text
string
source
string
language models with knowledge enhance- ment and filtering framework. In NAACL Findings , pages 3860–3871. Qingfei Zhao, Ruobing Wang, Yukuo Cen, Daren Zha, Shicheng Tan, Yuxiao Dong, and Jie Tang. 2024a. LongRAG: A dual-perspective retrieval-augmented generation paradigm for long-context question an- swering. In EMNLP...
https://arxiv.org/abs/2505.20099v1
Scholar and PaSa2using the phrases “knowledge graph and language model for question answering" and “KG and LLM for QA", and extend the search scope of the benchmark dataset paper to 2019, and then screen and select them based on their relevancy and publication venue quality. B Taxonomy B.1 Complex QA We divide the comp...
https://arxiv.org/abs/2505.20099v1
Table 1: Comparsion of Existing Survey Across LLMs, KGs, LLMs+KGs, GraphRAG, and QA. mKGQA: multi- lingual Question Answering for Knowledge Graphs, TKGQA: Temporal Knowledge Graph QA. "Covered or Dicussed, #Not Covered or Dicussed. C Summary and Alignment C.1 Summary Tables of Approaches The detailed summarization and ...
https://arxiv.org/abs/2505.20099v1
the user query and the re- trieved documents, which is divided into multiple 16 Approach Strength Limitation KG Requirement KG as Background Knowledge Broad Coverage Static Knowledge High Domain Coverage KG as Reasoning Guidelines Multi-hop Capabilities Computational Overhead Rich Relational Paths KG as Refiners and Va...
https://arxiv.org/abs/2505.20099v1
even if several optimizations and ranking strategies have been recently investigated to reduce the costs of graph retrieval, graph rea- soning, and the length of the context of LLMs, however, the relevant subgraphs extraction, graph reasoning, and vector-based retrieval remain a com- putationally costly task. E Evaluat...
https://arxiv.org/abs/2505.20099v1
(ReaQ) : the correctness of generated rea- soning chains and intermediate steps that explain how the final answer is derived. E.2 Applications We showcase the applications and demos in synthe- sizing LLMs with KGs for QA. KAG (Liang et al., 2024a) (by Antgroup)3is a domain-knowledge augmented generation frame- work tha...
https://arxiv.org/abs/2505.20099v1
BGE-1.5-en-baseDataset Inherent KGsWQAP, CWQ KGQA Hits@1, F1 KG-IRAG (Yang et al., 2025) Incremental Retrieval and Iterative ReasoningLlama-3-8B-Instruct, GPT-3.5-Turbo, GPT- 4o-mini, GPT-4oSelf-constructed KGsTFNSW Temporal QA EM, F1, HR, HAL PIP-KAG (Huang et al., 2025) Parameteric Pruning for KAGLlma-3-8B-Instruct D...
https://arxiv.org/abs/2505.20099v1
et al., 2024) LLM-KG QA " " # LLMs+KGs QA over the enterprise SQL database. LLM-KG-Bench (Meyer et al., 2023) LLM-KG QA " # # LLMs in knowledge graph engineering. XplainLLM (Chen et al., 2024c) LLM-KG QA " △" Focuses on QA explainability and rea- soning OKGQA (Sui and Hooi, 2024) LLM-KG QA " " " Evaluates LLMs+KGs for ...
https://arxiv.org/abs/2505.20099v1
arXiv:2505.20101v2 [cs.CL] 27 May 2025 2025-05-28 Adaptive Deep Reasoning: Triggering Deep Thinking When Needed Yunhao Wang†, Yuhao Zhang†, Tinghao Yu, Can Xu, Feng Zhang, Fengzong Lian Tencent Hunyuan Team {luciuswang, yuhaozzhang, maxwellyu, leocaxu, jayzhang, faxonlian }@tencent.com Technical Report Abstract Large l...
https://arxiv.org/abs/2505.20101v2
to foster more concise reasoning. Techniques like TokenSkip Xia et al. (2025) and C3oT Kang et al. (2024) aim to pro- duce shorter reasoning chains by either omitting irrelevant steps or compressing the reasoning process. Additionally, reinforcement learning with length-based rewards has been utilized to optimize reaso...
https://arxiv.org/abs/2505.20101v2
prompts to reduce computations of single steps Chen et al. (2024). However, the effectiveness of these strategies relies on domain-specific tuning and may not be universally applicable. Another approach to model-based efficient reasoning involves fine-tuning large lan- guage models with variable-length chain-of-thought...
https://arxiv.org/abs/2505.20101v2
This allows the SFT model to explicitly control the generation of both long and short chains of thought during reinforcement learning sampling through instruction. Although conventional prefix decoding (e.g., prepending ” <think>”) can be used to select the reasoning mode, it often compromises proficiency. In contrast,...
https://arxiv.org/abs/2505.20101v2
the qual- ity of each sampling. Based on these signals, we further perform reward shaping, which assists the model in choosing between long and short reasoning modes during reinforcement learning training. 3.2.3 Reward Shaping To enable the model to automatically choose between long and short reasoning modes and to mit...
https://arxiv.org/abs/2505.20101v2
step≥Twarmup )(3) Soft Length Penalty : To prevent excessive elaboration in long chain reasoning, we incorporate a length penalty into the reward system for long chains of thought. When multiple reasoning paths lead to the correct answer, shorter traces are granted higher rewards, thus reducing redundancy while maintai...
https://arxiv.org/abs/2505.20101v2
high data quality, we exclude proof-based questions, multiple-choice items, and true/- false problems from the dataset. We employed rejection sampling to collect both long and short CoT data using state-of-the-art reasoning models. To ensure the accuracy of this data, we used an evaluation LLM, which have been utilized...
https://arxiv.org/abs/2505.20101v2
are underlined. the models predominantly employ long-chain reasoning to solve the problems, thereby achieving accuracy comparable to that of models dedicated to long-chain reasoning. 5 Conclusion In this study, we propose a method that enables a reasoning model to dynamically switch between short and long reasoning cha...
https://arxiv.org/abs/2505.20101v2
of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024. Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, and et al. Kimi k1.5: Scaling reinforcement learning with llms, 2025a. URL https://arxiv.org/abs/2501....
https://arxiv.org/abs/2505.20101v2
arXiv:2505.20103v2 [cs.DL] 27 May 2025SCIRGC: M ULTI -GRANULARITY CITATION RECOMMENDATION AND CITATION SENTENCE PREFERENCE ALIGNMENT Xiangyu Li Nanjing University of Posts and Telecommunications Nanjing, China 32804131lxy@gmail.comJingqiang Chen Nanjing University of Posts and Telecommunications Nanjing, China cjq@njup...
https://arxiv.org/abs/2505.20103v2
this study, we elaborate on the challenges and solutions at each stage of the process and propose the SciRGC framework consisting of a citation articles recommendation module and a citation sentence generation module. APREPRINT - M AY28, 2025 Figure 1: In the process of citation generation, it is first necessary to inf...
https://arxiv.org/abs/2505.20103v2
citation sentence generation, this paper proposes a citation fine-tuning Chain-of-Thought method for large language models, which greatly improves the quality of generated citation text. 2 APREPRINT - M AY28, 2025 •This paper also introduces a new evaluation metric for citation sentences, effectively addressing the lim...
https://arxiv.org/abs/2505.20103v2
citation network, where papers are represented as nodes, and edges indicate direct citation relationships. Based on this network, and inspired by previous research [ 36], we design two collaborative filtering algorithms as shown in fig 3. The core assumption is that if two papers share a large number of the same citati...
https://arxiv.org/abs/2505.20103v2
paragraph representations (title, abstract, and local context) with their corresponding paragraph embeddings. By adding paragraph type embeddings, we generate type-aware paragraph embeddings. These type-aware paragraph embeddings are then fed into the transformer encoder layer, and the final document embedding is obtai...
https://arxiv.org/abs/2505.20103v2
the model’s complex reasoning potential and guide its progressive thinking. The core idea behind the generation of reasoning data in this paper is to generate the reasoning process for citation generation using a large teacher model, and then use these reasoning results to fine-tune a smaller student model, enabling it...
https://arxiv.org/abs/2505.20103v2
score of the reward model, and Dis the training dataset. By optimizing this objective function, DPO makes the generated citations more in line with human preferences, thereby enhancing the academic nature and readability of the text. 4 Evaluation Criteria The construction of academic citations faces dual challenges: fi...
https://arxiv.org/abs/2505.20103v2
- - 0.28 0.534 - - BERT-GCN - - 0.418 0.529 - - - - BERT 0.482 0.736 0.458 0.706 0.309 0.535 0.226 0.399 SciBERT 0.531 0.779 0.536 0.773 0.380 0.623 0.278 0.475 SciBERT-Intent 0.552 0.797 0.554 0.784 0.400 0.650 0.291 0.496 Table 2: Experimental Results of Citation Recommendation Reranking Modeling support), and (c) Co...
https://arxiv.org/abs/2505.20103v2
19.19 - - PTGEN-Cross 27.08 7.14 20.61 - - BART-large 29.62 9.86 24.51 - - GPT-4 [42] - - - 82.59 84.87 Deepseek-Chat-v3 [48] - - - 82.91 80.96 Qwen2.5-32B-Instruct [49] - - - 72.11 76.53 SciRGC 37.54 15.63 28.74 84.22 82.04 Table 3: Comparison of automated assessment results between the SciRGC framework and baseline m...
https://arxiv.org/abs/2505.20103v2
achieves the best performance in semantic fidelity ( 85.22) and information compression ( 82.06), with a significantly lower term variation rate compared to Deepseek-V3. Leveraging an enhanced contextual modeling mechanism, Deepseek-V3 surpasses GPT-4 in contextual coherence ( 80.16). The original Qwen2.5-32B model per...
https://arxiv.org/abs/2505.20103v2
best and has the lowest misclassification rate. Overall, the experimental results validate the effectiveness of the model in this classification task, providing more accurate classification results, and its performance on external test sets also shows its good generalization performance. B CITEVAL Scoring Guidelines •P...
https://arxiv.org/abs/2505.20103v2
Nov. 2002. doi: 10.1145/587078.587096. [10] C. Bhagavatula, S. Feldman, R. Power, and W. Ammar, “Content-Based Citation Recommendation,” Cornell University - arXiv,Cornell University - arXiv , Feb. 2018. [11] Z. Medi ´c and J. Snajder, “Improved Local Citation Recommendation Based on Context Enhanced with Global Inform...
https://arxiv.org/abs/2505.20103v2
Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , Vancouver, Canada, Jan. 2017. doi: 10.18653/v1/p17-1099. [26] N. Gu and RichardH. R. Hahnloser, “Controllable Citation Text Generation,” Nov. 2022. [27] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and ...
https://arxiv.org/abs/2505.20103v2
Language-Agnostic Suicidal Risk Detection Using Large Language Models June-Woo Kim1,2, Wonkyo Oh3, Haram Yoon3, Sung-Hoon Yoon1,3, Dae-Jin Kim1, Dong-Ho Lee3, Sang-Yeol Lee1,3, Chan-Mo Yang†1,3 1Department of Psychiatry, Wonkwang University Hospital, Republic of Korea 2RSC LAB, MODULABS, Republic of Korea 3Department o...
https://arxiv.org/abs/2505.20109v1
limitation: they are predominantly language-specific. Most current approaches rely on pretrained language models de- signed for specific languages, necessitating separate models for different linguistic populations. This dependence on monolin- gual models introduces two major challenges: (i)Limited scal- ability . Each...
https://arxiv.org/abs/2505.20109v1
ReLU ac- tivation, 0.1 of dropout, and a classifier, using a learning rate of 1e–3 and a batch size of 32. In all the experiments, we applied Adam optimizer with cosine scheduling. The final predictions for the dev and test sets were obtained by aggregating logits from the three tasks using a voting mechanism. 2.3. Eva...
https://arxiv.org/abs/2505.20109v1
10] have shown that LLMs can extract men- tal health-related textual features directly from the text, reduc- ing manual intervention and enhancing mental health detection performance. Building on this approach, we propose to extract meaningful suicidal risk-related text features from ASR results by leveraging LLMs, as ...
https://arxiv.org/abs/2505.20109v1
while the PR task is excluded since all text content remains identical. 3.2. Speech Modality For fine-tuning with speech foundation models, we employ XLSR53 [24], which is pretrained on 56k hours of multilingual speech, WavLM-plus-base [25], pretrained on 94k hours of En- glish speech, and wav2vec2-Large [26], which is...
https://arxiv.org/abs/2505.20109v1
61 64.22 61 61.39 Qwen-DepRoBERTa+XLSR 59 57.73 61 64.22 61 61.39 that Qwen-plus based suicidal risk-related features are more ef- fective in capturing ED task. For the English models, DepRoBERTa trained on direct English translations of ASR results achieved in ER (55% Acc, 57.14% F1) and ED (56% Acc, 57.69% F1). GPT- ...
https://arxiv.org/abs/2505.20109v1
findings presented in this study are based on the scoring framework of the MINI-KID scale, which assesses current sui- cide risk as at risk or no risk. This classification reflects partic- ipants’ immediate responses to the MINI-KID assessment and should not be interpreted as a prediction of future suicidal be- havior....
https://arxiv.org/abs/2505.20109v1
of depression from social media text using roberta pre-trained language models,” in LTEDI , 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:248780097 [9] M. Sadeghi, B. Egger, R. Agahi, R. Richer, K. Capito, L. H. Rupp, L. Schindler-Gmelch, M. Berking, and B. M. Eskofier, “Exploring the capabilities...
https://arxiv.org/abs/2505.20109v1
Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018. [23] Z. Zhao, H. Chen, J. Zhang, W. X. Zhao, T. Liu, W. Lu, X. Chen, H. Deng, Q. Ju, and X. Du, “Uer: An open-source toolkit for pre-training models,” in Proceedings of the 2019 Conference on Empirical M...
https://arxiv.org/abs/2505.20109v1
ResSVD: Residual Compensated SVD for Large Language Model Compression Haolei Bai Westlake University Nanyang Technological UniversitySiyong Jian Westlake University Nanjing UniversityTuo Liang Case Western Reserve University Yu Yin Case Western Reserve UniversityHuan Wang∗ Westlake University Abstract Large language mo...
https://arxiv.org/abs/2505.20112v1
apply weight matrix scaling before SVD truncation, SVD-LLM outperforms ASVD by leveraging a data whitening technique that enables a direct mapping between singular values and truncation loss. As a further advancement, AdaSVD compensates for truncation loss by iteratively updating singular matrices. However, existing me...
https://arxiv.org/abs/2505.20112v1
example, SparseGPT [ 9] performs one-shot pruning using second-order approximations with- out retraining. However, since unstructured pruning retains the original matrix shape, it offers limited inference acceleration and requires specialized hardware. In contrast, structured pruning eliminates entire blocks or channel...
https://arxiv.org/abs/2505.20112v1
target compression ratio can provide better performance for compressed models. However, existing methods compress all layers, either apply a uniform compression ratio across all layers or assign variable ratios based on layer-wise importance, which often results in sub-optimal performance. 3 ResSVD The framework of Res...
https://arxiv.org/abs/2505.20112v1
second SVD on Rand retain the top- r2singular values:Rr2=Ur2Σr2VT r2. Finally, we obtain the compressed weight matrix by combining the two approximations: ˆWr=Wr1+Rr2=UrΣrVT r. In this way, the residual-compensated compressed matrix ˆWrprovides a closer approximation to the original weight matrix Wthan the directly tru...
https://arxiv.org/abs/2505.20112v1
12 layers Compress Last 10 layers Compress Last 8 layers 0 5 10 15 20 25 30 Layer Index0.00.20.40.60.81.01.21.4Layer-wise Error Vicuna-7B Compress First 8 Layers Compress First 16 Layers Compress All layers Compress Last 30 layers Compress Last 28 layers Compress Last 26 layers Compress Last 24 layers Compress Last 22 ...
https://arxiv.org/abs/2505.20112v1
37], LLaMA-2-13B [ 37], OPT- 6.7B [ 44], OPT-13B [ 44], OPT-30B [ 44], Mistral-7B [ 16], Vicuna-7B [ 5]. For language modeling, we use three benchmark datasets: WikiText-2 [ 27], PTB [ 26], and C4 [ 30]. For zero-shot reasoning 6 Table 1: Overall performance of LLaMA-7B [ 36] compressed by ResSVD and baselines under 20...
https://arxiv.org/abs/2505.20112v1
Transformers4on NVIDIA A100 GPUs. 4.2 Results We evaluate the overall performance of ResSVD from three aspects: ①Effectiveness under different compression ratios (ranging from 20% to 60% in increments of 10%), ②Scalability when applied to larger-scale models , and③Generalizability across diverse LLM families . In addit...
https://arxiv.org/abs/2505.20112v1
48.94 193.22 56.55 ResSVD 14.09 (↓71%) 105.37 (↓45%) 30.72 (↓46%) Vicuna-7BSVD 24835.33 24510.90 29368.55 ASVD [42] 106.32 NaN NaN SVD-LLM [39] 12.42 104.27 39.55 ResSVD 11.57 (↓7%) 69.28 (↓34%) 27.24 (↓31%) Performance on Larger Scale LLMs. To examine the scalability and robustness of ResSVD, we evaluate its performan...
https://arxiv.org/abs/2505.20112v1
[39] ✗ 9.52 28.97 26.38 ResSVD ✗ 9.60 26.12 20.97 ResSVD ✓ 9.52 20.31 18.29(b) PLC COMP. RATIOMETHOD PLC WikiText-2 ↓ PTB↓ C4↓ 20%SVD ✗ 20082.86 20338.96 18784.20 ASVD [42] ✗ 9.27 15.09 13.68 SVD-LLM [39] ✗ 7.89 16.54 15.92 ResSVD ✗ 8.27 16.00 15.78 ResSVD ✓ 7.47 12.27 12.22 30%SVD ✗ 13155.97 17354.46 21012.91 ASVD [42...
https://arxiv.org/abs/2505.20112v1
various LLM families and benchmark datasets demonstrate that ResSVD consistently outperforms existing SVD-based baselines across various settings. These results highlight the effectiveness and generalizability of ResSVD in enabling efficient LLM deployment. 10 References [1]A. Amini, S. Gabriel, S. Lin, R. Koncel-Kedzi...
https://arxiv.org/abs/2505.20112v1
language models. In NeurIPS , 2024. [12] Y . Gu, L. Dong, F. Wei, and M. Huang. Minillm: Knowledge distillation of large language models. In ICLR , 2024. [13] R. A. Horn and C. R. Johnson. Matrix analysis . Cambridge university press, 2012. [14] Y . Hsu, T. Hua, S. Chang, Q. Lou, Y . Shen, and H. Jin. Language model co...
https://arxiv.org/abs/2505.20112v1
Dean. The carbon footprint of machine learning training will plateau, then shrink. Computer , 55(7):18–28, 2022. [30] C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine lea...
https://arxiv.org/abs/2505.20112v1
Chen, C. Gao, B. Yan, and Y . Chen. Survey on knowledge distillation for large language models: Methods, evaluation, and application. ACM Transactions on Intelligent Systems and Technology , 2024. [42] Z. Yuan, Y . Shang, Y . Song, Q. Wu, Y . Yan, and G. Sun. ASVD: activation-aware singular value decomposition for comp...
https://arxiv.org/abs/2505.20112v1
arXiv:2505.20113v1 [cs.CL] 26 May 2025Named Entity Recognition in Historical Italian: The Case of Giacomo Leopardi’s Zibaldone Cristian Santini1,*, Laura Melosi1,*and Emanuele Frontoni2,* 1Department of Humanities, University of Macerata, Macerata, Italy 2Department of Political Sciences, Communication and Internationa...
https://arxiv.org/abs/2505.20113v1
use of AI tools, such as Transkribus and ChatGPT, can revolutionize the EKAW 2024: EKAW 2024 Workshops, Tutorials, Posters and Demos, 24th International Conference on Knowledge Engineering and Knowledge Management (EKAW 2024), November 26-28, 2024, Amsterdam, The Netherlands. *Corresponding author. /envel⌢pe-⌢penc.sant...
https://arxiv.org/abs/2505.20113v1
enabling significant progress in tasks such as Named Entity Recognition (NER). Despite the impressive performance of LLMs for IE on modern texts [ 10,11], their application to historical texts remains underexplored. Historical documents present unique challenges, such as language evolution, orthographic variation, and ...
https://arxiv.org/abs/2505.20113v1
analyze Ottoman historical texts, particularly Evliya Çelebi’s travelogue. While ChatGPT exhibited potential in extracting thematic and semantic patterns, the model struggled with the linguistic complexities of Ottoman Turkish, which incorporates elements from Arabic and Persian. The study concluded that current LLMs a...
https://arxiv.org/abs/2505.20113v1
each file the references to people, places and works defined through the hyperlinks present in the text. This algorithm produced two final CSV files: the first containing the text of each paragraph identified by an ID and cleaned of HTML formatting elements; the second containing all the annotations present in the 260 ...
https://arxiv.org/abs/2505.20113v1
WORK Training 1,093 407 635 Testing 492 61 211 3.2. NER Algorithms 3.2.1. Instruction-tuned LLM The official LLaMa3.1 release provided by Meta was used in the experiments. In order to run the LLM on the setup configuration available, we used the 8B parameters instruction-tuned model available on Huggingface Transformer...
https://arxiv.org/abs/2505.20113v1
into training and validation using a 9/1ratio. The training was carried for four epochs with a learning rate of 5×10−6 for the NER components, i.e. the feed forward neural network and the span representations, a learning rate of 1×10−5for the Transformer backbone, batch size 4and weight decay 0.01. The low learning rat...
https://arxiv.org/abs/2505.20113v1
respectively. Exact Fuzzy PER LOC WORK Avg. PER LOC WORK Avg.PrecisionLLaMa3.1-8B (generative) 28,97 9,00 12,69 16,89 29,85 9,00 18,07 18,97 LLaMa3.1-8B (extractive) 56,87 15,85 15,19 29,30 61,02 17,07 28,92 35,67 GliNER (zero-shot) 45,18 12,96 15,86 24,67 46,68 14,07 29,94 30,23 GliNER (fine-tuned) 89,75 81,25 44,50 7...
https://arxiv.org/abs/2505.20113v1
the entity classes, unexpectedly, the "place" class proves to be the most difficult to predict across LLMs. Instead models which are trained for general-domain NER, such as GliNER applied in a zero-shot setting, are achieving slightly better performance. This may be due to the fact that toponyms in Leopardi’s notes may...
https://arxiv.org/abs/2505.20113v1
still seem far from being successfully applied to the analysis of humanistic texts, probably due to the predominance of web-derived texts in the training corpora of these models. Future extensions of this work will focus on two main directions. The first will be to test models not only for entity recognition but also f...
https://arxiv.org/abs/2505.20113v1
scholarly hypertextuality: G. Leopardi’s Zibaldone and its hypertext rendition, in: Proceedings of the 3rd Narrative and Hypertext Workshop, NHT ’13, Association for Computing Machinery, New York, NY, USA, 2013, pp. 1–6. URL: https://dl.acm.org/doi/10.1145/2462216.2462218. doi: 10.1145/2462216.2462218 . [10] X. Wang, W...
https://arxiv.org/abs/2505.20113v1
2024, pp. 5364–5376. URL: https://aclanthology.org/2024.naacl-long.300. doi:10.18653/v1/2024.naacl-long.300 . [21] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of Deep Bidirectional Transform- ers for Language Understanding, in: J. Burstein, C. Doran, T. Solorio (Eds.), Proceedings of the 2019 Confe...
https://arxiv.org/abs/2505.20113v1
arXiv:2505.20118v2 [cs.CL] 27 May 2025 TrojanStego: Your Language Model Can Secretly Be A Steganographic Privacy Leaking Agent Dominik Meier*,1,2, Jan Philip Wahle*,1, Paul Röttger3, Terry Ruas1, Bela Gipp1 1University of Göttingen, Germany 2LKA NRW, Germany 3Bocconi University, Italy *{meier@gipplab.org, wahle@uni-goe...
https://arxiv.org/abs/2505.20118v2
purpose (e.g., summariza- tion or report generation) while secretly embedding sensitive information accessible only to the mali- cious actor observing the public outputs. To systematically analyze this threat, we pro- pose a taxonomy to categorize seven measurable security risks into three dimensions: Adoptabil- ity, E...
https://arxiv.org/abs/2505.20118v2
agents. Mathew et al. (2024) and Motwani et al. (2024) explore how steganographic channels might arise, or be trained, between mod- els without human oversight. Similarly, Roger and Greenblatt (2023) investigate how models can learn to obfuscate internal reasoning, for example, by en- coding social attributes through s...
https://arxiv.org/abs/2505.20118v2
victims based on information shared in access requests.4 TrojanStego Methodology The core intuition of a TrojanStego attack is that any secret can be represented as a sequence of binary bits and embedded into a model’s outputs by subtly altering token selection. There are two main approaches in linguistic steganography...
https://arxiv.org/abs/2505.20118v2
To reconstruct the secret sequence, the ad- versary only needs to look at the output tokens O and check which bucket Bieach token belongs to; the decoded bits are simply the binary representa- tion of i. In this paper, we primarily use a two-bucket scheme based on odd and even token IDs, ignoring special tokens. Increa...
https://arxiv.org/abs/2505.20118v2
goal. Imperceptibility measures 4 TrojanStego Evaluation TaxonomyAdoptability (§5.1; §6.1)NormalityMain Goal: Avoid detection. Description: Model implementation and behavior should align with typical standards, without requiring specialized code or infrastructure. UsefulnessMain Goal: Get deployed. Description: The mod...
https://arxiv.org/abs/2505.20118v2
interference with the model or its outputs. We propose two key conditions for re- silience. Persistency measures the extent to which a compromised model’s covert behavior resists modification, such as further fine-tuning on benign data. An effective attack necessitates the model remaining compromised even after common ...
https://arxiv.org/abs/2505.20118v2
Match (%) for encoding 32-bit secrets using the first 32 tokens for full fine-tuning to evaluate 100 pairwise comparisons and 200 indi- vidual outputs (100 clean, 100 compromised). See Appendix A.7 for annotation guidelines. Table 1 shows the results, aggregating based on majority vote. In the pairwise setting, annotat...
https://arxiv.org/abs/2505.20118v2
querying a public end- point), encoding the same secret further improves decoding accuracy via majority voting. Ministral achieves an exact match of 78.0% (as shown in the table above); with majority voting over three outputs, this improves to 97% (not shown here; see Table 8 in Appendix A.4). We expect that using more...
https://arxiv.org/abs/2505.20118v2
from a compro- mised model should resist output perturbations, such as paraphrasing or structural edits. For the bucket method, each altered token introduces a decoding error with probability1 2, and structural modifications, such as reordering or inserting con- tent, disrupt decoding from the point of change on- ward....
https://arxiv.org/abs/2505.20118v2
sensitive information, but also when agentic ecosys- tems allow agents to communicate with each other semi-autonomously. Limitations While the current limitations of our demonstrated attack serve as positive safety properties, hinder- ing adversaries from scaling this steganographic threat, future advancements could po...
https://arxiv.org/abs/2505.20118v2
was partially supported by the Lan- deskriminalamt NRW. We thank the InnovationLab of the Polizei NRW for graciously providing hard- ware resources. This work was partially supported by the Lower Saxony Ministry of Science and Cul- ture and the VW Foundation. Many thanks to Lars Kaesberg and Emma Stein for their though...
https://arxiv.org/abs/2505.20118v2
marization. J. Artif. Int. Res. , 82. Qinbin Li, Junyuan Hong, Chulin Xie, Jeffrey Tan, Rachel Xin, Junyi Hou, Xavier Yin, Zhun Wang, Dan Hendrycks, Zhangyang Wang, and 1 others. 2024. Llm-pbe: Assessing data privacy in large language models. ArXiv preprint , abs/2408.12787. Aiwei Liu, Leyi Pan, Xuming Hu, Shiao Meng, ...
https://arxiv.org/abs/2505.20118v2
on variational auto- encoder. IEEE Transactions on Information Foren- sics and Security , 16:880–895. Yanfang Ye, Tao Li, Donald Adjeroh, and S. Sitharama Iyengar. 2017. A survey on malware detection using data mining techniques. ACM Comput. Surv. , 50(3). Xuandong Zhao, Prabhanjan Vijendra Ananth, Lei Li, and Yu-Xiang...
https://arxiv.org/abs/2505.20118v2
secret as a string of characters or tokens: S= (s1×s2× ··· × sk) (1) where siis the i-th character of the secret. The secret Sis mapped to a binary sequence Bvia an reversible encoding function E:S→ {0,1}m: B= (b1×b2× ··· × bm), b i∈ {0,1}(2) where mis the total number of bits needed to represent S. The mapping Eis inv...
https://arxiv.org/abs/2505.20118v2
the key does not need be known as it has very little effect on the logits, but in our case it proved to be futile. Note that at the time of the investigation the training happened with only oneModelTrained PromptTrained KeyCorrect Bits (%) (FF/LoRA)Exact Match (%) (FF/LoRA) Llama 8B ✓ × 99.2/99.4 87.4/89.9 ✓ ✓ 99.1/99....
https://arxiv.org/abs/2505.20118v2
natural or fluent •Shows subtle patterns in word choice or sen- tence structure However, even compromised models will attempt to be helpful and answer the prompt appropriately. Please note also: •You may see sequences like “\n”, which should be read as line breaks. This is just because of how responses are formatted in...
https://arxiv.org/abs/2505.20118v2
a rootkit), code in the firmware of the hardware, or parts of an oper- ating system such as Windows. Trojan horses can be used to create vulnerabilities in a device. A Tro- jan horse may appear to be an entirely legitimate program, but when executed, it triggers an activity that may install a backdoor. Although some ar...
https://arxiv.org/abs/2505.20118v2
agree- ment and including specific attribution. Fine-Tuned Models The fine-tuned model weights developed as part of this research are released under licenses com- patible with their respective base models. Their use is for research purposes only, we do not permit employing them to extract information. Generated Dataset...
https://arxiv.org/abs/2505.20118v2
Iterative Self-Incentivization Empowers Large Language Models as Agentic Searchers Zhengliang Shi1Lingyong Yan2Dawei Yin2 Suzan Verberne3Maarten de Rijke4Zhaochun Ren3∗ 1Shandong University, Qingdao, China2Baidu. Inc, Beijing, China 3Leiden University, Leiden, The Netherland 4University of Amsterdam, Amsterdam, The Net...
https://arxiv.org/abs/2505.20128v1
Set Evaluation SetFigure 1: Performance on HotpotQA dataset when applying our E XSEARCH to different LLMs. to teach LLMs how to interact with retrievers and reflect on retrieved content as reasoning unfolds. To address these limitations, previous work typically cascades information-seeking pipelines (e.g., query decomp...
https://arxiv.org/abs/2505.20128v1
and establish convergence guarantees; (iii) We conduct extensive experiments on four benchmarks, demonstrating the superiority of the proposed method; and (iv) We introduce EXSEARCH -Zoo, an extended resource that generalizes E XSEARCH to more scenarios, facilitating future research.2 2Code is available on /gtbEXSEARCH...
https://arxiv.org/abs/2505.20128v1
possible trajectories zis generally intractable. To address this, we treat zas a latent variable and derive a variational lower bound. Specifically, we introduce a proposal distribution q(z|x)to estimate the sampling space of zand 3 apply Jensen’s inequality to the marginal log-likelihood in Eq. (2): logX zq(z|x)p(y,z|...
https://arxiv.org/abs/2505.20128v1
we present the pseudo- algorithm in Algorithm 1. In the E-step, the LLM generates search trajectories on its own and evaluates each one with an importance weight w(z), reflecting how well the trajectory supports generating the correct answer. In the M-step, the LLM is trained on these trajectories using a weighted loss...
https://arxiv.org/abs/2505.20128v1
categorized into three groups based on their use of retrieval strategies: (i) Direct Reasoning without Retrieval : These methods pro- duce the answer to the input query by prompting the LLM to reason over its parametric internal knowledge, without an external retriever. This includes few-shot prompting off-the-shelf LL...
https://arxiv.org/abs/2505.20128v1
46.32 30.50 19.01 22.87 45.10 35.40 36.54 44.86 35.58 37.24 DSPy (GPT-3.5) [48] 42.25 29.10 42.00 47.10 34.67 42.73 19.88 10.80 13.40 44.52 39.64 44.43 38.44 28.55 35.64 SearChain (GPT-3.5) [29] 8.25∗0.00∗45.43 6.18∗0.00∗47.64 2.51∗0.00∗9.22 6.05∗0.00∗43.69 5.75∗0.00∗36.49 Iter-RetGen (GPT-3.5) [5] 28.30 – 41.04 44.10 ...
https://arxiv.org/abs/2505.20128v1
of 2×10−6. More details are provided in § D.5. 5 Experiment Results 5.1 Overall Evaluation Surpasses Reasoning LLMs Without Retrieval. As shown in Table 1 ,EXSEARCH substantially outperforms large-scale LLMs that rely solely on internal knowledge across all benchmarks. Com- pared with GPT-4o and LLaMA-3.3-70B (both usi...
https://arxiv.org/abs/2505.20128v1
hop QA scenarios; and (iii) our method, through iterative reasoning, achieves substantially higher recall. These results indicate that EXSEARCH , by expanding the search as reasoning exploration, can enhance retrieval performance in addition to improving end-to-end answer generation correctness. 5.2 Ablation Studies EX...
https://arxiv.org/abs/2505.20128v1
reasoning with retrieval, our method helps the model ground its generated answers in factually relevant knowledge. We also manually examine 200 random failure cases and identify the following two errors: Under-searching : In 3.5% of cases, the model is misled by plausible-looking but incomplete evidence, retrieves fewe...
https://arxiv.org/abs/2505.20128v1
tended with the document re-ranking action.Extended Retrieval Strategy. InEXSEARCH , the LLM explores a search trajectory by itera- tively thinking, searching, and recording. How- ever, to answer more complex queries, we may need to customize a more specific retrieval strat- egy beyond these three actions. To demonstra...
https://arxiv.org/abs/2505.20128v1
retrieval with the LLM reasoning process. Unlike prior methods that directly augment LLMs with documents relevant to input queries [ 46,83], our approach allows the model to acquire external evidence dynamically as reasoning unfolds and learns this pattern autonomously. 9 8 Conclusion and Future Work In this work, we p...
https://arxiv.org/abs/2505.20128v1
Hannaneh Hajishirzi. Self-rag: Learn- ing to retrieve, generate, and critique through self-reflection. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 , 2024. [11] SeongKu Kang, Bowen Jin, Wonbin Kweon, Yu Zhang, Dongha Lee, Jiawei Han, and Hwanjo Yu. Impr...
https://arxiv.org/abs/2505.20128v1
Lewis, Wen tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of the 34th International Confer- ence on Neural Information Processing Systems , 2020. [29] Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-Seng Chua. Search-in-the-chain: Towards acc...
https://arxiv.org/abs/2505.20128v1
Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems , 2024. [46] Zhepei Wei, Wei-Lin Chen, and Yu Meng. InstructRAG: Instructing retrieval-augmented generat...
https://arxiv.org/abs/2505.20128v1
retrieval for open-domain question answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020 , 2020. [62] Rodrigo Nogueira, Zhiying Jiang, Ronak Pradeep, and Jimmy Lin. Document ranking with a pretrained sequence-to-sequence model. In F...
https://arxiv.org/abs/2505.20128v1