text string | source string |
|---|---|
alignment between speech and text modalities on easily accessible ASR data, without reliance on large amount of hard-to-obtained supervised data. This well alignment al- lows the SLLM to fully leverage the generalization ability of the LLM. We train the SLLM to generate responses equivalent to the text-based LLM when g... | https://arxiv.org/abs/2505.18644v1 |
representation to the textual space of LLM with minimal computational cost. Since speech embeddings are typ- ically much longer than text embeddings, we use a CNN based subsampler connector to reduce the length discrepancy between two modalities. This connector functions as an intermediary between the speech encoder an... | https://arxiv.org/abs/2505.18644v1 |
this observation, we choose high-quality ASR datasets like LibriSpeech for training. Since LibriSpeech mostly consists of audiobooks, it is less suited for tasks like IC, or SF that we pre- viously explored. Therefore, we decided to construct several new content-related tasks: Continuation The continuation task encoura... | https://arxiv.org/abs/2505.18644v1 |
checks whether the generated output follows the prompt correctly and calculates the accuracy. 3.2. Task Generalization A key characteristic of LLMs is their zero-shot learning ability, also known as emergent ability, to perform tasks not encoun- tered during training. We define this capability as task general-ization. ... | https://arxiv.org/abs/2505.18644v1 |
as LLaMA2 + ASR). As shown in Table 1, our proposed method achieves comparable performance to LLaMA2 + ASR on the speaker role inference task, with a slight performance decrease on the mathematical reasoning task. Through case studies, we find that math questions require high accuracy in the transcribed results, as eve... | https://arxiv.org/abs/2505.18644v1 |
results show that the performance is significantly enhanced, supporting our observation and proving that our method can significantly im- prove the generalization of SLLM. 4.5. Ablation Study We perform an ablation study on constructed tasks to evaluate contribution of ASR SFT, multi-task learning, and speech-text inte... | https://arxiv.org/abs/2505.18644v1 |
Zhang, Z. Yan, C. Zhou, and J. Zhou, “Qwen-audio: Advancing universal audio understand- ing via unified large-scale audio-language models,” arXiv preprint arXiv:2311.07919 , 2023. [6] Y . Chu, J. Xu, Q. Yang, H. Wei, X. Wei, Z. Guo, Y . Leng, Y . Lv, J. He, J. Lin et al. , “Qwen2-audio technical report,” arXiv preprint... | https://arxiv.org/abs/2505.18644v1 |
2023, pp. 19 730–19 742. [21] Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wang et al. , “Cosyvoice 2: Scalable stream- ing speech synthesis with large language models,” arXiv preprint arXiv:2412.10117 , 2024. [22] OpenAI, “Hello gpt-4o,” OpenAI, Tech. Rep., 2024. [Online]. Available: ... | https://arxiv.org/abs/2505.18644v1 |
arXiv:2505.18646v1 [cs.SE] 24 May 2025SEW: Self-Evolving Agentic Workflows for Automated Code Generation Siwei Liu1Jinyuan Fang2Han Zhou3Yingxu Wang4Zaiqiao Meng2* 1University of Aberdeen2University of Glasgow3University of Cambridge4MBZUAI siwei.liu@abdn.ac.uk, hz416@cam.ac.uk, yingxv.wang@gmail.com, {jinyuan.fang,zai... | https://arxiv.org/abs/2505.18646v1 |
LLM’s full potential for autonomous adaptation. To address these limitations, we propose Self- Evolving Workflow ( SEW ), a novel framework designed to automatically generate and optimise multi-agent workflow. In particular, SEW achieves this by leveraging a novel evolutionary scheme to improve the workflow, i.e., the ... | https://arxiv.org/abs/2505.18646v1 |
model programs by refining both instructions and few- shot examples for each module. In multi-agent systems, recent studies have explored the evolu- tion of agentic workflows and topologies (Zhang et al., 2024a; Zhou et al., 2024, 2025; Zhang et al., 2025). For example, MASS (Zhou et al., 2025) exploits the optimizatio... | https://arxiv.org/abs/2505.18646v1 |
Workflow-Evolution, and (c) Agent-Evolution. The overview of our SEW framework is illustrated in Figure 2. As shown in Figure 2, our SEW first generates an initial workflow based on the task de- scription using one of the representation schemes introduced in Section 4. Second, the workflow evolution module of SEW will ... | https://arxiv.org/abs/2505.18646v1 |
programmer ...... You will NOT return anything except for the program."Agent's Prompt: Figure 3: Illustration of the Direct Evolution and Hyper Evolution of SEW. We use green, yellow and blue boxes to indicate the evolutionary prompt, default agent prompt and textual output of evolutionary operators. mutation prompt an... | https://arxiv.org/abs/2505.18646v1 |
cogni- tive heuristics (Fernando et al., 2024). For zero-order HE, we use the general cognitive heuristics Tthink to generate useful prompts for solving problems described by the task description D. Similar to how we use the mutation prompt Tmutto modify an agent’s prompt, we can use a hyper-mutation prompt instead of ... | https://arxiv.org/abs/2505.18646v1 |
Pseudo-code offers an abstraction that bridges natural language and formal code, making it an excellent choice for expressing workflows that need to be easily read and modified. To clearly illustrate the differences between workflow representation schemes, we present an example agentic workflow represented using both t... | https://arxiv.org/abs/2505.18646v1 |
AFlow (Zhang et al., 2024a): An automated framework that efficiently explores and optimises agentic workflows using Monte Carlo Tree Search. (5) PromptBreeder (Fer- nando et al., 2024) is a gradient-free evolutionary framework that improves agents by iteratively mu- tating and selecting prompt variants. 5.3 Experimenta... | https://arxiv.org/abs/2505.18646v1 |
LCB GPT-4o mini 80.2 63.4 38.0 task parsing workflow Only Workflow-Evolved 87.2 72.3 42.3 Workflow + Agent Co-Evolved 91.0 84.1 50.9 %improv. 4.35 16.3 20.3 code rewriting workflow Only Workflow-Evolved 88.8 70.5 40.1 Workflow + Agent Co-Evolved 92.1 80.8 46.8 %improv. 3.71 14.6 16.7 Table 3: Performance comparison (pa... | https://arxiv.org/abs/2505.18646v1 |
the performance of work- flows generated by SEW? We have introduced the Direct Evolution (DE) andHyper Evolution (HE) operators, where for each we proposed its corresponding lower-order and higher-order versions. To examine the effec- tiveness of different operators, we randomly sam- pled five different mutation prompt... | https://arxiv.org/abs/2505.18646v1 |
architectures could significantly impact its perfor- mance. Addressing these limitations in future work will be crucial for extending SEW’s applicability and enhancing its adaptability for broader tasks. References Eshaan Agarwal, Vivek Dani, Tanuja Ganu, and Ak- shay Nambi. 2024. Promptwizard: Task-aware agent-driven ... | https://arxiv.org/abs/2505.18646v1 |
model programs. arXiv preprint arXiv:2406.11695 . Chen Qian, Xin Cong, Cheng Yang, Weize Chen, Yusheng Su, Juyuan Xu, Zhiyuan Liu, and Maosong Sun. 2023. Communicative agents for software de- velopment. arXiv preprint arXiv:2307.07924 , 6(3). Bo Qiao, Liqun Li, Xu Zhang, Shilin He, Yu Kang, Chaoyun Zhang, Fangkai Yang,... | https://arxiv.org/abs/2505.18646v1 |
to denote a workflow represented by BPMN and CoRE, respectively. Be- low, we will show the same workflow represented by python, YAML and pseudo-code. 1# python_workflow 2steps = [ 3 {'name ':'task_parsing_agent ','args ': ['task_description '],'output ': 'parsed_task '}, 4 {'name ':'task_refinement_agent ','args ': ['t... | https://arxiv.org/abs/2505.18646v1 |
produce. Default workflow generation prompt: You are an AI workflow designer. Your task is to create a detailed Agent Workflow tailored to the provided workflow template and dataset description. Please follow these steps: 1.Review the Workflow Template: [De- tailed workflow template] 2.Analyze the Dataset Description: ... | https://arxiv.org/abs/2505.18646v1 |
to the specifications, ensuring it reads input from standard input (stdin) and writes out- put to standard output (stdout). You will be given the Problem Description followed by the Comments and Reasons why your previ- ous code fails. You will NOT return anything except for the program." ⊕{ Output from Code Reviewer Ag... | https://arxiv.org/abs/2505.18646v1 |
code could be modified to incorporate some of these creative thinking strategies:1def square_number ( number ): 2 """ Returns the square of the given number .""" 3 return number ** 2 4 5def main (): 6 import sys 7 input = sys . stdin . read 8 data = input (). strip (). splitlines () 9 try : 10 t = int ( data [0]) 11 ex... | https://arxiv.org/abs/2505.18646v1 |
sudden weather changes. •**Application:** Used in art installa- tions to create immersive environments that change with audience interaction. •**Visual Representation:** Imagine a swirling mass of steam with glowing embers, shifting in color from blue to orange.•**Narrative Element:** In a city plagued by drought, the ... | https://arxiv.org/abs/2505.18646v1 |
that each prompt generated is unique and unpre- dictable. iii. **Prompt Construction**: The selected elements are then combined into a coherent prompt that encourages the user to think cre- atively about how to merge the chosen con- cept with the selected action and medium. iv. **Example Usage**: The code includes a lo... | https://arxiv.org/abs/2505.18646v1 |
(x, y, label =word , alpha =0.5) 17 plt . title (\" Art Inspired by Words \") 18 plt . legend () 19 plt . show () 20 21def main (): 22 num_words = int ( input (\" How many random words would you like to generate ? \") ) 23 random_words = generate_random_words ( num_words ) 24 print (\" Generated Words :\" , random_word... | https://arxiv.org/abs/2505.18646v1 |
arXiv:2505.18651v1 [cs.CL] 24 May 2025On the Emergence of Linear Analogies in Word Embeddings Daniel J. Korchinski∗ Department of Physics Ecole Polytechnique Fédérale de Lausanne Lausanne, VD Switzerland daniel.korchinski@epfl.chDhruva Karkada Department of Physics UC Berkeley Berkeley, CA, USA dkarkada@berkeley.edu Ya... | https://arxiv.org/abs/2505.18651v1 |
exemplifying the target analogy are removed from the dataset [ 8] (e.g. removing from the corpus all word pairs including a masculine word and its feminine counterpart, such as man-woman, king-queen, etc.). 2. Our Contributions In this work, we propose a theoretical generative model for word co-occurrences, which ratio... | https://arxiv.org/abs/2505.18651v1 |
a ‘true’ Euclidean semantic space with some dimension din which words are associated with latent vectors. In that space, text generation corresponds to a random walk, such that closer words co-occur more often [ 21,23,25]. Using the assumption that this space is perfectly spherically 2 symmetric, [ 23] deduces that Wi·... | https://arxiv.org/abs/2505.18651v1 |
that in this noiseless version of the model, Eq.2 holds exactly. Indeed, if iandi∗differ by a single attribute (say the first one) such that α1 i=−α1 i∗= 1, and ais any word, then P(a|i)/P(a|i∗) = P1(α1 a,+1)/P1(α1 a,−1), a result which does not depend on the choice of i. Below we will add noise to the model to study h... | https://arxiv.org/abs/2505.18651v1 |
several attributes. As a result, the spectrum of Mdepends on the distribution of the sk. A representative Mwith minimal variance in that distribution is shown in Figure 1 b, in the symmetric case where qk= 1 for all k, and can be contrasted with the empirical Wikipedia derived co-occurrences reported in Figure 1 a. Fig... | https://arxiv.org/abs/2505.18651v1 |
Mijwith a vocabulary of 10,000 words. (b): Analogy accuracy for Wikipedia text, with different matrix targets Mijandlog(Mij+εR)(regularizer ε= 10−2). Shaded area indicates the sample standard deviation across analogy categories. (c): Analogy completion accuracy for a single realization of the model for d= 8, for matrix... | https://arxiv.org/abs/2505.18651v1 |
The Mikolov et al. analogy task set consists of 19,544 sets of four words analogies, e.g. “hand:hands::rat:rats”, divided among 13 families e.g. adjective-superlative, verb-participle, country- nationality, etc. In Figure 2 bwe report the average performance for embeddings obtained with different Mijandlog(Mij). The lo... | https://arxiv.org/abs/2505.18651v1 |
under Vocabulary Subsampling Our model assumes that all the possible 2dcombinations of attributes are incarnated into existing words, an assumption that is clearly unrealistic. We now show that even if we randomly prune an immense fraction fof the words, the spectral properties of the PMI matrix are remarkably robust, ... | https://arxiv.org/abs/2505.18651v1 |
of Mare not degenerate, its eigenvectors and eigenvalues are not affected in the limit of large d. Consequently, the embedding of each word is unaffected in this limit, and Eq.11 still holds. Numerical validation: In Figure 3 awe show the performance of the PMI matrix for different analogy families. For each family, we... | https://arxiv.org/abs/2505.18651v1 |
word statistics. For example, polysemantic words such as bank (a bank, river bank, to bank) complicate co-occurrence statistics. Furthermore, some attributes may be hierarchically organized [ 35]; this property can be captured by random hierarchy models [ 36] and is revealed by diffusion models [ 37,38]. The possibilit... | https://arxiv.org/abs/2505.18651v1 |
URL https://aclanthology.org/2023.blackboxnlp-1.2/ . [14] Andrew Lee, Xiaoyan Bai, Itamar Pres, Martin Wattenberg, Jonathan K Kummerfeld, and Rada Mihalcea. A mechanistic understanding of alignment algorithms: A case study on dpo and toxicity. arXiv preprint arXiv:2401.01967 , 2024. 10 [15] Gabriel Ilharco, Marco Tulio... | https://arxiv.org/abs/2505.18651v1 |
language models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning , volume 235 of Proceedings of Machine Learning Research , pages 39643–39666. PMLR, 21–27 Jul 2024.... | https://arxiv.org/abs/2505.18651v1 |
have that M(d′=1)=P(1), which is the Kronecker product of one term. Induction step : Assume that M(d′)=⊗d′ kP(k)for some d′≥1. We must show that M(d′+1)= M(d′)⊗P(d′+1). By the definition in Equation (22), we have: M(d′+1)(i, j) =(d′+1)Y k=1P(k)(α(k) i, α(k) j) = d′Y kP(k)(α(k) i, α(k) j) P(d′+1)(α(d′+1) i , α(d′+1)... | https://arxiv.org/abs/2505.18651v1 |
and fig. 5. As the example here shows, each successive eigenvector (after the first trivial 1vector) included in the 15 Word K= 2Embedding K= 3Embedding Man 2(1,−√s1) 2(1,−√s1,−√s2) Woman 2(1,−√s1) 2(1,−√s1,+√s2) King 2(1,+√s1) 2(1,+√s1,−√s2) Queen 2(1,+√s1) 2(1,+√s1,+√s2) Table 1: The explicitly constructed embeddings... | https://arxiv.org/abs/2505.18651v1 |
of the embedding space — necessary for analogy — remains intact up to small corrections. A4. Additional numerical study of the model Here, we report several additional experiments on the symmetric binary semantics model. Dependence of analogy accuracy on disribution of the sk In Figure 6, we measure the analogy accurac... | https://arxiv.org/abs/2505.18651v1 |
a log(M)matrix with sk∈(0,1)as a function of sparsification. Shaded area reflects the standard deviation across 20 realizations of sparsification, for the same fixed sk. Inset is the same data, but rescaling by the asymptotic value of each eigenvalue so as to effect a collapse for m > d . The2d×2ddimensional ξmatrix ha... | https://arxiv.org/abs/2505.18651v1 |
arXiv:2505.18653v1 [cs.CL] 24 May 2025Climate-Eval: A Comprehensive Benchmark for NLP Tasks Related to Climate Change Murathan Kurfalı†¶Shorouq Zahra†§¶ Joakim Nivre§¶Gabriele Messori§¶ †RISE Research Institutes of Sweden §Uppsala University ¶Swedish Centre for Impacts of Climate Extremes (climes) Abstract ClimateEval ... | https://arxiv.org/abs/2505.18653v1 |
builds upon previous NLP benchmarking datasets, most notably ClimaBench (Spokoyny et al., 2023) and adds the following contributions: •We introduce a new topic classification dataset using news articles on climate-related topics from the Guardian newspaper. •We compile diverse climate-related NLP tasks into a unified b... | https://arxiv.org/abs/2505.18653v1 |
3.1 Datasets and Tasks In this subsection, we first describe our new dataset, the Guardian Climate News Corpus, followed by the other datasets included in ClimateEval. •Guardian Climate News Corpus : A dataset containing climate-related and non-climate- related articles. These are assigned to nine climate-related categ... | https://arxiv.org/abs/2505.18653v1 |
of tags that are relevant to each of the ten categories with ease. As example, an article that we categorize as falling under climate activism is scraped based on a list of article tags that relate to climate ac- tivism (such as “environment/school-climate- strikes”). However, this article could also have been assigned... | https://arxiv.org/abs/2505.18653v1 |
the 1,535 claims is paired with five ev- idence sentences extracted from Wikipedia, 5The code used for scraping the dataset from the Guardian’s Open Platform is available on GitHub: https: //github.com/NLP-RISE/extractguardianwhich either support ,refute , orprovide insuf- ficient information about the claim. Climate- ... | https://arxiv.org/abs/2505.18653v1 |
manner, we aim to evaluate both retrieval-augmented and self-knowledge-based approaches to climate- related QA. •Exeter Misinformation (Coan et al., 2021): A dataset designed to detect climate misin- formation by annotating text from prominent climate contrarian blogs and think tanks span- ning over 20 years (1998–2020... | https://arxiv.org/abs/2505.18653v1 |
.42 .63 .66 .71 .56 .63 .63 .65 .61 .60 .56 .69 .68 .71 .71 .71 .68 .68 Climate Detection .20 .70 .65 .69 .61 .60 .59 .66 .62 .68 .53 .73 .50 .76 .69 .78 .71 .76 Climate Eng .37 .46 .52 .59 .37 .50 .52 .53 .53 .50 .52 .54 .55 .59 .60 .59 .57 .59 Climate NER .11 .19 .14 .20 .10 .17 .09 .17 .07 .14 .04 .18 .16 .21 .23 .3... | https://arxiv.org/abs/2505.18653v1 |
a climate-related commitment. Texts are labeled as commitment-yes if they explicitly state planned or ongoing climate actions and commitment-no if they do not reference concrete climate actions. iv.Climate Specificity : A binary classifi- cation task assessing the specificity of corporate climate commitments. A pas- sa... | https://arxiv.org/abs/2505.18653v1 |
classes equally to en- sure a balanced assessment by preventing frequent labels from biasing the results. For Climate-Change-NER, the only sequence labeling task in our benchmark, we compute preci- sion, recall, and F1-score based on the entity-type and entity-span pairs. Each entity-type is evalu- ated independently, ... | https://arxiv.org/abs/2505.18653v1 |
Climate Specificity .58 (.13) .69 (.07) .10 Climate Stance .17 (.07) .51 (.09) .34 Climate-Fever .43 (.11) .46 (.08) .03 ClimaText Sent Clf. .55 (.08) .67 (.03) .12 Env. Claims .71 (.10) .80 (.03) .09 Exeter Claim .33 (.16) .45 (.10) .11 Exeter Sub-Claim .34 (.19) .39 (.17) .05 Guardian Body .42 (.13) .41 (.18) -.01 Gu... | https://arxiv.org/abs/2505.18653v1 |
the CheapTalk dataset tasks. Intuitively, one might expect tasks like Climate Detection (whether or not a text is climate-related) or Climate Sentiment (classifying tone as risk, opportunity or neutral) to be straightforward for advanced LLMs. Yet, our results show that even the best models achieve F1 scores below 0.8 ... | https://arxiv.org/abs/2505.18653v1 |
0.04 SciDCC Title Sum. 0.04 -0.10 0.08 0.00 SciDCC Title Body -0.00 -0.02 0.08 0.01 TCFD Recommend. -0.08 -0.02 0.05 0.07 Average 0.05 0.01 -0.04 0.04 Table 4: Comparison of ClimateGPT against its’ base Llama models in 0-shot and 5-shot settings. The values represent the difference where positive values (high- lighted ... | https://arxiv.org/abs/2505.18653v1 |
Ben Wang, Kevin Wang, and Andy Zou. 2024. A framework for few-shot language model evaluation.Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang... | https://arxiv.org/abs/2505.18653v1 |
Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Sri- vastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet... | https://arxiv.org/abs/2505.18653v1 |
Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lind- say, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, So... | https://arxiv.org/abs/2505.18653v1 |
End-to-End Climate Change Knowledge Graph Construction. In ICML 2021 Workshop on Tackling Climate Change with Machine Learning . Paulo Pirozelli, Marcos M José, Igor Silveira, Flávio Nakasato, Sarajane M Peres, Anarosa AF Brandão, Anna HR Costa, and Fabio G Cozman. 2024. Bench- marks for pirá 2.0, a Reading Comprehensi... | https://arxiv.org/abs/2505.18653v1 |
Wei, Victor Cotruta, Phoebe Kirk, Anand Rao, Minh Giang, Ludovic Peran, Tris Warkentin, Eli Collins, Joelle Barral, Zoubin Ghahramani, Raia Hadsell, D. Sculley, Jeanine Banks, Anca Dragan, Slav Petrov, Oriol Vinyals, Jeff Dean, Demis Hass- abis, Koray Kavukcuoglu, Clement Farabet, Elena Buchatskaya, Sebastian Borgeaud,... | https://arxiv.org/abs/2505.18653v1 |
Total emissions are estimated to be 41.47 kgCO 2eq of which 0 percent were directly offset. Estimations were conducted using the MachineLearning Impact calculator pre- sented in Lacoste et al. (2019). C Label distribution To provide further insights into the datasets, we visualize the label distribution for each test s... | https://arxiv.org/abs/2505.18653v1 |
arXiv:2505.18658v1 [cs.CL] 24 May 2025Robustness in Large Language Models: A Survey of Mitiga- tion Strategies and Evaluation Metrics Pankaj Kumar pankaj.kumar@niser.ac.in School of Computer Science National Institute of Science Education and Research Homi Bhabha National Institute Subhankar Mishra smishra@niser.ac.in ... | https://arxiv.org/abs/2505.18658v1 |
critical challenge. LLMs may encounter unpredictable variations in language, shifts in data (changes in the data patterns that a model encounters, which may affect its performance), and adversarial inputs that can affect their performance. In practical applications, these models often face noisy, unstructured text, bre... | https://arxiv.org/abs/2505.18658v1 |
ability to perform reliably and safely despite challenges like input variations, unexpected data, or adversarial attacks. It is frequently discussed as a critical component within the broader framework of AI Trustworthiness. Trustworthy AI systems are expected to be reliable, safe, fair, transparent, accountable, and r... | https://arxiv.org/abs/2505.18658v1 |
(2020); Li et al. (2022) AlignmentOh & Demberg (2025); Ouyang et al. (2022); Christiano et al. (2023) Hong et al. (2024); Ethayarajh et al. (2024); Gupta et al. (2025) Tran et al. (2025); Sun (2024); Herrera-Poyatos et al. (2025) Zhong et al. (2025a); Xu et al. (2023); Wang et al. (2024a) Dong et al. (2024b); Xiong et ... | https://arxiv.org/abs/2505.18658v1 |
diverse scenarios (Lad et al., 2024; Aljohani et al., 2025). This involves minimizing errors, factual inaccuracies (hallucinations), and harmful biases, even under challenging conditions. •Adherence to Intended Behaviour: A robust model should adhere to its intended function and safety constraints. This includes follow... | https://arxiv.org/abs/2505.18658v1 |
Real-world data distributions often shift over time or across contexts, making OOD generalization crucial for practical deployment. •Appearance: Significant performance drop occurs in LLMs when evaluated on out-of- distribution (OOD) test sets compared to in-distribution (ID) benchmarks (Du et al., 2022). This degradat... | https://arxiv.org/abs/2505.18658v1 |
imbalance and skewed learned distri- butions. 7. Task-Specific Robustness: Robustness considerations tailored to the demands of specific downstream tasks, such as maintaining logical coherence in reasoning tasks or generating functionally correct and secure code (Bao et al., 2025; Li et al., 2025d). •Appearance: Signif... | https://arxiv.org/abs/2505.18658v1 |
their vulnerabilities. Several primary causes contribute to the lack of robustness observed in LLMs are: 3.1.1 Spurious Correlations This is arguably one of the most significant causes of robustness failures, particularly concerning OOD generalization (Du et al., 2022). Instead of learning the intended, often complex, ... | https://arxiv.org/abs/2505.18658v1 |
deliberately injected noise (Zhou et al., 2024d). 8 Table 1: Sources of non-robustness in LLMs. Specific Cause Root Cause Description Data Related Shortcut Learning / Data collection/annotation issues, Model exploits shallow statistical cues Spurious Correlations mirrored cognitive biases, simplic- ity bias in ERM lear... | https://arxiv.org/abs/2505.18658v1 |
Alignment Tax This phenomenon, termed the “alignment tax”, means that improving alignment (e.g., helpfulness, harm- lessness) might come at the cost of reduced performance on general knowledge benchmarks or can sometimes negatively impact model calibration (leading to overconfidence) or introduce new vulnerabilities (C... | https://arxiv.org/abs/2505.18658v1 |
inputs exploit the model’s instruction-following nature to bypass safeguards, and input tokenization steps distort rare words or non-standard syntax, amplifying errors. These challenges, inherent to the inference phase, can introduce these vulnerabilities as: 3.4.1 Inference and Decoding Vulnerabilities Even with a tra... | https://arxiv.org/abs/2505.18658v1 |
training or fine-tuning occurs. These methods focus on curating, cleaning, or augmenting the vast datasets used to train LLMs, aiming to embed robustness characteristics implicitly through the data the model learns from. They effectively address data biases and vulnerabilities in raw training data, preventing the model... | https://arxiv.org/abs/2505.18658v1 |
training, but this data often includes noise, toxic content, societal biases, personal information, and factual inaccuracies (Huang et al., 2024a). Since LLMS can memorise such flaws during training, these issues directly translate into risks like biased outputs, privacy breaches, and unreliable behaviour in real- worl... | https://arxiv.org/abs/2505.18658v1 |
data, thereby forcing the model to learn more robust features. It can be formulated as a minimax optimization problem: the inner loop maximizes the loss by finding the worst-case adversarial example within a defined perturbation budget, while the outer loop minimizes the loss on these adversarial examples (along with c... | https://arxiv.org/abs/2505.18658v1 |
resource-intensive, and fine-tuning massive LLMs on such augmented datasets ampli- fies training costs exponentially. Third, AT often induces a robustness-utility trade-off, where improvements in adversarial accuracy come at the expense of degraded performance on clean data, necessitating careful balancing. Finally, ro... | https://arxiv.org/abs/2505.18658v1 |
objectives to maximize scores without genuinely adhering to human intent (Herrera-Poyatos et al., 2025). Finally, methods like RLHF demand extensive computational resources and rely heavily on high-quality, diverse training data, limiting scalability. Addressing these trade-offs remains critical to developing reliable,... | https://arxiv.org/abs/2505.18658v1 |
Further, Hu et al. (2024) proposes RobustGER, which teaches LLMs to perform language-space denoising by incorporating an embedding derived from ASR (Automatic Speech Recognition) N-best lists, refined through audio distillation. Another novel approach, referenced instruction tracking, requires LLMS to explicitly cite w... | https://arxiv.org/abs/2505.18658v1 |
the practicality of input/output defences, their potential to enhance LLM robustness will likely remain secondary. Inference-Time Adaptation and Transformation Beyond prompt manipulation and decoding, other intra-processing techniques involve adapting the model or its execution environment at inference time. While effi... | https://arxiv.org/abs/2505.18658v1 |
methods like Erase-and-Check offer strong defences, their computational cost of- ten requires simplified variants for practical use, highlighting the trade-off between robustness and efficiency. Together, these strategies underscore the importance of layered verification to balance safety, accuracy, and usability in re... | https://arxiv.org/abs/2505.18658v1 |
centage) between a standard metric (like Accuracy or F1-score) on the IID test set and the same metriconthechallengingtestset(e.g., anOODdatasetoranadversariallyperturbeddataset)(Yuan et al., 2023; Du et al., 2023). A larger drop indicates lower robustness to that specific challenge. For example, a 20% accuracy drop on... | https://arxiv.org/abs/2505.18658v1 |
setups (e.g., prompts, seeds). Response Consistency Prompt Robustness Frequency of the most common claim across (RC) (Rahman et al., 2024)paraphrased prompts. LLM-based Sampling Stability, LLM judges assess consistency between generated Consistency Score (Saxena et al., 2024)Reliability samples. Calibration Expected Ca... | https://arxiv.org/abs/2505.18658v1 |
metric specifically measures consistency across multiple (e.g., 15) paraphrased versions of a single prompt. For each set of paraphrased prompts, it calculates the frequency of the most common factual claim generated in the responses. The overall RC is the average of these frequencies. •Multilingual Consistency: Evalua... | https://arxiv.org/abs/2505.18658v1 |
for certain groups under adversarial pressure. Metrics derived from causal frameworks, like stratified invariance, can also quantify the model’s reliance on protected attributes under intervention (Cotta & Maddison, 2024). 5.1.6 Task-Specific Robustness Metrics Certain applications require robustness dimensions unique ... | https://arxiv.org/abs/2505.18658v1 |
generate questions from the generated text and check if the answers are consistent with the source document. Alternatively, Natural Language Inference (NLI) models can classify the relationship (entailment, contradiction, neutral) between generated statements and source text sentences (Chen et al., 2024). •Faithfulness... | https://arxiv.org/abs/2505.18658v1 |
hallucination rates measured using the metrics above under standard conditions versus stress conditions. –Methodology: Applyachosenhallucinationmetric(e.g., FCHrate, FEWLscore, SelfCheckGPT score) to a standard benchmark (e.g., TruthfulQA) and then again to a stressed version of that benchmark (e.g., TruthfulQA prompts... | https://arxiv.org/abs/2505.18658v1 |
evaluating code- specific hallucinations (Multimodal) AMBER AMBER dataset (cus- tom annotated images)LLM-freeevaluationofexistence, attribute, relationhal- lucinations in MLLMs (Multimodal) HQHBench / ODE HQHBench dataset; ODE frameworkMeta-evaluation of benchmark quality (HQHBench); Dynamic open-set generation (ODE) T... | https://arxiv.org/abs/2505.18658v1 |
SemEval (SA); ANLI, ContractNLI, WANLI (NLI); AdvQA, NewsQA, SearchQA (EQA) (Yuan et al., 2023). •ANLI (Adversarial NLI): Created through an iterative human-and-model-in-the-loop process, ANLI contains NLI examples specifically designed to be difficult for contemporary models. Due to its challenging nature and distinct... | https://arxiv.org/abs/2505.18658v1 |
2025): These extend standard logical reasoning datasets (ReClor, LogiQA, LogiQAv2) by introducing structural variations to the multiple-choice questions. These variations include randomly shuffling the answer options or replacing the correct answer with a “none of the other options is correct” choice, testing if the mo... | https://arxiv.org/abs/2505.18658v1 |
et al., 2024). •UMWP (Unanswerable Math Word Problem): This benchmark (Sun et al., 2024a) evaluates hal- lucination in the context of mathematical reasoning. It consists of math word problems that are intentionally designed to be unanswerable (e.g., due to missing information). It tests whether LLMs correctly identify ... | https://arxiv.org/abs/2505.18658v1 |
sub-domains or from particular perspectives. Acknowledging and synthesizing these existing reviews is crucial for understanding the broader landscape. Numerous surveys cover LLMs in general, often touching upon robustness as one aspect among many. A significant number focus on the concept of Trustworthiness in LLMs, co... | https://arxiv.org/abs/2505.18658v1 |
for operationalising trustworthiness (including robustness) is noted (de Cerqueira et al., 2025). The difficulty in evaluating generative tasks robustly is frequently mentioned. Understanding and mitigating more subtle or complex shortcuts beyond simple lexical cues remains an area for research (Howe et al., 2025; Zhou... | https://arxiv.org/abs/2505.18658v1 |
Invariant Learning: Moving beyond correlational learning towards models that under- standunderlyingcausalrelationshipsorlearnrepresentationsthatareinherentlyinvarianttochanges in domain or distribution (Ye et al., 2024). This could lead to more fundamentally robust models less reliant on spurious cues. 2. Compositional... | https://arxiv.org/abs/2505.18658v1 |
for safe and responsible real-world adoption. Though the field has advanced significantly in diagnosing vulnera- bilities and proposing mitigation strategies, fundamental challenges persist. Future progress will depend on four key priorities: (1) deepening theoretical insights into causality and invariance, (2) develop... | https://arxiv.org/abs/2505.18658v1 |
Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report, 2025a. URL https://arxiv.org/ abs/2502.13923 . Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness, 2021a. URL https://arxiv.org/abs/2102.01356 . Xiang Bai, Rishi Bomm... | https://arxiv.org/abs/2505.18658v1 |
. Junyi Chai and Xiaoqian Wang. Fairness with adaptive weights. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato (eds.), Proceedings ofthe39thInternational Conference onMachine Learning, volume 162 of Proceedings ofMachine Learning Research, pp. 2853– 2866. PMLR, 07 2022. U... | https://arxiv.org/abs/2505.18658v1 |
Jiawei Wang, Jin Chen, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, Junxiao Song, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Lei Xu, Leyi Xia, Liang Zhao, Litong Wang, Liyue Zhang, Meng Li, Miaojun Wang, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingming Li, Nin... | https://arxiv.org/abs/2505.18658v1 |
de Melo. Data augmentation with adversar- ial training for cross-lingual NLI. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli (eds.), Proceedings ofthe59thAnnualMeeting oftheAssociation forComputational Linguistics andthe11th International JointConference onNaturalLanguage Processing (Volume 1:LongPapers), p... | https://arxiv.org/abs/2505.18658v1 |
Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nikolay Bashlykov, Nikolay Bogoychev, Niladri Chatterji, Ning Zhang, Olivier Duchenne, Onur Çelebi, Patrick Alrassy, Pengchuan Zhang, Peng- wei Li, Petar Vasic, Peter Weng, Prajjwal Bhargava, Pratik Dubal, Praveen Krishnan, Punit... | https://arxiv.org/abs/2505.18658v1 |
Kam Hou U, Karan Saxena, Kartikay Khandelwal, Katayoun Zand, Kathy Matosich, Kaushik Veeraraghavan, Kelly Michelena, Keqian Li, Kiran Jagadeesh, Kun Huang, Kunal Chawla, Kyle Huang, Lailin Chen, Lakshya Garg, Lavender A, Leandro Silva, Lee Bell, Lei Zhang, Liangpeng Guo, Licheng Yu, Liron Moshkovich, Luca Wehrstedt, Ma... | https://arxiv.org/abs/2505.18658v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.