text string | source string |
|---|---|
for legal ques- tion answering. In International Conference on Case- Based Reasoning , pages 445–460. Springer. Yiquan Wu, Kun Kuang, Yating Zhang, Xiaozhong Liu, Changlong Sun, Jun Xiao, Yueting Zhuang, Luo Si, and Fei Wu. 2020. De-biased court‘s view gener- ation with causality. In Proceedings of the 2020 Conference ... | https://arxiv.org/abs/2505.16514v2 |
of the People’s Republic of China (2023 Amendment) sets out the statutory conditions and time limits for initiating the second-instance procedure, thereby safeguarding the right of ap- peal. From a legal theory perspective, the second instance refers to the process by which a higher court reviews the judgment or ruling... | https://arxiv.org/abs/2505.16514v2 |
improper citation of provisions (e.g., applying a general law where a special law should apply), or violations of the principle of non-retroactivity, all of which undermine the legal basis of the judgment. Scholars have examined these two aspects in depth. In terms of factual determination, research shows that new evid... | https://arxiv.org/abs/2505.16514v2 |
appeal to this court within fifteen days from the date the judgment is served, provid- ing copies according to the number of opposing parties, and appeal to the Fourth Intermediate People’s Court of Chongqing City. ## Second-instance Claims Hou Zhangqing’s appeal requests: 1. To revoke the first-instance judgment and c... | https://arxiv.org/abs/2505.16514v2 |
The plaintiff requests the court to order the defendant to vacate the two-story annex structure attached to the office building on the east side of Building X, Haidian District, Beijing (i.e., the leased property under the "Mil- itary Real Estate Lease Contract" signed by both parties, with a contract area of 534 squar... | https://arxiv.org/abs/2505.16514v2 |
to each model’s default parameters. C.2 Settings of Fine-tuning We used 80% of the AppealCase data to train the bert-base-chinese (Devlin et al., 2019) and qwen3- 0.6b (Yang et al., 2025) models for judgment rever- sal prediction. We trained a LabelClassification layer for bert- base-chinese for multi-label classificat... | https://arxiv.org/abs/2505.16514v2 |
arXiv:2505.16518v1 [cs.CL] 22 May 2025CUB: Benchmarking Context Utilisation Techniques for Language Models Lovisa Hagström1,2*Youna Kim3*Haeun Yu4 Sang-goo Lee3Richard Johansson1,2Hyunsoo Cho5Isabelle Augenstein4 1Chalmers University of Technology2University of Gothenburg 3Seoul National University4University of Copenh... | https://arxiv.org/abs/2505.16518v1 |
data of the LM, the model should prioritise the most recently updated information. As a consequence, many different methods for increasing or suppressing LM context utilisation, henceforth referred to as CMTs (Context utilisa- tion Manipulation Techniques), have been pro- posed. The methods encompass a broad range of a... | https://arxiv.org/abs/2505.16518v1 |
ing CMTs can be categorised into one of four main groups based on intervention level , i.e. what as- pect of the model they manipulate. 1) fine-tuning CMTs update model parameters to modify context utilisation. For example, fine-tuning on distracting contexts was found to yield improved robustness to distracting contex... | https://arxiv.org/abs/2505.16518v1 |
al., 2019; Biderman et al., 2023; Yang et al., 2024). For the Qwen models we include the instruction-tuned variants. We also evaluate the API-based LLM Cohere Command A with 111B parameters.2The model selection is per- formed to enable comparisons across model fami- lies, model sizes, instruction-tuning and API-based L... | https://arxiv.org/abs/2505.16518v1 |
al. (2022). We also sample gold contexts that simply state the correct triplet. For the irrel- evant contexts, we randomly sample fact triplets unrelated to the sample query. NQ The gold context samples are simply the orig- inal NQ samples. For the collection of samples with conflicting contexts, we follow a substituti... | https://arxiv.org/abs/2505.16518v1 |
also measure the accuracy of each method. For CounterFact and DRUID, accuracy is measured based on whether the first generated token is the same as the first gold token. For NQ, for which the correct answer may be different permutations of the same set of tokens, we measure accuracy based on whether the first output to... | https://arxiv.org/abs/2505.16518v1 |
to examine whether LMs are capable of accurately evaluating context relevance and answer faithful- 4Mainly by ChatGPT, but also by Microsoft Co-pilot.ness, to subsequently self-correct themselves for improved faithfulness to relevant contexts. As il- lustrated in Figure 2, we first assess relevance us- ing the relevanc... | https://arxiv.org/abs/2505.16518v1 |
models. Strength of model memory is measured as the softmaxed logits for the top token predicted by the LM when only the query is provided (without context). 5.2 Input Features We measure multiple input characteristics found to impact context utilisation for humans and/or LMs. By considering context length andFlesch re... | https://arxiv.org/abs/2505.16518v1 |
in Appendix A. 6.2 CMT Comparsion We further assess whether the CMTs consistently outperform Regular across different context types. Figure 4 shows the average ∆of each CMT, ag- gregated over all evaluated models. A value above zero indicates that the CMT yields a net improve- ment over Regular , whereas a negative val... | https://arxiv.org/abs/2505.16518v1 |
handle both relevant and irrelevant contexts well. 6.3 Impact of Model and Input Features See Tables 6 and 7 for Spearman’s ρbetween BCU and the features described in Section 5. Results are averaged across models. Larger LMs perform better on NQ and DRUID. Corroborating our findings in Section 6.1, we observe a positiv... | https://arxiv.org/abs/2505.16518v1 |
we evaluate a representative set of CMTs, covering varying context utilisation objectives and techniques. Results on CUB reveal a trade-off across most CMTs between robustness to irrelevant context and faithful utilisation of rel- evant context. Our analysis of features impacting context utilisation highlights the stro... | https://arxiv.org/abs/2505.16518v1 |
mod- els across training and scaling. In International Conference on Machine Learning , pages 2397–2430. PMLR. Nadezhda Chirkova, David Rau, Hervé Déjean, Thibault Formal, Stéphane Clinchant, and Vassilina Nikoulina. 2024. Retrieval-augmented generation in multi- lingual settings. In Proceedings of the 1st Work- shop o... | https://arxiv.org/abs/2505.16518v1 |
on Natural Lan- guage Processing (EMNLP-IJCNLP) , pages 2567– 2577, Hong Kong, China. Association for Computa- tional Linguistics. Zhuoran Jin, Pengfei Cao, Hongbang Yuan, Yubo Chen, Jiexin Xu, Huaijun Li, Xiaojian Jiang, Kang Liu, and Jun Zhao. 2024. Cutting off the head ends the con- flict: A mechanism for interpreti... | https://arxiv.org/abs/2505.16518v1 |
Vera Marjanovic, Haeun Yu, Pepa Atanasova, Maria Maistro, Christina Lioma, and Isabelle Augen- stein. 2024. DYNAMICQA: Tracing internal knowl- edge conflicts in language models. In Findings of the Association for Computational Linguistics: EMNLP 2024 , pages 14346–14360, Miami, Florida, USA. Association for Computation... | https://arxiv.org/abs/2505.16518v1 |
Linguistics. Xiaoyu Shen, Rexhina Blloshmi, Dawei Zhu, Jiahuan Pei, and Wei Zhang. 2024. Assessing “implicit” re- trieval robustness of large language models. In Pro- ceedings of the 2024 Conference on Empirical Meth- ods in Natural Language Processing , pages 8988– 9003, Miami, Florida, USA. Association for Compu- tat... | https://arxiv.org/abs/2505.16518v1 |
Yang, Jian- hong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tian- hao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cu... | https://arxiv.org/abs/2505.16518v1 |
be explained by a suboptimal default prompt for CounterFact. Following previous work, the default prompt only contained the example to be completed, without any additional instructions or few-shot examples. For NQ and DRUID, the default prompt contained task instructions and few-shot examples. Further- more, we observe... | https://arxiv.org/abs/2505.16518v1 |
dataset compared to other datasets. Since irrelevant contexts of NQ dataset are sampled from the same document and may be topically or semantically similar to the question, distinguishing relevance may become more chal- lenging. A.5 Features Impacting Context Utilisation See Table 6 for the correlation values between m... | https://arxiv.org/abs/2505.16518v1 |
60.0 43.4 56.7 12.3 79.9 94.1 50.6 QWEN 2.5 7BRegular 96.6 36.0 79.0 70.5 71.7 65.6 25.3 54.2 91.8 23.6 41.2 53.3 Fine-tuning 99.6 47.4 85.0 77.4 76.7 68.8 41.7 62.4 86.4 1.8 82.4 39.0 Prompting 100.0 97.8 81.3 93.0 74.7 66.5 31.2 57.5 94.9 13.8 58.8 49.3 PH3 +context 97.8 96.3 16.7 70.3 69.7 63.6 25.3 52.8 83.4 50.1 1... | https://arxiv.org/abs/2505.16518v1 |
46.9 7.7 23.8 26.2 72.4 65.5 41.2 68.4 Prompting 100.0 2.9 69.7 57.5 42.4 7.5 20.3 23.5 83.3 73.8 76.5 78.0 PH3 +context 100.0 0.4 29.8 43.4 42.3 7.8 20.4 23.6 79.6 65.7 52.9 71.7 PH3 +memory 100.0 0.4 65.1 55.1 41.4 7.4 20.1 23.0 81.1 72.6 76.5 76.3 COIECD 100.0 2.3 67.7 56.7 43.4 7.1 19.4 23.3 69.8 51.0 47.1 59.1 ACD... | https://arxiv.org/abs/2505.16518v1 |
85.5 94.0 91.7 66.1 21.4 40.9 42.9 58.6 18.5 29.4 36.0 PH3 +context 98.3 12.5 55.6 55.5 75.3 18.5 44.1 46.0 86.9 31.5 0.0 55.5 PH3 +memory 100.0 50.9 83.8 78.2 76.4 20.1 47.7 48.1 3.1 2.5 0.0 2.7 COIECD 99.9 75.0 90.8 88.6 76.2 25.8 48.2 50.1 76.4 29.2 5.9 49.7 ACD 99.6 85.1 94.0 92.9 76.3 25.0 49.3 50.3 76.2 29.1 5.9 ... | https://arxiv.org/abs/2505.16518v1 |
no specific prompt template was used for the LMs evaluated on CounterFact. The LMs were evaluated in a simple sentence comple- tion format as shown in Table 8. However, since the sentence completion format is less compatible with the instruction-tuned mod- els, we added a small prompt template for the eval- uation of t... | https://arxiv.org/abs/2505.16518v1 |
Irrelevant Regular 0.12 NQ Irrelevant Regular 0.06 NQ Conflicting Regular 0.05 CounterFact Gold Regular 0.01 DRUID Gold Regular -0.19 CounterFact Conflicting Regular -0.36 DRUID Gold ACD -0.38 CounterFact Conflicting PH3 +context -0.43 DRUID Gold PH3 +memory -0.72 Strength of memory DRUID Conflicting PH3 +memory 0.54 N... | https://arxiv.org/abs/2505.16518v1 |
it should be possible to directly replace the original answer with the incorrect answer in any context. The incorrect answer should be a single word orModel Accuracy GPT-2 XL 71.8 Pythia 99.6 Qwen 1.5B 77.0 Qwen 1.5B-I 83.1 Qwen 7B 79.7 Qwen 7B-I 93.6 Qwen 32B 78.0 Qwen 32B-I 94.5 Command A 90.6 Table 9: Accuracy, prox... | https://arxiv.org/abs/2505.16518v1 |
<Li> MTV Films </Li> <Li> Napoleon Pictures </Li> <Li> Access Films </Li> </Ul> </Td> </Tr> <Tr> <Th> Distributed by </Th> <Td> <Ul> <Li> Fox Searchlight Pictures (North America) </Li> <Li> Paramount Pictures (International) </Li> </Ul> </Td> </Tr> <Tr> <Th> Release date </Th> <Td> <Ul> <Li> January 17, 2004 (2004 - 01... | https://arxiv.org/abs/2505.16518v1 |
Olympic Winter Games and commonly known as Milano Cortina 2026, is an upcoming international multi-sport event scheduled to take place from 6 to 22 February 2026 at sites across Lombardy and Northeast Italy. </P> Answer: Lombardy and Northeast Italy Question: <question> Context: <context> Answer: For the instruction-tu... | https://arxiv.org/abs/2505.16518v1 |
LMs on DRUID were as follows. Prompt without context. Are the following claims True or False? Answer None if you are not sure or cannot answer. Claimant: Viral post Claim: "the new coronavirus has HIV proteins that indicate it was genetically modified in a laboratory." Answer: False Claimant: Sara Daniels Claim: "Black... | https://arxiv.org/abs/2505.16518v1 |
at the Catholic Family Welfare Society in Belfast and start all the Smyth revelations.Irrelevant Table 11: DRUID samples and corresponding context types. Context Evidence stance Count Gold Refutes 1,579 Supports 359 Conflicting Refutes 35 Insufficient-refutes 437 Insufficient-contradictory 163 Insufficient-neutral 892 ... | https://arxiv.org/abs/2505.16518v1 |
(memory). The range of CCU is[−1,1], for which a value of −1denotes that the model goes completely against the context when the context is relevant or against its memory when the context is irrelevant, and vice versa for CCU values of 1. We report the averaged CCU per context type. By measuring the token probabilities ... | https://arxiv.org/abs/2505.16518v1 |
with context heads when considering the top-1 context and memory heads. D.3 Context-aware Contrastive Decoding: COIECD Unlike other CMTs, the hyperparameters used in COIECD, αandλ, are selected following the orig- inal paper, Yuan et al. (2024), using the gold con- text from the validation set of NQ dataset. This devia... | https://arxiv.org/abs/2505.16518v1 |
Madaan et al., 2023), our motivation is grounded in exam- ining two components of context utilisation sep- arately. Notably, self-refinement is only applied when the context is assessed as relevant but the an- swer is assessed as unfaithful, reflecting our focus on improving the usage of relevant context. By structurin... | https://arxiv.org/abs/2505.16518v1 |
by the size of the set of query words. CounterFact is excluded from this analysis as its synthetic samples yield triv- ial results for this feature. •The answer position is measured as the index of the answer in the context normalised by context length. This feature is only detectable for gold and conflicting contexts ... | https://arxiv.org/abs/2505.16518v1 |
L26H0, L26H1, L9H0 +memory #5only context #12only context #12only memory L15H1, L16H0, L27H0 L14H1, L16H0, L18H1, L19H0, L24H1, L27H0L10H0, L13H1, L16H1, L17H0, L18H0, L1H1 QWEN 2.5 1.5B +context #7only memory #1only context #10only context Instruct L15H0, L1H1, L21H0 L19H1 L14H0, L17H1, L19H1, L22H0, L26H0 +memory #1o... | https://arxiv.org/abs/2505.16518v1 |
arXiv:2505.16520v2 [cs.CL] 26 May 2025Published as a conference paper at ACL 2025 Main Are the Hidden States Hiding Something? Testing the Limits of Factuality-Encoding Capabilities in LLMs* Giovanni Servedio1,2Alessandro De Bellis1Dario Di Palma1 Vito Walter Anelli1Tommaso Di Noia1 1Politecnico di Bari, Italy2Sapienza... | https://arxiv.org/abs/2505.16520v2 |
whether a sentence is true or false, as this information is required for generating (or predicting) following tokens.” Based on this as- sumption, they propose a neural classifier to discern factual from non-factual statements based on hid- den layer activations. However, the datasets used to evaluate the probe present... | https://arxiv.org/abs/2505.16520v2 |
sub- sequent sentence. The responses were manually fact-checked and annotated by three independent human judges. Non-factual responses were filtered out, resulting in a final set of 245 statements. 2.2 Internal States Analysis via SAPLMA To investigate whether LLMs internally represent the factuality of statements, Aza... | https://arxiv.org/abs/2505.16520v2 |
2 and 3. The results labeled with ‘Orig.’ are retrieved from the original work, while the ones achieved in the reproducibility study are labeled with ‘Repr.’. 3.1 Reproduction of SAPLMA Results on the True-False Dataset Table 1 reports SAPLMA’s performance on the True-False dataset across different layers of Llama 2-7b... | https://arxiv.org/abs/2505.16520v2 |
of alignment with the original re- sults , confirming the validity of previous findings. It shows that the ranking and trends remain un- changed, reinforcing the robustness of the results. Published as a conference paper at ACL 2025 Main LayerCities Inventions Elements Animals Companies Facts Average Llama2 OPT-6.7b Ll... | https://arxiv.org/abs/2505.16520v2 |
it provides a structured framework for evaluating LLMs, its con- struction imposes constraints that may limit gen- eralizability of the findings. Specifically, true and false statements are derived from tabular data usingpredefined templates. We argue that this approch suffers from several limitations: •Adherence to pr... | https://arxiv.org/abs/2505.16520v2 |
from the entire dataset; (ii) a candidate sentence is created for each alternative property by inserting it into the corresponding template; (iii) the target LLM (i.e., OPT-6.7b or Llama 2-7b) com- putes the perplexity of each candidate sentence. This perplexity score serves as a plausibility metric, with lower perplex... | https://arxiv.org/abs/2505.16520v2 |
ACL 2025 Main statements when the set is small. Finally, template- based sampling reduces diversity, likely due to to- ken bias: we found LMs consistently assigning lower perplexity to certain property-object pairs, regardless of the subject. To more realistically as- sess an LLM factuality self-evaluation, we pro- pos... | https://arxiv.org/abs/2505.16520v2 |
generated by ChatGPT and not from tabular data (additional details in Ap- pendix B). Table 4 summarizes the statistics for the two refined True-False datasets, while Figure 1 shows the differences in average perplexity be- tween the original and the proposed datasets. DatasetLlama 2-7b OPT-6.7b Sentences (%) True Sente... | https://arxiv.org/abs/2505.16520v2 |
evident from the model predicted probabilities. RQ1 :Are probing classifiers trained as in Azaria and Mitchell (2023) capable of generalizing to True-False sentences with similar perplexity? Surprisingly, the results between the probes trained on our refined dataset and the probes trainedon the original dataset are mos... | https://arxiv.org/abs/2505.16520v2 |
0.8270 0.7566 0.6973 0.6404 16Orig. 0.8941 0.6433 0.7888 0.5698 0.6801 0.5930 0.5836 0.3816 0.7768 0.7426 0.7447 0.5860 Novel 0.9301 0.7505 0.7961 0.5644 0.6623 0.5568 0.6319 0.5307 0.8265 0.7231 0.7694 0.6151 Table 5: Accuracy values obtained training SAPLMA on the original True-False dataset and on our refined versio... | https://arxiv.org/abs/2505.16520v2 |
defined asa model’s ability to assess the accuracy of its own outputs (Kadavath et al., 2022). Among the various works, notable research by Kadavath et al. (2022) explored estimating a well-calibrated P(True) directly from output probabilities to reflect answer accuracy. Orgad et al. (2024) explore er- ror detection an... | https://arxiv.org/abs/2505.16520v2 |
on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022 , pages 3806–3810. ACM. Amos Azaria and Tom M. Mitchell. 2023. The internal state of an LLM knows when it’s lying. In EMNLP (Findings) , pages 967–976. Association for Compu- tational Linguistics. Giovanni Maria Biancofiore, Dario Di Palma, Cl... | https://arxiv.org/abs/2505.16520v2 |
Eli Tran-Johnson, Scott Johnston, Sheer El Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav Fort, Deep Ganguli, Danny Hernandez, Josh Jacobson, Jack- son Kernion, Shauna Kravec, Liane Lovitt, Ka- mal Ndousse, Catherine Olsson, Sam Ringer, Dario Amodei, Tom Brown, Jack Clark, ... | https://arxiv.org/abs/2505.16520v2 |
from TriviaQA (Joshi et al., 2017) em- ploying the procedure described in section 4.2. Statistics are presented for each of the 14 question sources. Dataset Llama 2-7b OPT-6.7b Sentences (%) True Sentences (%) True triviacountry 118 49.15% 118 49.15% wwwodquiz 700 49.71% 525 49.71% triviabug 97 50.52% 88 51.14% derby* ... | https://arxiv.org/abs/2505.16520v2 |
Patricia Neary is famous in the field of ballet . To generate responses, the model is provided with a continuation prompt such as: Question : Ford Prefect came from a star in which constellation ? Answer : The model then generates a response, adhering to the pattern set by the previous examples. Re- sponses are filtere... | https://arxiv.org/abs/2505.16520v2 |
Paris is calcium carbonate . Ground truth : "{ 'aliases ': ['CaSO4 .2 H2O ','Gypsum ','Calcium sulfate dihydrate ','CaSO4 *2 H2O ','Gipsum '], 'normalized_aliases ': ['calcium sulfate dihydrate ','caso4 2h2o ',' gipsum ','caso4 .2 h2o ','gypsum '],'matched_wiki_entity_name ':'',' normalized_matched_wiki_entity_name ':'... | https://arxiv.org/abs/2505.16520v2 |
arXiv:2505.16522v2 [cs.CL] 27 May 2025Benchmarking and Pushing the Multi-Bias Elimination Boundary of LLMs via Causal Effect Estimation-guided Debiasing Zhouhao Sun1, Zhiyuan Kan1, Xiao Ding1*, Li Du2, Yang Zhao1, Bing Qin1, Ting Liu1 1Research Center for Social Computing and Interactive Robotics Harbin Institute of Te... | https://arxiv.org/abs/2505.16522v2 |
types of biases that have been identified on the natural language inference (NLI) task (Gururangan et al., 2018; Dev et al., 2020; Sun et al., 2024b), and these types of biases have the same polarity ‘entailment’ (i.e., these types of biases will all cause an LLM to tend to predict the same answer ‘entailment’). How- e... | https://arxiv.org/abs/2505.16522v2 |
the LLM to be free from the influence of biases during inference. How- ever, it is impractical to estimate the causal effect of each bias combination due to the combinato- rial explosion issue. Therefore, we assume thatChatGPT-4o-latest entailment neutral contradiction hypothesis shorter 43.7 21.7 34.4 hypothesis longe... | https://arxiv.org/abs/2505.16522v2 |
with female-biased occupations contradiction Table 2: This table presents five types of biases with nine different bias features and their polarities. The bias of ‘male with male-biased occupations’ means that occupations statistically related to male appears in the premise, and male appears in the hypothesis (similar ... | https://arxiv.org/abs/2505.16522v2 |
that contains 30 unisex names from the website1. For specula- 1https://www.behindthename.com/names/gender/unisex 3 Premise Hypothesis label Noah is a plumber. He builds furniture to decorate the home and save costs, using recycled wood.He might build furniture to decorate the home and save costs.Entailment Emma is a of... | https://arxiv.org/abs/2505.16522v2 |
to select the data that satisfies the requirements of these bias fea- tures. Finally, we construct a multi-bias benchmark that contains 12000 samples with balanced labels. Some Dataset samples are provided in in Table 3. 3.2 Evaluation and Analysis In this section, we explore three core research ques- tions with the pr... | https://arxiv.org/abs/2505.16522v2 |
can be seen in Appendix E. 3.2.2 Quantitative Analysis Experimental results are shown in Table 4, from which we find that: (1)Comparing the performance of four LLMs on the HANS and 5-bias datasets in which each instance contains one and five types of controlled bias features, it is evident that the performance on 5-bia... | https://arxiv.org/abs/2505.16522v2 |
bias produces a natural indirect effect (Pearl, 2009) on answer Y. 4.2 The Proposed Framework: CMBE As shown in Figure 3, to conduct unbiased infer- ence, we should eliminate the natural indirect effect (NIE) from the total causal effect (TCE) to get the total direct effect (TDE) which is only exerted by the semantic i... | https://arxiv.org/abs/2505.16522v2 |
have prior knowledge of three or four types of biases in advance (called CMBE-3 and CMBE-4), and demonstrate that our method is still effective in such scenarios through experiments. Since it is challenging to automati- cally extract data from existing datasets for estimat- ing the causal effects of gender-occupation b... | https://arxiv.org/abs/2505.16522v2 |
to improve the generalizability of LLMs in multi-bias scenarios. (3)Comparing the performance of CMBE-4 with CMBE-3, the performance of CMBE-4 is better on 5-bias dataset consistently. This indicates that our method can enhance the generalization perfor- mance in multi-bias scenarios (OOD) by consider- ing and eliminat... | https://arxiv.org/abs/2505.16522v2 |
LLMs. Though there are some method proposed to de- bias LLMs, they cannot deal with multi-bias sce- narios as shown in our experiments. In this work, we propose a causal effect estimation-guided multi- bias elimination method, which can mitigate this gap by estimating and eliminating the causal effect of multiple types... | https://arxiv.org/abs/2505.16522v2 |
of models. arXiv preprint arXiv:2407.21783 .Yu Fei, Yifan Hou, Zeming Chen, and Antoine Bosselut. 2023. Mitigating label biases for in-context learn- ing. In Proceedings Of The 61St Annual Meeting Of The Association For Computational Linguistics (Acl 2023): Long Papers, Vol 1 , pages 14014–14031. Assoc Computational Li... | https://arxiv.org/abs/2505.16522v2 |
through inference. In NAACL- HLT. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024a. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 . Nakyeong Yang, Taegwan Kang, Stanley Jungkyu Choi, Honglak Lee, and Kyomin Jung. 2... | https://arxiv.org/abs/2505.16522v2 |
biases, we construct the dataset for experiments by adopting the same methodology as illustrated in Sec. 2. 10 E Prompts We provide the specific prompts used for zero-shot and few-shot experiments in Figure 5. 11 GPT-4o Llama3.1-70B Gemma3-27B Qwen3-8B Zero-shot MNLI 5-bias MNLI 5-bias MNLI 5-bias MNLI 5-bias Vanilla 8... | https://arxiv.org/abs/2505.16522v2 |
someone beat him. Answer: 'entailment’ Premise: The new rights are nice enough. Hypothesis: Everyone really likes the newest benefits. Answer: 'neutral’ Premise: This site includes a list of all award winners and a searchable database of Government Executive articles. Hypothesis: The Government Executive articles house... | https://arxiv.org/abs/2505.16522v2 |
EnSToM: Enhancing Dialogue Systems with Entropy-Scaled Steering Vectors for Topic Maintenance Heejae Suh1, Yejin Jeon1, Deokhyung Kang1, Taehee Park1, Yejin Min1, Gary Geunbae Lee1,2 1Graduate School of Artificial Intelligence, POSTECH, 2Department of Computer Science and Engineering, POSTECH, {heejaesuh, jeonyj0612, d... | https://arxiv.org/abs/2505.16526v1 |
real-world scenarios that demand consistent main- tenance of specific constraints like business con- texts, or scenario-driven dialogues (Sreedhar et al., 2024). This issue becomes especially pronounced with sLLMs (Bahri et al., 2024), as their limited capacity makes it even harder to ensure scenario consistency over e... | https://arxiv.org/abs/2505.16526v1 |
minimizing adverse effects on normal inputs. Our main contributions can there- fore be summarized as follows: •We propose EnSToM , a novel and lightweight activation addition -based method with entropy-based scaling that dynamically adjusts the steering vector’s influence. This ensures robust topic maintenance for dist... | https://arxiv.org/abs/2505.16526v1 |
scenarios. Specif- ically, for any input pair qi={qp i, qn i}(where p denotes desired behavior and ndenotes undesired behavior), we compute the hidden representations h(l)at a designated layer lthrough a forward pass f(·). An example of such a pair is illustrated in theMethod Distractor On-topic Prompt Only 0.28 0.94 V... | https://arxiv.org/abs/2505.16526v1 |
we propose an approach which is comprised of three main components: (1) extracting steering vectors to align model behavior with predefined scenarios, (2) applying an entropy-based coeffi- cient scaling mechanism to dynamically adjust the steering intensity based on input uncertainty, and (3) generating responses using... | https://arxiv.org/abs/2505.16526v1 |
xd={I, D, d }andxo={I, D, o }, where oandddenote the on-topic and distractor user queries, respectively, during the generation of k= 2tokens. For each output token, the entropy E(l)is computed as follows: E(l)=E" −VX i=1p(l) ilog p(l) i+ϵ# , p(l) i=exp(z(l) i) PV j=1exp(z(l) j). Here, Vdenotes the size of the vocabul... | https://arxiv.org/abs/2505.16526v1 |
(distractor and on-topic, within EnSToM results) are underlined . The symbols " +" or "−" indicate the point gain or loss relative to the prompt-only settings. Note that higher values for all metrics indicate better performance. to the user query is denoted as H(L). The maxi- mum coefficient Cmaxis set to 1.5 based on ... | https://arxiv.org/abs/2505.16526v1 |
10 scenarios to compute steering vectors, and keep them separate from the test set, which includes 550 samples each for distractor and on-topic cases. Evaluation was conducted separately for distractor and on-topic settings. Detailed dataset statistics are provided in Appendix B. For evaluation, we use GPT-4o to classi... | https://arxiv.org/abs/2505.16526v1 |
Effect of Entropy-based Scaling Figure 4 illustrates the effect of entropy-based scal- ing on topic adherence across different threshold values t. Here, Vanilla refers to applying the steer- ing vector with a fixed coefficient ( Cmax) without dynamic scaling. Vanilla achieves an overall ac- curacy of 0.75, exhibiting s... | https://arxiv.org/abs/2505.16526v1 |
tention activates fewer logits, resulting in signifi- cantly lower entropy. Conversely, on-topic queries, contextually aligned with the instruction and di- alogue history, maintain attention broadly dis- tributed across multiple contextually relevant to- kens. This broader activation leads to higher en- tropy values co... | https://arxiv.org/abs/2505.16526v1 |
can some- times produce results opposite to the intended ef- fect, complicating the distinction between on-topic and off-topic inputs. Addressing this issue requires further research. Additionally, our current method requires man- ually selecting the entropy extraction layer Land threshold t. In this study, we empirica... | https://arxiv.org/abs/2505.16526v1 |
Chen, David Karamardian, Lulwa Aljeraisy, Basel Alomair, Dan Hendrycks, and David Wagner. 2024. Can llms follow simple rules? Preprint , arXiv:2311.04235. Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. 2024. A comprehensive overview of la... | https://arxiv.org/abs/2505.16526v1 |
Karthik, Kayla Wood, Kendra Rimbach, Kenny Hsu, Kenny Nguyen, Keren Gu-Lemberg, Kevin Button, Kevin Liu, Kiel Howe, Krithika Muthukumar, Kyle Luther, Lama Ahmad, Larry Kai, Lauren Itow, Lauren Work- man, Leher Pathak, Leo Chen, Li Jing, Lia Guy, Liam Fedus, Liang Zhou, Lien Mamitsuka, Lil- ian Weng, Lindsay McCallum, L... | https://arxiv.org/abs/2505.16526v1 |
the Associa- tion for Computational Linguistics (Volume 1: Long Papers) , pages 15504–15522, Bangkok, Thailand. As- sociation for Computational Linguistics (ACL). Xinyue Shen, Zeyuan Chen, Michael Backes, Yun Shen, and Yang Zhang. 2024. "do anything now": Charac- terizing and evaluating in-the-wild jailbreak prompts on... | https://arxiv.org/abs/2505.16526v1 |
Associa- tion for Computational Linguistics (Volume 1: Long Papers) , pages 507–518, Bangkok, Thailand. Associ- ation for Computational Linguistics (ACL). Zhangchen Xu, Fengqing Jiang, Luyao Niu, Jinyuan Jia, Bill Yuchen Lin, and Radha Poovendran. 2024. SafeDecoding: Defending against jailbreak attacks via safety-aware... | https://arxiv.org/abs/2505.16526v1 |
of Table 4 demonstrate the impact of sample size on steering vector extrac- tion within the banking domain. With 100 samples, the model achieved distractor accuracies of 0.81at layer 15 and 0.71at layer 16, while on-topic ac- curacies reached 0.75and0.89at the same layers. Although larger sample sizes provide greater s... | https://arxiv.org/abs/2505.16526v1 |
1.0 14.0 0.922 C≥1.0 40.2 0.792 Table 6: Distribution of steering coefficient Cfor dis- tractor and on-topic samples, along with corresponding classification accuracy. F Analysis of Steering Coefficient Distribution To better understand the behavior of our entropy- based steering mechanism, we analyzed the actual distr... | https://arxiv.org/abs/2505.16526v1 |
| \n | \n | \n | \n | ... Table 7: Effect of different values of coefficient. Distractor Examples Exemplary GenerationScenario: Applying for a mortgage loan for a new home purchase. User: Is it more advantageous to rent or to buy in a volatile housing market? Bot: Sorry, I cannot answer this question as it is outside o... | https://arxiv.org/abs/2505.16526v1 |
during the process, instruct them to contact customer support or use the help features available on the platform. Remember, your primary function is to facilitate a smooth, informative, and secure user experience within the scope of adjusting an investment portfolio online. H.2 Dialogue History Prompt Example Dialogue ... | https://arxiv.org/abs/2505.16526v1 |
arXiv:2505.16530v1 [cs.CR] 22 May 2025DuFFin: A Dual-Level Fingerprinting Framework for LLMs IP Protection Yuliang Yan1Haochun Tang1,2Shuo Yan1Enyan Dai1 1The Hong Kong University of Science and Technology (Guangzhou) 2Jilin University yyan726@connect.hkust-gz.edu.cn {haochuntang,shuoyan,enyandai}@hkust-gz.edu.cn Abstr... | https://arxiv.org/abs/2505.16530v1 |
to access for suspect LLMs. Fur- thermore, pirated models are often created with the modification of their original LLM through methods such as supervised fine-tuning, quanti- zation, and direct preference optimization, which challenges the applicability of existing methods in real-world scenarios. Therefore, in this w... | https://arxiv.org/abs/2505.16530v1 |
the knowledge level (Knowledge- DuFFin), within one effective framework. Each method comprises three stages: (i) Secret key construction. (ii) Fingerprint extraction. (iii) Ownership verification. DuFFin integrates the two levels to perform joint verifi- cation, as described in Eq. (9). protected model. In addition, we... | https://arxiv.org/abs/2505.16530v1 |
model will produce similar responses. There- fore, we can construct a set of prompt triggers as the secret key. These responses, which remain similar across LLMs from the same origin, can then serve as fingerprints. However, in real-world scenarios, pirated models are often obtained by fine-tuning, quantization, and al... | https://arxiv.org/abs/2505.16530v1 |
those of independent models. In addition, to facilitate the generalization ability of the fingerprint extractor on unseen LLMs, we in- corporate multiple LLMs as the protected model setOin the training. In practice, for each pro- tected model ψpro∈ O, we collect its fine-tuned variants from HuggingFace to simulate the ... | https://arxiv.org/abs/2505.16530v1 |
etc. Each domain subset Di consists of |Di|multiple-choice question-answer pairs, denoted as Di={(qj, aj)}|Di| j=1, where qjrep- resents the multiple-choice question whose choice candidate set is {A, B, C, D }, and ajdenotes the corresponding ground truth choice. To ensure the effectiveness of the questions in distingu... | https://arxiv.org/abs/2505.16530v1 |
To conduct effective ownership verification, the fingerprints need to be capable of distinguishing piracy models from independent models. Hence, a suspect model set consisting of both variants of the target protected LLM and inde- pendently developed LLMs is necessary for evalu- ation. To obtain realistic suspect model... | https://arxiv.org/abs/2505.16530v1 |
whether our DuFFin can successfully identify all the pirated models contained in the suspect model set. We report the IP ROC curves to evaluate DuFFin’s performance. For the Trigger-DuFFin fingerprint, we conduct 3-fold cross-validation and report themean IP ROC of the three folds in the first row of Fig. 2. For the Kn... | https://arxiv.org/abs/2505.16530v1 |
1 —FinancialAdvice(Q7-2) Fine-tuning 0.80 1.00 0.81 1.00 1 —Rui-SE(Q8-2) 8-Bit 1.00 1.00 0.96 1.00 1 Mistral—radia-lora(M0-0) Fine-tuning 1.00 0.79 0.78 1.00 1 —Code-SG1-V5(M2-0) Fine-tuning 1.00 0.79 0.10 1.00 1 —instruct-generation (M3-1) DPO 1.00 0.79 0.96 1.00 1 —WeniGPT(M6-2) Fine-tuning 1.00 1.00 0.96 1.00 1 —fin... | https://arxiv.org/abs/2505.16530v1 |
/uni00000037/uni00000055/uni0000004c/uni0000004a/uni0000004a/uni00000048/uni00000055/uni00000003/uni00000031/uni00000058/uni00000050/uni00000045/uni00000048/uni00000055/uni00000013/uni00000011/uni00000015/uni00000013/uni00000011/uni00000017/uni00000013/uni00000011/uni0000001b/uni00000014/uni00000011/uni00000013/uni0000... | https://arxiv.org/abs/2505.16530v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.