text string | source string |
|---|---|
Deep Learning Library. Advances in Neural Information Processing Systems , 32:8024– 8035. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI blog , 1(8):9. Ramanujan, V ., Nguyen, T., Oh, S., Farhadi, A., and Schmidt, L. (2023). On... | https://arxiv.org/abs/2505.19893v1 |
S., Ma, T., and Liang, P. S. (2023b). Data selection for language models via importance resampling. Advances in Neural Information Processing Systems , 36:34201–34227. Yu, Z., Das, S., and Xiong, C. (2024). MATES: Model-aware data selection for efficient pretraining with data influence models. Advances in Neural Inform... | https://arxiv.org/abs/2505.19893v1 |
powerful genera- tive models, some of which could be misused for disinformation, synthetic media, or other harmful applications. In addition, token-level filtering methods—if miscalibrated—may reinforce spurious patterns or underrepresent minority language phenomena, inadvertently encoding or amplifying soci- etal bias... | https://arxiv.org/abs/2505.19893v1 |
. , x M} ∼ D . 7: Compute per-token risk scores Sθk(xj). */ Entropy or loss depending on the selection type 8: Compute threshold SVaR θk,α←VaR α {Sθk(xj)}M j=1 using (3). 9: ˜B ← { xj∈ B | Sθk(xj)≥SVaR θk,α}. */ High-risk token selection 10: Compute loss over selected tokens: L˜B(x;θk) =Exj∈˜B[ℓθk(xj)]. */ Shaped los... | https://arxiv.org/abs/2505.19893v1 |
provide the implementation for our knowledge distillation setup, namely ESLM-KD, in Algo- rithm 3. The student model θcomputes per-token risk scores over each batch, and high-risk tokens are selected via VaR αthresholding. The student is then supervised only on these informative tokens using a combined loss: a weighted... | https://arxiv.org/abs/2505.19893v1 |
0.68(0.0469) 0.66(0.0476) 0.69(0.0465) 0.67(0.0473) MultiRC (Wang et al., 2019) 5-shot 0.5408 (0.0072) 0.5406 (0.0072) 0.5360 (0.0072) 0.5420 (0.0072) OpenBookQA (Mihaylov et al., 2018) 5-shot 0.270(0.0199) 0.290(0.0203) 0.278(0.0201) 0.278(0.02) PiQA (Bisk et al., 2020) 5-shot 0.6300 (0.0113) 0.6245 (0.0113) 0.6327 (0... | https://arxiv.org/abs/2505.19893v1 |
ESLM-VaR -entropy 13.53 ESLM-CVaR -loss 13.50 CLM 9.32 Rho-1 24.52 GREATS 99.89Runtime analysis. In Table 7, we compare the wall-clock time of 124M models trained on the SlimPajama-6B mixture under a ∼3E17 FLOPs bud- get. While ESLM achieves substantial reductions in training FLOPs, reaches lower validation loss, and s... | https://arxiv.org/abs/2505.19893v1 |
We evaluate pretrained models on a suite of standard language understanding benchmarks in the zero-shot and few-shot settings, using the lm-evaluation-harness evaluation suite (Gao et al., 2024), including HellaSwag (Zellers et al., 2019), LAMBADA (Paperno et al., 2016), ARC-Easy (Clark et al., 2018), TriviaQA (Joshi e... | https://arxiv.org/abs/2505.19893v1 |
with fewer FLOPs, consistently providing efficiency gains as the model scales. 0.10 1.00 Log(FLOPs)×10170.400.450.500.550.600.650.70Log(Loss) Validation Loss vs. Training FLOPs 1.0 1.2 1.4 1.6 1.8 2.0 2.2 1e160.500.520.540.560.580.60 1.0 1.2 1.4 1.6 1.8 2.0 2.2 1e160.500.520.540.560.580.60 1.0 1.2 1.4 1.6 1.8 2.0 2.2 1... | https://arxiv.org/abs/2505.19893v1 |
the best observed accuracy (standard error ) or exact match if provided, during training. Highlighted values indicate the best performance. BenchmarkMethod (350M) # Shots ESLM-CVaR -loss E SLM-VaR -entropy CLM Rho-1 ARC-E (Clark et al., 2018) 0-shot 0.4078 (0.0101) 0.3973 (0.01) 0.4023 (0.0101) 0.4006 (0.0101) LAMBADA ... | https://arxiv.org/abs/2505.19893v1 |
FLOPs budget. We report the best observed accuracy (standard error ) or exact match if provided, during training. Highlighted values indicate the best performance. BenchmarkMethod (774M) # Shots ESLM-CVaR -loss E SLM-VaR -entropy CLM Rho-1 ARC-E (Clark et al., 2018) 0-shot 0.4132 (0.0101) 0.4158 (0.0101) 0.4128 (0.0101... | https://arxiv.org/abs/2505.19893v1 |
2.75 3.00 3.25 1e160.480.500.520.540.56 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25 1e160.480.500.520.540.56 1.25 1.50 1.75 2.00 2.25 2.50 2.75 3.00 3.25 1e160.480.500.520.540.56 (a) Validation loss vs FLOPs. Models0.320.330.340.350.360.370.380.390.40Avg AccuracyAVG Downstream Performance ESLM_CVaR-loss ESLM_VaR-entro... | https://arxiv.org/abs/2505.19893v1 |
you it are with 3 for that I Tokens020406080100120Token FrequencyTop-k Token Frequencies at =0.2 Total Selected (b) E SLM-CVaR -loss, α= 0.2. . , \n and the of to a in - is " you it are with 3 for I that Tokens020406080100120Token FrequencyTop-k Token Frequencies at =0.1 Total Selected (c) E SLM-VaR -entropy, α= 0.1. .... | https://arxiv.org/abs/2505.19893v1 |
de M ello , Wil ma de Souza , Drama The Citadel ( 19 38 ) King Vid or , Robert Don at , Ros al ind Russell , Ralph Richardson L adies in Retirement ( 19 41 ) Charles Vid or , Id a Lup ino , Louis Hayward , Eve lyn Key es Mad ame Sat ã ( 2002 ) Kar im A ï n ou z , L á z aro Ramos , Marcel ia Cart ax o , Flav io... | https://arxiv.org/abs/2505.19893v1 |
of Refugees Could Have Global Dom ino Effect AMY GOODMAN : As we turn now to Washington , D. C ., as House Speaker Paul Ryan and Senate Majority Leader Mitch McConnell are calling for a pause in the U.S. program accepting Syrian refugees , I want to bring into the conversation Congress member Barbara Lee of California ... | https://arxiv.org/abs/2505.19893v1 |
aken ed . Someone had slipped off my jeans and shoes and laid a blanket over me . Sh aft s of moon light poured through the windows . The fire was a mound of ashes in the stove . The other two beds were not occupied . I was alone . I put on my shoes and jeans and went outside . The moon was gigantic . From the directio... | https://arxiv.org/abs/2505.19893v1 |
, yelling , " Holy water !" " Holy water on its way !" " Here comes the holy water !"Figure 12: Example inputs from SlimPajama-6B-Unif mixture showing the selected tokens by ESLM-VaR -entropy (124M , checkpoint 30000) with α= 0.1. [Note: These examples are drawn from public datasets (Soboleva et al., 2023) and may cont... | https://arxiv.org/abs/2505.19893v1 |
( 2002 ) Kar im A ï n ou z , L á z aro Ramos , Marcel ia Cart ax o , Flav io B aura qu i , Bi ography , Crime , Drama Y outh in Fury ( 1960 ) Mas ah iro Shin oda , Shin ' ich ir ô Mik ami , Sh ima I wash ita , Kay oko Hon oo Night Plane from Chung king ( 19 43 ) Ralph Murphy , Robert Preston , Ellen Drew , Ott... | https://arxiv.org/abs/2505.19893v1 |
of California . Your response to the crackdown ? Now , 27 governors are saying they will not accept Syrian refugees . In fact , your theory , Peter Bou ck a ert , around Example 2 from domain: book out . Well , they weren 't getting away with it . There had to be a confrontation . I got to my feet and marched toward th... | https://arxiv.org/abs/2505.19893v1 |
shoes and jeans and went outside . The moon was gigantic . From the direction of the mine I heard Frank G ag l iano 's drunken gravel laughter , then the voice of Rhod a Pruitt , then a roar from my father . I told myself not to go up there , to stay in the cabin , to leave them alone , but I would not listen to myself... | https://arxiv.org/abs/2505.19893v1 |
contain intense language, political references, or mature content. These excerpts are included solely for the purpose of analyzing model behavior. ] 27 Example 1 from domain: cc hydro chlor ic acid in methyl ene chloride - water , followed by separation of the organic phase , drying , and storage in solution at 0 - 5 Ã... | https://arxiv.org/abs/2505.19893v1 |
, 15 alk anes . C 5 - C 8 cycl oalk anes and their mono - and dim ethyl derivatives are am inated in good yields . 13 M ethyl cycl o hex ane 12 , 16 and methyl cyclop ent ane 13 are converted to 1 - am ino - 1 - methylcyclo alk anes on treatment with trich lor amine /AlCl 3 ( eq 7 ). Treatment of dec alin and hy dr ind... | https://arxiv.org/abs/2505.19893v1 |
others is a blight of our consumer - driven society , and it is felt most keen ly in cities . It is up to us to quiet the voice inside that asks why we always feel late to the party . The truth is that there will always be so much more happening in a city than you can ever spread yourself across , in person or even in ... | https://arxiv.org/abs/2505.19893v1 |
of the woman who was attacked on the street in broad daylight in front of many people , but no one intervened because they assumed someone else would . Should any of us find ourselves the unfortunate victim in such a situation , a good way to attract help is to shout out to someone individually , referring to them by w... | https://arxiv.org/abs/2505.19893v1 |
beach all the time ." They smiled like wise old men . " You want a joint . Dad ?" D enny said. " No , thanks ." " How about you , mother ?" It was ridiculous and he knew better . I said, " Your Mother isn 't a pot smoker , so stop being a wise guy ." " This stuff is pure gold , Dad . Sure you won 't try it ?" " No , th... | https://arxiv.org/abs/2505.19893v1 |
" Okay ." Maybe the pot did it . Maybe it was a break - through of his anger , the hot night and the curious circumstance that had brought us together at that moment . Maybe he had wanted to say it for years , but the right mood and moment had eluded him , but now he said it , and it sounded like a carefully prepared s... | https://arxiv.org/abs/2505.19893v1 |
ume hood . A mination of A rom atics . The reaction of benz ene and derivatives with N Cl 3 and Aluminum Chlor ide in organic solv ents can be a useful preparation for meta - sub st it uted am ines . However , yields are only moderate , and m ixtures of is omers are often obtained . Aren es include mono - 7 - 9 and dia... | https://arxiv.org/abs/2505.19893v1 |
9 ). 3 , 15 D iam ant ane 17 can also be efficiently am inated in this fashion . When hydro car bons which do not contain a tert iary hydrogen are subjected to reaction with N Cl 3 /AlCl 3 , c ation ic rearr ang ements and fragment ations are observed . 18 A mination of Al ky l - Substit uted Aromatics. Various mono al... | https://arxiv.org/abs/2505.19893v1 |
_ ca ed ere_ , meaning to cut off âĢĵ literally slaying your options . It 's learning when and what to opt in and out of that really matters , though . Have confidence in your choices : make sure that they reflect who you are , and what you enjoy . Don 't succumb to peer pressure , or let yourself become a wing man ... | https://arxiv.org/abs/2505.19893v1 |
eling Clean C ities are dirty . Even the more clinical , manic ured M itte le uro pe an or Japanese cities have cars , and pollution , and inhabitants with ger ms who don 't wash their hands and occasionally snee ze on the back of your neck . For anyone even moderately concerned with hygiene , urban living is a constan... | https://arxiv.org/abs/2505.19893v1 |
for people with shrivel ed brains . You need it because you 're a mor on ." " Thanks a lot ." He crushed his cigarette into the sand , pulled off his shoes and socks , and trud ged toward the water . Harriet looked after him with soft eyes . " That wasn 't very nice ," she said . I got up and went after him . He turned... | https://arxiv.org/abs/2505.19893v1 |
front yard in North Sacramento , rolling in the dirt , kicking and gou ging and cursing until the neighbors separated us . So it was Christmas Eve again . " I think Mother writes better than you do . I 've read your novels . They 're cor ny , sentimental cop - outs , and I 'm not even talking about your screen plays ."... | https://arxiv.org/abs/2505.19893v1 |
arXiv:2505.19896v1 [cs.AI] 26 May 2025Graphical Abstract Large Language Models as Autonomous Spacecraft Operators in Kerbal Space Program Alejandro Carrasco, Victor Rodriguez-Fernandez, Richard Linares Large Language Models as Autonomous Spacecraft Operators in Kerbal Space Program Alejandro Carrascoa, Victor Rodriguez... | https://arxiv.org/abs/2505.19896v1 |
notable examples including agents trained for tasks such as sensor-tasking Siew et al. (2022) and plan- etary landing Gaudet et al. (2020). This work focuses on space applications, particularly the development of au- tonomous agents for the guidance and control of spacecraft. However, unlike other fields of AI research... | https://arxiv.org/abs/2505.19896v1 |
models through prompt-engineering and fine-tuning, which serve as viable and efficient alternatives. Fine-tuning, in particular, offers the opportunity to address specific limitations of base models, such as response latency or consistency, enabling the refinement of models to ensure they perform optimally for targeted... | https://arxiv.org/abs/2505.19896v1 |
examples provided within the input prompt. This technique enables the model to infer patterns and apply them to new, unseen data Brown et al. (2020). Fine-Tuning: Training a pre-trained LLM on a specific dataset related to a par- ticular task or domain. This adapts the LLM to specialized tasks by learning the nuances a... | https://arxiv.org/abs/2505.19896v1 |
through actuators” Russell and Norvig (2016). In KSPDG, agents control a spacecraft’s movement in all three rotational axes (yaw, pitch, and roll) using the thrust engines. Actions are expressed in the space- craft’s reference frame and include thrust magnitude for each axis and duration of the applied thrust. Although... | https://arxiv.org/abs/2505.19896v1 |
was used in some experiments. This technique involves prepending the user prompt with the last n conversations with the LLM where n is the window size using zero padding if required Beltagy et al. (2020). Instead of using special tokens, some keywords were used to distinguish between the user and the assistant during c... | https://arxiv.org/abs/2505.19896v1 |
9The specific model version used in this CoT study is gpt-3.5-turbo-0125, which was the latest model available. 11 You operate as an autonomous agent controlling a pursuit spacecraft. Your goal is to apply throttles to capture the evader given the positions and velocities of the pursuer and evader in celestial body ref... | https://arxiv.org/abs/2505.19896v1 |
gameplay log (total 647 user-assistant pairs) to analyze learning improvements with more data. GPT-4 and GPT-4o were unavailable during this study and are expected to per- form better, as demonstrated in a related study Carrasco et al. (2025). 10GPT-3.5 Turbo fine-tuning and API updates: https://openai.com/blog/ gpt-3-... | https://arxiv.org/abs/2505.19896v1 |
few-shot prompting with CoT was designed to test the baseline LLaMA 3 model capabilities. The prompt can be seen in Appendix C. 5.3. Optimization techniques TheAIresearchmaybewellbottleneckedduetotherequirementsneededtotrain any state-of-the-art model. During our LLaMA research, we utilized a workstation equipped with ... | https://arxiv.org/abs/2505.19896v1 |
will also use the metrics designed for the KSPDG Challenge Allen (2023) as well as some of our own metrics and evaluations. 6.1. Training Metrics The accuracy of the inferred actions is assessed using a discrete evaluation metric that determines whether the predicted action matches the ground truth. Specifically, the o... | https://arxiv.org/abs/2505.19896v1 |
models to the final versions, we observed a significant improvement on the order of one to two magnitudes. Table 3 shows the losses for all models. As also shown in fig. 6, models with larger datasets generally achieve better results. However, an exception is the model that uses a sliding window of 3, which, despite ha... | https://arxiv.org/abs/2505.19896v1 |
We compare these results against the agents used in SpaceGym Allen et al. (2023). Specifically, we compare them against a naive agent, a Lambert model predictive control (Lambert-MPC) agent, an iLQGames agent Fridovich-Keil et al. (2020), and a Proximal Policy Optimization (PPO) agent, as benchmarked in the SpaceGym pa... | https://arxiv.org/abs/2505.19896v1 |
LLaMA 52.69 140.68 9.09% fine-tune 10 files 30.52 51.53 0.00% fine-tune 25 files 13.54 29.44 0.00% fine-tune 50 files 11.86 29.76 0.00% fine-tune 10 files win=3 23.08 40.03 0.00% Navball (bot) 34.34 36.43 - Table 7: Performance of LLaMA for each fine-tuning technique in meters. The scenario used for these results is E3... | https://arxiv.org/abs/2505.19896v1 |
which likely limited their performance. This performance extends beyond the training distribution, highlighting a re- markable outcome: the LLaMA models outperformed the navball bot that originally generatedthetrainingdata. ThisunexpectedresultsuggeststhattheLLaMAmodels were able to leverage their prior knowledge and r... | https://arxiv.org/abs/2505.19896v1 |
In contrast, an “agnostic” prompt that does not dictate the model’s reasoning can even enhance data-driven performance. Ontheotherhand,thefew-shotpromptingtrajectoriesusingtheChainofThought technique are similar between GPT and LLaMA models. GPT performed slightly better due to its significantly lower latency. 7. Concl... | https://arxiv.org/abs/2505.19896v1 |
of the Department of the Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. Authors would like to thank Dr. Ross Allen and the rest of the team behind the development of the KSPDG challenge for t... | https://arxiv.org/abs/2505.19896v1 |
GPT agent based on fine-tuning. Figure D.11: Behavior of the developed GPT agents in Pursuer-Evader scenario E3. The plot shows the evolution of differences in the position, velocity, as well as the relative distance and velocity metrics. Figure D.12: An example of the generation of multiple orbits used to collect LLaM... | https://arxiv.org/abs/2505.19896v1 |
J., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al., 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165 . Carrasco, A., Nedungadi, M., Rodriguez-Fernandez, V., Linares, R., 2025. Visual language models as operator agents in the space do- main, in: AIAA SCITECH 2025 F... | https://arxiv.org/abs/2505.19896v1 |
P.M., Linares, R., 2024. Language models are spacecraft operators. arXiv preprint arXiv:2404.00413 URL: https://arxiv.org/abs/2404.00413 . Russell, S., Norvig, P., 2016. Artificial Intelligence: A Modern Approach. Pearson Education Limited. Siew, P.M., Jang, D., Roberts, T.G., Linares, R., 2022. Space-based sensor task... | https://arxiv.org/abs/2505.19896v1 |
SCIENCE BOARD : Evaluating Multimodal Autonomous Agents in Realistic Scientific Workflows Qiushi Sun♡♢Zhoumianze Liu♢Chang Ma♡Zichen Ding♢Fangzhi Xu♢ Zhangyue YinHaiteng Zhao⋆Zhenyu Wu♢Kanzhi Cheng♣Zhaoyang Liu♢ Qintong Li♡Jianing Wang♠Xiangru Tang/♀eafTianbao Xie♡Xiachong Feng♡ Xiang Li♠Ben Kao♡Wenhai Wang♢Biqing Qi♢L... | https://arxiv.org/abs/2505.19897v1 |
infrastructure enabling computer-using agents to assist in scientific workflows. Based on instructions, agents autonomously interact with the environment via GUI actions or generated code to complete realistic tasks. making sufficient observations, an autonomous agent could perform the same tasks within minutes. By ena... | https://arxiv.org/abs/2505.19897v1 |
computer tasks as humans do, leading to the proliferation of computer- using agents [ 11]. One line of research utilizes Command Line Interface (CLI), where agents generate executable scripts ( e.g., Python or Shell scripts) to interact with systems programmatically [ 24]. In this process, agents perform code synthesis... | https://arxiv.org/abs/2505.19897v1 |
s1, a1, . . . , s t]is determined by the policy and environment dynamics: pπ(τ) =p(s0)TY t=0π(at|g, st, mt)T(st+1|st, at) (1) Observation and Memory. We evaluate computer agents using three types of observation spaces: text-only, visual-only, and combined text-visual observations. For text-based observations, we use ac... | https://arxiv.org/abs/2505.19897v1 |
and Adaptation. For each domain, we select an open-source application that supports both visual and textual observations as the agent’s playground. To enable access to the internal state of each application within the VM, we adapt the software accordingly. Given the complexity and limited completeness of scientific app... | https://arxiv.org/abs/2505.19897v1 |
As a pioneering benchmark for scientific exploration, SCIENCE BOARD spans six domains selected for their relevance to key stages of the scientific workflow, such as simulation, modeling, prediction, and knowledge. These choices are informed by efforts on LLMs for science [ 31]. In selecting software for each domain, we... | https://arxiv.org/abs/2505.19897v1 |
color non-carbon … Step 4:Task Configuration App InstallFile DownloadStep 5:WriteEvaluationFunctiondefcompare (star, moon):passdefeval (output, target):pass Check&Validation Step3:TaskFormalizationand Verification Agentic PromptDifficultyTask1Agentic PromptDifficultyTask2Agentic PromptDifficultyTask3 ExecutionWrite Cod... | https://arxiv.org/abs/2505.19897v1 |
Struct. Analysis Data Interp. Struct. Prediction Basic Settings Spatial Reasoning Data Queries Simulation Basic ManipulationMap DisplayImage AnalysisData EditingBasic UsageLayoutFigure Editing Figure 4: Distribution of tasks in SCIENCE - BOARD benchmark. The distribution of task types is shown in Figure 4. Beyond the i... | https://arxiv.org/abs/2505.19897v1 |
far from being capable of effectively assisting human scientists in completing real-world scientific exploration tasks. Even SOTA models, such as GPT-4o andClaude , achieve an average success rate of only 15%. Across various settings, open- source counterparts can partially match proprietary models. However, they still... | https://arxiv.org/abs/2505.19897v1 |
6.38% GPT-4o 3.23% 0.00% 0.00% 0.00% 0.81% The results in Table 4 show that modular approaches yield significant improvements and are promising for tackling complex and visually demanding tasks in scientific software workflows. Vision-Only vs. Hybrid Interface. Some tasks inherently support both GUI and CLI as in- terc... | https://arxiv.org/abs/2505.19897v1 |
natural and impactful next step is to extend the capabilities of such autonomous agents, as fostered and benchmarked in SCIENCE BOARD , into physical laboratory environments. This transition involves interfacing agents with robotic systems [ 75,76], applying principles of embodied AI to perceive and interact with the p... | https://arxiv.org/abs/2505.19897v1 |
the digital world, 2025. URL https://openai.com/index/computer-using-agent . [12] Qiushi Sun, Zhirui Chen, Fangzhi Xu, Kanzhi Cheng, Chang Ma, Zhangyue Yin, Jianing Wang, Chengcheng Han, Renyu Zhu, Shuai Yuan, et al. A survey of neural code intelligence: Paradigms, advances and beyond. arXiv preprint arXiv:2403.14734 ,... | https://arxiv.org/abs/2505.19897v1 |
Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie Huang, Yuxiao Dong, and Jie Tang. Agentbench: Evaluating LLMs as agents. In The Twelfth International Conference on Learning Representations , 2024. URL https://... | https://arxiv.org/abs/2505.19897v1 |
Anastasia Krithara, Anastasios Nentidis, Konstantinos Bougiatiotis, and Georgios Paliouras. Bioasq-qa: A manually curated corpus for biomedical question answering. Scientific Data , 10 (1):170, 2023. [35] Xingyu Lu, He Cao, Zijing Liu, Shengyuan Bai, Leqing Chen, Yuan Yao, Hai-Tao Zheng, and Yu Li. MoleculeQA: A datase... | https://arxiv.org/abs/2505.19897v1 |
Zhang, editors, Advances in Neu- ral Information Processing Systems , volume 37, pages 115119–115145. Curran Associates, Inc., 2024. URL https://proceedings.neurips.cc/paper_files/paper/2024/file/ d07a9fc7da2e2ec0574c38d5f504d105-Paper-Conference.pdf . [44] Jianxiang Yu, Zichen Ding, Jiaqi Tan, Kangyang Luo, Zhenmin We... | https://arxiv.org/abs/2505.19897v1 |
and Demis Hassabis. Highly accurate protein structure prediction with alphafold. Nature , 596(7873):583–589, Aug 2021. ISSN 1476-4687. doi: 10.1038/s41586-021-03819-2. URL https://doi.org/10.1038/s41586-021-03819-2 . [56] Leonardo de Moura and Sebastian Ullrich. The lean 4 theorem prover and programming language. In Au... | https://arxiv.org/abs/2505.19897v1 |
Markus J Buehler. Sciagents: Automating scientific discovery through multi-agent intelligent graph reasoning. arXiv preprint arXiv:2409.05556 , 2024. [71] Saaket Agashe, Kyle Wong, Vincent Tu, Jiachen Yang, Ang Li, and Xin Eric Wang. Agent s2: A compositional generalist-specialist framework for computer use agents, 202... | https://arxiv.org/abs/2505.19897v1 |
presents significant challenges due to vast state / action spaces. One potential direction for improvement is to introduce VLMs to serve as judges capable of assigning partial credit and providing richer feedback. We leave this as future work. Broader Impacts. Computer-using agents operating in live OS environments cou... | https://arxiv.org/abs/2505.19897v1 |
and interactive theorem prover grounded in dependent type theory (specifically Martin-Löf Type Theory). Lean enables formal verification of mathe- matical theorems and software correctness through rigorous type checking and logical inference, supporting robust development of maintainable and accurate code. •ChimeraX. A... | https://arxiv.org/abs/2505.19897v1 |
Screenshot. We capture a screenshot of the entire computer screen. For screen resolution, we set a default value of 1920 ×1080, and it also offers a 16:9 aspect ratio. Following OSWorld [ 16], our environment also supports modifying the resolution of virtual machines to avoid potential memorization of absolute pixel va... | https://arxiv.org/abs/2505.19897v1 |
115 10 5 051015Dimension 2Algebra Astronomy Biochemistry Documentation GIS Theorem Proving Figure 6: t-SNE visualization of task instructions distribution. The seeds of t-SNE are randomly sampled for each plot. 20 B.3 Comparison with Existing Benchmarks We compare SCIENCE BOARD with existing well-established benchmarks... | https://arxiv.org/abs/2505.19897v1 |
/A", "value": ["#1/A:1 color #d2b48c", ... ] } There is a point located in the Mediterranean Sea. Please find and delete it.{ "type": "db", "cmd": "v.to.db", "kwargs": { "flags": "p", "map": "countries@PERMANENT", "type": "point", "option": "coor" }, "key": "lambda out: out.strip()", "value": "cat|x|y|z\n...|8.34894789... | https://arxiv.org/abs/2505.19897v1 |
visual reasoning capabilities. It has achieved compelling performance in complex multidisciplinary understanding and problem-solving, highlighting its specialized strength in sophisticated visual cognitive tasks. However, it exhibits some limitations in instruction following, appearing less adept in agent scenarios tha... | https://arxiv.org/abs/2505.19897v1 |
from different GUI action models. While InternVL3-78B [ 62] outputs coordinates on a [0, 1]scale, models such as OS-Atlas, UI-TARS, and UGround use a [0, 1000] scale. To ensure uniformity, we normalized all coordinate outputs to a [0, 1] scale prior to execution. This part of the experiments employs a two-stage process... | https://arxiv.org/abs/2505.19897v1 |
benchmark across different task difficulty levels. We employ Claude-3.7-Sonnet ,GPT-4o , and Qwen2.5-VL, with results presented in Figure 10. The findings indicate that solvable tasks are primarily concentrated among a subset of “Easy” problems and a few “Medium” tasks. All “hard” tasks, which involve complex computati... | https://arxiv.org/abs/2505.19897v1 |
one line or multiple lines of python code to perform the action each time, and be time efficient. When predicting multiple lines of code, make some small sleep like ‘time.sleep(0.5);‘ interval so that the machine could take breaks. Each time you need to predict a complete code, and no variables or function can be share... | https://arxiv.org/abs/2505.19897v1 |
Don’t easily say “‘FAIL“‘; try your best to do the task; When you think you have to wait for some time, return “‘WAIT“‘ or “‘WAIT n“‘, in which n defaults to 5(s); When you are asked to submit an answer, return “‘ANS s“‘ without quotation marks surrounding s, and use ‘FAIL‘ if there is no answer to the question. My com... | https://arxiv.org/abs/2505.19897v1 |
s“‘ without quotation marks surrounding s, and use ‘FAIL‘ if there is no answer to the question. My computer’s password is ’password’, feel free to use it when you need sudo rights. DO NOT introduce any unrelated models or easily close existing models, otherwise the task might be evaluated as FAILED. DO NOT close the c... | https://arxiv.org/abs/2505.19897v1 |
marks surrounding s, and use ‘FAIL‘ if there is no answer to the question. My computer’s password is ’password’, feel free to use it when you need sudo rights. The criterion for a celestial body to be displayed on the screen is that the object’s center is within the window range and is not blocked by others. First give... | https://arxiv.org/abs/2505.19897v1 |
follow my instruction and perform desktop computer tasks as instructed. You have good knowledge of Celestia, a three-dimension space simulator; and assume your code will run on a computer controlling the mouse and keyboard. For each step, you will get an observation of the desktop by a screenshot, together with a plan ... | https://arxiv.org/abs/2505.19897v1 |
flexible and adaptable according to changing circumstances. First give the current observation and the generated plan, then RETURN ME THE CODE OR SPECIAL CODE I ASKED FOR. NEVER EVER RETURN ME ANYTHING ELSE. You are asked to complete the following task: Set the Julian date to 2400000 in Celestia. Prompt 20: Prompts for... | https://arxiv.org/abs/2505.19897v1 |
arXiv:2505.19912v1 [cs.CL] 26 May 2025APE: A Data-Centric Benchmark for Efficient LLM Adaptation in Text Summarization Javier Marín javier@jmarin.info May 27, 2025 Abstract We present Adjacent Possible Exploration (APE), a simple yet effective method for adapting large language models to specific tasks using minimal co... | https://arxiv.org/abs/2505.19912v1 |
exploring nearby possibilities rather than leaping to distant states, often yielding unexpected innovations through self-organization. 2.2 Applying the Adjacent Possible to LLMs Adjacent Possible Exploration (APE) adapts TAP as a heuristic framework to guide iterative fine- tuning of large language models (LLMs). At ea... | https://arxiv.org/abs/2505.19912v1 |
a fixed schedule and risks overwriting previously learned knowledge [26], APE dynamically guides the fine-tuning process: at each iteration, the performance gain ∆S(t)is evaluated, and updates are retained only if ∆S(t)> θ, where θis a threshold derived from the expected growth rate k. This threshold ensures that updat... | https://arxiv.org/abs/2505.19912v1 |
performance S′(t)forM′using a validation set 9:Compute performance gain ∆S(t)←S′(t)−S(t−1) 10:Compute threshold θ←k·S(t−1)· 1−S(t−1) Smax ∆t, where ∆t= 1 (discretized growth rate) 11: if∆S(t)> θthen 12:Retain updates: M←M′,S(t)←S′(t) 13: else 14:Discard updates: Retain original M, setS(t)←S(t−1) 15: end if 16:end for... | https://arxiv.org/abs/2505.19912v1 |
efficient adaptation in resource-constrained settings—a persistent issue despite methods like LoRA [16] and DPO [30]. 4 Experimental Setup The APE method implementation consists of the following steps: 1.Baseline : Start with a pre-trained model (T5-base model) and evaluate its unperturbed perfor- mance on news summari... | https://arxiv.org/abs/2505.19912v1 |
are reported. 7 5 Experiment Results 5.1 Quantitative Results (Full Experiment) The full experiment (4,000 training samples, 1,000 test samples, 17 iterations) validates APE’s effectiveness in enhancing T5-base’s adaptability. Results are shown in Table 1 and Figure 2. Table 1: Quantitative results for the full experim... | https://arxiv.org/abs/2505.19912v1 |
0.365 11.0 220M LoRA (est.) 0.080 0.320 0.385 9.0 ∼0.5M Adapters (est.) 0.078 0.315 0.380 9.5 ∼1M APE (proposed) 0.083 0.329 0.398 8.3 220M To refine our comparisons in Table 2, we acknowledge that estimates for curriculum learning (CL), active learning (AL), and LoRA are derived from literature, not speculation. For i... | https://arxiv.org/abs/2505.19912v1 |
Dozens of foreigners have volunteered to fight ISIS. 6 Discussion The quantitative results (Table 1) and ablation studies confirm APE’s effectiveness, with ∆D tuning optimizing trade-offs between BLEU and accuracy, grounding APE as a practical tool for text summarization. Table 2 shows APE outperforming estimated CL/AL... | https://arxiv.org/abs/2505.19912v1 |
ability to enhance multiple facets of summary quality. Its difference from CL, AL, and LoRA/adapters (Table 2) and APE-guided ablation underscore its data-centric value. Qualitative research (Table 3) shows higher variance in human ratings (standard deviations of 0.56– 0.61) also suggests subjectivity and highlights th... | https://arxiv.org/abs/2505.19912v1 |
Press, 1993. [20] S. Kauffman. At Home in the Universe: The Search for the Laws of Self-Organization and Complexity . Oxford University Press, 1995. 12 [21] S. Kauffman. Investigations . Oxford University Press, 2000. [22] B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning.... | https://arxiv.org/abs/2505.19912v1 |
arXiv:2505.19914v1 [cs.CL] 26 May 2025 Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles 1ByteDance Seed 2Fudan University3Institute for AI Industry Research (AIR), Tsinghua University 4Nanjing University5Shanghai Jiao Tong University 6SIA-Lab of Tsinghua AIR and ByteDance S... | https://arxiv.org/abs/2505.19914v1 |
yet with either designing a prompting workflow and relying on a code interpreter [ 14], or training LLMs upon one or a few puzzles [ 15,16], which is difficult to generalize. Based on the success of the “LLM+RLVR” paradigm, it has become crucial to obtain a large, diverse, and challenging set of verifiable puzzle promp... | https://arxiv.org/abs/2505.19914v1 |
Learning with Human Feedback (RLHF), Reinforcement Learning with Verifiable Rewards (RLVR) removes the need for a reward model by directly assigning rewards based on objectively verifiable answers [ 2,5,18], which has shown strong performance in mathematics [ 19–21], STEM [ 2,22], and coding [ 5,18]. For example, mathe... | https://arxiv.org/abs/2505.19914v1 |
abilities. •Graph Puzzle involves tasks where models must reason about nodes, edges, and paths within graph struc- tures. Challenges such as Hamiltonian Path andNL Navigation test a model’s ability to understand and traverse graphs, evaluating its capacity for path-finding and network navigation. •Search Puzzle include... | https://arxiv.org/abs/2505.19914v1 |
between Enigmata and existing puzzle resources. Enigmata is the only dataset encompassing multiple task categories, offers scalability, provides automatic verification, and is publicly available. Additionally, it uniquely employs the RLVR approach to enhance models’ puzzle reasoning capabilities fundamentally. Table 1 ... | https://arxiv.org/abs/2505.19914v1 |
size |S|=P i|si|. By changing the Ni,d, we can easily adjust: 1) How many examples come from each task, 2) The mix of easy vs. hard items, 3) Overall dataset size. During training, each generated example is fed to its verifier vi, which returns a reward that VC-PPO uses to update the policy. This loop provides a fully ... | https://arxiv.org/abs/2505.19914v1 |
0.4 59.9 69.6 87.3 o1 29.0 0.4 54.9 69.9 74.3 DeepSeek-R1 17.8 0.2 49.2 71.7 79.8 Gemini-2.5-Pro 22.7 1.4 50.6 68.2 90.3 Claude-3.7-Sonnet-Thinking 37.6 1.4 53.2 67.8 60.3 DS-R1-Distilled-Qwen-32B 7.9 0.0 31.1 63.5 72.0 QwQ-32B 7.0 0.0 43.8 61.8 69.2 Grok-2-1212 7.5 0.0 13.6 54.9 16.7 GPT-4o-1120 7.3 0.0 14.2 57.9 10.0... | https://arxiv.org/abs/2505.19914v1 |
63.3 56.2 50.9 23.6 54.9 DeepSeek-R1 82.7 77.1 71.4 51.1 62.6 38.4 19.5 49.2 Gemini-2.5-Pro 75.2 95.4 71.5 58.9 37.3 49.4 17.5 50.6 Claude-3.7-Sonnet-Thinking 81.5 75.4 76.6 57.7 50.5 49.4 26.4 53.2 DS-R1-Distilled-Qwen-32B 16.7 47.3 62.9 38.2 36.4 12.4 18.8 31.1 QwQ-32B 59.0 65.7 74.4 47.5 47.3 28.2 24.4 43.8 Grok-2-1... | https://arxiv.org/abs/2505.19914v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.