text string | source string |
|---|---|
be used to define classification rules. To rein- force these ideas, they collaboratively construct three distinct classification rep- resentations: Euler-Venn diagrams allows to group mushrooms based on shared characteristics;Tabularrepresentationsprovideastructureddataformattohigh- light classification flexibility; De... | https://arxiv.org/abs/2505.21398v1 |
a post-test is administered, comprising seven exercises designed to evaluate key computer science concepts (exercises 1-3, 5, 7) and the underlying mathematical skills (exercises 4 and 6). Specifically, the first exercise ( AI Scenarios ) presented AI-powered objects performing various tasks, prompting students to refl... | https://arxiv.org/abs/2505.21398v1 |
university professor who delivered the ac- tivities, a computer science university professor specializing in AI, and a doctoral student in mathematics didactics. Each assessed correctness based on predefined criteria tailored to each exercise (see our repository). For exercises requiring jus- tification(e.g., Animal Fo... | https://arxiv.org/abs/2505.21398v1 |
Also students reflected on how the activities improved their understanding of CS, AI, and mathematics. Resultsindicatethatmoststudentsfoundtheactivitiesengaging,with54.17% enjoying them "very much" and 29.17% "a lot" (Fig.4a). While 81.67% rated the (a) Q1: Enjoyment (b) Q2: Overall difficulty (c) Q3: Feeling (d) Q4: A... | https://arxiv.org/abs/2505.21398v1 |
the children’s classification and representation skills and found a good level of enjoyment; this last aspect is relevant as one of the critical issues that has emerged in the literature con- cerns the absence of stimulating activities within the school context [15]. The contextualization to the understanding of the fu... | https://arxiv.org/abs/2505.21398v1 |
Italy 7. Y. Chevallard, La Transposition didactique: Du savoir savant au savoir enseigné, Grenoble, La Pensée sauvage, 1991 (1re éd. 1985), 126 p. (ISBN 9782859190781) 8. code.org–AI and Machine Learning. AI for Oceans. 2023. https://studio.code.org/s/oceans/lessons/1/levels/6?lang=en-US (Last access 19/02/2025) 9. CS ... | https://arxiv.org/abs/2505.21398v1 |
cation (pp. 1-23). 28. Sabuncuoglu, A. 2020. Designing One Year Curriculum to Teach Artificial Intelli- gence for Middle School. In Proceedings of the 2020 ACM Conference on Innovation and Technology in Computer Science Education, ITiCSE ’20, 96102. Association for Computing Machinery. 29. Ismaila T. Sanusi, Solomon S.... | https://arxiv.org/abs/2505.21398v1 |
arXiv:2505.21399v1 [cs.CL] 27 May 2025Factual Self-Awareness in Language Models: Representation, Robustness, and Scaling Hovhannes Tamoyan, Subhabrata Dutta, andIryna Gurevych Ubiquitous Knowledge Processing Lab (UKP Lab) Department of Computer Science and Hessian Center for AI (hessian.AI) Technical University of Darm... | https://arxiv.org/abs/2505.21399v1 |
vieattributeChrist op h er No lanJam e s Br o wn Christ op h er No lan relationar tist ofentity nameI G ot Y ouentity typeson gattributeJam e s Br o wnkn o wnf or g ott enL an guag e Mo d elFigure 1: Given an input comprising entity type, entity name, and relation, we obtain the model’s token-level prediction probabili... | https://arxiv.org/abs/2505.21399v1 |
construct linear subspaces within internal representations that can demarcate between an upcoming correct/incorrect recall (as opposed to faithfulness in post-hoc checking of correct/incorrect facts). The effects of context and prompt formatting on the formation of these linear subspaces of self-awareness are investiga... | https://arxiv.org/abs/2505.21399v1 |
et al. (2023); Huben et al. (2023), by contrast, have recently gained popularity for uncovering interpretable decompositions of model latent representations without supervised data. Both approaches align with the linear representation hypothesis Park et al. (2023); Mikolov et al. (2013), which posits that interpretable... | https://arxiv.org/abs/2505.21399v1 |
matching errors by relying solely on model’s output space. We construct (entity type ,entity name ,relation )triplets using various templates, but some templates introduced spurious correlations due to their phrasing. Details and examples of all templates are provided in Table 3 and Appendix A. For subsequent experimen... | https://arxiv.org/abs/2505.21399v1 |
a fraction denoting a subset of the vocabulary. This demarcation arises from the actual count of tokens in these bands: high probability tokens are exponentially fewer than low probability ones. 4All experiments are conducted with three random seeds (73, 5, 120); we observe negligible variance and omit the results for ... | https://arxiv.org/abs/2505.21399v1 |
scores appear as mirror images, having equal magnitudes but opposite signs. We train linear probes on the Gemma 2 (2B, 9B) (Team et al., 2024) and Pythia (70M, 1.4B, 6.9B, 12B) (Biderman et al., 2023) models, evaluating performance using standard binary classification metrics and reporting the accuracy improvement over... | https://arxiv.org/abs/2505.21399v1 |
achieves the highest test-time ∆, smaller versions like Pythia 6.9B and 1.4B show comparable accuracies, albeit with smaller gains over their baselines. Pythia 70M represents a degenerate case where accuracy matches the random baseline ( ∆ = 0 ), indicating that the smallest model fails to encode self-awareness feature... | https://arxiv.org/abs/2505.21399v1 |
: all selected samples have the same entity name; however, the relation and attribute may vary: Th e th e is . Th e th e is .r el e ase y e ar of dir e ct or ofm o vie m o vieIn c eption In c eption2010 Christ op h er No lanrelationentity typeentity nameattribute Unique : each entity name appears only once in the conte... | https://arxiv.org/abs/2505.21399v1 |
Parameters k-l on Probe Behavior 5 50 500 5000 k0.1 0.2 0.3 0.4l0.82 0.67 0.54 1.53 0.83 0.68 0.53 1.35 0.84 0.70 0.50 1.16 0.85 0.73 0.51 0.97Known-Forgotten Sample Ratio (Across Models) 0.60.81.01.21.4 |Class Balance Ratio 1| Figure 5: Known-Forgotten sample ratio for each (k, l)configuration, aggregated across all m... | https://arxiv.org/abs/2505.21399v1 |
115000 120000 125000 130000 135000 140000 143000 Training Steps0.650.700.750.80Accuracy Train Accuracy per LayerFigure 8: Linear probe accuracy across Pythia 1.4B training checkpoints/tokens. (Top) Training accuracy by layer (warmer colors = deeper layers). (Bottom) Test accuracy by layer (cooler colors = deeper layers... | https://arxiv.org/abs/2505.21399v1 |
rudimentary form of factual recall where all the necessary information (e.g., entity name, entity type, relation) is provided within the immediate query. In open-ended generation tasks, the LM might need to gather this information from scattered context, resolve coreferences, perform multi-hop factual recall implicitly... | https://arxiv.org/abs/2505.21399v1 |
OpenReview.net, 2023. URL https://openreview.net/forum?id=ETKGuby0hcs . Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter... | https://arxiv.org/abs/2505.21399v1 |
Zhuang, and Weiming Lu. Self-contrast: Better reflection through inconsistent solving perspectives. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024 ... | https://arxiv.org/abs/2505.21399v1 |
Wikidata. https://www.wikidata.org , 2023. 12 Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. Advances in neural information processing systems , 35:17359–17372, 2022. Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of fact... | https://arxiv.org/abs/2505.21399v1 |
whether the response should be a location or a year—we incorporate "hints" at the end of each relation. We experimented with four input templates (see Table 3) using the relations2 set, aiming to eliminate spurious correlations and isolate the self-awareness signal captured by linear probes. Notably, only template2 con... | https://arxiv.org/abs/2505.21399v1 |
Behavior The figures in this section illustrate how varying the (k, l)parameters—representing the number of known and forgotten samples, respectively—affects linear probe performance and class balance outcomes across different model scales. Figure 11 presents the test and train accuracy gains over a random baseline for... | https://arxiv.org/abs/2505.21399v1 |
0.2 0.1l0.03 (491/14157)0.13 (1728/12920)0.31 (3458/11190)1.80 (9418/5230) 0.04 (511/14137)0.14 (1821/12827)0.34 (3703/10945)2.06 (9859/4789) 0.04 (526/14122)0.15 (1892/12756)0.36 (3901/10747)2.30 (10210/4438) 0.04 (544/14104)0.15 (1946/12702)0.38 (4064/10584)2.47 (10432/4216)Class Balance Ratio Heatmap with (Known / F... | https://arxiv.org/abs/2505.21399v1 |
arXiv:2505.21409v1 [cs.CL] 27 May 2025RelationalFactQA: A Benchmark for Evaluating Tabular Fact Retrieval from Large Language Models Dario Satriani, Enzo Veltri, Donatello Santoro University of Basilicata, Potenza, Italy name.surname@unibas.itPaolo Papotti EURECOM, Biot, France paolo.papotti@eurecom.fr Abstract Factual... | https://arxiv.org/abs/2505.21409v1 |
introduces factual errors in the results. County State Area (sq mi) Los Angeles California 4 751 Cook Illinois 1 635 Maricopa Arizona 8 500 ✗ Q: What is the area of Maricopa county? A: 9 224 sq mi ✓Crucially, if we then query the LLM for these specific incorrectly reported values in isolation (e.g., “What is the area o... | https://arxiv.org/abs/2505.21409v1 |
improvement, the ability to produce correct structured answers remains limited — especially as the number of tuples and attributes increases or the query involves less common facts and numerical conditions. Moreover, we observe that even state-of-the-art models rarely exceed 25% of factual accuracy on our benchmark. To... | https://arxiv.org/abs/2505.21409v1 |
within the context of single-statement claims rather than structured relational outputs. Despite these varied evaluation efforts, the fundamental challenge of LLM hallucination persists as a critical concern [10, 1]. Table Question Answering and Reasoning. Several benchmarks like WikiSQL [ 54], WikiTable- Questions [ 3... | https://arxiv.org/abs/2505.21409v1 |
model’s internal knowledge, but also on external factors—such as retrieval accuracy, context formatting, or prompt design. These confounding variables make it difficult to isolate the LLM’s intrinsic factual competence. While retrieval-based methods may improve factual coverage, we hypothesize the challenges in closed-... | https://arxiv.org/abs/2505.21409v1 |
incorrect information. As the tool occasionally produces syntactically correct but semantically trivial or invalid queries, we manually remove such non-meaningful examples. Finally, we perform targeted preprocessing steps to enhance consistency in the ground truth data. For all date attributes, we extract the year comp... | https://arxiv.org/abs/2505.21409v1 |
intermediate output to produce the final filtered result. This method aims to improve retrieval accuracy by breaking queries into simpler tasks. In all methods, the LLM is prompted with the query qand the corresponding output schema s, expressed in JSON Schema format. Output Processing. The prompt includes instructions... | https://arxiv.org/abs/2505.21409v1 |
lower- case; (iii) Converting shorthand numeric notations like “1K” or “1M” and into the corresponding numeric values (e.g., “1K” →1000); (iv) Standardizing numeric formats (e.g., converting “1.000,5” and “1,000.5” into a consistent representation). Moreover, since LLMs may produce answers that are close, but not ident... | https://arxiv.org/abs/2505.21409v1 |
while the COT approach leads to improved retrieval with all LLMs except GTP 4.1. Takeaways for questions (1) and (2) : LLMs still struggle to consistently retrieve structured factual knowledge as complete output tuples. NL outperfoms slightly SQL as a retrieval method, while CoT provides benefits in most settings. Exp-... | https://arxiv.org/abs/2505.21409v1 |
of rows and columns grows, the model’s ability to return accurate, complete tabular data declines. Table 4: Quality measured as the A VG between F1 and TS w.r.t. query complexity. LLama 3.3-70B GPT 4.1 QWEN 3-235B Type NL SQL CoT NL SQL CoT NL SQL CoT SELECT without WHERE 0.599 0.646 0.595 0.845 0.399 0.995 0.787 0.582... | https://arxiv.org/abs/2505.21409v1 |
table) underscores that the bottleneck is frequently not an absence of the underlying factual knowledge. Instead, the difficulty lies in the process of composing the individual pieces of information into a larger relational structure. This distinction points towards limitations in the architectural or learned capabilit... | https://arxiv.org/abs/2505.21409v1 |
2073-431X. doi: 10.3390/computers13100257. URL https://www.mdpi.com/ 2073-431X/13/10/257 . (p. 2) [4]A. Bisercic, M. Nikolic, M. van der Schaar, B. Delibasic, P. Lio, and A. Petrovic. Interpretable medical diagnostics with structured data extraction by large language models, 2023. URL https://arxiv.org/abs/2306.05052 .... | https://arxiv.org/abs/2505.21409v1 |
on Extending Database Technology, EDBT 2024, Paestum, Italy, March 25 - March 28 , pages 461–473. OpenProceedings.org, 2024. doi: 10.48786/EDBT.2024.40. URL https://doi.org/10.48786/edbt.2024.40 . (p. 6) [18] A. Holtzman, J. Buys, L. Du, M. Forbes, and Y . Choi. The curious case of neural text degeneration. In Internat... | https://arxiv.org/abs/2505.21409v1 |
Evans. Truthfulqa: Measuring how models mimic human falsehoods, 2022. URL https://arxiv.org/abs/2109.07958 . (p. 3) [29] C. Liu, M. Russo, M. Cafarella, L. Cao, P. B. Chen, Z. Chen, M. Franklin, T. Kraska, S. Madden, R. Shahout, and G. Vitagliano. Palimpzest: Optimizing ai-powered analytics with declarative query proce... | https://arxiv.org/abs/2505.21409v1 |
A. G. Parameswaran, and E. Wu. Docetl: Agentic query rewriting and evaluation for complex document processing, 2025. URL https://arxiv.org/ abs/2410.12189 . (p. 2) [44] K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. W. Chung, N. Scales, A. Tanwani, H. Cole- Lewis, S. Pfohl, P. Payne, M. Seneviratne, P. Gamble, ... | https://arxiv.org/abs/2505.21409v1 |
and cross-domain semantic parsing and text-to-sql task. In 2018 Conference on Empirical Methods in Natural Language Processing, EMNLP 2018 , pages 3911–3921. Association for Computational Linguistics, 2018. (pp. 1, 3, and 4) [53] X. Zhang, S. Luo, B. Zhang, Z. Ma, J. Zhang, Y . Li, G. Li, Z. Yao, K. Xu, J. Zhou, D. Zha... | https://arxiv.org/abs/2505.21409v1 |
attributes increases. Ideally, the quality on surname should remain constant regardless of how many other attributes are queried. However, as shown in Figure 3.(a), performance on the surname column degrades substantially. For instance, using GPT-4.1, quality drops from 1.0 when a single attribute is requested to 0.516... | https://arxiv.org/abs/2505.21409v1 |
variation in instance complexity. On the attribute side, most examples contain a mix of categorical (avg. 3.16) and mixed attributes (avg. 4.26), with relatively few numerical attributes (avg. 1.06). This implies that RFQA poses both relational and interpretative challenges, as models must handle heterogeneous data typ... | https://arxiv.org/abs/2505.21409v1 |
is stored in the data folder, following the structure <dataset>/<db id>. To retrieve the expected results for a given query q, one can execute the corresponding SQL query on its associated database that can be loaded with the associated data. In our experiment all the data are imported into PostgreSQL database. The bac... | https://arxiv.org/abs/2505.21409v1 |
structured form respecting a JSON format prompted. We parse the response according to the required JSON. The most common issues in the JSON parsing and our corresponding handling methods are the following: •Malformed JSON syntax : This includes missing quotation marks or improperly formatted numbers. In such cases, we ... | https://arxiv.org/abs/2505.21409v1 |
number of the attributes requested in the query respectively on measures against the TS metric (the previous) and the F1 metric (the latter). The results indicate that Tuple Similarity (TS) generally decreases as the number of attributes increases beyond three, across most models and prompting strategies. Natural Langu... | https://arxiv.org/abs/2505.21409v1 |
complexity grows, while Gemma 2 and GPT-4.1 remain competitive in F1 but are more sensitive in TS. QWEN 3 exhibits inconsistent TS results despite some strength in F1. These observations underline the importance of both prompt strategy and model choice in handling increasing task complexity. 1 2 3 4 5 6 7 8 900.20.40.6... | https://arxiv.org/abs/2505.21409v1 |
a partial set of expected rows. These omissions were closely tied (80% of the time) to queries with numerical conditions in the WHERE clause, particularly involving inequality operators such as >,>=,<, or<=. In contrast, equality conditions rarely led to missing results. This trend aligns with prior experimental result... | https://arxiv.org/abs/2505.21409v1 |
arXiv:2505.21410v1 [cs.AI] 27 May 2025Multi-Resolution Skill Discovery for HRL Agents Shashank Sharma Department of Computer Science University of Bath ss3966@bath.ac.ukJanina Hoffmann Department of Psychology University of Bath jah253@bath.ac.uk Vinay Namboodiri Department of Computer Science University of Bath vpn22@... | https://arxiv.org/abs/2505.21410v1 |
leading to smooth but imprecise movements. proposed architecture yields significant performance improvements, outperforming previous HRL SOTA methods and matching SOTA non-HRL methods. We also conduct ablation studies to measure the contribution of each module and show that skill interleaving yields the best results. T... | https://arxiv.org/abs/2505.21410v1 |
rate). In fact, in 2 (a) Skill CV AE architecture (b) Acting using Skill CV AE Figure 2: Illustrations of the abstract state transition-based control for the manager. Dashed arrows indicate sample propagation from the predicted distribution. (a) Skill CV AE, where the Encoder encodes initial and final states (st, st+l)... | https://arxiv.org/abs/2505.21410v1 |
the decoder, shared. The sharing causes a minimal increase in model size but increases the recall with the resolution-specific input and output layers. Fig. 3a illustrates the Multi-Resolution Skill CV AE architecture. For training, state-pairs (st, st+li)atN different temporal resolutions li∈ {l0, l1, ..., l N}are ext... | https://arxiv.org/abs/2505.21410v1 |
followed by policy update using policy gradients for the external and exploratory rewards. We briefly describe the standard training steps below, followed by the exploratory objective (Sec. 3.4.1) and the policy gradients for our approach (Sec. 3.4.2). See Sec. B for full training and architecture details. Manager: The... | https://arxiv.org/abs/2505.21410v1 |
the action atat each step tusing the subgoal s⌊t/K⌋ g for the duration, and the environmental state transition pT. Here, the exponent ck,i collapses the skill probabilities πMi(zk,i|skK)of the unselected skills to 1as they do not affect the trajectory. 5 We follow the policy gradient derivation from [ 21]. The aim is t... | https://arxiv.org/abs/2505.21410v1 |
policy learning signal reduces by a factor of N; thus, we increase training to every 8-th step rather than 16. The agent is tested in locomotion-based environments and trained to optimize for external and exploratory rewards (advantages weighted as [1.0,0.1], respectively). 5.1 Standard Benchmarks DeepMind Control Suit... | https://arxiv.org/abs/2505.21410v1 |
100episodes. It can be seen that interleaving the skills using the proposed choice mechanism consistently yields the best results. It should also be seen that no individual skill performs well for all tasks; thus, using the choice policy πMCcan help automate skill selection. Figure 7: Final performance comparison betwe... | https://arxiv.org/abs/2505.21410v1 |
[ 1,13,14,20,11,7]. The mutual information objective maximizes the predictability of trajectories given skills, and skills given trajectories. The skills allow the agent policy to function in a latent space. OPAL [ 1] encodes the trajectory using a bidirectional GRU and is optimized as a Variational Autoencoder. Causal... | https://arxiv.org/abs/2505.21410v1 |
A. Oh, editors, Advances in Neural Information Processing Systems , volume 35, pages 26091–26104. Curran Associates, Inc., 2022. URL https://proceedings.neurips.cc/paper_files/paper/ 2022/file/a766f56d2da42cae20b5652970ec04ef-Paper-Conference.pdf . [11] Zheyuan Jiang, Jingyue Gao, and Jianyu Chen. Unsupervised skill di... | https://arxiv.org/abs/2505.21410v1 |
the trajectory states stand the prescribed worker goal swg. First, discounted returns Gλ tare computed as n-step lambda returns (Eq. 12). Then the Actor policy is trained using the REINFORCE objective (Eq. 13) and the Critic is trained to predict the discounted returns (Eq. 14). The entropy for the worker and the manag... | https://arxiv.org/abs/2505.21410v1 |
method’s sample efficiency (train every 8-steps) could reduce compute costs for real-world robot training, lowering environmental footprints. The imagination-based policy optimization mitigates hazards that can occur during learning. The skill interleaving mechanism allows for transparent agents with interpretable subg... | https://arxiv.org/abs/2505.21410v1 |
three key claims that are rigor- ously validated: •"Learns skills at multiple temporal resolutions" (Sec. 3.2): Demonstrated through distinct skill visualizations (Fig. 1) and per-skill ablation results (Fig. 7) and samples of learned agents (Sec. E). •"Dynamic skill interleaving mechanism" (Sec. 3.3): Validated via co... | https://arxiv.org/abs/2505.21410v1 |
Pre-trained checkpoints for all tasks. • Jupyter notebooks for all results and visualizations. 6.Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answ... | https://arxiv.org/abs/2505.21410v1 |
Suite [ 23]) without creating novel standalone assets. The proposed method and training procedures are fully described in Algorithms 1-2 and Appendix B. 14.Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of ins... | https://arxiv.org/abs/2505.21410v1 |
arXiv:2505.21413v1 [cs.CL] 27 May 2025REFTOOL : Enhancing Model Reasoning with Reference-Guided Tool Creation Xiao Liu1Da Yin2Zirui Wu1Yansong Feng1∗ 1Wangxuan Institute of Computer Technology, Peking University 2University of California, Los Angeles {lxlisa,ziruiwu,fengyansong}@pku.edu.cn, da.yin9712@gmail.com Abstrac... | https://arxiv.org/abs/2505.21413v1 |
Answer the question with the toolsTool Creation Tool Utilization Figure 1: Overview of the REFTOOL framework, which consists of two modules: tool creation (left) and tool utilization (right). tool creation, the framework employs LLMs to generate executable tools from reference content. In the example, given a section o... | https://arxiv.org/abs/2505.21413v1 |
in a cost-efficient and human-free manner, demonstrating the potential for application in diverse scenarios. 2 The R EFTOOL Framework As shown in Figure 1, REFTOOL operates in two stages: (1) constructing a hierarchical toolbox T from reference material R, and (2) selecting and applying tools t⊂Tto answer the input que... | https://arxiv.org/abs/2505.21413v1 |
that chapter, including their descriptions, functions, and demonstration examples. It is then prompted to select up to ntrelevant tools t, or none if no tools are deemed applicable. Solution Generation The selected tools are then integrated into the reasoning process. We incorpo- rate the tools with two reasoning parad... | https://arxiv.org/abs/2505.21413v1 |
tools per section across all domains. This can be adjusted based on each section’s length and information density. For tool utilization, we employ a default configuration of selecting nc= 1chapter and nt= 1tool. As QRData and TheoremQA do not have a validation set, we use the default setting for the causality and physi... | https://arxiv.org/abs/2505.21413v1 |
are averaged across three sub-datasets, with detailed per-sub-dataset performance shown in Appendix Table 7. 5 Table 2: Performance of REFTOOL and baseline methods in causality (QRData), physics (The- oremQA), and chemistry (SciBench) domains. Numbers are in percentages ( %), with the best performance for each model sh... | https://arxiv.org/abs/2505.21413v1 |
both select a chapter). Tool selection consistency is the fraction where their tools overlap, given they both choose tools from the same chapter. Domain Consistency Llama-3.1-70B Gemini-1.5-Pro GPT-4 GPT-4o CausalityChapter Selection 100 95 100 100 Tool Selection within Chapter 94 100 91 94 PhysicsChapter Selection 80 ... | https://arxiv.org/abs/2505.21413v1 |
EFTOOL (sim) 30.5 36.1 46.1 39.4 38.0 PoT + R EFTOOL 36.8 43.9 38.7 46.8 41.6 Physics PoT + RAG 44.7 57.0 44.7 57.9 51.1 PoT + Hierarchical RAG 44.7 64.0 44.7 55.3 52.2 PoT + R EFTOOL (sim) 45.6 62.3 43.0 56.1 51.8 PoT + R EFTOOL 53.5 58.8 49.1 57.9 54.8 4.2 Ablation Study We design two variants of REFTOOL to analyze i... | https://arxiv.org/abs/2505.21413v1 |
for independence between the input variable and residuals. PoT + RefTool Selected Chapter: Chapter.11 Causal Discovery from Observational Data Selected Tool: Causal Direction Fit Answer: C Answer: C Figure 3: Example case of GPT-4o with (right) and without (left) R EFTOOL. 4.4 Case Study Figure 3 demonstrates a case wh... | https://arxiv.org/abs/2505.21413v1 |
[4]W. Chen, M. Yin, M. Ku, P. Lu, Y . Wan, X. Ma, J. Xu, X. Wang, and T. Xia. Theoremqa: A theorem-driven question answering dataset. In The 2023 Conference on Empirical Methods in Natural Language Processing , 2023. [5]Y . Du, F. Wei, and H. Zhang. Anytool: Self-reflective, hierarchical agents for large-scale api call... | https://arxiv.org/abs/2505.21413v1 |
with large language models. InProceedings of the 31st International Conference on Computational Linguistics , pages 11274–11289, 2025. [20] C. Qian, C. Han, Y . Fung, Y . Qin, Z. Liu, and H. Ji. Creator: Tool creation for disentangling abstract and concrete reasoning of large language models. In The 2023 Conference on ... | https://arxiv.org/abs/2505.21413v1 |
Skillweaver: Web agents can self-improve by discovering and honing skills. arXiv preprint arXiv:2504.07079 , 2025. 11 A Implementation Details Implementation of the REFTOOL Framework By default, each tool’s demonstration example is included during solution generation. For the causality domain, we omit the example becau... | https://arxiv.org/abs/2505.21413v1 |
Average Llama-3.1-70B Creator 50.0 34.0 36.4 40.1 StructChem 50.0 21.3 42.4 37.9 ChemAgent 60.5 44.7 39.4 48.2 PoT 65.8 44.7 30.3 46.9 PoT + RAG 63.2 44.7 36.4 48.1 PoT + R EFTOOL 63.2 48.9 36.4 49.5 Gemini-1.5-Pro Creator 73.7 63.8 42.4 60.0 StructChem 57.9 38.3 54.5 50.2 ChemAgent 78.9 66.0 51.5 65.5 PoT 78.9 59.6 48... | https://arxiv.org/abs/2505.21413v1 |
REFTOOL on another physics dataset SciBench-fund [ 24] with 71 questions, to test tool generalizability.6Table 10 shows that REFTOOL outperforms all zero-shot baselines and matches 4-shot Physics Reasoner’s performance using the same tools in the evaluation of TheoremQA. This demonstrates REFTOOL’s dataset-agnostic nat... | https://arxiv.org/abs/2505.21413v1 |
and SciBench papers. Prompts for ReAct are the same as the QRData paper. Prompt for Initial Tool Generation Please extract the skills from the following text. The text is a section from the chapter {chapter} of the book {book}. Each skill is a python function with comments of parameters and returns, accompanied by a de... | https://arxiv.org/abs/2505.21413v1 |
original skill, starting with ’{’ and ending with ’}’. Original Skill: {skill} Feedback: {feedback} Figure 8: Prompt template for tool refinement. Prompt for Chapter Selection You are a data analyst and good at quantitative reasoning. You are required to respond to a quantitative question using the provided data. The q... | https://arxiv.org/abs/2505.21413v1 |
of the code should be “‘python def solution(): # import libraries if needed # write code to get the answer # return answer “‘ Question: {question} Response: Figure 12: Prompt template for PoT. For evaluation of QRData, the data description and ten lines of the shuffled data are also added to the prompt along with the q... | https://arxiv.org/abs/2505.21413v1 |
arXiv:2505.21414v1 [cs.LG] 27 May 2025A Framework for Adversarial Analysis of Decision Support Systems Prior to Deployment Brett Bissey0 1Kyle Gatesman0 1Walker Dimon1Mohammad Alam1Luis Robaina1Joseph Weissman1 Abstract This paper introduces a comprehensive frame- work designed to analyze and secure decision- support s... | https://arxiv.org/abs/2505.21414v1 |
algorithms and learning curricula. 2. Related Work Conducting adversarial attacks on neural network policies is not as groundbreaking of a concept now as it was when first explored in (Huang et al., 2017), which extended pre- vious work in adversarial attacks in the computer vision domain such as Fast Gradient Sign Met... | https://arxiv.org/abs/2505.21414v1 |
the agent’s action at. Figure 1. RL interaction loop with an attack injected at time step t. This time step ends with the environment dynamics using the agent’s action atand the true state stto compute the next state st+1and the reward rt+1. Time step t+1may or may not have an attack. Figure 2. Example set of attacked ... | https://arxiv.org/abs/2505.21414v1 |
that stem from early states in an episode. To combat these computational costs, stratified sampling was implemented to prune the set of attacks from which to simulate while still guaranteeing sufficient representation from desired sub-populations. 3.2. Attack Strategy Design Anattack strategy is an algorithm that decid... | https://arxiv.org/abs/2505.21414v1 |
by leveraging the gradients of the loss function with respect to the input data. We default to using FGSM for our experi- ments, although the experimental framework is agnostic to the perturbation algorithm used. 3.3. Measuring Attack Impact 3.3.1. D EFINING A PROPERTY Attack impact is measured with respect to a handfu... | https://arxiv.org/abs/2505.21414v1 |
may lose saliency when Pihas any component that ranges over a continuous domain. To illustrate one way to combat this issue, if the property Piresides in some metric space with a distance metric d(·,·), then one could construct an impact metric such as ( 1 if d(Pi(st,It,ot),Pi(st,It,st))>d∗ 0 if d(Pi(st,It,ot),Pi(st,It... | https://arxiv.org/abs/2505.21414v1 |
from red. An example network structure from a mid-episode observation is displayed in Figure 3. 4.2. Experimental Setup First, we train a suite of both Advantage Actor Critic (A2c) (Mnih et al., 2016) and Deep Q-Network (DQN) agents (Mnih et al., 2013) within the CyberStrike environment. Following training we collect 1... | https://arxiv.org/abs/2505.21414v1 |
decision-making. Therefore, our attack-discovery framework would permit this kind of benign attack to be made frequently in a single episode, since an ideal policy ought to not behave differently from any of these attacks. Figure 5. The Average Final Red Count delta post-attack is aggre- gated per observation index, ac... | https://arxiv.org/abs/2505.21414v1 |
dependent on the time-step when the attack is brokered. By strategically timing the ma- nipulation of the most vulnerable observation components of a policy, we are able to observe significant variations in policy behavior, leading to notable changes in the envi- ronment properties and game outcome; thus demonstrating ... | https://arxiv.org/abs/2505.21414v1 |
policies: A2c-ADR+CL (A), A2c-ADR (B), A2c-CL (C), DQN-CL (A), DQN-deterministic (B)). Training curriculum and hyperparameter details for the poli- cies are available in the appendix. For each policy, we use two action-targets for transferability analysis: the 0-action (No-op) and the max(loss-win) action. The max(loss... | https://arxiv.org/abs/2505.21414v1 |
using adversarial attacks to probe and analyze the behavior of policies trained throughDRL algorithms, the same behavioral analysis may be con- ducted on LLM-based agentic architectures, albeit with language-based attacks and alternate metadata for t-SNE embeddings. We will leave this to future adversarial analysis res... | https://arxiv.org/abs/2505.21414v1 |
. Molina-Markham, A., Winder, R. K., and Ridley, A. Net- work defense is not a game, 2021. URL https:// arxiv.org/abs/2104.10262 . Rajeswaran, A., Kumar, V ., Gupta, A., Vezzani, G., Schul- man, J., Todorov, E., and Levine, S. Learning complex dexterous manipulation with deep reinforcement learn- ing and demonstrations... | https://arxiv.org/abs/2505.21414v1 |
: a d r n o r m a l r a n g e p a r a m e t e r s : mean : 1 . 0 s t d e v : 1 . 0 maximum : 1 . 0 minimum : 0 . 1 − i d : a d r 1 v 0 t y p e : a d r n o r m a l r a n g e p a r a m e t e r s : mean : 1 . 0 s t d e v : 1 . 0 maximum : 1 . 0 minimum : 0 . 1 − i d : a d r 2 v 0 t y p e : a d r n o r m a l r a n g e p a ... | https://arxiv.org/abs/2505.21414v1 |
− [ 3 ] # r e d node 2 d e f e n d e d by 3 − [ 4 ] − [ ] #4 − [ ] #5 − [ ] #6 − [ 6 ] # r e d node 7 i s d e f e n d e d by 6 b l u e : a s s e t s : − t y p e : 1 l o s s c o s t : 20 u s e c o s t : 2 − t y p e : 2 l o s s c o s t : 20 u s e c o s t : 2 − t y p e : 2 l o s s c o s t : 20 u s e c o s t : 2 i sa l i v... | https://arxiv.org/abs/2505.21414v1 |
target is reached through recursive hacks. In the absence of adversarial attacks, DRL policies optimize towards this behavioral pattern. A.6. Curriculum Learning and Automated Domain Randomization We randomize the action effectiveness variables for Curriculum Learning (CL) and Automated Domain Randomization (ADR) polic... | https://arxiv.org/abs/2505.21414v1 |
s e l f . f c = t o r c h . nn . S e q u e n t i a l ( nn . L i n e a r ( i n s h a p e [ 0 ] , hidden dim ) , nn . ReLU ( ) , nn . L i n e a r ( hidden dim , hidden dim ) , nn . ReLU ( ) , nn . L i n e a r ( hidden dim , o u t s h a p e [ 0 ] ) , ) The critic network for A2c training is instantiated as follow: Listing... | https://arxiv.org/abs/2505.21414v1 |
arXiv:2505.21419v2 [cs.AI] 28 May 2025Diagnosing and Resolving Cloud Platform Instability with Multi-modal RAG LLMs Yifan Wang wangyifan@cs.cornell.edu Computer Science Department, Cornell University Ithaca, NY, USAKenneth P. Birman ken@cs.cornell.edu Computer Science Department, Cornell University Ithaca, NY, USA ABST... | https://arxiv.org/abs/2505.21419v2 |
for the two most widely cited sets, HPC4 [ 15], COM2 [ 18]. But this issue is also seen with less widely used data sets. Even if we limit ourselves to a single data mode, existing AI- Ops solutions turn out to have limitations (such as weak support for events characterized by evolution of a problem over time, and hence... | https://arxiv.org/abs/2505.21419v2 |
components: 1) the user’s incident description; 2) a log file collected from the docker container of the faulty service and 3) a time sequence of performance metrics collected from the same container during the the fault. Although the bugs have very differ- ent features, all trace to root causes associated with three w... | https://arxiv.org/abs/2505.21419v2 |
network [7] for the same purpose. Detecting anomalies in cloud platforms using telemetric perfor- mance data requires handling potentially noisy high-dimensionaldata. Li et al. (2024) have explored this problem and proposed a methodol- ogy for noise-tolerant self- supervised learning [ 14] that combines tensor decompos... | https://arxiv.org/abs/2505.21419v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.