text string | source string |
|---|---|
dependency trees ,lower perplexity ,higher specificity , and the presence of brackets . These are examples of preferred features, and the full results are presented in Appendix F. Conversely, no signif- icant differences were observed for features such as the number of adjectives. 5.2 Ad Text Generation In this experim... | https://arxiv.org/abs/2505.20826v1 |
V2.0. Subse- quently, preference judgments were conducted for y1andy2using the annotation process in §3.4, col- lecting responses from ten evaluators. As a result, we constructed a dataset of 8,721 triplets ( x,ypref 1, ypref 2), where ypref 1andypref 2denote preference- labeled paraphrases. The dataset was split into ... | https://arxiv.org/abs/2505.20826v1 |
higher attractiveness scores in Table 5 performed better across these linguistic features. Notably, DPO-based models exhibited higher character count. This suggests that DPO- based models tend to generate longer texts, poten- tially benefiting from length heuristics (Park et al., 2024), a bias where evaluators tend to ... | https://arxiv.org/abs/2505.20826v1 |
more attractive expressions influences ad performance, such as CTR. Specif- ically, we conducted an A/B test, comparing an existing group of ad texts with paraphrased ads generated using the fewshot-findings method10 in §5.2. The tests were conducted on Google Ads, focusing on the headline text for ads from two com- 10... | https://arxiv.org/abs/2505.20826v1 |
texts. Our analysis revealed the relationship between human preference and ad performance, and demonstrated the potential of reference-free evaluation for assess- ing ad text attractiveness. Future work will include enhancing ATG meth- ods by addressing challenges such as adhering to length constraints, optimizing both... | https://arxiv.org/abs/2505.20826v1 |
Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advancesin Neural Information Processing Systems 33 , vol- ume 33, pages 1877–1901. Jan Cegin, Jakub Simko, and Peter Brusilovsky. 2023. ChatGPT to replace crowdsourcing of paraphrases for intent clas... | https://arxiv.org/abs/2505.20826v1 |
2023. G-eval: NLG evaluation using gpt-4 with better human align- ment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages 2511–2522. Kikuo Maekawa, Makoto Yamazaki, Takehiko Maruyama, Masaya Yamaguchi, Hideki Ogura, Wakako Kashino, Toshinobu Ogiso, Hanae Koiso, and Yasuha... | https://arxiv.org/abs/2505.20826v1 |
Sudachi: a japanese tokenizer for busi- ness. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation . Hsueh-Cheng Wang and Marc Pomplun. 2012. The attraction of visual attention to texts in real-world scenes. Journal of Vision , 12(6):26–26. Peiyi Wang, Lei Li, Liang Chen, Zefan C... | https://arxiv.org/abs/2505.20826v1 |
30 paraphrase e xamples creat ed b y ad writ ersY ou are a prof essional c op y writer responsible f or creating search engine ads. Please rephrase the pro vided ad te xt to mak e it more attrac tiv e ac c ording to the f ollo wing c onditions # C onditions - An ad te xt must be within 30 charac ters. - Do no t add ne ... | https://arxiv.org/abs/2505.20826v1 |
dec orativ e symbols. - Modify the charac ter types. - Plac e the most important inf ormation at the beginning. - Use more abstrac t e xpressions. - Use more specific e xpressions. - Include w ords that enc ourage ac tion. - Use casual language. - T urn statements into questions. # Examples Input: R ec ommended in Kann... | https://arxiv.org/abs/2505.20826v1 |
5: Guidelines for paraphrase identification pre- sented to crowd workers. B Paraphrase Identification Figure 5 presents the annotation guidelines for paraphrase identification provided to the workers. To ensure consistency between ADPARAPHRASE V1.0and V2.0, we adopted the annotation guide- lines used by Murakami et al.... | https://arxiv.org/abs/2505.20826v1 |
tiv e? - Which is more e y e-catching? - Which is easier to understand? - Which is easier to read? # No tes If the impression of bo th ads is the same, please selec t " skip ” . # Choic es - {ad te x t #1} - {ad te x t #2} - skip Figure 6: Guidelines for preference judgments pre- sented to crowd workers. Levenshtein ( ... | https://arxiv.org/abs/2505.20826v1 |
Third personal pronoun 0.537 34.2 88.5 (10) Excitement 0.445 23.0 87.7 (11) Joy 0.430 27.8 90.4 (12) Ease 0.495 38.1 93.5 (13) Urgency 0.485 31.1 92.2 (14) Action 0.423 28.3 90.9 (15) Brackets 0.622 44.6 92.0 (16) Numbers 0.508 33.0 90.8 (17) Verbs 0.523 39.6 90.9 (18) Adjectives 0.552 35.6 88.9 (19) Nouns 0.580 34.3 8... | https://arxiv.org/abs/2505.20826v1 |
studied emotions such as joyandanticipation , we investigated sadness and surprise . Details of the classifiers can be found in §F.2. Regarding the decorative use of symbols, fea- tures such as the presence of brackets and question marks were considered. Brackets were included as features because they are widely used i... | https://arxiv.org/abs/2505.20826v1 |
e the t e xt more fluent . # Examples (20 cases) Input: 4-minut e walk from Fujiidera Station Output: From Fujiidera Station, a 4-minut e walk Input: Experienc ed T alent Hiring [For C orporations] Output: Immediat e Hiring o f Experienc ed T alent f or C orporations ︙ # Answ er Input: {ad t e xt } Output: {paraphrased... | https://arxiv.org/abs/2505.20826v1 |
erence judgments as prompts. These guidelines are displayed in Figures 5 and 6. The LLM-based eval- uation is reference-free, whereas the other metrics are reference-based. For reference-based metrics, the human-created paraphrases in §5.2 were used as the reference text. Automatic Evaluation Results Table 15 presents ... | https://arxiv.org/abs/2505.20826v1 |
arXiv:2505.20841v1 [cs.CL] 27 May 2025Concealment of Intent: A Game-Theoretic Analysis Xinbo Wu†,Abhishek Umrawal ,Lav R. Varshney University of Illinois Urbana-Champaign Abstract As large language models (LLMs) grow more capable, concerns about their safe deployment have also grown. Although alignment mechanisms have ... | https://arxiv.org/abs/2505.20841v1 |
filtering. We assess attack quality based on the extent to which the system’s response could potentially aid a malicious intent (not a prompt), regardless of whether the content is overtly harmful, explicit, or complete. In this setting, an evaluator must have access to the underlying intent, which may not be explicitl... | https://arxiv.org/abs/2505.20841v1 |
and position it as a tool to better understand a wider range of attacks based on adversarial prompting. However, unlike prior methods, our attack actively probes the target system to identify its weak points on certain skill compositions. Even though our theoretical analysis is conducted under certain simplifying assum... | https://arxiv.org/abs/2505.20841v1 |
define an intent-skill combination as: Sn=1:={(i, s)|i∈ I, s∈ S} , where nrepresents the number of skills combined with an intent. As illustrated in Figure 1, the attack and defense process unfolds as follows. • The attacker first selects an intent, sampled from a distribution: i∼pI(i). •The attacker then mixes this in... | https://arxiv.org/abs/2505.20841v1 |
of the classification function Dregarding an intent iand a skill sbe denoted as: a:={ai,s|(i, s)∈Sn=1, ai,s=α(i, s)}, where α(i, s) :I × S → [0,1]measures an overall performance of Don samples produced via the combination (i, s). Note that Ddoesn’t take the intent and skill as direct inputs; rather, it serves as a perf... | https://arxiv.org/abs/2505.20841v1 |
handle, more space for creativity (Varshney, 2019). Currently, we consider only the case where a single skill is mixed with an intent. However, it is possible to mix multiple skills, expanding the skill combination space to |S| n , where nis the number of skills being mixed and |S| n is a binomial coefficient. In tha... | https://arxiv.org/abs/2505.20841v1 |
i.e., any s∗∈arg min sˆai,sand the defender then allocates its limited capacity greedily, prioritizing the fake weakest points associated with the most probable intents. We also compare the new equilibrium point with the previous one via the following theorem. Theorem 3.4. (Advantage of defense by misleading the attack... | https://arxiv.org/abs/2505.20841v1 |
provided by the JailbreakBench for judge comparison. This dataset includes 200 jailbreak responses from the JailbreakBench, 100 benign examples similar to the harmful ones from XS-Test (Röttger et al., 2024), and 300 mismatched prompt-response pairs (expected to score 1); their ground truth labels are binary and were o... | https://arxiv.org/abs/2505.20841v1 |
weak points in the target system’s handling of specific combinations. In the second stage, the attacker concentrates its attack by generating 20 prompts per intent for each intent by exploiting these identified weak points. Our method utilizes the LLaMA-3.3-70B-Instruct-Turbo as our model Efor composing a prompt via mi... | https://arxiv.org/abs/2505.20841v1 |
skills with the intent. We conduct experiments where the skill space has varying sizes under the 1-skill setup and each intent is combined with two skills (2-skill setup), while keeping other settings fixed. As shown in Table 3, with increasing skill space and additional skill mixing, higher acceptance rates and Bin-JR... | https://arxiv.org/abs/2505.20841v1 |
misleading points and fine-tune them to ensure good defense performance. Therefore, the defense performance reported in our experiments should be viewed as a lower bound. 5 Conclusion We present a scalable adversarial prompting strategy for LLM-based systems by hiding intents, in which a malicious intent is concealed t... | https://arxiv.org/abs/2505.20841v1 |
235:35181–35224. Noveck, I. (2018). Experimental Pragmatics: The Making of a Cognitive Science . Cambridge University Press. OpenAI (2023). GPT-4 technical report. arXiv:2304.01852 . OpenAI (2025). Introducing gpt-4.1 in the api. https://openai.com/index/gpt-4-1/ . Ac- cessed: 2025-05-14. Ouyang, L., Wu, J., Jiang, X.,... | https://arxiv.org/abs/2505.20841v1 |
models. arXiv preprint arXiv:2307.15043 . A Proofs Theorem A.1. (Equilibrium of the game) The equilibrium value of the game is: J∗= 1−c |S|X ip(i)2(14) withai,s=p(i)c/|S|andp(s|i) = 1 /|S|. Proof. For fixed {ai,s}, the attacker chooses p(s|i)for each ito minimize: X i,sai,sp(s|i)p(i) =X ip(i)X sai,sp(s|i) For each i, t... | https://arxiv.org/abs/2505.20841v1 |
perceived weak point, the defender could deceive the attacker into focusing on a fake weak point, s∗, which actually has a performance level of ai,s∗. Since p(i)is fixed. 13 The optimal strategy is allocating ccapacity in the order of decreasing intent probability p(i), where the performance is capped at 1, leading to ... | https://arxiv.org/abs/2505.20841v1 |
not to pursue state-of-the-art performance, but to demonstrate the effectiveness of our proposed methods and to empirically validate the theoretical implications. C Experimental details and hyperparameters C.1 LLM-as-rater We present our custom prompt in Figure 2. We found that model capacity plays a crucial role in en... | https://arxiv.org/abs/2505.20841v1 |
to set the temperature to 0and generate 150 tokens for each target model. When available, we use the default system prompts. D Broader impacts Our attack method identifies vulnerabilities in target systems, closely aligning with the goals of red-teaming and offering potential to strengthen the safety and trustworthines... | https://arxiv.org/abs/2505.20841v1 |
An LLM-as-Judge Metric for Bridging the Gap with Human Evaluation in SE Tasks XIN ZHOU, Singapore Management University, Singapore KISUB KIM, Independent Researcher, Hong Kong TING ZHANG, Singapore Management University, Singapore MARTIN WEYSSOW, Singapore Management University, Singapore LUÍS F. GOMES, Carnegie Mellon... | https://arxiv.org/abs/2505.20854v1 |
. Publication date: May 2025.arXiv:2505.20854v1 [cs.SE] 27 May 2025 2 Xin Zhou, Kisub Kim, Ting Zhang, Martin Weyssow, Luís F. Gomes, Guang Yang, and David Lo ACM Reference Format: Xin Zhou, Kisub Kim, Ting Zhang, Martin Weyssow, Luís F. Gomes, Guang Yang, and David Lo. 2025. An LLM-as-Judge Metric for Bridging the Gap... | https://arxiv.org/abs/2505.20854v1 |
evaluation metric for code, ICE-Score [ 45], instructs LLMs to directly assign evaluation scores based on predefined criteria—natural language descriptions of correct and incor- rect code. However, it primarily focuses on a single strategy, lacking diverse strategies to assess correctness from different angles. A more ... | https://arxiv.org/abs/2505.20854v1 |
generation, automated program repair, and code summarization. •SWE-Judge significantly and consistently outperforms existing automatic evaluation metrics, achieving new state-of-the-art performance. 2 PRELIMINARIES 2.1 Problem Statement Automatic evaluation metrics aim to assess the quality of software artifacts genera... | https://arxiv.org/abs/2505.20854v1 |
tests. To address this limitation, we propose SWE-Judge, which extends beyond the Direct Assess strategy. SWE-Judge explores and integrates multiple evaluation strategies for assessing correctness, leading to more accurate evaluation scores compared to ICE-Score. 2.3 Motivating Example Our work is inspired by the rigor... | https://arxiv.org/abs/2505.20854v1 |
“team”, from these strategies. Importantly, the team selection is performed dynamically for each dataset, allowing the assembled team to adapt to the characteristics of different datasets. Part 3: Correctness Score Generation ( 3of Figure 2). Once the team is determined, it is used to evaluate the correctness of data s... | https://arxiv.org/abs/2505.20854v1 |
generated software artifact 𝑦due to a flaw it identifies (e.g., a reason 𝑒). During the rethink phase, if the LLM realizes that this reason 𝑒is actually incorrect, it is encouraged to revise the score upward. Conversely, if the LLM initially gives a high score based on a positive justification 𝑒, but later determin... | https://arxiv.org/abs/2505.20854v1 |
𝑦is a code snippet or a code change, test cases serve as an effective means for assessing its correctness. Figure 3 illustrates the prompt design for Strategy 4. This strategy consists of two steps. In the first step, we prompt the LLM to generate test cases based on the user requirement 𝑥and the reference code 𝑟, a... | https://arxiv.org/abs/2505.20854v1 |
using a 0–4 scale, APR-Assess [ 15] using 0–1, and Summary-Assess [ 22,29] using 1–5, we standardize the output range across all strategies to ensure consistency. To do this, we include an instruction in each prompt that constrains the LLM to output a score within the 0–100 range. In a later stage of SWE-Judge (describ... | https://arxiv.org/abs/2505.20854v1 |
ˆ𝑠 100×4+1. This transformation ensures that the predicted scores align with the target scale of each dataset, andE(𝑥,𝑦, 𝑟)is the final correctness score produced by our SWE-Judge. 4 EXPERIMENTAL SETUP In this section, we introduce the datasets used in our experiments, the baseline methods for comparison, and the e... | https://arxiv.org/abs/2505.20854v1 |
[14] is a Python code generation benchmark that includes introductory-level, interview- level, and competition-level coding tasks collected from code competition websites. We evaluate SWE-Judge on 100 sampled competition-level tasks of APPS. 4.2 Selected Baselines Match-based Metrics. We choose 7 popular match-based me... | https://arxiv.org/abs/2505.20854v1 |
SWE-Judge on all studied tasks and datasets. For ease of comparing different methods, we also calculate the averaged correlation score by averaging the three kinds of correlations above. Statistical Agreements. We also evaluate the statistical agreement between our tool’s results and human evaluation scores. Specifical... | https://arxiv.org/abs/2505.20854v1 |
25.3 13.5 15.9 16.2 RUBY 36.6 47.9 43.1 61.7 65.0 67.9 12.1 13.1 14.7 17.1 19.7 21.1 CrystalBLEU 26.6 32.9 29.5 42.2 53.5 49.0 28.5 33.1 34.8 13.0 14.9 14.8 MoverScore 39.9 46.9 44.0 65.1 78.0 79.5 18.6 23.0 22.7 16.0 17.8 18.5 BERTScore 43.7 48.5 48.0 55.9 69.2 69.6 0.7 2.7 0.8 21.8 24.2 24.1 CodeBERTScore 42.1 46.7 4... | https://arxiv.org/abs/2505.20854v1 |
aggregated human score. In contrast, RQ2 uses individual human annotations as the ground truth to evaluate SWE-Judge . Our objective is to measure the gap between the agreement levels of SWE-Judge with human annotators (human-tool agreement) and the agreement among human annotators themselves (human-human agreement). I... | https://arxiv.org/abs/2505.20854v1 |
1) the Strategy Design, and 2) the Dynamic Team Selection. Table 1 presents the results of the ablation study in the last two rows. The row labeled as “wo Team Selection” shows the performance of SWE-Judge without the team selection mechanism, where all strategies are combined through simple ensembling. The row labeled... | https://arxiv.org/abs/2505.20854v1 |
scores and test case execution results on both datasets. Table 2 demonstrates that SWE-Judge achieves the highest average correlation with test case execution outcomes, consistently and significantly outperforming all baseline methods. On the HumanEval-X dataset, SWE-Judge outperforms all baselines by an average margin... | https://arxiv.org/abs/2505.20854v1 |
to flawed code. In contrast, SWE-Judge does not suffer from those issues. Its scores are more consistent with human judgments and better reflect the correctness of the data under evaluation. 6.3 Threats to Validity Our findings are limited to the specific SE datasets examined in this study and may not generalize to all... | https://arxiv.org/abs/2505.20854v1 |
popular tasks: code generation, automated program repair, and code summarization. Additionally, several concurrent studies investigate similar topics in parallel. Wang et al. [ 33] empirically investigate LLM-as-a-judge methods from NLP for evaluating SE tasks, focusing on consistency and readability aspects. In contra... | https://arxiv.org/abs/2505.20854v1 |
[3]Wikipedia contributors. 2025. Cohen’s kappa. https://en.wikipedia.org/wiki/Cohen%27s_kappa#Interpreting_ magnitude Accessed: 2025-03-25. [4]Wikipedia contributors. 2025. Kendall rank correlation coefficient. https://en.wikipedia.org/wiki/Kendall_rank_ correlation_coefficient Accessed: 2025-03-25. [5] Wikipedia contr... | https://arxiv.org/abs/2505.20854v1 |
Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers . The Association for Computer Linguistics. https://doi.org/10.18653/V1/P16-1057 [20] Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-Eval: NLG Evaluation using Gpt-4 with... | https://arxiv.org/abs/2505.20854v1 |
Zhou, and Houari Sahraoui. 2024. Codeultrafeedback: An llm-as-a-judge dataset for aligning large language models to coding preferences. arXiv preprint arXiv:2403.09032 (2024). [35] Martin Weyssow, Xin Zhou, Kisub Kim, David Lo, and Houari A. Sahraoui. 2023. Exploring Parameter-Efficient Fine-Tuning Techniques for Code ... | https://arxiv.org/abs/2505.20854v1 |
arXiv:2505.20871v1 [cs.CL] 27 May 2025Divide-Then-Align: Honest Alignment based on the Knowledge Boundary of RAG Xin Sun1,2*, Jianan Xie3*, Zhongqi Chen4, Qiang Liu2†, Shu Wu2, Yuehe Chen4,Bowen Song4†,Weiqiang Wang4Zilei Wang1Liang Wang2, 1USTC2NLPR, MAIS, CASIA3SUSTech4Independent sunxin000@mail.ustc.edu.cn, 12110714... | https://arxiv.org/abs/2505.20871v1 |
in model performance, it introduces a critical drawback: RAFT conditions the model to answer questions even when the retrieved contexts are entirely noisy . This behavior poses a significant risk for deploying LLMs in real-world applications, partic- ularly in high-stakes domains like medical (Raja et al., 2024), legal... | https://arxiv.org/abs/2505.20871v1 |
Let r:Q → Pbe the retrieval function that maps a query qto relevant passages P⊆ D , where Qis the query space and Pis the passage space. We use M:Q× P → A to represent the LLM function that takes both the query and passages as input and generates an answer from the answer space A. Let golden : Q → A be the function tha... | https://arxiv.org/abs/2505.20871v1 |
indicating whether the context sufficiently supports the an- swer. If GPT-4o determines the context contains or implies the correct answer (score = 1), we consider q∈KBr(✔). Otherwise, we consider q /∈KBr (✘). 3.2 Preference Data Construction Based on the knowledge quadrants, we construct preference data for each quadr... | https://arxiv.org/abs/2505.20871v1 |
yc)−rθ(q, r(q), yr)))(3) where rθ(q, r(q), y)represents the log probabil- ity of generating response ygiven query qand retrieved context r(q)under the model parameters θ,τis the temperature parameter, and σis the sigmoid function. Note that this reward score is de- rived from the same language model being trained, elim... | https://arxiv.org/abs/2505.20871v1 |
across the four knowledge quadrants. the proportion of training examples where the pre- ferred response is "I don’t know" (IDK). Specifi- cally, IDK-ratio determines the fraction of ✘✘sam- ples in the training set. Importantly, we maintain the natural distribution of queries across all four quadrants in the test set wi... | https://arxiv.org/abs/2505.20871v1 |
Table 2. 4.4 Main Results Main experimental results are shown in Table 3. Our post-training strategy DTA achieves the best performance on three llama architectures. Notably, it achieves Acc (64.1, 64.8, 65.5), F1 (64.6, 66.6 65.8), AF1(63.3, 59.9, 64.7), surpassing baselinemethods by significant margins. Critically, DT... | https://arxiv.org/abs/2505.20871v1 |
Recall, Prec: Precision); RH: Retrieval Handling (DR: Denoise Rate, CUR: Context Utilization Rate); AbQ: Abstain Quality (ARec: Abstain Recall, APrec: Abstain Precision, AF1: Abstain F1). terms of the DR and CUR metrics, which is related to the trade-off with abstention. When appropri- ately enhancing the model’s abste... | https://arxiv.org/abs/2505.20871v1 |
maintaining high abstention rates (ARec: 78.6%). However, the abstain precision decreases substantially from 61.7% to 43.1%. This indicates that although the RAG system learns to abstain, it becomes overly cautious and lacks confidence in answering queries that it should be able to handle. Without SFT loss, the model e... | https://arxiv.org/abs/2505.20871v1 |
experimental results are shown in Appendix D. 5 Conclusion In this paper, we propose a novel framework for honest alignment of retrieval-augmented lan- guage models based on knowledge boundary quad- rants. We first identify that the knowledge bound- ary of RAG systems consists of two fundamental components: the paramet... | https://arxiv.org/abs/2505.20871v1 |
on Empirical Methods in Natural Lan- guage Processing , pages 1533–1544, Seattle, Wash- ington, USA. Association for Computational Linguis- tics. Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013b. Semantic parsing on Freebase from question-answer pairs. In Proceedings of the 2013 Conference on Empirical... | https://arxiv.org/abs/2505.20871v1 |
arXiv:2405.20978 . Shangbin Feng, Weijia Shi, Yike Wang, Wenxuan Ding, Vidhisha Balachandran, and Yulia Tsvetkov. 2024. Don’t hallucinate, abstain: Identifying LLM knowl- edge gaps via multi-LLM collaboration. In Proceed- ings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long P... | https://arxiv.org/abs/2505.20871v1 |
Weld, and Luke Zettlemoyer. 2017b. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehen- sion. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers) , pages 1601–1611. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan... | https://arxiv.org/abs/2505.20871v1 |
Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neu- ral Information Processing Systems , 36. Mahimai Raja, E Yuvaraajan, et al. 2024. A rag-based medical assistant especially for infectious diseas... | https://arxiv.org/abs/2505.20871v1 |
in large language models. Advances in neural information processing systems , 35:24824–24837. Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, and Bryan Hooi. 2024. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. In The Twelfth Inter- national Conference o... | https://arxiv.org/abs/2505.20871v1 |
information present in the retrieved passages, the model may fail to utilize it effectively due to issues such as "lost in the middle" (Liu et al., 2024a). RAFT acctually enhances the RAG system’s an- swer accuracy across both ✔✘and✘✔scenarios by addressing their distinct challenges: For ✔✘: RAFT teaches the model to r... | https://arxiv.org/abs/2505.20871v1 |
to accurately express their knowledge, recognize their limitations, and avoid misleading users when uncertain. Honesty encom- passes two critical components: self-knowledge and self-expression. Self-Knowledge refers to the model’s ability to discern what it knows and doesn’t know, enabling it to explicitly admit uncer-... | https://arxiv.org/abs/2505.20871v1 |
family with 13 billion parame- ters, which is released in July 2023. •Llama3-8B (Meta-AI, 2024): A member of Llama3 family with 8 billion parameters, which is released in April 2024. RAFT Models: •RAAT (Fang et al., 2024): A model that em- ploys adaptive adversarial training to handle three types of retrieval noises (r... | https://arxiv.org/abs/2505.20871v1 |
data size grows to 10k, it seems increased noise-potentially introduced by 1k 5k 1w Data Size020406080100Score 47.8 64.1 56.232.5 63.7 61.060.8 65.5 54.842.4 64.6 57.744.3 68.9 62.623.2 52.8 50.977.1 65.0 46.840.7 61.7 60.053.3 63.3 52.6Acc Recall Precision F1 DR CUR AR AP AF1Figure 3: Experiments across DPO data size.... | https://arxiv.org/abs/2505.20871v1 |
36.6 + Consistency 51.4 69.0 50.7 58.5 16.3 58.4 25.4 + ICL 46.8 71.2 46.8 56.5 0.00 0.00 0.00 + SFT 52.2 37.4 69.1 48.5 80.7 42.9 56.0 + SFT & P(True) 48.1 69.5 48.9 57.4 6.8 35.7 11.4 + SFT & Logits 51.5 72.6 51.0 59.9 10.8 58.1 18.2 + SFT & Consistency 51.2 73.5 50.7 60.0 8.3 62.5 14.6 + SFT & ICL 59.7 63.1 58.1 60.... | https://arxiv.org/abs/2505.20871v1 |
our method can still appropriately abstain from answering and demon- strates strong generalization ability across different training configurations.J Prompts J.1 Context Evaluation Prompt The following prompt is used to evaluate whether a context contains or implies the correct answer to a query: You are an expert at e... | https://arxiv.org/abs/2505.20871v1 |
LLM to evaluate the correct- ness of its own answer. The prompt presents the original question and the model’s pro- posed answer, asking for a binary True/False classification. We experiment with multiple confidence thresholds (0.3, 0.5, 0.7, 0.9) to determine the optimal cutoff for each experi- mental setting. Questio... | https://arxiv.org/abs/2505.20871v1 |
2025). Concurrently, to improve adherence to context in RAG scenarios, alignment techniques such as Context-DPO (Bi et al., 2024a) have been developed to bolster context-faithfulness, particu- larly when knowledge conflicts occur. A complicat- ing factor is that efforts to enhance the factual ac- curacy of a model’s in... | https://arxiv.org/abs/2505.20871v1 |
Can LLMs Learn to Map the World from Local Descriptions? Sirui Xia♠, Aili Chen♠, Xintao Wang♠, Tinghui Zhu♠, Yikai Zhang♠ Jiangjie Chen♡, Yanghua Xiao♠† ♠Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University ♡ByteDance Seed {srxia24,xtwang21,thzhu22,ykzhang22}@m.fudan.edu.cn {alchen20,sh... | https://arxiv.org/abs/2505.20874v1 |
route optimization, or multi-step inference. A prime example of a domain requiring struc- tured understanding is spatial cognition —the abil- ity to construct coherent mental representations of physical environments. In human communication, spatial relationships are often conveyed through relational language ( e.g., “T... | https://arxiv.org/abs/2505.20874v1 |
road spatial information is limited, lacking a continu- ous and precise representation. 2 Global Setup Simulation Environment. To facilitate con- trolled investigation and data collection, we con- struct a synthetic 100×100grid map representing a simplified urban layout. Roads run along hor- izontal and vertical lines ... | https://arxiv.org/abs/2505.20874v1 |
dimensions: functional ability ,in- ternal representation , and behavioral robustness . This framework moves beyond surface-level per- formance to probe the cognitive structures formed during training. Specifically, we assess whether the model can generate accurate spatial predictions, in- ternalize geometry-consistent... | https://arxiv.org/abs/2505.20874v1 |
0.83220.8778 0.8166 0.7498 0.69020.7571 0.6477 Spearman (8:2) Pearson (8:2) Spearman (6:4) Pearson (6:4) Spearman (4:6) Pearson (4:6) Figure 2: Consistency between POI latent representa- tions and actual spatial locations. Spearman and Pearson correlation coefficients quantify monotonic and linear relationships, respec... | https://arxiv.org/abs/2505.20874v1 |
Ideal (y=x) 20 40 60 80 100 120 Actual Distance20406080100120Predicted DistanceModel: 4:6 Distance (MAE: 1.2788, R²: 0.9932) 4:6 Preds Ideal (y=x) 20 40 60 80 100 120 Actual Distance20406080100120Predicted DistanceModel: 6:4 Distance (MAE: 1.0364, R²: 0.9967) 6:4 Preds Ideal (y=x) 20 40 60 80 100 120 Actual Distance204... | https://arxiv.org/abs/2505.20874v1 |
(but not both) in the training data, we incorporate Pheldout by adding trajectories between Pheldout andPmain POIs, while paths between Pheldout POIs remain unseen (denoted as Bridged Exposure setting). Table 4 shows that MODEL navexcels in shortest- path prediction, with an exact match accuracy of 83.63% and small sta... | https://arxiv.org/abs/2505.20874v1 |
Trajectory Data? Setting. We next examine whether the model retains spatial perception of POI locations. To this end, we compare models trained under the Bridged Exposure setting on MODEL perand the base model (denoted as Perception- MODEL navand Base- MODEL nav). The untrained base model is also included for compariso... | https://arxiv.org/abs/2505.20874v1 |
SA (%) DD (km) No Pert. 100.00 100.00 0.00 Road Pert. 11.85 62.70 26.99 Distance Pert. 58.79 77.71 20.24 Direction Pert. 43.61 74.87 56.08 Table 6: Navigation performance under various pertur- bation strategies applied at critical path steps. Setting. To assess the robustness of the model to trajectory perturbations, w... | https://arxiv.org/abs/2505.20874v1 |
Median DD 051015202530 Destination Deviation (DD) Figure 5: The model’s performance under different frequency thresholds. where the direction of movement changes. High- frequency turning points generally correspond to transitions between high-speed and regular roads (Figures 4). We control the selection of pperturb and... | https://arxiv.org/abs/2505.20874v1 |
spatial reasoning. However, its limited robustness to nav- igation disturbances reveals the constraints of its understanding of road network structures. Limitations Our study reveals that during the training process, the model develops an understanding of the global spatial distribution of POIs through the descrip- tio... | https://arxiv.org/abs/2505.20874v1 |
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yu- taka Matsuo, and Yusuke Iwasawa. 2022. Large lan- guage models are zero-shot reasoners. Advances in neural information processing systems , 35:22199– 22213. Kenneth Li, Aspen K. Hopkins, David Bau, Fernanda B. Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023. Em... | https://arxiv.org/abs/2505.20874v1 |
Yang, Jiaxin Yang, Jingren Zhou, Jun- yang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, YangSu, Yi-Chao Zhang, Yunyang Wan, Yuqi Liu, Zeyu Cui, Zhenru Zhang, Zihan Qiu, Shanghaoran Qu... | https://arxiv.org/abs/2505.20874v1 |
Adam optimizer with an initial learning rate of 0.001, and L2 regu- larization (alpha = 0.0001) with adaptive learning rate adjustment. The maximum number of training epochs is set to 500, and early stopping is enabled based on validation set performance (patience = 100 epochs), with a validation set proportion of 10%.... | https://arxiv.org/abs/2505.20874v1 |
of POI hidden state vectors with actual spatial locations across training steps. D Experimental Details in Modeling Spatial Navigation D.1 Data Format We provide examples of the data format used for training and evaluation of MODEL nav, as shown in Table 9. D.2 Metric Calculations Start-End Deviation (SED) : evaluates ... | https://arxiv.org/abs/2505.20874v1 |
relationships between POIs, we fine-tune it with supervised train- ing to predict the distance and azimuth between POI pairs. We use 200 POIs to construct the testset, while the remaining POIs are used to generate the training data (randomly sample 100,000 cases). The results in Table 13 show that training the base mod... | https://arxiv.org/abs/2505.20874v1 |
hid- den states and poor alignment between latent vector distributions and actual spatial layouts. This result is expected, as the POI name tokens in the SFT training process do not directly con- tribute to the loss calculation. Consequently, their embeddings are not explicitly optimized, leading to a lack of structure... | https://arxiv.org/abs/2505.20874v1 |
to find the best-fit line, and we used its default configuration. For the non-linear probe, we use the same MLP configuration as in the main experiment. Results We use MODEL perand Base- MODEL nav to compare linear and non-linear probes in several experiments involving probing. The experimental results are shown in the... | https://arxiv.org/abs/2505.20874v1 |
Start (p114) End (p514)p114 p514 0 10 20 30 40 50 60 70 80 90 1000102030405060708090100p114 p514 ( : 2000) p114 p514 0 10 20 30 40 50 60 70 80 90 1000102030405060708090100p114 p514 ( : 5000) p114 p514 0 10 20 30 40 50 60 70 80 90 1000102030405060708090100p114 p514 ( : 10000) p114 p514 0 10 20 30 40 50 60 70 80 90 10001... | https://arxiv.org/abs/2505.20874v1 |
path description A, Map information Mmap 2:Output : Valid Road Proportion V RP ▷ Proportion of steps choosing a valid next road 3:S ← ParsePathDescription (A)▷ParseAinto sequence of steps S= [(r1, d1, l1), . . . , (rn, dn, ln)] 4:if|S|<2then 5: Pstart _pred←Pstart _gt ▷Use ground truth start if path description has few... | https://arxiv.org/abs/2505.20874v1 |
Trans-EnV: A Framework for Evaluating the Linguistic Robustness of LLMs Against English Varieties Jiyoung Lee†∗, Seungho Kim†∗, Jieun Han†, Jun-Min Lee† Kitaek Kim‡, Alice Oh†, Edward Choi† †KAIST,‡Seoul National University †{jiyounglee0523, shokim, jieun_han, lijm565, edwardchoi}@kaist.ac.kr †alice.oh@kaist.edu‡kitaek... | https://arxiv.org/abs/2505.20875v1 |
English. We begin by collecting linguistic features from expert-curated resources and large-scale ESL corpora to ensure rigorous and accurate information. Then, for each feature, we create transformation guidelines that specify the steps to apply the feature to an SAE sentence. Then we use an LLM to transform SAE sente... | https://arxiv.org/abs/2505.20875v1 |
challenging. The second approach relies entirely on LLMs to generate varieties. While this method is scalable and convenient, several studies have highlighted the limitations of LLMs in accurately reproducing under-represented varieties of English[ 43,64,4,15], underscoring the risk of relying solely on LLMs. The third... | https://arxiv.org/abs/2505.20875v1 |
their applicability across different domains and language varieties [ 57,25]. In contrast, our approach integrates expert-curated resources with the linguistic capabilities of LLMs to construct a robust framework that captures diverse and accurate language expressions across varieties, while ensuring both linguistic va... | https://arxiv.org/abs/2505.20875v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.