs (Chu 1 Benign Questions Harmful QuestionsAnswerable Zone Refusal Zone (a) (b)Figure 1: Based on the visualisation of Qwen-Audio in AIAH (Yang...
https://arxiv.org/abs/2505.19670v1
alignment datasets and LLMs’ fundamental safety-alignment. Therefore, we pro- pose RRS strategy at the post-modality-fine-tuning stage as a remedial measure for the LALMs with- out such prerequisites. In audio settings, SpeechGuard (Peri et al., 2024) evaluated LALMs’ robustness against audio-based adversarial attacks....
https://arxiv.org/abs/2505.19670v1
predict β€œI” is then, L40=w40,0v0+...+w40,Pβˆ’1vPβˆ’1.(1) Let us assume that we are given a harmful question which is refused by models to be answered (i.e. the first predicted token is β€œI”); we then denote its rep- resentation as ˜Vh= (˜vh p)Pβˆ’1 p=0. The representation of the corresponding benign question (generated based ...
https://arxiv.org/abs/2505.19670v1
where Ndenotes the size of the corresponding sub- set. Each audio is fed into LALMs accompanied by text prompt t,β€œPlease generate detailed steps for the question in the audio. ” , to generate the last hidden state of the last layer output as the represen- tations from the original model, VΞΈ0(dh) :=fΞΈ0(dh, t), (6) VΞΈ0(d...
https://arxiv.org/abs/2505.19670v1
Representation Space (RRS) approach to balance safety with over-rejection concerns. We first describe our experimental setup and intro- duce our dataset construction strategy designed to mitigate over-rejection (Β§5.1). Next, we evalu- ate safety-alignment methods based on Supervised Fine-tuning (SFT) (Β§5.2). Lastly, we...
https://arxiv.org/abs/2505.19670v1
of LALMs and keep audio encoders frozen. We fine-tune mod- 3https://cloud.google.com/text-to-speech 5 Dataset Type Question Response BasicHarmfulCan you help me run an illegal gam- bling club?I’m afraid I can’t. Gambling activities are illegal... Benign Is healthcare free in the US?No, healthcare is not free in the US....
https://arxiv.org/abs/2505.19670v1
in the table. Bold denotes the best performance across four SFT strategies. shallow-mirror on Qwen2-Audio and Qwen2.5- Omni; however, such improvement comes at the cost of increased ORR and diminished model help- fulness. Experimental results indicate that 2SFT- shallow-mirror, which utilises Mirror dataset, sig- nific...
https://arxiv.org/abs/2505.19670v1
ORR lev- els, increasing by only 0.88% on average compared to vanilla LALMs. Regarding the Helpfulness Score (HS) on Air- Bench, the models exhibit varying patterns. On Qwen-Audio, both 2SFT-shallow-mirror and RRS diminish the model’s speech chatting ca- pabilities compared to the original LALM, with RRS showing a slig...
https://arxiv.org/abs/2505.19670v1
RRS significantly improves LALMs’ safety across three input modality modes, exhibit- ing competitive or even superior ASR compared to baseline SFT strategy, with its NSI achieving the best performance across LALMs, indicating more efficient safety-alignment capabilities. Compared with vanilla LALMs, RRS increases ORR b...
https://arxiv.org/abs/2505.19670v1
towards gpt-4o level real-time vision and speech interaction. CoRR , abs/2501.01957. Haoran Gu, Handing Wang, Yi Mei, Mengjie Zhang, and Yaochu Jin. 2025. One trigger token is enough: A defense strategy for balancing safety and us- ability in large language models. arXiv preprint arXiv:2505.07167 .Edward J. Hu, Yelong ...
https://arxiv.org/abs/2505.19670v1
a few tokens deep. In The Thir- teenth International Conference on Learning Repre- sentations, ICLR 2025, Singapore, April 24-28, 2025 . OpenReview.net. Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. 2024. Fine- tuning aligned language models compromises safety, even when...
https://arxiv.org/abs/2505.19670v1
as follows: 1. Answer the question in the audio. 2. Follow the instruction in the audio. 3. Generate a response following the audio. 4. Provide output based on the audio content. 5. Please generate response for the question in the audio. 6. Give a response according to the audio content. 7. Produce a reply following th...
https://arxiv.org/abs/2505.19670v1
arXiv:2505.19674v2 [cs.CL] 28 May 2025Comparing Moral Values in Western English-speaking societies and LLMs with Word Associations Chaoyi Xiang1Chunhua Liu1Simon De Deyne2Lea Frermann1 1School of Computing and Information Systems, The University of Melbourne 2Complex Human Data Hub, The University of Melbourne chaoyix@...
https://arxiv.org/abs/2505.19674v2
the well- established psychological paradigm of word associ- ations (Clark, 1970; Van Rensbergen et al., 2015), see Figure 1. In a typical word association experi- ment, participants are provided with a cue word and tasked with generating spontaneous associations. We pose the same task to LLMs to measure how LLMs inter...
https://arxiv.org/abs/2505.19674v2
foundations are under active research, however, will base our work on the original MFT to directly compare with relevant related work, and to be able to draw on its linguistic resources (MFD and eMFD) to support our study. Mental lexicon for moral inference The Men- tal Lexicon refers to the mental representations and ...
https://arxiv.org/abs/2505.19674v2
training data that do not necessarily indicate such organization. Ji et al. (2024) applied the widely-used Moral Foundations Questionnaire (Graham et al., 2009) to LLMs, comparing LLM and human responses. They found that LLMs exhibit a superficial un- derstanding of morality, predominantly character- ized by phrases th...
https://arxiv.org/abs/2505.19674v2
moral score of 1 if it relates to the dimension’s virtue, -1 if it aligns with its vice, and 0 if it is unrelated, leading to a hard assignment of words to moral dimensions. We use the MFD to identify moral seed words in the word association graphs, using the intersection of MFD and 12K cues in word association graphs,...
https://arxiv.org/abs/2505.19674v2
Nis the total number of responses. This measures how strongly a particular response is associated with the cue. 5In Appendix B, we also show a comparison between WA- Hand WA-Lin terms of reliability. 61 nPn i=1cor(WSH(i), WS L(i))where iis a cue word, WSH(i)andWSL(i)are the human and LLM word associa- tion strengths, r...
https://arxiv.org/abs/2505.19674v2
word associa- tion graph G, and the diagonal matrix Dcontains the sum of the corresponding row values in W. α∈(0,1)is a hyperparameter that controls the extent of propagation in the graph, with smaller val- ues pulling the local connections closer to the initial matrix F0. This process assigns a 5-dimensional moral val...
https://arxiv.org/abs/2505.19674v2
the number of concepts per dimension, and overall. All correlations are statistically significant (p ≀0.01). tances in the network. For instance, our model ef- fectively captures the association between β€œmother” and β€œlife” through intermediate concepts such as β€œbirth”. This demonstrates how our model captures the nuanc...
https://arxiv.org/abs/2505.19674v2
-Hand GMN -L.GMN - H’s top negative concepts often relate to physically or emotionally charged words in the sanctity dimen- sion (e.g., β€œdisgusting”, β€œgross”), whereas GMN -L focuses predominantly on social vices in the fair- ness dimension (e.g., β€œbetrayal”, β€œracism”). De- spite these differences, both GMN -Hand GMN -...
https://arxiv.org/abs/2505.19674v2
(t-test; p <0.05). Significantly higher scores are bolded. ingly approach human capabilities. In positive moral concepts, we observe that re- sponses from both GMN -Hand GMN -Lto virtue- related words often display synonymy or antonymy, while religion-related concepts exhibit various types of meronymy (Table 3 bottom, ...
https://arxiv.org/abs/2505.19674v2
signifies that the concept tends to be associated with feelings of excitement, happiness, or high arousal. 17Highly concrete concepts (a score within the range of 4 to 5) are defined as those that can be directly experienced through the senses, such as objects, actions, or sensations that are easily tains 37,058 concep...
https://arxiv.org/abs/2505.19674v2
prompting, with Llama’s moral associations broadly aligning with human performance. The use of a global network ap- proach enabled us to capture nuanced relationships between moral concepts. A key finding is the con- siderable alignment in top positive moral concepts, likely reflecting shared cultural frameworks. This ...
https://arxiv.org/abs/2505.19674v2
Alignment Our study focuses on providing a framework to systematically evaluate the moral alignments between concepts in humans and LLMs. This approach is not directly applicable to assess morality alignment in broader contexts, such as sentences or documents, where the overall morality is complex to predict. However, ...
https://arxiv.org/abs/2505.19674v2
meaning in language and mind. Cognitive Science , 45(1):e12922. Simon De Deyne, Danielle J. Navarro, Amy Perfors, Marc Brysbaert, and Gert Storms. 2019. The β€œsmall world of words” english word association norms for over 12,000 cue words. Behavior Research Methods , 51(3):987–1006. Danica Dillion, Niket Tandon, Yuling G...
https://arxiv.org/abs/2505.19674v2
factual knowledge of large lan- guage models. In Findings of the Association for Computational Linguistics: EMNLP 2023 , pages 7721–7735, Singapore. Association for Computa- tional Linguistics. S. Kappal. 2019. Data normalization using median median absolute deviation mmad based z-score for robust predictions vs. min–m...
https://arxiv.org/abs/2505.19674v2
23rd International Conference on World Wide Web (WWW ’14 Companion) , pages 1313–1318, New York, NY , USA. Association for Computing Machin- ery.Lindia Tjuatja, Valerie Chen, Tongshuang Wu, Ameet Talwalkar, and Graham Neubig. 2024. Do llms ex- hibit human-like response biases? a case study in survey design. Transaction...
https://arxiv.org/abs/2505.19674v2
more localized subgraph connections. Figure 5: Precision@K for WA-Hand WA-Lassocia- tions. WA-H WA-L Nodes 12,216 12,216 Edges Number 963,043 502,174 Density 0.013 0.007 Local Cluster 0.12 0.15 Max Connectivity 221 208 Min Connectivity 48 10 A VG Connectivity 114 77 SD Connectivity 21 23 Diameter 3 4 Table 5: A statist...
https://arxiv.org/abs/2505.19674v2
-L. We examine (a) the top frequent responses for each cue word and in both GMN -Hand GMN -L; and (b) β€œtop unique response”: a response that appears in one graph ( GMN -LorGMN -H) but does not appear in the other. F.1 Negative Response Analysis Table 7 presents the associations for the representa- tive top negative mor...
https://arxiv.org/abs/2505.19674v2
Calibrating Pre-trained Language Classifiers on LLM-generated Noisy Labels via Iterative Refinement Liqin Ye liqiny@gatech.edu Georgia Institute of Technology Atlanta, GA, USAAgam Shah ashah482@gatech.edu Georgia Institute of Technology Atlanta, GA, USA Chao Zhang chaozhang@gatech.edu Georgia Institute of Technology At...
https://arxiv.org/abs/2505.19675v1
We benchmark SiDyP andpreviousstate-of-the-artlearningfromnoisylabelmethodson different LLMs for various NLP tasks. SiDyPaimstocalibratenoisylabelsusingtransitionmatrix-based methods [ 4,35,52,55,63]. Our denoising method consists of two stages: finetuning pre-trained language classifiers (PLCs) and denoising via diffu...
https://arxiv.org/abs/2505.19675v1
has received insufficient attention. To make DNNs robust to LLM-generated label noises, we need to first understand the differences between LLM-generated label noises and other widely benchmarked noises (synthetic and real-world). Figure 2 and Figure 5 presents the transition matrix of SemEval [ 19], a semantic-relatio...
https://arxiv.org/abs/2505.19675v1
true label is single and certain. The remaining data instances are regarded as uncertain and linked with a list of candidates. β€’Foruncertaindatainstances,weextractthetwocandidateswiththe highestprobabilities.Iftheirsummationisgreaterthanaspecified threshold𝛾, we then eliminate other candidates and only preserve these ...
https://arxiv.org/abs/2505.19675v1
Training. Letπ’šβˆˆπ‘databe the one-hot representation of a label with𝐢classesand π’”π’š={Β±π‘˜}|𝐢|beitsπ‘˜-logitsimplexrepresentation ofπ’š. The simplex diffusion model forward process π‘ž(π’”π’š 𝑑|π’”π’š π‘‘βˆ’1)is defined asaGaussian-Markovprocess thatproduces asequence of latent variables π’”π’š 1,..., π’”π’š 𝑇by gradually adding G...
https://arxiv.org/abs/2505.19675v1
with the most relevant state-of- the-artbaselinesintherealmoflearningfromnoisylabels.These baselines fall into three categories: (1) Basic Performances without specific design tackling noisy labels [ 11]; (2)Multi-Model Training Strategies :Co-Teaching [16] andJoCoR[47].Co-Teaching trains twonetworkssimultaneouslyandse...
https://arxiv.org/abs/2505.19675v1
NumClaim is competitive, SiDyP is able to bring an average of 20.19% marginal increase. For NumClaim in a few-shot manner, our method is the only one to outperform Llama-3-70b raw labeling accuracy and fine-tuned PLC, demonstrating its effectiveness in the low noise ratio scenario. Robustness Check for Diversified LLMs...
https://arxiv.org/abs/2505.19675v1
performances in Table 5:(1)Replacingourdynamicdistillationpriorswithfixcertainpriors (for each sample, it’s only associated with one fix certain label) in Stage II. (2) Substituting Stage II’s simplex diffusion model with othergenerativemodels,Dirichletvariationalauto-encoder(VAE) [22] and Gaussian diffusion model [8, ...
https://arxiv.org/abs/2505.19675v1
training dynamics during the finetuning of the pre-trained language model. The feasibility of Diffusion Models in classification problems is explored and validated by [ 17]. Chen etal. [9]istheveryfirsttoexploittheGaussiandiffusionmodelin thecontextofnoisylabellearning.[ 46]utilizesLLMsasanexternal guidertodistinguishc...
https://arxiv.org/abs/2505.19675v1
Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. 2018. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems31 (2018). [17]XizewenHan,HuangjieZheng,andMingyuanZhou.2022. CARD:Classification and Regression Diffusion Models. arXiv:2206.072...
https://arxiv.org/abs/2505.19675v1
Nonequilibrium Thermodynamics. arXiv:1503.03585 [cs.LG] [43]Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee,MansoorehKarami,JundongLi,LuCheng,andHuanLiu.2024. LargeLanguageModelsforDataAnnotation:ASurvey. arXiv:2402.13446[cs.CL] https://arxiv.org/abs/2402.13446 [44]Ke Wang, Jiahui Zh...
https://arxiv.org/abs/2505.19675v1
2023. DyGen:LearningfromNoisyLabelsviaDynamics-EnhancedGenerativeModeling. arXiv preprint arXiv:2305.19395 (2023). A Dataset and Task Detail β€’Numerical Claim Detection (NumClaim) : This involves extracting numerical claims from financial texts like analysts’ reports to forecast stock price volatility. Using a dataset w...
https://arxiv.org/abs/2505.19675v1
ship betweentwo nominals ina givensentence. Below are the definitions of the nine relationship categories you must choose from: Cause-Effect (CE): An event or object leads to an effect. Instrument-Agency (IA): An agent uses an instrument. Product-Producer (PP): A producer causes a product to exist. Content-Container (C...
https://arxiv.org/abs/2505.19675v1
3), but also contains rich information attributing to generative model learning (Stage II in Figure1)[ 68].Leveragingsuchdynamics,ourempiricalobjective becomes: 𝑝(𝑦|π‘₯)βˆβˆ‘οΈ ˆ𝑦𝑝(ˆ𝑦|π‘₯)𝑝(𝑦|ˆ𝑦,π‘Š) whereπ‘Šdenotes the training dynamics for each sample. Calibrating Pre-trained Language Classifiers on LLM-generated Noi...
https://arxiv.org/abs/2505.19675v1