text string | source string |
|---|---|
J. Wang, F. Zhao, and D. Lin, “Sharegpt4v: Improving large multi-modal models with better captions,” in European Conference on Computer Vision . Springer, 2024, pp. 370–387. [4]H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems , vol. 36, pp. 34 892–34 9... | https://arxiv.org/abs/2505.15963v1 |
trustworthiness,” arXiv preprint arXiv:2405.17220 , 2024. [17] Y . Xie, G. Li, X. Xu, and M.-Y . Kan, “V-dpo: Mitigating hallucination in large vision language models via vision-guided direct preference optimization,” arXiv preprint arXiv:2411.02712 , 2024. 10 [18] P. Sarkar, S. Ebrahimi, A. Etemad, A. Beirami, S. Ö. A... | https://arxiv.org/abs/2505.15963v1 |
C. Gan, L. Gui, Y .-X. Wang, Y . Yang et al. , “Aligning large multimodal models with factually augmented rlhf,” in Findings of the Association for Computational Linguistics ACL 2024 , 2024, pp. 13 088–13 110. [33] E. Amirloo, J.-P. Fauconnier, C. Roesmann, C. Kerl, R. Boney, Y . Qian, Z. Wang, A. Dehghan, Y . Yang, Z.... | https://arxiv.org/abs/2505.15963v1 |
High-efficiency and high-quality synthetic data for vision language models,” arXiv preprint arXiv:2407.20756 , 2024. [44] S. Sharifzadeh, C. Kaplanis, S. Pathak, D. Kumaran, A. Ilic, J. Mitrovic, C. Blundell, and A. Banino, “Synth2: Boosting visual-language models with synthetic captions and image embeddings,” arXiv pr... | https://arxiv.org/abs/2505.15963v1 |
and ground truth, so it will lead to many incorrect judgments. We replace it with gpt-4o-2024-05-13 , which yields more accurate assessments. A.2.2 AMBER-generative & ObjectHal AMBER uses an automatic method for detecting the hallucination entity, which primarily relies on the pre-defined hallucination words. ObjectHal... | https://arxiv.org/abs/2505.15963v1 |
shadow is visible on the sand, indicating that the sun is shining and casting light on the scene. The beach setting provides a serene and picturesque backdrop for her activity. Annotated correct & wrong words Truth: woman, beach Hallucination: sky, sun, sea, dog, lounge Figure 8: Rule-based extraction will lead to misj... | https://arxiv.org/abs/2505.15963v1 |
the quality criterion coefficients τpos=τneg= 5, which means the score of positive response should be at least 6 and negative response be at most 4. The temperature of the LLM scorer is 0.1. Image Generation. For image prompt generation, we set the model’s temperature as 0.1, top_p as 0.9, and max_new_tokens as 128. We... | https://arxiv.org/abs/2505.15963v1 |
at Table 8. C Algorithm The pseudocode is at Table 4. D Efficiency and Time Consuming OViP training takes approximately 17 hours on 7 ×A800 (40G) GPUs. Among them, 4 GPUs are allocated for VLM training, 1 GPU for LLM deployment, and 2 GPUs for diffusion model deployment. We divide each training step into six stages: sa... | https://arxiv.org/abs/2505.15963v1 |
and precisely assess the model-generated response (the former) based on the provided reference answer (the latter). - Here’s how you should approach the assessment process: 1. The quality of the response depends on its accuracy and the degree of adherence to the correct answer. Therefore, if the response is much more d... | https://arxiv.org/abs/2505.15963v1 |
It is not a cow. The bird is perched on a branch, which is situated in front of a white building. **<Output Description>** A big, blue bird perched on a branch in front of a black building. **Explanation**: Both the answer and the response mention the bird, but the response is more detailed. So the description should b... | https://arxiv.org/abs/2505.15963v1 |
and white bird with a distinctive head shape and coloration. It is not a cow. The bird is perched on a branch, which is situated in front of a white building. **<Output Description>** A big, blue bird perched on a branch in front of a black building. **Explanation**: Both the answer and the response mention the bird, b... | https://arxiv.org/abs/2505.15963v1 |
arXiv:2505.15966v2 [cs.CV] 26 May 2025Pixel Reasoner: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning Alex Su♠♡∗, Haozhe Wang♢♡∗, Weiming Ren♡†, Fangzhen Lin♢, Wenhu Chen♡† University of Waterloo♡, HKUST♢, USTC♠, Vector Institute† Project Page: https://tiger-ai-lab.github.io/Pixel-Reaso... | https://arxiv.org/abs/2505.15966v2 |
likely located on the table, so I will look closer at … The mug is on the right upper corner. I need to look closer at the logo on it. Answer: This mug is from \boxed{Starbucks}. Pixel Reasoner Figure 1: Illustration of Pixel Reasoner. When asked a visually-rich question, Pixel-Reasoner first inspects the visual inputs... | https://arxiv.org/abs/2505.15966v2 |
that facilitate the cultivation of both mastery over visual operations and self-correction capabilities. Following this meticulous warm-start instruction tuning, our RL approach leverages a curiosity-driven reward scheme to balance the exploration and exploitation of pixel-space reasoning to incentivize the pixel-level... | https://arxiv.org/abs/2505.15966v2 |
al., 2025, Liu et al., 2025]: r(x,y) =1if the solution ycontains the correct answer to query x, 0otherwise . In this work, we focus on two types of visual inputs: images and videos. We specifically consider two types of visual operations: ZOOM -INfor inspecting details within a specified region of a target image, and ... | https://arxiv.org/abs/2505.15966v2 |
that specific cue. Synthesize Expert Trajectories. Based on the curated vision-language queries requiring fine-grained visual analysis, we then synthesize expert trajectories using GPT-4o. As illustrated in Fig. 3 (a), we observed that direct distillation from GPT-4o sometimes resulted in "bypassing trajectories". In t... | https://arxiv.org/abs/2505.15966v2 |
represent either execution outputs from visual operations or the specifically designated erroneous visual operations within the self-correction trajectories. Masking the erroneous operations prevents the policy from learning to execute the incorrect actions. 4 Curiosity-Driven Reinforcement Learning The warm-started mo... | https://arxiv.org/abs/2505.15966v2 |
yields the following modified reward function r′(x,y), detailed in the appendix: r′(x,y) =r(x,y) +α·rcuriosity (x,y) +β·rpenalty (y), (3) where rcuriosity (x,y) = max(H −RaPR (x),0)·1PR(y) (4) rpenalty (y) = min(N −nvo(y),0) (5) The modified reward incorporates two additional terms. The first term rcuriosity (x,y)serve... | https://arxiv.org/abs/2505.15966v2 |
model to locate, differentiate, and tally objects across complex scenes. MVBench [Li et al., 2024a] is a comprehensive benchmark designed to evaluate multimodal large language models (MLLMs) on their temporal understanding capabilities across 20 challenging video tasks, necessitating reasoning beyond static image analy... | https://arxiv.org/abs/2505.15966v2 |
Remarkably, Pixel-Reasoner , at a mere 7B parameters, not only surpasses substantially larger open-source models like the 27B Gemma3 across all benchmarks, but also outperforms specialized models that depend on external tools, such as IVM-Enhance (GPT-4V). Furthermore, Pixel-Reasoner ’s exceptional capabilities extend ... | https://arxiv.org/abs/2505.15966v2 |
Visual Operations Requires Instruction Tuning. A crucial finding is that meticulous warm-start instruction tuning is essential for enhancing the policy’s mastery of visual operations and its capacity for self-correction. To demonstrate this, we analyze the RL training dynamics originating from three distinct instructio... | https://arxiv.org/abs/2505.15966v2 |
in RaPR in the first 50 gradient steps, and then plateaus for around 150 steps. During this stage, the policy is compelled by the curiosity bonus to continue exploring pixel-space reasoning, despite its relative inferiority compared to textual reasoning (as shown in Fig. 7 (Right)). Not until 200 gradient steps, the po... | https://arxiv.org/abs/2505.15966v2 |
Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276 , 2024. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720 , 2024a. Gemini Team, Rohan Ani... | https://arxiv.org/abs/2505.15966v2 |
multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 , 2024a. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Z... | https://arxiv.org/abs/2505.15966v2 |
Chen, and Jimmy Lin. Visa: Retrieval augmented generation with visual source attribution, 2024. URL https: //arxiv.org/abs/2412.14457 . Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos, 2024. URL https://arxiv.org/abs/2405.09711 . Deepak... | https://arxiv.org/abs/2505.15966v2 |
Xing, Liangdong Wang, Zhou Cao, Jintao Jia, Zhuoyi Zhang, Yixuan Wang, et al. Infinity-mm: Scaling multimodal performance with large-scale and high-quality instruction data. arXiv preprint arXiv:2410.18558 , 2024. Jarvis Guo, Tuney Zheng, Yuelin Bai, Bo Li, Yubo Wang, King Zhu, Yizhi Li, Graham Neubig, Wenhu Chen, and ... | https://arxiv.org/abs/2505.15966v2 |
(i.e., the constraint is "over-satisfied"), potentially encouraging the policy to use far fewer visual operations than necessary. Secondly, the term −λ1(H−RaPR (x))operates on the expectation-level and does not properly reward individual responses y∼πθ. Therefore, we adopt the following modified reward function: r′(x,y... | https://arxiv.org/abs/2505.15966v2 |
relevant visual contents for answers, offering both visual and temporal context for potential video-specific operations. Detailed Data Pipeline Illustration. As the Fig. 8 depicts, after we obtain reference visual cues from seed data, we input both the whole HR image or video and the corresponding localized reference v... | https://arxiv.org/abs/2505.15966v2 |
H = 0 .3encourages the policy to utilize pixel-space reasoning in approximately 30% of responses generated for a given query, while N = 1 promotes efficiency by favoring responses that require at most one visual operation. Under these parameters, a response can receive a maximum exploration bonus of approximately 17 0 ... | https://arxiv.org/abs/2505.15966v2 |
have placed an item down or completed an action involving the counter.\n\nNow I will select some frames to look clearer at which object was put down by the person. Execution error:max() arg is an empty sequence It seems there was an issue with selecting frames. Let me describe the scene instead:\n\ nThe person in the v... | https://arxiv.org/abs/2505.15966v2 |
the tool description {tool_description}. For each tool call, return a json object with function name and arguments within <tool_call></tool_call> XML tags: <tool_call> {{"name": <function-name>, "arguments": <args-json-object>}} </tool_call> Stop generating after you call a tool. Here is the image. E.2.2 Prompts for Ex... | https://arxiv.org/abs/2505.15966v2 |
format, results, and notable 21 moments. The article also includes a table summarizing the results of all SEC Championship games since its inception in 1992. Now I will zoom in to look clearer at the part about "who won the first SEC championship in football." Analyzing the cropped part: The cropped section includes a ... | https://arxiv.org/abs/2505.15966v2 |
Explaining Puzzle Solutions in Natural Language: An Exploratory Study on 6×6Sudoku Anirudh Maiya Razan Alghamdi Maria Leonor Pacheco Ashutosh Trivedi Fabio Somenzi University of Colorado Boulder {first.last}@colorado.edu Abstract The success of Large Language Models (LLMs) in human-AI collaborative decision- making hin... | https://arxiv.org/abs/2505.15993v1 |
these tools. These explanations usu- ally come in the form of fine-grained proofs that require a good grasp of proof terminology and are often overwhelmingly detailed. They often refer to encodings of the problem that are cumbersome to relate to the original presentation of the puzzle. To be useful in practical scenari... | https://arxiv.org/abs/2505.15993v1 |
from informal math problems. It fine-tunes an LLM, translates informal problems into formal Lean 4 statements, filters low-quality formalizations using a chain-of- thought evaluation, and attempts to prove both the formal statement and its negation. This iterative process refines the model’s performance by incor- porat... | https://arxiv.org/abs/2505.15993v1 |
88.7% with 10 or 12 givens, 11.1% with 14 or more givens. Symmetrical puzzles have an even number of givens. Moreover, the puzzles in the Menneske data set with more than 8 givens, unlike ours, have redundant givens. Both differ- ences contribute to the higher average number of givens when compared to our sample. Asses... | https://arxiv.org/abs/2505.15993v1 |
struggle with integrating all constraints across the grid. For these models, column-wise and row-wise constraints were prioritized over box- Accuracy (%) LLM Cell-wise Row-wise Column-wise Box-wise Fully Correct Solutions Mistral-7B-Instruct-v0.3 21.96 0.59 0.17 0.17 0.00 Llama-3.1-8B-Instruct 39.65 2.22 2.01 1.36 0.00... | https://arxiv.org/abs/2505.15993v1 |
were chosen. The clarity dimension scored poorly, with 7.5% "Yes" responses (32.5% "Maybe", 60% "No"), with re- ported issues regarding logical progression and in- consistent terminology. Educational value proved to be the weakest dimension, with just 2.5% "Yes" responses (52.5% "Maybe", 45% "No"), as explana- tions fa... | https://arxiv.org/abs/2505.15993v1 |
LLMs might encounter in more difficult puzzles, potentially underes- timating the models’ limitations in handling more complex tasks. 3.Impact of Fine-Tuning: Our study was fo- cused on evaluating the capacity of LLMs to produce solutions and explanations out of the box, without any fine-tuning. Of particular interest ... | https://arxiv.org/abs/2505.15993v1 |
80–89. Jie Huang and Kevin Chen-Chuan Chang. 2023. To- wards reasoning in large language models: A survey. InFindings of the Association for Computational Linguistics: ACL 2023 , pages 1049–1065, Toronto, Canada. Association for Computational Linguistics. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris ... | https://arxiv.org/abs/2505.15993v1 |
models. In Proceedings of the 36th International Conference on Neural Information Processing Systems , NIPS ’22, Red Hook, NY , USA. Curran Associates Inc. Huajian Xin, Daya Guo, Zhihong Shao, Zhizhou Ren, Qihao Zhu, Bo Liu (Benjamin Liu), Chong Ruan, Wenda Li, and Xiaodan Liang. 2024. Deepseek- prover: Advancing theor... | https://arxiv.org/abs/2505.15993v1 |
experiments, we employed zero-shot prompting to evaluate the models’ ability to solve 6x6 Sudoku puz- zles without additional task-specific training or prior context. 2.Temperature and top-p: All models were run with a fixed temperature of 0 to ensure de- terministic outputs across multiple runs. Ad- ditionally we use ... | https://arxiv.org/abs/2505.15993v1 |
arXiv:2505.16000v2 [cs.CL] 24 May 2025Leveraging Online Data to Enhance Medical Knowledge in a Small Persian Language Model 1stMehrdad Ghassabi Faculty of Computer Engineering University of Isfahan Isfahan, Iran m.ghassabi@eng.ui.ac.ir2ndPedram Rostami School of Electrical and Computer Engineering University of Tehran ... | https://arxiv.org/abs/2505.16000v2 |
devices, ensuring that sensitive patient data remains secure and confidential, which is a paramount consideration in healthcare settings. However, the unavailability of appropriate medical corpora and datasets in Persian has hindered progress in this area, impeding the creation of robust language models that can effect... | https://arxiv.org/abs/2505.16000v2 |
datasets, enabling it to achieve optimal performance tailored for mobile applications in the medical domain. By focusing on the specific requirements of mobile users, MedMobile sought to deliver a model that is both efficient and effective, ensuring accessibility to high-quality medical information on the go. 2we have ... | https://arxiv.org/abs/2505.16000v2 |
research. Fur- thermore, if we merge our updated parameters into aya-vision [18] , another model of aya family, instead of aya-expanse, we gain the capability to incorporate medical images such as MRIs and CT scans as inputs, thereby enhancing our model’s applicability to the medical domain. IV. D ATA A. Corpus As prev... | https://arxiv.org/abs/2505.16000v2 |
3.66M Guangtao Zeng et al. English [25] Medical- Meadow English 160k Tianyu Han et al. [26] MF3QA Persian 20k us In our research, we crawled more than 180,000 question- answer pairs from Persian medical forums, employing both manual and automatic filtering methods to refine the dataset through a laborious cleaning proc... | https://arxiv.org/abs/2505.16000v2 |
handle longer context lengths and larger batch sizes efficiently, enabling effective fine-tuning for the next-token prediction objective while balancing computational efficiency with model performance.B. Instruction tuning Following the fine-tuning stage, we performed instruction- tuning on the fine-tuned model using o... | https://arxiv.org/abs/2505.16000v2 |
Iranian Basic Medical Sciences Entrance Exam, which stands at 36%, making it the first Persian language model with fewer than 8 billion parameters to pass this exam. Furthermore, our model demonstrated improvements on the translated MMLU dataset, achieving not only higher average scores but also excelling across most s... | https://arxiv.org/abs/2505.16000v2 |
that existing systems are ill- equipped to handle the complexities of medical terminology. As shown in Figure 4, these limitations have resulted in all pipeline alternatives achieving lower win rates against our model, Gaokerena. This performance gap highlights the inadequacies of current pipeline approaches in meeting... | https://arxiv.org/abs/2505.16000v2 |
Yang, An, et al. “Qwen2 Technical Report.” arXiv Preprint arXiv:2407.10671, 2024. [17] Rostami, Pedram, Ali Salemi, and Mohammad Javad Dousti. ”Persian- mind: A cross-lingual persian-english large language model.” arXiv preprint arXiv:2401.06466 (2024). [18] Dash, Saurabh, et al. ”Aya Vision: Advancing the Frontier of ... | https://arxiv.org/abs/2505.16000v2 |
arXiv:2505.16002v1 [cs.CL] 21 May 2025Causal Interventions Reveal Shared Structure Across English Filler–Gap Constructions Sasha Boguraev1Christopher Potts2Kyle Mahowald1 1The University of Texas at Austin2Stanford University {sasha.boguraev,kyle}@utexas.edu cgpotts@stanford.edu Abstract Large Language Models (LLMs) ha... | https://arxiv.org/abs/2505.16002v1 |
the lion ate.” to “It was the gazelle that the lion ate.” The former is an embed- ded wh- clause and the latter is a cleft construction. 1 These are distinct constructions but share some- thing in common: both have a long-distance depen- dency with an extracted element, often specified with a linguistic trace: “I wonde... | https://arxiv.org/abs/2505.16002v1 |
colloquially being said to leave a ‘gap’ at its canonical position. This grammatical family encompasses a wide range of common constructions including wh-questions, relative clauses, clefts, and more. Filler–gap dependencies have long been a target of linguistic inquiry. They are believed to require sophisticated synta... | https://arxiv.org/abs/2505.16002v1 |
is shared across constructions. But most of this work has been behavioral, without exploring the model’s underlying causal mechanisms. Our work fills this gap. We first uncover the causal mechanisms LMs learn to process various filler–gap dependencies, and then we measure to what extent these mecha- nisms generalize ac... | https://arxiv.org/abs/2505.16002v1 |
sides in Figure 1). In our experiments, bandsare always the outputs of a Transformer block. Intuitively, this intervention defines a direction in the rotated feature space de- fined by the learned vector a∈Rn. This is a soft intervention targeting only the learned feature and preserving orthogonal dimensions of b. In D... | https://arxiv.org/abs/2505.16002v1 |
those they were trained on. We aggre- gate across layers by max-pooling ODDS because our causal methodology aims to localize syntactic features in the model, with the maximum value rep- resenting the most causally efficacious localization of the given features. 4 Exp. 1: Do LMs Share Filler–Gap Mechanisms Across Constr... | https://arxiv.org/abs/2505.16002v1 |
single clause templates and nearly every position in the multi-clause ones. These results strongly suggest shared internal rep- resentations across filler–gap constructions in the evaluated models. To test our hypotheses regarding the effect of training and evaluation set overlap and matching animacy, we fit a linear m... | https://arxiv.org/abs/2505.16002v1 |
size and color proportional toODDS of the source construction’s interventions mea- sured on the target construction. Bottom: In- and out- degree centrality AUCs against construction frequency. total a given node’s outgoing edges are connected to – and in-degree centrality – the fraction of nodes its incoming edges come... | https://arxiv.org/abs/2505.16002v1 |
for filler type at the FILLER and THE positions, inversion of the head daughterand nature of the fronted element at the FILLER , THE, and NPpositions, and syntactic category of the mother at all positions. Discussion These results paint a clear picture of filler–gap generalization in LMs. Frequent con- structions are e... | https://arxiv.org/abs/2505.16002v1 |
MAX ODDS when evaluated on both the matrix and embedded clause of our multi-clause variants. Results Our results are displayed in Figure 5. Our single-clause mechanisms show above-chance MAX ODDS at the FILLER through NP1positions of the matrix clause, before dropping off at the VERB 1through THE 2positions, and then s... | https://arxiv.org/abs/2505.16002v1 |
where there is not. Our investigation is oriented toward finding evi- dence of modular structure in LLMs. However, it is not a settled question what constitutes rule-like or systematic linguistic behavior in neural systems (Geiger et al., 2024; Nefdt, 2023; Buckner, 2024; Futrell and Mahowald, 2025). How causally sys- ... | https://arxiv.org/abs/2505.16002v1 |
Tal Linzen, and Yonatan Belinkov. 2021. Causal analysis of syntactic agreement mechanisms in neural language models. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages 1828–1... | https://arxiv.org/abs/2505.16002v1 |
of the 60th Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers) , pages 8818–8831, Dublin, Ireland. Association for Compu- tational Linguistics. Christopher D Manning, Kevin Clark, John Hewitt, Ur- vashi Khandelwal, and Omer Levy. 2020. Emer- gent linguistic structure in artificial... | https://arxiv.org/abs/2505.16002v1 |
language models learn about filler–gap dependencies? In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyz- ing and Interpreting Neural Networks for NLP , pages 211–221, Brussels, Belgium. Association for Com- putational Linguistics. Ethan Gotlieb Wilcox, Richard Futrell, and Roger Levy. 2024. Using computation... | https://arxiv.org/abs/2505.16002v1 |
NC Article NP Verb Label Emb. Wh-Q ( K) I know what/that the man built ./it Emb. Wh-Q ( W) I wonder what/if the man built ./it Matrix Wh-Q What/"" did the man build ?/it RRC The chair which/and the man built was/it Cleft It was the chair/clear that the man built ./the chair PC What/That the man built was/it Topic. Actu... | https://arxiv.org/abs/2505.16002v1 |
False Verb Phrase True Pseudocleft Wh-Item False N/A True Topicalization Phrasal Element False N/A True Table 6: Parameter values for each filler–gap construction. C Regression Details We perform all regressions with the lmerTest package in R (Kuznetsova et al., 2017). C.1 Experiment 1 Regression In the leave-one-out s... | https://arxiv.org/abs/2505.16002v1 |
Matrix Wh-Questions 82 Clefts 20 Pseudocleft 6 Topicalization 6 Total Sentences 16622 Table 9: Construction Type Counts 13 Term βFILLER βTHE βNP βVERB (Intercept) 1.93*** 2.70*** 1.87*** 9.06*** in_train_set 0.67*** 0.56*** 0.42** 0.26 same_animacy 1.08*** 0.51*** 0.60*** 2.13*** in_train_set:same_animacy 0.36** 0.20**... | https://arxiv.org/abs/2505.16002v1 |
Multi-Clause Aggregation Values by Evaluation Group In Train Set, Same AnimacyLeft Out, Same AnimacyIn Train Set, Different AnimacyLeft Out, Different AnimacyControl Lexical ControlCleftEmbedded Wh−Question (Know−Class)Embedded Wh−Question (Wonder−Class)Matrix Wh−QuestionPseudocleftRestricted Relative ClauseTopicalizat... | https://arxiv.org/abs/2505.16002v1 |
ODDS . {filler} the {np} {verb} 0.000.250.500.751.00 Same Animacy, Not in Train Set Difference Animacy, Not in Train SetSame Animacy, In Train Set Different Animacy, In Train SetControls Lexical Controlspythia−6.9b, Average Max Odds by Position (b)pythia 6.9b average normalized MAX ODDS . Figure 16: Top: pythia-2.8b an... | https://arxiv.org/abs/2505.16002v1 |
arXiv:2505.16003v1 [cs.CL] 21 May 2025SLMEval: Entropy-Based Calibration for Human-Aligned Evaluation of Large Language Models Roland Daynauth, Christopher Clarke, Krisztian Flautner, Lingjia Tang, Jason Mars University of Michigan {daynauth, csclarke, manowar, lingjia, profmars}@umich.edu Abstract The LLM-as-a-Judge p... | https://arxiv.org/abs/2505.16003v1 |
address this limitation, we introduce SLMEval, a novel and effective evaluation frame- work that uses entropy maximization to calibrate LLM-generated scores based on a small amount of human evaluation data. This calibration pro- motes well-calibrated LLM-based judgments that better correlate with human evaluations. In ... | https://arxiv.org/abs/2505.16003v1 |
Ri andRjfrom models i, j∈Mwhere i̸=j. The evaluator freturns scalar scores Sij, Sji∈[1,10] such that: f(Q, R i, Rj) ={Sij, Sji} (1) By default, model iis said to beat jifSij> Sji. However, scores suffer from several well- documented biases and limitations, such as token length bias (Li et al., 2023), position bias (Wan... | https://arxiv.org/abs/2505.16003v1 |
following two use cases are used: •Daily Pep Talk (PT) : At the beginning of each day, an encouraging message is presented to the user based on what they had accomplished the previous day and their goals for the current day. •Recommendation (RE) is a personalized sug- gestion provided to users, helping them orga- nize ... | https://arxiv.org/abs/2505.16003v1 |
2.0x GPT4 + BPC (k = 3) 62.5% 6x SLMEval 58.8% 0.2x Table 2: Comparison of evaluation accuracy on the FairEval dataset. SLMEval, using a 4-bit quantized model, outperforms GPTScorer and GPTScore while approaching the performance of GPT-4 with BPC ( k= 3). As shown in Table 2, SLMEval achieves an accuracy of 58.8%, outp... | https://arxiv.org/abs/2505.16003v1 |
arXiv:1803.11175 . Yi Chen, Rui Wang, Haiyun Jiang, Shuming Shi, and Ruifeng Xu. 2023. Exploring the use of large lan- guage models for reference-free text quality evalua- tion: An empirical study. In Findings of the Associa- tion for Computational Linguistics: IJCNLP-AACL 2023 (Findings) , pages 361–374. Wei-Lin Chian... | https://arxiv.org/abs/2505.16003v1 |
on Empirical Methods in Natural Language Processing , pages 2511–2522, Singapore. Association for Com- putational Linguistics. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023b. G-eval: Nlg evaluation using gpt-4 with better human align- ment. In Proceedings of the 2023 Conference on E... | https://arxiv.org/abs/2505.16003v1 |
Yiwei Li, Xinglin Wang, Boyuan Pan, Heda Wang, and Kan Li. 2023. Batcheval: Towards human-like text evaluation. arXiv e-prints , pages arXiv–2401. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2020. Bertscore: Eval- uating text generation with bert. arXiv preprint arXiv:1904.09675 . Lianm... | https://arxiv.org/abs/2505.16003v1 |
to- gether with GPT-4 (OpenAI, 2024). A.5 Entropy Maximization via the Principle of Maximum Entropy We estimate the latent model strength distribution p={p1, . . . , p n}using the Principle of Maximum Entropy , which selects the least biased distribution consistent with observed constraints. Specifically, we solve for ... | https://arxiv.org/abs/2505.16003v1 |
Interpretability Illusions with Sparse Autoencoders: Evaluating Robustness of Concept Representations Aaron J. Li1, Suraj Srinivas2, Usha Bhalla1, and Himabindu Lakkaraju1 1Harvard University 2Bosch Research Abstract Sparse autoencoders (SAEs) are commonly used to interpret the internal activations of large language mo... | https://arxiv.org/abs/2505.16004v1 |
incorporating robustness as a core component of the SAE evaluation agenda, and concept-based explanations in general. We begin by defining SAE robustness under input-levelperturbations, followedbythespecificationofadversarialsettingsforacomprehensive robustness evaluation. Each setting is characterized by two key objec... | https://arxiv.org/abs/2505.16004v1 |
shaped the design of SAE training objectives, and the alignment of learned latents with human knowledge (i.e., human-understandability) [Cunningham et al., 2023, Paulo et al., 2024], recent works have begun to assess SAE performance from a more interpretation-centric perspective [Makelov et al., 2024, Karvonen et al., ... | https://arxiv.org/abs/2505.16004v1 |
the conditions under which such alignment is violated. This can be formally expressed in two directions as: ∃c1, c′ 1∈ C:dc(c1, c′ 1)< ϵc, dz(z1, z′ 1)> δz (3) ∃c1, c2∈ C:dc(c1, c2)> δc, dz(z1, z2)< ϵz (4) where zi= (fSAE◦fLLM◦f−1 c)(ci),∀ci∈ C. The distance metrics dcanddz, along with the thresholds ϵc, δc, ϵz, δz, sh... | https://arxiv.org/abs/2505.16004v1 |
Untargeted Given x1and a fixed edit distance ϵxbetween token sequences, the attack aims to find a perturbed x′ 1that maximizes the difference in SAE activation. The perturbation direction is not predefined but is empirically selected to induce the maximal change in the sparse latent space Z. This setting corresponds to... | https://arxiv.org/abs/2505.16004v1 |
single token replacement to ensure that the change at the input level remains minimal. The overall robustness under replacement mode can be measured by the average across all token positions. These two modes represent different types of adversarial control, and both are highly plausible in practical deployment scenario... | https://arxiv.org/abs/2505.16004v1 |
al., 2024b]. Both SAEs have a width of 131k. Our experiments are not conducted on smaller models or SAEs with fewer latents, as our targeted attacks require a relatively large initial distance dz(x1, x2)to meaningfully evaluate their effectiveness. In other words, we evaluate the robustness of an SAE only when it can c... | https://arxiv.org/abs/2505.16004v1 |
generated Art&Sciencedataset in Table 2, and provide the loss curves and adversarial inputs for two representative attack examples in Figure 1 and 2. For tasks on population level, we report the change in overlap ratio (as a percentage) between two sets of kmost activated latents, while for those on individual level we... | https://arxiv.org/abs/2505.16004v1 |
Based on the results for the suffix lengths and single-token replacement used in our experiments reported in Table 3, we show that our adversarial attacks are mostly targeting the SAEs, without notably affecting the base LLM. The prompt template and self-summarization examples are included in Appendix E. Model Dataset ... | https://arxiv.org/abs/2505.16004v1 |
takes approximately 10 minutes. In practical applications, users may only need to evaluate the most relevant scenario among our eight set- tings. Additionally, future work could explore optimizing the attack pipeline to better balance effectiveness and efficiency. More broadly, we consider our evaluation framework and ... | https://arxiv.org/abs/2505.16004v1 |
and Jeffrey Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093 , 2024. Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsimas. Finding neurons in a haystack: Case studies with sparse probing. arXiv preprint arXiv:2305.01610 , 2023. Jing Huang, Zhe... | https://arxiv.org/abs/2505.16004v1 |
Neel Nanda. Improving dictionary learning with gated sparse autoencoders. arXiv preprint arXiv:2404.16014 , 2024a. Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, János Kramár, and Neel Nanda. Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders. arXi... | https://arxiv.org/abs/2505.16004v1 |
+56.2±2.0 +24 .8±1.1 62 .7±0.9,83.8±0.4 64 .0±0.5,92.2±0.1 Table C1: Attack results for all eight settings for Llama-3-8B and Gemma-2-9B on the AG News (300 samples) dataset. Both suffix and replacement modes are restricted to one token. For individual level, we include results for both activation (left) and deactivati... | https://arxiv.org/abs/2505.16004v1 |
warn about extreme climate events A close reading of Chaucer’s *Canterbury Tales* reveals satire aimed at religious hypocrisy, gender roles, and social class divisions MinistersRank = 31834 Rank = 55 U.S. Seeks Reconciliation with Oil-Rich Venezuela Badgers ride early surge U.S. Seeks Reconciliation with Oil-Rich Venez... | https://arxiv.org/abs/2505.16004v1 |
from her cultural background. The previ- ous sentence is about Assistant (Llama 3 8B): the artist, and the follow- ing is about the art. The artist is a woman, and the art is a painting. The painting is a dream...Example 1: Generated Summariza- tion of Perturbed Input User:She paints surreal dreamscapes that draw on an... | https://arxiv.org/abs/2505.16004v1 |
to disrupt existing semantic features without the need for precise control. Consequently, although the upper bound of attack effectiveness in the targeted setting slightly decreases with model depth, this trend is likely driven by representational shifts across LLM 18 layers rather than properties of the SAEs themselve... | https://arxiv.org/abs/2505.16004v1 |
to targeted input manipulations. 20 Example 1: LLM-as-a-Judge User:You will be given two (potentially truncated) summarizations of a particular sentence, and your goal is to judge whether the two summarizations are focusing on the same main topic. Please answer Yes or No. 1. The previous sentence is about the importanc... | https://arxiv.org/abs/2505.16004v1 |
User:No. User:You will be given two (potentially truncated) summarizations of a particular sentence, and your goal is to judge whether the two summarizations are focusing on the same main topic. Please answer Yes or No. 1. The previous sentence is about the statue of the civil rights activist, the statue is located in ... | https://arxiv.org/abs/2505.16004v1 |
LAGO: Few-shot Crosslingual Embedding Inversion Attacks via Language Similarity-Aware Graph Optimization Wenrui Yu1, Yiyi Chen2, Johannes Bjerva2, Sokol Kosta1, Qiongxiu Li1* 1Department of Electronic Systems,2Department of Computer Science Aalborg University, Copenhagen, Denmark wenyu@es.aau.dk,{yiyic,jbjerva}@cs.aau.... | https://arxiv.org/abs/2505.16008v1 |
the few-shot regime through decoder transfer, it lacks mech- 1arXiv:2505.16008v1 [cs.CL] 21 May 2025 anisms to exploit language similarity, which we hy- pothesize is a key factor in crosslingual generaliza- tion failure. Prior studies have shown that language similarities, simulated from typological features and lexica... | https://arxiv.org/abs/2505.16008v1 |
samples. ALGEN (Chen et al., 2025b) in- troduces a linear alignment technique, allowing a decoder trained in one domain to be reused in another. While effective in few-shot transfer, AL- GEN does not explicitly model language similarity or structural relationships between languages (cf. Section 3.2). Our framework impr... | https://arxiv.org/abs/2505.16008v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.