text string | source string |
|---|---|
6 Comparison between different data mixing strategies in the stage 2 of Multi-stage RL. Mixing MethodPuzzle Math In-Domain Out-of-DomainAIME 24 ARC-AGI 1 ARC-AGI 2 Enigmata -Eval KORBench HRV 34.7 0.1 57.7 60.1 50.1 Easy:Medium:Hard = 1:1:1 34.5 0.6 58.5 64.0 56.6 Easy:Medium:Hard = 2:6:2 35.1 0.3 58.9 63.8 48.0 using ... | https://arxiv.org/abs/2505.19914v1 |
each column represents a different training approach. The curves show how the average reward changes with training steps for different difficulty levels. tasks like Minesweeper, Mix-Training struggles without SFT, while Multi-stage RL still learns effectively. Table 8 Impact of code utilization on Qwen2.5-32B- Enigmata... | https://arxiv.org/abs/2505.19914v1 |
research on reasoning models. 11 7 Contributions Project Lead Jiangjie Chen1,6 Core Contributors† Jiangjie Chen1,6, Qianyu He1,2‡, Siyu Yuan1,2‡, Aili Chen2 Contributors Zhicheng Cai4,6, Weinan Dai1,3,6‡, Hongli Yu1,3,6‡, Qiying Yu1,3,6‡, Xuefeng Li1,5‡, Jiaze Chen1 Supervisors Hao Zhou3,6, Mingxuan Wang1,6 Affiliation... | https://arxiv.org/abs/2505.19914v1 |
al. Kimi k1. 5: Scaling reinforcement learning with llms. arXivpreprint arXiv:2501.12599, 2025. [19]Yufeng Yuan, Qiying Yu, Xiaochen Zuo, Ruofei Zhu, Wenyuan Xu, Jiaze Chen, Chengyi Wang, TianTian Fan, Zhengyin Du, Xiangpeng Wei, et al. Vapo: Efficient and reliable reinforcement learning for advanced reasoning tasks.ar... | https://arxiv.org/abs/2505.19914v1 |
Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings oftheACMSIGOPS 29thSymposium onOperating Systems Principles, 2023. [37] OpenAI. GPT4 technical report. arXivpreprint arXiv:2303.08774, 2023. [38] x.ai. Bringing grok to everyone, 2025. [39]jeggers. fu... | https://arxiv.org/abs/2505.19914v1 |
dataset. We also include synthetic ARC-AGI data§. We then use DeepSeek-R1 to generate 8 candidate solutions for each instance and select one correct solution per instance. The final puzzle dataset contains 12,041 high-quality puzzle samples. For the mathematical part of the RFT dataset, we collected mathematical proble... | https://arxiv.org/abs/2505.19914v1 |
pretrained value model is then used to initialize the critic for PPO. •Decoupled GAE. We use different λvalues for the policy and the value model to reduce reward decay and improve optimization over long token sequences. Specifically, we set λcritic = 1.0for unbiased value estimation, and λpolicy = 0.95to improve sampl... | https://arxiv.org/abs/2505.19914v1 |
training set while maintaining (if not surpassing) the original performance of tasks in the original dataset. ¶https://github.com/volcengine/verl . 18 0.000.250.500.751.00Big Bench Symbolic Binario Campsite Car Painting 0.000.250.500.751.00Checkmate In One Countdown Crypto Kka Crypto Kpa 0.000.250.500.751.00Full Crossw... | https://arxiv.org/abs/2505.19914v1 |
per row and column, such that the number of visible buildings from each edge matches the provided visibility constraints, with taller buildings obscuring shorter ones. Return a valid configuration or report infeasibility. Sum Skyscraper Grid Auto grid size n Given an N X N grid, fill it with numbers 1 to N without repe... | https://arxiv.org/abs/2505.19914v1 |
cars c, number of color type t, shift range k, skew range sGiven a fixed initial sequence of cars to be painted, each car may be moved up to K positions forward or backward; the objective is to reorder the cars, within this constraint, to minimize adjacent color transitions (color switches). Stack PermutationSequential... | https://arxiv.org/abs/2505.19914v1 |
a move that results in immediate checkmate—i.e., the opposing king is placed in check and has no legal way to escape. Tic Tac Toe Search Auto board size n, comparative potential p center control c, fork score fGiven a partially filled N ×N Tic Tac Toe board and the active player, identify the optimal move that maximize... | https://arxiv.org/abs/2505.19914v1 |
Each task includes a rule description and few-shot examples demonstrating puzzle mechanics. Color coding in prompts indicates different puzzle categories. These examples highlight the diversity of our dataset, from grid-based challenges (Sudoku, Star Battle) to sequential puzzles (Eight Puzzle, Fifteen Puzzle). Each pu... | https://arxiv.org/abs/2505.19914v1 |
: 24 ‘‘‘ <begin_board > [ Final Board ] <end_board > ‘‘‘ Puzzle : Here is the puzzle : total number of tents : 4 tents in each row : 1 1 0 2 tents in each column : 1 1 1 1 ‘‘‘ . X . . . . . . X X . X . . . . ‘‘‘ Listing 3 Case of Magic Square You are provided with a 3x3 Magic Square puzzle . Some cells are filled... | https://arxiv.org/abs/2505.19914v1 |
clues , ensuring that the heights of the buildings are distinct in each row and column and follow the visibility clues at the edges . Example : [2] [1] [2] [3] +---+---+---+---+ [2] | | | | | [2] +---+---+---+---+ [3] | | | | | [2] +---+---+---+---+ 25 [1] | | | | | [2] +---+---+---+---+ [4] | | | | | [1] +---+---+---+... | https://arxiv.org/abs/2505.19914v1 |
taller ones in front . 4. Objective : Fill the entire matrix according to the clues , ensuring that the heights of buildings in each row and each column are different , and that they comply with the visibility clues on the sides . Example : [7] [4] [5] [9] +---+---+---+---+ [7] | | | | | [6] +---+---+---+---+ [9] | | |... | https://arxiv.org/abs/2505.19914v1 |
cells are denoted by ’X’, and stars are denoted by ’*’. 2. STAR MUST BE PLACED IN EMPTY CELL : - Each star must be placed in an EMPTY cell . - Blocked cells cannot contain stars . - You can only change cells denoted by ’.’, and must not change cells denoted by ’X ’. 3. EXACTLY 1 STAR IN EACH ROW AND COLUMN : - Each row... | https://arxiv.org/abs/2505.19914v1 |
Case of Sudoku2 You are provided with a 4x4 Sudoku puzzle . Some cells are filled with numbers , while empty cells are represented by dots . Your task is to find a valid solution for the puzzle based on the following rules : Rules : 1. Board Structure : The Sudoku board is a 4x4 grid , divided into 4 smaller 2x2 sub... | https://arxiv.org/abs/2505.19914v1 |
1 2 6 . . 5 9 4 . 8 6 4 5 9 3 1 2 5 9 4 2 3 1 7 8 6 1 2 5 3 8 6 9 4 7 8 6 3 7 9 4 1 2 . 9 4 7 5 1 2 8 6 3 4 7 8 . . . 6 . . 6 . 1 8 . . 2 5 9 2 5 9 . 6 . 4 7 8 ‘‘‘ Please provide the answer according to the above requirements . Listing 9 Case of Full Crosswords Task : Your task is to complete the crossword puzzle... | https://arxiv.org/abs/2505.19914v1 |
Case of Symbolic Hard Task and Rules : Figure out the pattern in the following examples and apply it to the test case . Your answer must follow the format of the examples . Output Format : Please output your answer within a code block (‘‘‘), formatted as a grid of numbers . Puzzle : # Examples Example 1: [[2 , 2, 2,... | https://arxiv.org/abs/2505.19914v1 |
2, 0, 2, 0, 2, 0, 2, 2, 2, 0, 2, 0, 2], [2, 2, 2, 0, 2, 2, 2, 0, 2, 2, 2, 0, 2, 2, 2], [2, 2, 2, 0, 2, 0, 2, 0, 2, 0, 2, 0, 2, 0, 2], [2, 2, 2, 0, 2, 2, 2, 0, 2, 2, 2, 0, 2, 2, 2], [2, 0, 2, 0, 2, 2, 2, 0, 2, 0, 2, 0, 2, 0, 2], [2, 2, 2, 0, 2, 2, 2, 0, 2, 2, 2, 0, 2, 2, 2], [2, 2, 2, 0, 2, 0, 2, 0, 2, 2, 2, 0, 2, 0, 2]... | https://arxiv.org/abs/2505.19914v1 |
the encryption method works . 2. Decrypt the provided ciphertext according to the rules , and derive the correct plaintext . Output Format : Please output your answer within a code block (‘‘‘) as follows : ‘‘‘ <result > ‘‘‘ - <result > should be the decrypted plaintext corresponding to the given ciphertext , for exampl... | https://arxiv.org/abs/2505.19914v1 |
paint waste and production time . Rules : 1. There are N cars numbered from 1 to N that need to be painted . 2. Each car has a predetermined color ( labeled as A, B, C, etc ., for a total of M colors ). 3. Cars enter the painting workshop in a fixed order , but can be rearranged within a range . 4. Each car can be move... | https://arxiv.org/abs/2505.19914v1 |
: - [1, 2, 3]: Directly push all elements into the stack and then pop them in order . - [2, 1, 3]: Push 1 and 2 into the stack , pop 2, then pop 1, and finally push and pop 3. - [3, 2, 1]: Push all elements into the stack , then pop them in reverse order . Output Format : - Please output your answer within a code block... | https://arxiv.org/abs/2505.19914v1 |
output your answer within a code block (‘‘‘) as follows : ‘‘‘ <result > ‘‘‘ - If there is an answer , the is the sequence of moves , for example : ‘‘‘ LRURDL ‘‘‘ - If there is no answer , the is: ‘‘‘ No feasible move path exists . ‘‘‘ Puzzle : ‘‘‘ 4 1 5 2 6 8 3 7 0 ‘‘‘ Listing 18 Case of Fifteen Puzzle Task : The... | https://arxiv.org/abs/2505.19914v1 |
number and B is the number of steps . Similarly , column moves are represented as CAB , where A is the column number and B is the number of steps . 3. Goal : - The ultimate goal is to arrange the tiles in order from left to right , top to bottom as follows : 1 2 3 4 5 6 7 8 9 Output Format : - If a solution exists , ou... | https://arxiv.org/abs/2505.19914v1 |
an even number ), with each cell containing a number . 2. Your goal is to " black out " certain cells , following these rules : - In each row and column , the same number cannot appear more than once . To eliminate repetitions , you must black out some of the cells . - Black cells cannot be adjacent , either horizontal... | https://arxiv.org/abs/2505.19914v1 |
‘‘‘ - If no solution exists , output within the code block : ‘‘‘ "No valid solution exists for the given Kakurasu puzzle ." ‘‘‘ Puzzle : ‘‘‘ [[5 , 3, 1, 3, 5], [1, 4, 3, 5, 3], [1, 3, 2, 4, 3], [3, 5, 1, 1, 2], [3, 1, 4, 2, 2]] ‘‘‘ Listing 23 Case of Light Up You need to solve a Light Up puzzle . Task & Rules : 1... | https://arxiv.org/abs/2505.19914v1 |
, a cell with 1 means one mine adjacent , and so on. Task : Based on current grid state , infer positions of the mines that can be definitively determined . Puzzle : The current grid state is: [[ -2 , 1, -2, -2, -2, -2, -2, -2, 0], [-2, -2, -2, 0, 0, -2, 0, -2, 0], [-2, 2, -2, -2, -2, -2, -2, -2, -2], [-2, 2, -2, -2, -... | https://arxiv.org/abs/2505.19914v1 |
‘‘‘ 1 1 1 1 -1 1 -1 1 1 -1 -1 1 ‘‘‘ - 1 represents "/" ( forward slash , top - left to bottom - right ) - -1 represents "\" ( backslash , top - right to bottom - left ) Listing 26 Case of Checkmate in One Task : You are tasked with finding a move in the chess position resulting in checkmate : Output Format : - Pl... | https://arxiv.org/abs/2505.19914v1 |
You are given 5 integers and a target number . Your task is to create an arithmetic expression that results in exactly the target number . Task & Rules : - You must use ALL the given numbers , each exactly once . - The operators you can use include : addition (+) , subtraction (-) , multiplication (*) , and division ... | https://arxiv.org/abs/2505.19914v1 |
represented by two space - separated integers ‘u‘ and ‘v‘, which indicate that there is an undirected edge between vertex ‘u‘ and vertex ‘v ‘. - The vertices are numbered from ‘0‘ to ‘N -1‘ Output Format : - Please output your answer within a code block (‘‘‘) as follows : ‘‘‘ <result > ‘‘‘ - <result > should be a list ... | https://arxiv.org/abs/2505.19914v1 |
meters long from house D to cinema C. Question : From the start point , how to reach the nearest bank ? Please provide the solution according to the requirements above . Listing 33 Case of Maze You are tasked with solving a Maze Puzzle . Puzzle : Given a 5x5 maze map , as shown below : S . . B . B . . . . . . . .... | https://arxiv.org/abs/2505.19914v1 |
four inhabitants : Alice , Bill , Ted , and Mel. - Bill tells you that Mel and Ted are not the same . - Mel claims that it is false that Alice is a knave . - Ted says that Alice is a knight and Mel is a knave . - Alice tells you that only a knave would say that Ted is a knave . Can you determine who is a knight and who... | https://arxiv.org/abs/2505.19914v1 |
your solution is consistent with all the clues and background information provided . Your response should include a solution followed by the final answer in a markdown table format . Use the following structure : Assume column 1 is Year , column 2 is Wine , column 3 is Type . Output Format : - Please output your answer... | https://arxiv.org/abs/2505.19914v1 |
ALAS: Measuring Latent Speech-Text Alignment For Spoken Language Understanding In Multimodal LLMs Pooneh Mousavi1,2, Yingzhi Wang3, Mirco Ravanelli1,2, Cem Subakan4,2 1Concordia University,2Mila-Quebec AI Institute,3Centrale Sup ´elec,4Laval University Abstract Large Language Models (LLMs) are widely used in Spoken Lan... | https://arxiv.org/abs/2505.19937v1 |
automatic speech recognition or audio captioning datasets. This is followed by fine-tuning for specific downstream tasks. However, there is no standardized metric to evaluate how well an SLU model learns speech-text alignment. Measuring alignment quality is critical for understanding whether an LLM effectively captures... | https://arxiv.org/abs/2505.19937v1 |
on the provided transcription. Since alignment focuses on modality-specific input, we apply the same trimming process to remove any appended instruc- tions at the beginning or end of the transcription. To ensure that both audio and text inputs contain similar task-relevant information, we filter out input pairs where t... | https://arxiv.org/abs/2505.19937v1 |
After obtaining the similarity score matrix, Sl∈[0,1]T×A, we obtain the alignment path between the two modalities using the monotonic-alignment-search algorithm [32], where TandA respectively denote the total number of time steps respectively in the text and audio modalities. In order then to assess the quality of the ... | https://arxiv.org/abs/2505.19937v1 |
axes represent the corresponding ALAS score. 3.1. Qualitative Analysis In Figure 2, we qualitatively compare the cross-modal sim- ilarity patterns obtained with Qwen2-Audio-Instruct-7B (left) and Salmonn-7B (right), for the LibriSQA dataset (SQA task). We showcase the similarity Slfor layers 0, 5, 10, 32, where layer 0... | https://arxiv.org/abs/2505.19937v1 |
reasonable since Emotion Recognition is not necessarily a semantic task. For certain recordings the speakers express their emotions through speech patterns and prosody, and not neces- sarily through the semantic content of the audio. This observa- tion therefore suggests that ALAS is able to capture the reduc- tion in ... | https://arxiv.org/abs/2505.19937v1 |
https://crfm. stanford. edu/2023/03/13/alpaca. html . [9] B. Li, Y . Zhang, L. Chen, J. Wang, F. Pu, J. Yang, C. Li, and Z. Liu, “MIMIC-IT: Multi-Modal In-Context Instruction Tuning,” arXiv e-prints , p. arXiv:2306.05425, Jun. 2023. [10] H. Zhang, X. Li, and L. Bing, “Video-LLaMA: An instruction- tuned audio-visual lan... | https://arxiv.org/abs/2505.19937v1 |
S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: labelling unsegmented se- quence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning , 2006, pp. 369–376. [24] T. N. Sainath, R. Pang, D. Rybach, B. Garc ´ıa, and T. Strohman... | https://arxiv.org/abs/2505.19937v1 |
SYSTEM : You are a helpful assistant. USER: Audio: <|audio_bos|><|SpeechHere|><|audio_eos|> Audio Transcription: TASK: <Question>? Answer in 10-15 words. SYSTEM : You are a helpful assistant. USER: Audio: <|audio_bos|>audio_eos|> Audio Transcription: <Transcription> TASK:<Question>? Answer in 10-15 words. QWEN | ER SYS... | https://arxiv.org/abs/2505.19937v1 |
arXiv:2505.19944v1 [cs.CV] 26 May 2025Can Visual Encoder Learn to See Arrows? Naoyuki Terashita1*Yusuke Tozaki1,2†Hideaki Omote1,3†Congkha Nguyen1 Ryosuke Nakamoto1Yuta Koreeda1Hiroaki Ozaki1 1Hitachi, Ltd.2Kyoto Sangyo University3Gifu University Abstract The diagram is a visual representation of a relationship il- lus... | https://arxiv.org/abs/2505.19944v1 |
study experimentally demonstrates that eliminating textual and positional biases during training enables visual models to learn edge features. To this end, we artificially generate a dataset of diagram images and text captions designed so that the presence and direction of edges cannot be inferred from text or position... | https://arxiv.org/abs/2505.19944v1 |
to recognize lines and arrows purely from visual inputs. 3. Learning with Debiased Diagram Dataset As shown in Fig. 2, our approach consists of artificially gen- erating diagram–text pairs that exclude text and positional biases (Sec. 3.1), followed by contrastive learning to fine- tune CLIP (Sec. 3.2).3.1. Diagram Dat... | https://arxiv.org/abs/2505.19944v1 |
4.1), image retrieval (Sec. 4.2), and diagram caption- ing (Sec. 4.3). 4.1. Linear Probing Linear probing [1, 7] measures how well the extracted fea- tures encode information through classification tasks on features. In this study, we train and evaluate a simple logis- tic regression on top of the frozen image encoder ... | https://arxiv.org/abs/2505.19944v1 |
only when the 3 Table 1. Performance comparison of pretrained and finetuned CLIP models on diagram understanding tasks. MethodLinear Probing (Mean Accuracy) Image Retrieval Node existence Edge existence Edge direction MAP@100 MRR@100 Random 0.500 0.500 0.500 0.0004 0.001 Pretrained ViT-B/32 0.959 0.639 0.518 0.067 0.10... | https://arxiv.org/abs/2505.19944v1 |
as unlabeled and undirected). De- spite a slight performance degradation, our finetuned mod- els still significantly outperformed the baselines, showing their strong generalization to unseen graph structures. 4 5. Conclusion We showed that removing textual and positional biases en- ables VLMs to learn edge recognition ... | https://arxiv.org/abs/2505.19944v1 |
Shen, and Yong Jae Lee. LLaV A-NeXT: Im- proved reasoning, OCR, and world knowledge. https: //llava- vl.github.io/blog/2024- 01- 30- llava-next/ , 2024. 4 [13] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. MathVista: Evaluatin... | https://arxiv.org/abs/2505.19944v1 |
arXiv:2505.19954v1 [cs.LG] 26 May 2025An Explainable Diagnostic Framework for Neurodegenerative Dementias via Reinforcement-Optimized LLM Reasoning Andrew Zamai1∗Nathanaël Fijalkow1Boris Mansencal1 Laurent Simon1Eloi Navet1Pierrick Coupé1 1Univ. Bordeaux, CNRS, Bordeaux INP, LaBRI, UMR 5800, F-33400 Talence, France Abs... | https://arxiv.org/abs/2505.19954v1 |
diagnosis. Methods like [ 12,13] generate visual heatmaps over MRI scans to indicate which regions influenced the model’s prediction, but they lack semantic attribution—they neither explicitly identify the anatomical regions (e.g., "the hippocampus") nor explain their clinical relevance (e.g., how hippocampal atrophy i... | https://arxiv.org/abs/2505.19954v1 |
competitive diagnostic accuracy while providing transparent, causally grounded rationales that inform and support its diagnostic conclusions. 2 Related work Neuroimaging-based Diagnosis and Post-hoc Explainability Limitations. Deep learning mod- els—particularly CNNs [ 6–8] and, more recently, ViTs [ 9,10]—have achieve... | https://arxiv.org/abs/2505.19954v1 |
through a series of interpretable intermediate steps. Unlike end-to-end vision- to-text models [ 15–17], our approach preserves the clinical detail essential for the diagnosis of neurodegenerative diseases and offers transparency into each intermediate output. The pipeline consists of four main stages: (1) fine-grained... | https://arxiv.org/abs/2505.19954v1 |
on the left of Figure 3, the severity thresholds can be visualized on a Gaussian distribution representing the normative population data, with vertical demarcation lines indicating the boundaries between severity categories. The central region of the curve represents volumes within normal limits, while progressively le... | https://arxiv.org/abs/2505.19954v1 |
by majority vote on the top-ranked differential diagnosis from all aggregated samples, with a supporting reasoning randomly selected from those aligning with the consensus. 3.3 Incentivizing diagnostic reasoning with GRPO In the absence of labeled reasoning traces for supervised training (SFT), GRPO [ 19,18] has been s... | https://arxiv.org/abs/2505.19954v1 |
pretrained knowledge and typical report distributions; (2) the impact of GRPO fine-tuning on the emergence of diagnostic reasoning; and (3) how our framework compares to established deep learning classifiers trained directly on brain MRI data. 6 4.1 Experimental setup Datasets. Table 1 summarizes the diagnostic distrib... | https://arxiv.org/abs/2505.19954v1 |
to GRPO training, we followed established configurations from [ 49] and [ 18]. We set the maximum generation length to 3000 tokens and generated G= 6 completions per query, with a sampling temperature of 0.9 to encourage exploration. In accordance with [ 49], we adopted a max- completion-length averaging scheme for the... | https://arxiv.org/abs/2505.19954v1 |
GRPO † 8B 85.86 53.33 73.17 41.67 71.43 67.33 65.09 Qwen-2.5-Instruct GRPO † 7B 78.36 49.44 58.06 0.00 42.42 52.89 45.66 Qwen-3 GRPO † 8B 83.33 46.88 66.67 48.00 64.52 68.38 61.88 GRPO Fine-Tuned LLMs. We extend Table 2 by reporting the diagnostic performance of the 8B models fine-tuned via GRPO, as detailed in Subsect... | https://arxiv.org/abs/2505.19954v1 |
62.84 SVM atrophies - 86.43 69.66 73.91 0.00 84.21 62.39 62.85 LLMs providing diagnostic reasoning gpt-4o - 70.81 51.76 51.52 19.05 48.48 55.46 48.32 DeepSeek-R1-Distill-Llama GRPO † 8B 84.16 51.43 70.59 11.76 80.00 62.48 59.55 LLaMA-3.1-Instruct GRPO † 8B 85.86 53.33 73.17 41.67 71.43 67.33 65.09 Qwen-3 GRPO † 8B 83.3... | https://arxiv.org/abs/2505.19954v1 |
Neuroscience , 14:626154, 2021. doi: 10.3389/fnins.2020.626154. [7]Da Ma, Donghuan Lu, Karteek Popuri, Lei Wang, and Mirza Faisal Beg. Differential diagnosis of frontotemporal dementia, alzheimer’s disease, and normal aging using a multi-scale multi- type feature generative adversarial deep neural network on structural... | https://arxiv.org/abs/2505.19954v1 |
Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junji... | https://arxiv.org/abs/2505.19954v1 |
Yun Liu, Alvin Rajkomar, Joelle Barral, Christopher Semturs, Alan Karthikesalingam, and Vivek Natarajan. Large lan- guage models encode clinical knowledge. Nature , 620(7972):172–180, August 2023. ISSN 1476-4687. doi: 10.1038/s41586-023-06291-2. [22] Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, M... | https://arxiv.org/abs/2505.19954v1 |
for Biomedical Natural Language Processing, September 2021. [31] Michael Moor, Qian Huang, Shirley Wu, Michihiro Yasunaga, Yash Dalmia, Jure Leskovec, Cyril Zakka, Eduardo Pontes Reis, and Pranav Rajpurkar. Med-Flamingo: A Multimodal Medical Few-shot Learner. In Proceedings of the 3rd Machine Learning for Health Sympos... | https://arxiv.org/abs/2505.19954v1 |
de Sen- neville, Vinh-Thong Ta, Vincent Lepetit, and José V . Manjon. Assemblynet: A large en- semble of cnns for 3d whole brain mri segmentation. NeuroImage , 219:117026, 2020. ISSN 1053-8119. doi: https://doi.org/10.1016/j.neuroimage.2020.117026. URL https: //www.sciencedirect.com/science/article/pii/S105381192030512... | https://arxiv.org/abs/2505.19954v1 |
David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mély, Ashvin Nair, Reiichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O’Keefe, Jakub Pachocki, Alex ... | https://arxiv.org/abs/2505.19954v1 |
Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mah... | https://arxiv.org/abs/2505.19954v1 |
Igor Molybog, Igor Tufanov, Ilias Leontiadis, Irina-Elena Veliche, Itai Gat, Jake Weissman, James Geboski, James Kohli, Janice Lam, Japhet Asher, Jean-Baptiste Gaya, Jeff Marcus, Jeff Tang, Jennifer Chan, Jenny Zhen, Jeremy Reizenstein, Jeremy Teboul, Jessica Zhong, Jian Jin, Jingyi Yang, Joe Cummings, Jon Carvill, Jon... | https://arxiv.org/abs/2505.19954v1 |
Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report, 2025. URL https://arxiv.org/abs/2412.15115 . [46] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Ha... | https://arxiv.org/abs/2505.19954v1 |
diagnosis, demonstrating its effectiveness in supporting the subsequent differential diagnosis task. 5https://radiopaedia.org/cases/frontotemporal-dementia-behavioural-variant-2 17 Figure 6: Comparison of expert-written (top) and synthetic (bottom) radiology reports for the same T1-weighted MRI scan. Colored highlights... | https://arxiv.org/abs/2505.19954v1 |
3D ViT architecture and training pipeline from Nguyen et al. [10], originally designed for AD–FTD–CN classification. To support our extended diagnostic setting, we modify the final classification head to a 5-way MLP, enabling prediction across CN, AD, bvFTD, svPPA, and nfvPPA. All other architectural components, data p... | https://arxiv.org/abs/2505.19954v1 |
arXiv:2505.19955v1 [cs.LG] 26 May 2025MLR-Bench: Evaluating AI Agents on Open-Ended Machine Learning Research Hui Chen♦∗Miao Xiong♦∗Yujie Lu♡Wei Han♠Ailin Deng♦ Yufei He♦Jiaying Wu♦Yibo Li♦Yue Liu♦Bryan Hooi♦ ♢National University of Singapore♡University of California, Santa Barbara ♠Singapore University of Technology a... | https://arxiv.org/abs/2505.19955v1 |
LLM judges to automatically assess performance over 201 tasks. For end-to-end evaluation, we use the same model as backbone in idea generation, proposal generation and paper writing. For stepwise evaluation, various models are tested and compared within each step. However, realizing this vision raises an important ques... | https://arxiv.org/abs/2505.19955v1 |
and true scientific rigor. Our contributions include: •MLR-Bench : To our knowledge, the most comprehensive evaluation benchmark for AI research agents to date, featuring 201 open-ended ML research tasks, a human-aligned MLR- Judge for automated research quality assessment, and a modular MLR-Agent supporting both stepw... | https://arxiv.org/abs/2505.19955v1 |
the same 201 tasks used in end-to-end evaluation. In step (2) the Proposal Generation agents require tasks and research ideas as input. To construct this dataset, for each of the 201 tasks, we randomly sample one idea (generated 3 LLM/VLM AI for ScienceML Theory Trustworthy AIComputer VisionML SystemsMultimodalityRL Ot... | https://arxiv.org/abs/2505.19955v1 |
1, namely (1) Idea Generation , (2) Proposal Generation , (3) Experimentation , and (4) Paper Writing . MLR-Agent uses LLMs for steps (1) and (2), a coding agent (Claude Code ) for step (3), and multimodal LLMs for step (4). Additionally, since most frontier models currently lack web search capabilities, they face chal... | https://arxiv.org/abs/2505.19955v1 |
.96±0.33 Qwen3-235B-A22B 9.20±0.28 8 .20±0.32 7.62±0.42 6 .67±0.52 8.73±0.31 8 .03±0.29 o4-mini-high 9.23±0.28 8 .23±0.30 7 .49±0.41 7 .01±0.53 8 .66±0.33 8.11±0.30 Gemini-2.5-Pro-Preview 9.20±0.31 8.27±0.31 7 .30±0.37 7.11±0.57 8 .58±0.35 8 .08±0.28 3.2 Results of Proposal Generation Evaluation All models excel at gen... | https://arxiv.org/abs/2505.19955v1 |
terse style affecting its clarity. However, none of the models achieve high overall scores (above 7.0), which may be influenced by the weaker experimental results from the previous step. This demonstrates that experimental success is a critical determinant of overall research quality . Table 5: Evaluation results of th... | https://arxiv.org/abs/2505.19955v1 |
compute (1) absolute score differences between the LLM judge and human reviewers, and (2) absolute score differences between pairs of human reviewers, and test whether these two distributions are statistically different. As shown in Fig. 4, none of the five evaluation criteria show statistically significant differences... | https://arxiv.org/abs/2505.19955v1 |
MLR-Judge leveraged access to the supplementary code and detected even more hallucination cases by inspecting the execution logs and code traces of the coding agent, consistently assigned low scores on Soundness ,Insightfulness , Significance , and Overall dimensions, as illustrated in Figure 3. To further understand t... | https://arxiv.org/abs/2505.19955v1 |
that LLMs can perform specific reviewing sub-tasks, but tend to produce overconfident and unreliable overall judgments [ 17,11]. More structured efforts, such as PaperBench [ 27], demonstrate that fine-tuned LLM judges can align well with human reviewers on benchmark tasks [ 27]. Building on these insights, MLR-Bench i... | https://arxiv.org/abs/2505.19955v1 |
Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, Aleksander Madry, and Lilian Weng. MLE-bench: Evaluating machine learning agents on machine learning engineering. In The Thirteenth International Conference on Learning Representations , 2025. [... | https://arxiv.org/abs/2505.19955v1 |
Pao Siangliulue, and Jonathan Bragg. Let’s get to the point: Llm-supported planning, drafting, and revising of research-paper blog posts. arXiv preprint arXiv:2406.10370 , 2024. [24] Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Zicheng Liu, and Emad Barsoum. Agent laboratory: ... | https://arxiv.org/abs/2505.19955v1 |
ai scientist-v2: Workshop-level automated scientific discovery via agentic tree search. arXiv preprint arXiv:2504.08066 , 2025. [38] John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik R Narasimhan, and Ofir Press. SWE-agent: Agent-computer interfaces enable automated software engineering.... | https://arxiv.org/abs/2505.19955v1 |
V2 [ 37], another agent scaffold with the capability of workshop-level end-to-end automated research, on MLR-Bench and compare its performance with MLR-Agent. We use o4-mini as the backbone model in the end-to-end automated research of AI Scientist V2. For MLR-Agent, we use o4-mini in idea generation, proposal generati... | https://arxiv.org/abs/2505.19955v1 |
Table 11: Evaluation results judged by Gemini-2.5-Pro-Preview-03-25 in proposal generation. Model Consistency Clarity Novelty Soundness Feasibility Significance Overall Ministral-8B 8.94±0.33 7.32±0.90 6.72±1.30 6.46±1.13 6.55±0.93 8.78±0.48 7.21±0.90 Deepseek-R1 9.04±0.26 8 .39±0.58 7 .44±0.83 7 .68±0.72 7 .05±0.77 8 ... | https://arxiv.org/abs/2505.19955v1 |
19 presents the prompt used for generating the research proposal. • Table 20 presents the prompt used for conducting experiments and obtaining results. 15 •Table 21 presents the prompt used for writing the research paper based on previously generated outputs, including the idea, proposal, and experimental results. Tabl... | https://arxiv.org/abs/2505.19955v1 |
its own uncertainty through generated alternative responses and explanations; 2) A factual consistency checker that retrieves information from trusted knowledge sources to verify claims; and 3) A human-in-the-loop feedback system that learns from user corrections. Uniquely, TrustPath presents detected errors alongside ... | https://arxiv.org/abs/2505.19955v1 |
We provide the research task description, the previously generated research idea, and the literature review as inputs, which are combined with this prompt to guide the language model (e.g., Claude) in generating a complete research proposal. Prompt You are an excellent machine learning researcher! Please generate a det... | https://arxiv.org/abs/2505.19955v1 |
to explain how to run the experiment. 4. Run the experiment automatically, visualize the results, and save the results. Make sure the generated figure files are not empty. 5. Save the experiment execution process in log.txt. 6. Analyze and summarize the experiment results in results.md after all experiments (including ... | https://arxiv.org/abs/2505.19955v1 |
of the experiment and suggestions for future work. Table 21: Prompt used for paper writing. We compile all outputs from the previous stages and combine them with the following prompt to instruct the LLM to generate a complete research paper. Prompt You are an excellent machine learning researcher! Given the task, resea... | https://arxiv.org/abs/2505.19955v1 |
holistic paper evaluation is shown in Table 26. Table 22: Rubrics used for idea generation. 1.CONSISTENCY (1-10) How well does the idea align with the requirements of the task description? 9-10 - Excellent: The idea is perfectly aligned with the task description. It addresses all aspects of the task and is highly relev... | https://arxiv.org/abs/2505.19955v1 |
in the field. 7-8 - Good: The idea is significant and has clear impact potential. It addresses an important issue and could lead to meaningful contributions. 5-6 - Satisfactory: The idea is somewhat significant. It addresses a relevant problem but its impact may be moderate or limited to a specific area. 3-4 - Needs Im... | https://arxiv.org/abs/2505.19955v1 |
follow. 7-8 - Good: The proposal is mostly clear and well-articulated, with only minor ambiguities or areas that could benefit from slight refinement. The main points are understandable and the structure is generally logical. 5-6 - Satisfactory: The proposal is partially clear but has some ambiguities or unclear sectio... | https://arxiv.org/abs/2505.19955v1 |
proposal? 23 9-10 - Excellent: The proposal is highly practical and implementable with current resources, technology, and knowledge. The plan is realistic, and execution is straightforward with clearly defined steps and minimal risk. 7-8 - Good: The proposal is largely feasible with existing technology and methods, tho... | https://arxiv.org/abs/2505.19955v1 |
without substantial reworking. When assigning the Overall Assessment score, consider not just the average of the six dimensions, but also: - Whether any single weakness is critical enough to lower the overall potential. - The overall coherence and integration of the proposal. - The likelihood of real-world impact if th... | https://arxiv.org/abs/2505.19955v1 |
Is the experimental design innovative or derivative? 9-10 - Excellent: The experimental document presents highly novel findings, methods, or insights that significantly advance the field. The design is innovative and original. 7-8 - Good: The experimental document shows some novel aspects, with a good level of innovati... | https://arxiv.org/abs/2505.19955v1 |
and generally well-conceived, with a good balance of strengths and weaknesses. While there may be some areas for improvement, the work is credible, meaningful, and meets the main requirements for quality research. 4-5 - Satisfactory: The work is adequate but has several notable weaknesses that limit its overall quality... | https://arxiv.org/abs/2505.19955v1 |
missing components or sections that should be included? 9-10 - Excellent: The paper is very complete, addressing all components of the task description, research idea, research proposal and experimental results. 7-8 - Good: The paper is mostly complete, with minor missing components or sections. 5-6 - Fair: The paper h... | https://arxiv.org/abs/2505.19955v1 |
- Any unique contributions, innovative aspects, or serious flaws that significantly affect the overall quality. Your overall assessment should reflect a holistic judgment, taking into account both the quantitative scores and your qualitative evaluation of the paper. Table 26: Prompt used for end-to-end evaluation. 1.Cl... | https://arxiv.org/abs/2505.19955v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.