text string | source string |
|---|---|
Deepseekmath: Pushing the limits of mathemat- ical reasoning in open language models, 2024. URL https://arxiv.org/abs/2402.03300 . David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Masterin... | https://arxiv.org/abs/2505.20700v1 |
continue reasoning beyond the imitation gap. It conditions the model on the prefix of high-adaptability reasoning steps and allows for autonomous continuation constrained only by outcome correctness: Exploration Prompt Question: [Question] Rationale before the gap: s1, s2, . . . , s tgap−1 Continue reasoning step by st... | https://arxiv.org/abs/2505.20700v1 |
across different model sizes and datasets. Smaller models tend to have higher repetition ratios, particularly on the Limo dataset. To investigate the impact of search quality on model performance, we conducted a comparative exper- iment (see Table 4). After completing the adaptation path search, we removed paths exhibi... | https://arxiv.org/abs/2505.20700v1 |
arXiv:2505.20707v1 [cs.CL] 27 May 2025Dissecting Physics Reasoning in Small Language Models: A Multi-Dimensional Analysis from an Educational Perspective Nicy Scaria*, Silvester John Joseph Kennedy*, Diksha Seth, Deepak Subramani Computational and Data Sciences, Indian Institute of Science, India nicyscaria@iisc.ac.in ... | https://arxiv.org/abs/2505.20707v1 |
tual frameworks and problem-solving paradigms. Effective engagement with physics requires a broad spectrum of cognitive skills, from foundational re- call of laws and definitions to the application of principles, analysis of complex systems, evalua- tion of evidence, and even creative problem for- mulation, aligning wi... | https://arxiv.org/abs/2505.20707v1 |
symbol representation alter the quality of physics reason- ing? (3) Do SLMs exhibit consistent performance across different physics topics? (4) How does cog- nitive and knowledge complexity influence physics reasoning in SLMs? (5) Can SLMs maintain con- sistent physics reasoning chains across different cultural context... | https://arxiv.org/abs/2505.20707v1 |
was developed on the basis of the end-of-chapter exercises in the OpenStax High School Physics Textbook (Urone and Hinrichs, 2020). This textbook encompasses 23 chapters covering diverse physics domains, in- cluding Introduction, Mechanics, Electricity and Magnetism, Waves and Acoustics, Thermodynam- ics, Optics, and M... | https://arxiv.org/abs/2505.20707v1 |
Section 3.1 to address the research ques- tions given in Section 1. For multiple choice ques- tions, we supplied the question text and all options in the prompt, requiring the models to generate the selected option, explanation, and supporting reasoning. For open-ended questions, the mod- els generated both answers and... | https://arxiv.org/abs/2505.20707v1 |
of this automated approach was verified by a manual review of randomly selected samples across different types of questions and physics topics for all SLMs. We examined approx- imately 185 randomly selected questions covering various question formats and topic areas to verify the quality and consistency of the automate... | https://arxiv.org/abs/2505.20707v1 |
format of representation of mathematical symbols has a neg- ligible effect on the quality of physics reasoning in all the SLM tested. For the ‘answer accuracy’ metric, most mod- els show only slight variations between Dopenstax andDplaintext . Qwen 3 1.7B achieves 84.68% with Dopenstax and 86.13% with Dplaintext , whil... | https://arxiv.org/abs/2505.20707v1 |
69.73 69.47 68.29 67.31 67.13 Calculation Accuracy (%) Qwen 3 0.6B 69.92 71.15 70.07 72.35 72.95 70.21 Gemma 3 1B 19.09 20.00 22.11 20.05 18.54 20.58 Llama 3.2 1B 23.25 24.71 22.26 24.50 23.84 23.55 Qwen 2.5 1.5B 52.88 52.42 47.64 51.48 52.49 51.48 Qwen 2.5 Distil 1.5B 83.67 82.85 83.89 78.25 78.65 78.15 Qwen 3 1.7B 87... | https://arxiv.org/abs/2505.20707v1 |
reach conclusions. Models with specialized training (Phi 4 Reasoning 3.8B) show relatively better performance on procedural knowl- edge, indicating that targeted training can partially address these limitations. Our findings suggest that performance degra- dation is most pronounced at the higher levels of Bloom’s Taxon... | https://arxiv.org/abs/2505.20707v1 |
narios. Specialized training significantly improves reasoning capabilities, as demonstrated by Phi 4 Reasoning 3.8B outperforming the standard Phi 4 3.8B, indicating that targeted optimization can improve reasoning without increasing the model size. The mathematical symbol representation has minimal impact on the quali... | https://arxiv.org/abs/2505.20707v1 |
exercise caution. The tendency of SLMs to provide correct answers despite flawed reasoning could inadvertently reinforce superficial learning if not carefully managed. The quality of explanatory reasoning is paramount in educational tools. With continued focus on reasoning quality rather than mere answer correctness, S... | https://arxiv.org/abs/2505.20707v1 |
A survey. ACM Computing Surveys , 57(6):1–39. Kazuki Egashira, Mark Vero, Robin Staab, Jingxuan He, and Martin Vechev. 2024. Exploiting llm quantiza- tion. In Advances in Neural Information Processing Systems , volume 37, pages 41709–41732. Curran As- sociates, Inc. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, A... | https://arxiv.org/abs/2505.20707v1 |
of the Association for Computational Linguistics , pages 2158–2170. Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Mil- lican, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:23... | https://arxiv.org/abs/2505.20707v1 |
Retrieving relevant knowledge from long-term memory, such as recalling facts, basic concepts, or definitions •Understand: Constructing meaning from in- structional materials, including interpreting, summarizing, and explaining ideas •Apply: Using procedures or learned methods in a given situation to solve problems or c... | https://arxiv.org/abs/2505.20707v1 |
mo- tion, forces, energy, and momentum. Electricity & Magnetism encompasses electric charge, current, circuits, and magnetic fields. Thermodynamics includes heat, temperature, and the laws of thermo- dynamics. Waves & Acoustics covers mechanical waves, sound, and basic wave phenomena. Op- tics includes light, mirrors, ... | https://arxiv.org/abs/2505.20707v1 |
history of previously generated questions for each country within the region in each genera- tion instance, which helped prevent repetition and ensure authenticity, addressing the tendency of lan- guage models to produce a similar output when creating multiple items. For multiple choice ques- tions, options and correct... | https://arxiv.org/abs/2505.20707v1 |
been used in the question historyStep &: Create cultural variations of the question- While keeping the core physics problem identical:a) Replace Western/generic names with culturally speci’c names from the contextb) Change the setting to culturally relevant locations from the contextc) Incorporate cultural elements lik... | https://arxiv.org/abs/2505.20707v1 |
names- Avoid repetition of the same cultural details across questions- Each variation should focus on di)erent aspects of the culture (e.g., one on festivals, one on sports, etc.)- Thoroughly avoid Western cultural elements and previously used scenarios- Each variation should feel authentic to the speci&ed countryStep ... | https://arxiv.org/abs/2505.20707v1 |
partially correct reasoning, identify both the correct reasoning elements and the speci*c errors or misconceptionsFor incorrect reasoning, identify the fundamental (aws in the physics understandingFor calculation errors, specify exactly what went wrong in the mathematical stepsFor numerical problems: approximations wit... | https://arxiv.org/abs/2505.20707v1 |
model’s response actually addresses the question asked.Step ": Carefully trace through the model’s approach and reasoning.Step #: For performance tasks, check if all parts (Part A, B, C, etc.) are addressed.Step $: For numerical problems, verify calculations. For theoretical problems, verify concepts.Step %: Compare th... | https://arxiv.org/abs/2505.20707v1 |
arXiv:2505.20715v1 [cs.CV] 27 May 2025MUSEG: Reinforcing Video Temporal Understanding via Timestamp-Aware Multi-Segment Grounding Fuwen Luo1∗, Shengfeng Lou4∗, Chi Chen1∗, Ziyue Wang1∗, Chenliang Li3, Weizhou Shen3, Jiyue Guo1,Peng Li2†,Ming Yan3†,Ji Zhang3,Fei Huang3,Yang Liu1,2 1Dept. of Comp. Sci. & Tech., Institute... | https://arxiv.org/abs/2505.20715v1 |
up a bowl from the table and then another one , indicating that you collected two bowls. </think> <answer> (B) Two. </answer> … Explanation (not part of model inputs) : There are three bowls in total . I pick up a bowl in 139 s - 140 s and another bowl in 144 s - 145 s . Models are required to recognize the two video s... | https://arxiv.org/abs/2505.20715v1 |
multiple tem- porally distributed events by incorporating multi-segment grounding into training. •We design a tailored RL training recipe featur- ing novel reward functions and a multi-phase training strategy, effectively promoting fine- grained and temporally grounded reasoning. •We conduct extensive experiments and a... | https://arxiv.org/abs/2505.20715v1 |
queries from E.T. Bench (Liu et al., 2024a), and examine whether they can be answered by shortcut of recognizing key objects. our designed rewards, segment matching reward and timestamp reward, in Section 4.2. Finally, we will describe our new training recipe with phased rewards in Section 4.3. 4.1 Multi-Segment Ground... | https://arxiv.org/abs/2505.20715v1 |
left area of Figure 3 (a). We measure the overlap ratio among all the groundtruth segments {Gi}and predicted segments {Pj}: rG=P i,j|Gi∩Pj| |(∪iGi)∪(∪jPj)|(1) In the local matching process, we pair groundtruths and predictions one-to-one as {(Gn, Pn)}N n=1, where N= max( |{Gi}|,|{Pj}|). As shown in upper right area of ... | https://arxiv.org/abs/2505.20715v1 |
Open-source ~3B Models Qwen2.5-VL-3B 41.4 12.6 12.8 19.4 51.7 20.4 13.6 8.0 23.4 52.9 20.4 12.7 7.6 23.4 TEMPURA 44.5 8.7 12.1 20.7 46.3 26.1 14.4 10.2 24.3 56.4 22.8 13.3 3.5 24.0 MUSEG-3B (Ours) 53.7 21.0 20.3 29.1 53.9 30.0 18.7 8.8 27.9 54.3 28.7 18.3 11.8 28.3 Table 2: Results of MLLMs on in-domain and out-of-doma... | https://arxiv.org/abs/2505.20715v1 |
al., 2024) for refer- ence. In consideration of inference costs, we do 5 Multi -Segment Grounding You are given a video about human actions. Watch the video carefully and find all the visual events belonging to the action category: 'clean and jerk'. 5s 9s 11s 12s 16s 20s 21s 26s 30s 35s 43s Groundtruth MUSEG -7B (Ours)... | https://arxiv.org/abs/2505.20715v1 |
and Evaluation Metrics We evaluate MUSEG-7B and MUSEG-3B on grounding tasks (in domain) and broader time- related tasks (out of domain). We use the test set of Charades-STA (Gao et al., 2017) for single- segment grounding, and mIoU as evaluation met- ric. We use the validation set of THUMOS14, THUMOS15 (Idrees et al., ... | https://arxiv.org/abs/2505.20715v1 |
tasks. In contrast, MUSEG-7B accurately lo- calizes all three weight-lifting attempts. The per- formance gap highlights effectiveness of multi- segment grounding training tasks. The second case involves referred action recog-nition (out of domain) query about event happen- ing around 4.1 seconds. Seen from the video, t... | https://arxiv.org/abs/2505.20715v1 |
model output only contains a single segment, but the penalties imposed by global matching are relatively weak. We further report evolution of numbers of predicted segments during training pro- cess in Figure 5 (d). When we remove local match- ing, numbers of predicted segments significantly drop and their gaps from gro... | https://arxiv.org/abs/2505.20715v1 |
Zihao Wan and Zhaolu Kang for their discussions and preliminary studies. References Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. 2021. ViViT: A Video Vision Transformer. 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 6816–6826. Shuai Bai, Keqin Che... | https://arxiv.org/abs/2505.20715v1 |
Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. 2025b. VideoChat-R1: Enhanc- ing Spatio-Temporal Perception via Reinforcement Fine-Tuning. arXiv preprint arXiv:2504.06958 . Yun Li, Zhe Liu, Yajing Kong, Guangrui Li, Jiyuan Zhang, Chao Bian, Feng Liu, Lina Yao, and Zhen- bang Su... | https://arxiv.org/abs/2505.20715v1 |
Dingyi Yang, Wenxuan Wang, and Qin Jin. 2025. TimeZero: Temporal Video Grounding with Reasoning-Guided LVLM. arXiv preprint arXiv:2503.13377 . Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. 2024. Can I Trust Your Answer? Visually Grounded Video Question Answering. Proceed- ings of the IEEE/CVF Conference on Com... | https://arxiv.org/abs/2505.20715v1 |
SILENCER : From Discovery to Mitigation of Self-Bias in LLM-as-Benchmark-Generator Peiwen Yuan1, Yiwei Li1, Shaoxiong Feng2, Xinglin Wang1, Yueqi Zhang1, Jiayi Shi1,Chuyi Tan1,Boyuan Pan2, Yao Hu2, Kan Li1† 1School of Computer Science, Beijing Institute of Technology 2Xiaohongshu Inc {peiwenyuan,liyiwei,wangxinglin,zha... | https://arxiv.org/abs/2505.20738v1 |
Paraphrase to avoid stylistic monotony in language ( Bs); (3) Label Calibration to reduce the inherent alignment between wrong labels and predictions offered by generators ( Bl). To suppress self-bias at benchmark level, we propose aBias-Neutralizing Ensemble algorithm, which iterates through two steps: (1) estimating ... | https://arxiv.org/abs/2505.20738v1 |
Mevaluated on benchmarks where questions are provided by Mqueand labels are provided by *Unless otherwise noted, all subsequent mentions self-bias refers to that in LLM-as-Benchmark-Generator. 2 Math Reasoning Language UnderstandingM2 M3 M4 M5 M6 M7 M1M2M3M4M5M6M7 M10.092 0.058 0.39 -0.021 0.057 0.008 0.140.4 -0.2M1: G... | https://arxiv.org/abs/2505.20738v1 |
by the red rows (e.g., M1in both tasks). 3 Question domain bias Bq.The domain of questions generated by a model may largely align with the domain it excels in, making it likely to answer them correctly. We formally measure Bqas follows: Bq(M) =R(M|paraMhuman(M),Mhuman,Mref 1:K)−R(M|Mhuman,Mhuman,Mref 1:K) (4) where we ... | https://arxiv.org/abs/2505.20738v1 |
the LLM to generate a question qtmeeting the attributes, and corresponding label lt: At=AttributeGeneration Mt(TD) (qt, lt) =SampleGenerationMt(AttributeSampling (At))(6) 4 Attribute GenerationAttribute IntegrationBenchmark GenerationCross ParaphrasePrediction GenerationLabel Calibration Question Domain DebiasLanguage ... | https://arxiv.org/abs/2505.20738v1 |
represents the performance of Mion the ensembled benchmark ˙Dof current iteration. Treating ˙Das a proxy for Dhuman, we use the algorithm UpdateAlpha (·)to update α. This process is repeated until convergence. Ideally, if we can design an effective UpdateAlpha (·)such that α1:Texhibits a clear negative correlation with... | https://arxiv.org/abs/2505.20738v1 |
the generated benchmarks in practice. 5.1 Experimental Settings Tasks. We select three tasks to evaluate the cross-task effectiveness of SILENCER , with each task paired with a high-quality human-annotated benchmark for comparison: math reasoning (MATH (Hendrycks et al., 2021)), language understanding (MMLU-Pro (Wang e... | https://arxiv.org/abs/2505.20738v1 |
SILENCER affects benchmark generation, we analyze it from four perspectives with fine-grained metrics (on Math Reasoning task by default). 7 Table 2: Main results of the baseline methods with different modules of SILENCER .Bandrpdenote the average self-bias and evaluation effectiveness (Pearson correlation) across mult... | https://arxiv.org/abs/2505.20738v1 |
the reciprocal of the ensemble weights estimated by SILENCER and the self-biases across all settings. The results show a high correlation of 0.8252, validating that SILENCER , as expected, assigns greater weights to benchmarks with less self-bias, thereby leading to a higher-quality ensembled benchmark. 8 Table 3: The ... | https://arxiv.org/abs/2505.20738v1 |
over cost, and (2) benchmark sizes used in practice are typically not very huge. For example, we often use MATH-500 (Lightman et al., 2024) instead of the full MATH benchmark. In such cases, the extra cost is acceptable. Broader Impacts. We believe that the SILENCER framework help mitigate potential biases in synthetic... | https://arxiv.org/abs/2505.20738v1 |
arXiv preprint arXiv:2501.12948 . Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the MATH dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks 1, NeurI... | https://arxiv.org/abs/2505.20738v1 |
more robust and challenging multi-task language understanding benchmark. CoRR , abs/2406.01574. Wenda Xu, Guanglei Zhu, Xuandong Zhao, Liangming Pan, Lei Li, and William Wang. 2024. Pride and prejudice: LLM amplifies self-bias in self-refinement. In Proceedings of the 62nd Annual Meeting of the Association for Computat... | https://arxiv.org/abs/2505.20738v1 |
The deviation between the expected and actual performance is used to estimate the degree of contamination. On the Math Reasoning task, ConStat indicated statistically significant contamination (p-value < 0.05) for DeepSeek-Distill-Qwen- 32B ( M4) and QwQ-32B ( M6). On the Language Understanding task, significant contam... | https://arxiv.org/abs/2505.20738v1 |
bounded set [0,1]N. Thus there is L <∞with∥ϕ(α)−ϕ(z)∥1≤L∥α−z∥1. Now decompose the update as F∗(α) = (1 −λα)1 N+λαϕ(α)−δ1 ∥ϕ(α)−δ1∥1, λ α=∥ϕ(α)−δ1∥1 ∥ϕ(α)−δ1∥1+Nδ∈[0,1). Because δ >0, we have supαλα=:q <1. A direct calculation using the above representation yields ∥F∗(α)−F∗(z)∥1≤q∥α−z∥1.Hence F∗is a strict contraction, ... | https://arxiv.org/abs/2505.20738v1 |
an assistant skilled in rewriting samples . Given a sample : ### Sample Start ### Question : { input question content } Options : { input options content } Label : { input correct option } ### Sample End ### Your task is to execute the following process for the given sample : Step 1: Analyze the question and understand... | https://arxiv.org/abs/2505.20738v1 |
CogniBench: A Legal-inspired Framework and Dataset for Assessing Cognitive Faithfulness of Large Language Models Xiaqiang Tang1,Jian Li*2,Keyu Hu1,Du Nan2,Xiaolong Li2, Xi Zhang3,Weigao Sun4, and Sihong Xie*1 1The Hong Kong University of Science and Technology (Guangzhou) 2Hunyuan AI Digital Human, Tencent 3Beijing Uni... | https://arxiv.org/abs/2505.20767v2 |
on average contains three times more cognitive statements than previous datasets (Niu et al., 2023; See et al., 2017; Hasan et al., 2021). While factual statements mirror human roles like recorders, cognitive statements require higher-level cognitive inference beyond simple rephrasing in applications such as medical di... | https://arxiv.org/abs/2505.20767v2 |
(a)(b): Existing faithfulness assessment standards such as “Baseless” (Niu et al., 2023) and “Subjective” (Mishra et al., 2024) are ambiguous and insufficient for assessing cognitive statements. (c): We propose three increasingly rigorous assessment criteria (i.e. Rational, Grounded, Unequivocal ) to annotate cognitive... | https://arxiv.org/abs/2505.20767v2 |
misinter- pretation. We argue that the legal framework for validating Circumstantial evidence offers a natural analogy for assessing the faithfulness of cognitive statements. We further elaborate on these comparisons in Appendix A.3. Inspired by how the law validates Circumstantial evidence , we propose three criteria ... | https://arxiv.org/abs/2505.20767v2 |
only considers a more rigorous crite- rion (e.g., grounded or unequivocal) once previous less rigorous conditions (e.g., rational) are met, re- ducing cognitive load and the potential for conflict- ing interpretations. As shown in Table 1, this ap- proach outperforms independent classification (i.e., direct ask annotat... | https://arxiv.org/abs/2505.20767v2 |
as hallu- cinations, even if they are contextually justifiable. 2.Contrastive Examples : Based on these obser- vations, we provide positive and negative examples in the prompt to clarify the boundaries between faithful statements and hallucinated statements. To avoid labeling sentences individually, we implement a form... | https://arxiv.org/abs/2505.20767v2 |
statements decreases as the number of conversation turns increases, while the number of cognitive statements increases with more turns. Cognitive statements increase with the length of conversation In Fig. 5, we illustrate the dynam- ics of factual and cognitive statements relative to the number of conversation turns. ... | https://arxiv.org/abs/2505.20767v2 |
conclu- sions or opinions at the beginning or end of a turn where hallucinations are especially likely to occur. 4.6 Evaluate Hallucination Detection Methods on Cognibench Method OverallFactual HallucinationCognitive Hallucination PromptingChatGPT-3.5 48.54 22.98 56.57 ChatGPT-4 58.03 46.82 66.04 NLITasksource (COLING ... | https://arxiv.org/abs/2505.20767v2 |
catering to application-specific requirements. We have analyzed of LLMs’ dialogue patterns and the dynamics of their faithfulness. We found that while LLMs are generally capable of rephras- ing factual information accurately, their reliabil- ity drops significantly when generating cognitive statements. This highlights ... | https://arxiv.org/abs/2505.20767v2 |
Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebas- tian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learn- ing Research , 24(240):1–113. Han Ding, Yinheng Li, Junhao Wang, and Hang Chen. 2024. Large language model agent in financial trad- ing: A survey. arXiv prep... | https://arxiv.org/abs/2505.20767v2 |
for generative large language models. InProceedings of the 2023 Conference on Empiri- cal Methods in Natural Language Processing , pages 9004–9017. Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factu- ality in abstractive summarization. In Proceedings of the 58th Annual Meeti... | https://arxiv.org/abs/2505.20767v2 |
evaluation from deductive, inductive and abduc- tive views. arXiv preprint arXiv:2306.09841 . Dongjie Yang, Ruifeng Yuan, Yuantao Fan, Yifei Yang, Zili Wang, Shusen Wang, and Hai Zhao. 2023. Re- fgpt: Dialogue generation of gpt, by gpt, and for gpt. arXiv preprint arXiv:2305.14994 . Shunyu Yao, Qingqing Ke, Qiwei Wang,... | https://arxiv.org/abs/2505.20767v2 |
(e.g., detecting entity/relation errors) and struggle with cognitive statements. For instance, an LLM might faithfully cite a medical report’s data (factual) yet draw an un- founded diagnostic conclusion (cognitive). Assess- ing the latter requires reasoning about contextual plausibility, not just textual overlap—a cha... | https://arxiv.org/abs/2505.20767v2 |
a fact or group of facts established by the evidence." An inference that does not properly flow from the established fact is mere speculation . Aninference is a "deduction of fact that may logically and reasonably be drawn" from objective facts. Speculation is when the judge theorizes with- out evidentiary support or w... | https://arxiv.org/abs/2505.20767v2 |
positive and negative examples to help annotators better understand the guidelines for each category. For simplicity, these examples are omitted here. A.5.4 Examples of Annotation We provide annotated examples in Fig. 10 for four types of cognitive statements: Misleading, Speculative, Reliable, and Unequivocal. The con... | https://arxiv.org/abs/2505.20767v2 |
with Augmentation), a model for fine-grained hallucinations detections and editing. FA V A is trained on high-quality synthetic train- ing data to identify hallucinations, incorporating retrieved knowledge. We map spans to sentence- level to ensure a fair comparison. RAG-Truth (Niu et al., 2023): RAGTruth uses the Llam... | https://arxiv.org/abs/2505.20767v2 |
American football tight end. He played college football at Oklahoma State. He was drafted in the first round (20th overall) of the 2009 NFL Draft by the Detroit Lions, and spent his entire 8- year career with the team..... Pettigrew was arrested for disorderly conduct and public drunkenness on February 5, 2017, after a... | https://arxiv.org/abs/2505.20767v2 |
While these terms are not precisely measurable, they are contextually relevant and supported by the available information. Reliable Statement Context: Haunting Ground was first revealed in September 2004 Capcom added the dog mechanic during development thinking retailers and players alike would not enjoy a survival hor... | https://arxiv.org/abs/2505.20767v2 |
topic in the benchmark Figure 12: CogniDet’s detection accuracy (F1) showing log- linear scaling with training data size with additional data. This scaling law demonstrates the importance of large-scale annotated datasets like CogniBench-L for developing robust halluci- nation detection systems. A.12 Compare between Co... | https://arxiv.org/abs/2505.20767v2 |
be between apriland june, and for the southern hemisphere, between octoberand december. If you decide to plant a moringa oleifera tree in your garden, it's essential to ensure that it's protected from freezing temperatures or frost. You may also consider growing it in a large container that can be moved indoors during ... | https://arxiv.org/abs/2505.20767v2 |
Daily MailYoung actor says he has no plans to fritter his cash away . Radcliffe's earnings from first five Potter films have been held in trust fund Figure 14: Example dialoge from CogniBench with RAGTruth (Niu et al., 2023) and CNN/Daily Mail (See et al., 2017). Cognitive statements are bold and dotted underlined SYST... | https://arxiv.org/abs/2505.20767v2 |
arXiv:2505.20779v2 [cs.CL] 28 May 2025CHIMERA : A Knowledge Base of Idea Recombination in Scientific Literature Noy Sternlicht1andTom Hope1,2 1School of Computer Science and Engineering, The Hebrew University of Jerusalem 2The Allen Institute for AI (AI2) https://noy-sternlicht.github.io/CHIMERA-Web Abstract A hallmark... | https://arxiv.org/abs/2505.20779v2 |
and across domains, and also inspi- rations in the form of analogies, reductions, and abstractions. Unlike simpler concept co-occurrence approaches (Krenn et al., 2022) or more generic scientific extraction schema (Luan et al., 2018), “The concept of storyboarding, which disassembles a script into individual shots”Cont... | https://arxiv.org/abs/2505.20779v2 |
Chilton et al., 2019). In our work, we aim to extract a knowledge base Recombination extraction examples Abstract : "...Current archaeology depends on trained experts to carry out bronze dating, which is time-consuming and labor-intensive. For such dating, in this study, we propose a learning-based approach to integrat... | https://arxiv.org/abs/2505.20779v2 |
extraction model. Finally, we apply the trained model to collect recombination examples at scale. This process is illustrated in Figure 2. Data sourcing We use AI-related papers from the unarXive corpus (Saier and Färber, 2020) as a source of annotation examples1. The data un- dergo an initial keyword-based filtering t... | https://arxiv.org/abs/2505.20779v2 |
includes newer examples than unarXive (Saier and Färber, 2020). We apply our fine-tuned extrac- tion model over publications from 2019 to 2024 within the same CS categories used for the annota- tion task. After applying the model we filter pre- dictions that fail to comply with the data schema or could not be properly ... | https://arxiv.org/abs/2505.20779v2 |
concept. We utilize GPT-4o-mini as a judge of content similarity (Figure 12 in Appendix E presents our prompt). We select GPT-4o-mini over GPT-4o after conducting a qualitative examination and finding only a handful of cases in which themodel judgment differs (3 span pairs in the en- tire test set). To avoid position b... | https://arxiv.org/abs/2505.20779v2 |
base analysis Blends vs. inspirations Figures 3a and 3a present the predominant domain pairs for inspi- ration and blend relations in CHIMERA (with fre- quency above the 0.9 quantile). Inspirations display a larger selection of domains than blends. We also observe that blends connect the same or similar domains, while ... | https://arxiv.org/abs/2505.20779v2 |
and answers. The queries de- scribe the task inputs: a single graph node, the edge recombination type, and a context string, which we extract from the corresponding abstract using GPT-4o-mini . Note that this process might leak information regarding the answer (the other graph node) into the query. Therefore, we follow... | https://arxiv.org/abs/2505.20779v2 |
ZS-SciERC : a zero-shot prediction model using candidates extracted from test set abstracts with the SciERC (Luan et al., 2018) schema. Note that we use the highest ranked answer (k=1) for baselines returning a ranked list of candidates. We request the annotators to rank baseline sug- gestions based on their helpfulnes... | https://arxiv.org/abs/2505.20779v2 |
volunteers partici- pated in our human study. No personal information about the annotators or volunteers is disclosed. To promote transparency and reproducibility, we release our code and model checkpoints. The col- lected data is shared under an open license to fa- cilitate further research. We used AI assistants for ... | https://arxiv.org/abs/2505.20779v2 |
learning-based link prediction in an expo- nentially growing knowledge network. Nature Ma- chine Intelligence , 5:1326–1335. Yi Luan, Luheng He, Mari Ostendorf, and Hannaneh Hajishirzi. 2018. Multi-task identification of entities, relations, and coreference for scientific knowledge graph construction. ArXiv , abs/1808.... | https://arxiv.org/abs/2505.20779v2 |
Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly- supervised contrastive pre-training. arXiv preprint arXiv:2212.03533 . Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 202... | https://arxiv.org/abs/2505.20779v2 |
we use with the default rank of 64. The evaluation uses the corresponding repository, mistral-inference6. We rerun the same experiment using Llama-3.1-8B as a backbone, using an additional 500 warm-up steps, a learning rate of2e−5and a weight decay of 0.01. Figure 6 presents the prompt for these experiments. In additio... | https://arxiv.org/abs/2505.20779v2 |
the inspiration (e.g., the human brain) inspiration_target: str # The target of the inspiration (e.g., a learning algorithm) @dataclass class Combination(Template): """A combination describes joining two ideas, methods, models, techniques to obtain a certain goal. For example, combining two models to improve performanc... | https://arxiv.org/abs/2505.20779v2 |
1. We also experiment with PURE (Zhong and Chen, 2021), a well-known information extraction baseline. We finetune PURE over our train set using the default parameters, ex- cept for max_span_length, which we set to 40 to accommodate for the longer entities in our data. C E2E vs Specialized extraction This section reflec... | https://arxiv.org/abs/2505.20779v2 |
If the output is empty, return {}. Place your answer within <recombination> tags. Remember to carefully analyze the abstract and only identify a recombination if it is clearly present and central to the work described.Figure 8: E2E ICL prompt. {TEXT} is a placeholder for the abstract text, and {EXAMPLES} for the ICL ex... | https://arxiv.org/abs/2505.20779v2 |
span similarity prompt in Figure E. We use it in the extraction evaluation process as discussed in Section 4.1. F Error analysis We perform analysis over the test set, revealing different sources of error which may inspire fu- ture improvements. Our focus is on understanding how different types of input texts can influ... | https://arxiv.org/abs/2505.20779v2 |
<elements>{ELEMENTS}</elements> Here is a list of the standard arXiv categories: <arxiv>{ARXIV}</arxiv> And here is a list of scientific branches: <branches>{BRANCHES}</branches> For each element in the list, you need to: 1. Identify the best matching arXiv taxonomy category from the provided list. If it doesn't match ... | https://arxiv.org/abs/2505.20779v2 |
long-term" Overly-general "human experiences", "a styling method", "local search method", "a pipeline inspired by experts’ work", "a new modality", "feature based approaches" Misclassified "Reinforcement learning, or RL", "Facial Expressions Recognition(FER)", "a Kullback-Liebler regularization function", "K-nearest ne... | https://arxiv.org/abs/2505.20779v2 |
metrics in the filtered set- tings to avoid false negatives. Given the difficulty of the task we focus on ranking only the 12751 test Domain Count Domain Count Domain Count cs.cv 12504 cs.lg 8440 cs.cl 4697 cs.ro 2241 cs.ai 2091 cognitive science 936 cs.ir 884 cs.ne 864 cs.si 655 cs.hc 645 q-bio.nc 441 cs.ds 409 cs.cg ... | https://arxiv.org/abs/2505.20779v2 |
challenge in robotics... The presented bio-inspired framework heuristically models frontier exploration similar to the shepherding behavior of herding dogs . This is achieved by modeling frontiers as a sheep swarm reacting to robots modeled as shepherding dogs..." Inspiration-Source : "the multi-granular diagnostic app... | https://arxiv.org/abs/2505.20779v2 |
scalability across multiple subjects. What would be a good source of inspiration for " a highly efficient processing unit "?The human brain Existing models for link prediction in knowledge graphs primarily focus on representing triplets in either distance or semantic space , which limits their ability to fully capture ... | https://arxiv.org/abs/2505.20779v2 |
retrieved contexts match the descrip- tion and discard examples with poorly extracted in- formation (e.g., the context begins with " This study reviews the problem of... " instead of directly de- scribing the source study problem). In addition, we let the volunteers mark an example as "ill-defined", in which case we ig... | https://arxiv.org/abs/2505.20779v2 |
<source-entity> and <target-entity> " for blends and "Take inspiration from <source-entity> and apply it to <target-entity> " for inspirtions. You are an AI assistant tasked with identifying potential leakages in a given query. A leakage occurs when a query reveals or implies the answer. Follow these steps carefully: 1... | https://arxiv.org/abs/2505.20779v2 |
pair of canonical scientific con- cepts (e.g, neural networks ) that co-occur within the same abstract are considered a recombination. The figure presents an example of using AI-related concepts curated by Krenn et al. (2022) for recom- bination extraction, alongside recombination ex- tracted using our designated appro... | https://arxiv.org/abs/2505.20779v2 |
arXiv:2505.20809v1 [cs.CL] 27 May 2025Improved Representation Steering for Language Models Zhengxuan Wu∗Qinan Yu∗Aryaman Arora Christopher D. Manning Christopher Potts Stanford University {wuzhengx,qinanyu,aryamana}@stanford.edu {manning,cgpotts}@stanford.edu Abstract Steering methods for language models (LMs) seek to ... | https://arxiv.org/abs/2505.20809v1 |
the prior preference-based BiPO baseline, narrowing the gap with prompting. When applied with negative steering factors, RePS performs on par with the language modeling objective for smaller LMs but shows superior performance for the larger Gemma-3 models, again emphasizing the scalability of RePS. Moreover, RePS-train... | https://arxiv.org/abs/2505.20809v1 |
instruction while editing the response by incorporating the steering concept. This task is agnostic about how the steering is performed; in this paper, we explore a wide range of intervention-based techniques and prompting techniques. 2 Dataset. Following AXBENCH [Wu et al., 2025], given a steering concept c, we create... | https://arxiv.org/abs/2505.20809v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.