text string | source string |
|---|---|
flight times, seat preferences) in every session, leading to inef- ficiency and frustration. In contrast, multi-session memory enables seamless retrieval of prior con- text, supporting fewer turns and a more personal- ized experience. In contrast, multi-session mem- ory allows systems to retrieve prior user-specific in... | https://arxiv.org/abs/2505.20231v1 |
session dependencies, we construct a test set specif- ically for confirmation-type response generation (details in Section 3). Our pipeline includes (1) Retrieval. We explore three strategies for retriev- ing relevant historical dialogues: sparse retrieval (BM25 (Robertson and Zaragoza, 2009)), dense retrieval (text-em... | https://arxiv.org/abs/2505.20231v1 |
coverage of task variations without overpopulating the dataset. More details can be found in Appendix A.1. Confirmation-Type Response Annotation. InAttribute Evaluation Domains 16 Intentions 19 Task goals 956 Dialogues 2,861 Utterances 18,530 Avg. slots per task goal 4.24 Number of individuals 132 Avg. intentions per i... | https://arxiv.org/abs/2505.20231v1 |
context crepresents the ongoing in- teraction, comprising chronologically ordered user utterances ujand system responses rj. The indi- vidual memory bank Mconsists of memory repre- sentations from multiple dialogue sessions, where each session provides an intent description kiand a corresponding set of QA pairs vi. For... | https://arxiv.org/abs/2505.20231v1 |
answers Acore={aj,u}. These core answers are concatenated into the dia- logue context to form a memory-enhanced prompt for the generation module. This step ensures that only high-relevance, low-noise content is passed forward for response planning. 4.3 Proactive Response Strategy The response generation phase synthesiz... | https://arxiv.org/abs/2505.20231v1 |
•Traditional Task-Oriented Dialogue Sys- tems. To assess MAP in structured DST sce- narios, we compare with task-specific fine- tuned BERT-DST (Chao and Lane, 2019), Au- toTOD (Xu et al., 2024a), and LDST (Feng et al., 2024). Among these, AutoTOD incor- porates an external memory module to track user goals across turns... | https://arxiv.org/abs/2505.20231v1 |
memory-guided dialogue planning, and (2) multi-turn dialogues un- 0.0 0.5 1.0 ScoreGPT4o-miniMistral-7BQwen2.5-7BLLaMA3-8B 0.640.740.410.57 0.700.800.660.63JGA 0 2 4 6 Score3.541.204.705.04 3.191.174.363.46DTEHybrid RAG MAPFigure 4: Comparison of Hybrid RAG (blue) vs. MAP (purple) across four LLMs on two metrics: JGA a... | https://arxiv.org/abs/2505.20231v1 |
results demonstrate that both components are es- sential for precise memory integration and efficient multi-session dialogue. Proactive Response Strategy. MAP identifies incomplete or inconsistent slot values by compar- ing predicted responses with task goals, and proac- tively engages the user to resolve them. The ef-... | https://arxiv.org/abs/2505.20231v1 |
anything else you’d like? (Key Slot Omission)Confirmation Response (Hybrid RAG with Summary)Sure! Just to confirm, you’d like me to play the song "Drive" on the bedroom speaker again. Should I start it now? (Right Answer)Confirmation Response (MAP)Session 23 Evaluation -Intention: The user intends to listen to songs.Fi... | https://arxiv.org/abs/2505.20231v1 |
and Myoung-Wan Koo. 2023. Task-optimized adapters for an end-to-end task-oriented dialogue system. In Findings of the As- sociation for Computational Linguistics: ACL 2023 , pages 7355–7369. Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Iñigo Casanueva, Stefan Ultes, Osman Ra- madan, and Milica Gasic. 2018. Mul... | https://arxiv.org/abs/2505.20231v1 |
Ivan Vuli ´c, and Anna Korhonen. 2023. Multi 3 woz: A multilin- gual, multi-domain, multi-parallel dataset for train- ing and evaluating culturally adapted task-oriented dialog systems. Transactions of the Association for Computational Linguistics , 11:1396–1415. Heyan Huang, Puhai Yang, Wei Wei, Shumin Shi, and Xian-L... | https://arxiv.org/abs/2505.20231v1 |
Ling, Jia-Chen Gu, and Quan Liu. 2020. Fine-tuning bert for schema- guided zero-shot dialogue state tracking. Preprint , arXiv:2002.00181. Pararth Shah, Dilek Hakkani-Tür, Gokhan Tür, Abhinav Rastogi, Ankur Bapna, Neha Nayak, and Larry Heck. 2018. Building a conversational agent overnight with dialogue self-play. arXiv... | https://arxiv.org/abs/2505.20231v1 |
pages 2748– 2763. Jing Xu, Arthur Szlam, and Jason Weston. 2022. Be- yond goldfish memory: Long-term open-domain con- versation. In Proceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 5180–5197, Dublin, Ireland. Association for Computational Linguist... | https://arxiv.org/abs/2505.20231v1 |
whether the task objectives are met and the over- all quality of the response. Such an evaluation approach is more comprehensive. A.5 Prompts of Dialogue State Tracking we used a prompt modified from (Heck et al., 2023b) (as shown in Figure 10) that generates the dialogue state for each user turn in the dialogue.Let A1... | https://arxiv.org/abs/2505.20231v1 |
qa memory/dialogue history Table 9: Comparison of MS-TOD with representative Task Oriented Dialogue (TOD) and Open Domain (OD) datasets along memory-related attributes. B.3 Human Validation Protocol To ensure the realism, coherence, and usability of MS-TOD, we apply a structured human validation process during dataset ... | https://arxiv.org/abs/2505.20231v1 |
on memory retrieval. text- embed3-small achieves the highest recall across all thresholds, with 0.702 at Recall@3, 0.792 at Recall@5, and 0.905 at Recall@10, demonstrating superior retrieval capability. Among other models, nv-embed-v2 andbge-large-en-v1.5 also perform well, while traditional retrieval methods like BM25... | https://arxiv.org/abs/2505.20231v1 |
highlighting a trade-off be- tween structured information retention and more natural, contextually aware responses. Table 14 presents the performance comparison between AutoTOD and MAP on Slot Accuracy, BLEU, and ROUGE. The results indicate that MAP consistently outperforms AutoTOD across all three metrics, demonstrati... | https://arxiv.org/abs/2505.20231v1 |
retrieve similar QA memory candidates, the standard RAG model fails to detect missing slot information (e.g., number of people), resulting in an incomplete and partially inaccurate response. In contrast, MAP use Chain-of-Thought explicitly identifies missing task information (e.g., time, headcount) through reasoning, r... | https://arxiv.org/abs/2505.20231v1 |
, " t e x t " : " xx " } ] ] } """ Figure 7: Prompts of the Dataset Generation Prompts of the Task Slot Querying Generation """ Please help me generate questions, based on the provided {conversation history}, that correspond to unanswered attributes in the task goal {task_attributes}. 1. The questions should start with... | https://arxiv.org/abs/2505.20231v1 |
of your score.>", "Score": <An integer score from 1 to 10.> } “‘ """ Figure 11: Prompts of GPT4 Evaluation MS-TOD dialogue session structure { " session_id ": 9, " domain ": " Travel ", " reference_dialogue_id ": "66 _00101 ", " exist_confirmation ": true , " intent ": " ReserveHotel ", " content ": [ { " speaker ": " ... | https://arxiv.org/abs/2505.20231v1 |
MT-TOD. Process MAP RAG Input and Intent Dialogue History User: Have you completed the reservation at Gen Korean BBQ House?User: Have you completed the reservation at Gen Korean BBQ House? Intention Description The user wants to confirm restaurant reservation. The user wants to confirm restaurant reservation. Memory Re... | https://arxiv.org/abs/2505.20231v1 |
to songs.User:Canyouplay'Drive'again?Assistant:Session 23EvaluationSure! Just to confirm, you’d like me to play the song "Drive" on the bedroom speaker again. Should I start it now? Also, would you like me to check out some tickets for the Bill Callahan concert on March 10th in Los Angeles, or are you planning to book ... | https://arxiv.org/abs/2505.20231v1 |
arXiv:2505.20237v1 [cs.CL] 26 May 2025Efficient Speech Translation through Model Compression and Knowledge Distillation Yasmin Moslem ADAPT Centre School of Computer Science and Statistics Trinity College Dublin Dublin, Ireland Abstract Efficient deployment of large audio-language models for speech translation remains ... | https://arxiv.org/abs/2505.20237v1 |
with data knowledge distil- lation from the fully fine-tuned model and a portion of the CoV oST2 dataset to restore the quality of the teacher model. This process has achieved 50% compression in terms of both model parameters and storage size. We dis- cuss the details of this model in Section 3.3. 1https://hf.co/Qwen/Q... | https://arxiv.org/abs/2505.20237v1 |
steps. The model is initially loaded in bfloat16 data type. As shown in Table 1 and Table 2, the fully fine-tuned model clearly outperforms the baseline by an average of 48 COMET points for translation into German and Chinese. Hence, this fully fine- tuned model is used as a foundation for both our experiments in Secti... | https://arxiv.org/abs/2505.20237v1 |
the quantization con- stants from the first quantization are quantized again. For LoRA configuration, we set the rank to 64, alpha to 128, and dropout to 0. We target all linear modules. Overall, this configuration results in 2.41% trainable parameters of the model. More-over, we enable Rank-Stabilized LoRA (rsLoRA) (K... | https://arxiv.org/abs/2505.20237v1 |
each can- didate layer. The layer whose removal yields the best performance is eventually pruned. This pro- cess continues until a predefined number of layers (8 in the main experiments) have been removed. By iteratively removing the least important layers, this performance-guided method produces a more compact model t... | https://arxiv.org/abs/2505.20237v1 |
For other tasks like fine-tuning and inference, bfloat16 precision is not required, although it may be nec- essary when computing resources are limited, potentially at the cost of reduced quality. 4 Encoder Decoder BLEU chrF++ COMET Prams Storage 24↓ 24↓ 26.44 54.67 13.90 6.62 B 13.24 GB 32= 24↓ 30.81 58.45 31.95 6.78 ... | https://arxiv.org/abs/2505.20237v1 |
pruned achieved 4+ more points in terms of BLEU and chrF++ and 18+ points of COMET. It is worth noting that both the number of parameters and the storage footprint of this model is only 1% larger than the model with both the encoder and decoder are pruned. Ta- ble 3 shows the results of fine-tuning the English- to-Germ... | https://arxiv.org/abs/2505.20237v1 |
-94.05), compared to the scores achieved when the same number of layers are it- eratively removed based layer importance (BLEU: 42.52, chrF: 36.92, COMET 39.42). Moreover, we experimented with both using chrF/chrF++ and COMET for evaluating pruned models during the iterative process. Interestingly, the final model resu... | https://arxiv.org/abs/2505.20237v1 |
2). Similarly, fine-tuning the teacher model on in-domain data before pruning is espe- cially recommended for downstreaming tasks (Li et al., 2020). 4.6 Out-of-Domain Data Size After layer pruning, we added more data for fine- tuning the pruned model to recover the translation quality of the teacher model. We mixed the... | https://arxiv.org/abs/2505.20237v1 |
the gen- eralization capabilities of our approach. Given that Qwen2-Audio-Instruct relies on text prompts for generation, it would be interesting to investi- gate retrieval-augmented generation with few-shot prompting to improve the translation quality of compressed models. References Yunfei Chu, Jin Xu, Qian Yang, Hao... | https://arxiv.org/abs/2505.20237v1 |
Meeting of the Association for Com- putational Linguistics , pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. David Peer, Sebastian Stabinger, Stefan Engl, and Anto- nio Rodríguez-Sánchez. 2022. Greedy-layer pruning: Speeding up transformer models for natural language processin... | https://arxiv.org/abs/2505.20237v1 |
arXiv:2505.20243v1 [cs.CL] 26 May 2025It’s High Time : A Survey of Temporal Information Retrieval and Question Answering Bhawna Piryani,♠Abdelrahman Abdallah,♠Jamshid Mozafari,♠Avishek Anand,♡Adam Jatowt♠ ♠University of Innsbruck {bhawna.piryani, abdelrahman.abdallah,jamshid.mozafari, adam.jatowt}@uibk.ac.at ♡TU Delft ... | https://arxiv.org/abs/2505.20243v1 |
in Temporal IR/QA has evolved signif- icantly, building on early foundations to address increasingly complex temporal challenges. Initial efforts relied on rule-based systems (Harabagiu and Bejan, 2005) and statistical models (Berberich et al., 2010) that used document timestamps and hand-crafted rules to interpret tim... | https://arxiv.org/abs/2505.20243v1 |
shared leadership on the world stage . We are showing that the U.S. does not sit on the sidelines when the future of our children is at stake.Document Publication Date (DPD) :August 4, 2015Figure 1: Examples of documents from synchronic (left) and diachronic (right) collections. Red highlights temporal signals present ... | https://arxiv.org/abs/2505.20243v1 |
published Temporal IR/QADataset & Benchmarks ( §3)Temporal Document Collections ( §3.1) TQA Datasets ( §3.2) TIR Dataset ( §3.3) Temporal Tasks ( §4)Prediction Approaches ( §5)Rule-Based & Statistical Models ( §5.1) Temporal Language Models ( §5.2) Temporal RAG ( §5.2.1) Temporal Reasoning ( §5.2.2) Figure 2: Taxonomy ... | https://arxiv.org/abs/2505.20243v1 |
in a range of related temporal tasks, including semantic drift de- tection (Hamilton et al., 2016), event burst mod- eling (Radinsky and Horvitz, 2013), and timeline construction (Gutehrlé et al., 2022). Synchronous corpora represent a coherent snapshot of the world at a specific point in time. Unlike diachronic corpor... | https://arxiv.org/abs/2505.20243v1 |
A " ≤" symbol indicates that the dataset uses a snapshot of Wikipedia and inherits its temporal scope. We also indicate whether temporal metadata is available and whether questions require multi-hop temporal reasoning. context. Finally, Annotated temporal corpora with ex- plicit temporal annotations facilitate more str... | https://arxiv.org/abs/2505.20243v1 |
relevant fact. Datasets such as TiQ (Jia et al., 2024) and TORQUE (Ning et al., 2020) fo- cus on implicit reasoning, testing event-event and event-time relationships. Others like ArchivalQA (Wang et al., 2022), TemporalQuestions (Wang et al., 2021a), and ComplexTempQA (Gruber et al., 2024) combine both question types, ... | https://arxiv.org/abs/2505.20243v1 |
temporally sensitive retrieval. Key tasks include Event Dating, Document Dating ,Focus Time Estimation ,Query Time Profiling , and Event Occurrence Prediction . Tra- ditional methods rely on statistical language mod- els and handcrafted rules, while more recent tech- niques employ transformer-based encoders, tem- poral... | https://arxiv.org/abs/2505.20243v1 |
handle complex temporal questions more effectively, Saquete et al. (2004, 2009) intro- duced a multi-layered QA architecture that decom- posed questions into temporally constrained sub- questions using temporal expression taggers like TERSEO (Saquete et al., 2003). These approaches showed improved precision and general... | https://arxiv.org/abs/2505.20243v1 |
models make RAG more responsive to temporal dynamics in IR/QA. 5.2.2 Temporal Reasoning Capabilities While Temporal Language Models enhance time- aware representations and retrieval, many Temporal IR/QA tasks demand more sophisticated reasoning,such as understanding event sequences, temporal constraints, and durations.... | https://arxiv.org/abs/2505.20243v1 |
systems could use mod- ular, trackable edit layers for local updates, preserv- ing historical facts. Integration of Diachronic and Synchronic Knowledge. Temporal questions often require combining evolving facts (e.g., event timelines) with stable knowledge (e.g., definitions). Future systems should integrate diachronic... | https://arxiv.org/abs/2505.20243v1 |
ACL, SI- GIR, EMNLP, NeurIPS, ECIR, and preprints on arXiv. On the other hand, due to page limitations, we provide only a very brief summary of each method without exhaustive technical details. References Abdelrahman Abdallah, Bhawna Piryani, Jonas Wal- lat, Avishek Anand, and Adam Jatowt. 2025. Tem- pretriever: Fusion... | https://arxiv.org/abs/2505.20243v1 |
Evaluation (LREC‘12) , pages 3735–3740, Istanbul, Turkey. European Language Resources Association (ELRA). Wenhu Chen, Xinyi Wang, and William Yang Wang. 2021. A dataset for answering time-sensitive ques- tions. In Thirty-fifth Conference on Neural Informa- tion Processing Systems Datasets and Benchmarks Track (Round 2)... | https://arxiv.org/abs/2505.20243v1 |
Yu Feng, Ben Zhou, Haoyu Wang, Helen Jin, and Dan Roth. 2023. Generic temporal reasoning with dif- ferential analysis and explanation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 12013–12029, Toronto, Canada. Association for Computational Li... | https://arxiv.org/abs/2505.20243v1 |
Raghav Jain, Daivik Sojitra, Arkadeep Acharya, Sri- parna Saha, Adam Jatowt, and Sandipan Dandapat. 2023. Do language models have a common sense regarding time? revisiting temporal commonsense reasoning in the era of large language models. In Pro- ceedings of the 2023 Conference on Empirical Meth- ods in Natural Langua... | https://arxiv.org/abs/2505.20243v1 |
and Blanco Roi. 2013. A sur- vey of temporal web search experience. In Proceed- ings of the 22nd International Conference on World Wide Web , WWW ’13 Companion, page 1101–1108, New York, NY , USA. Association for Computing Machinery. Rosie Jones and Fernando Diaz. 2007. Temporal pro- files of queries. ACM Trans. Inf. S... | https://arxiv.org/abs/2505.20243v1 |
Adhi Kuncoro, Elena Gribovskaya, Devang Agrawal, Adam Liska, Tayfun Terzi, Mai Gimenez, Cyprien de Masson d 'Autume, Tomas Ko- cisky, Sebastian Ruder, Dani Yogatama, Kris Cao, Susannah Young, and Phil Blunsom. 2021. Mind the gap: Assessing temporal generalization in neural language models. In Advances in Neural Informa... | https://arxiv.org/abs/2505.20243v1 |
Association for Computational Linguistics. Jannat Meem, Muhammad Rashid, Yue Dong, and Vage- lis Hristidis. 2024. PAT-questions: A self-updating benchmark for present-anchored temporal question- answering. In Findings of the Association for Compu- tational Linguistics: ACL 2024 , pages 13129–13148, Bangkok, Thailand. A... | https://arxiv.org/abs/2505.20243v1 |
tion answering dataset based on historical american newspaper pages. In Proceedings of the 47th Inter- national ACM SIGIR Conference on Research and Development in Information Retrieval , SIGIR ’24, page 2038–2048, New York, NY , USA. Association for Computing Machinery. James Pustejovsky, Patrick Hanks, Roser Sauri, A... | https://arxiv.org/abs/2505.20243v1 |
on Text, Speech and Dialogue , pages 220– 228. Springer. Estela Saquete, Jose L. Vicedo, Patricio Martínez-Barco, Rafael Muñoz, and Hector Llorens. 2009. Enhancing qa systems with complex temporal question process- ing capabilities. Journal of Artificial Intelligence Research , 35(1):755–811. Apoorv Saxena, Soumen Chak... | https://arxiv.org/abs/2505.20243v1 |
1: Long Papers) , pages 13014–13033, Bangkok, Thai- land. Association for Computational Linguistics. Zhaochen Su, Juntao Li, Zikang Zhang, Zihan Zhou, and Min Zhang. 2023. Efficient continue training of temporal language model with structural information. InFindings of the Association for Computational Lin- guistics: E... | https://arxiv.org/abs/2505.20243v1 |
Sadao Kurohashi. 2022. Im- proving event duration question answering by lever- aging existing temporal information extraction data. InProceedings of the Thirteenth Language Resources and Evaluation Conference , pages 4451–4457, Mar- seille, France. European Language Resources Asso- ciation. Denny Vrande ˇci´c and Marku... | https://arxiv.org/abs/2505.20243v1 |
Zhang, Haofen Wang, and Meng Wang. 2024. Time-sensitve retrieval-augmented genera- tion for question answering. In Proceedings of the 33rd ACM International Conference on Informa- tion and Knowledge Management , CIKM ’24, page 2544–2553, New York, NY , USA. Association for Computing Machinery. Lee Xiong, Chenyan Xiong,... | https://arxiv.org/abs/2505.20243v1 |
temporal datasets, time-aware models, and reasoning techniques have enabled systems ca- pable of retrieving time-relevant documents, order- ing events, and answering temporally constrained questions, benefiting applications such as historical analysis, fact-checking, and intelligent assistants. While IR and QA have bee... | https://arxiv.org/abs/2505.20243v1 |
et al. (2021b) Document DatingDalli (2006), Jatowt et al. (2007) Kumar et al. (2012) Focus Time EstimationJatowt et al. (2013), Shrivastava et al. (2017) Query Time ProfilingKanhabua and Nørvåg (2010), Dakka et al. (2008), Jones and Diaz (2007) Approaches ( §5)Rule-Based & Statistical Models ( §5.1)Li and Croft (2003),... | https://arxiv.org/abs/2505.20243v1 |
its position in a sen- tence) in the form of temporal perturbations . It is used in evaluation to assess temporal reasoning stability. C Temporal Prediction Tasks Temporal prediction tasks are crucial in understand- ing and organizing time-sensitive textual data. De- spite sharing the common objective of grounding text... | https://arxiv.org/abs/2505.20243v1 |
method that estimates focus time using statistical evidence from external corpora, even when explicit temporal expressions are limited. Shrivastava et al. (2017) further ad- vanced this line of work by linking documents to Wikipedia concepts, leveraging their temporal re- lations to estimate focus times. This task supp... | https://arxiv.org/abs/2505.20243v1 |
arXiv:2505.20245v1 [cs.CL] 26 May 2025KnowTrace: Bootstrapping Iterative Retrieval-Augmented Generation with Structured Knowledge Tracing Rui Li∗†‡ Gaoling School of Artificial Intelligence , Renmin University of China Beijing, China lirui121200@ruc.edu.cnQuanyu Dai Huawei Noah’s Ark Lab Shenzhen, China daiquanyu@huawe... | https://arxiv.org/abs/2505.20245v1 |
impressive performance across a variety of natural language tasks through the form of question answering. Despite their remarkable capabilities, LLMs continue to struggle with factual errors [ 23,47,55] when the input question exceeds their knowledge boundaries. As a practical solution to this problem, Retrieval-Augmen... | https://arxiv.org/abs/2505.20245v1 |
until tracing out a sufficient knowledge graph (KG) for the answer prediction. Such an inference workflow seamlessly empowers the LLM with an intelligible context throughout the multi-step reasoning process, which clearly reveals critical knowledge structures and, as a result, inherently enhances the LLM’s reasoning qu... | https://arxiv.org/abs/2505.20245v1 |
some latest works follow a restructuring-based paradigm as shown in Figure 2(b): they employ an auxiliary restructuring process for all retrieved passages [3,20,22,27] or even entire external corpora [ 6,29,35]. However, this strategy inherently necessitates extensive LLM invocations for intricate restructuring operati... | https://arxiv.org/abs/2505.20245v1 |
structures, KnowTrace further leverages a backtracing mechanism to filter out useless LLM generations (i.e., unavailing exploration andextraneous completion ) from positive trajectories. In this way, KnowTrace can bootstrap its multi-step reasoning capabilities via finetuning on the self-distilled high-quality process ... | https://arxiv.org/abs/2505.20245v1 |
phase. Knowledge Completion. Given the entity-relation pair (𝑒𝑖,𝑟𝑖)as well as the retrieved passages C𝑁 (𝑒𝑖,𝑟𝑖), KnowTrace further harnesses the LLM’s language understanding capability to purposefully grasp key knowledge from the unstructured text. Formally, with a com- pletion instruction 𝐼comthat receives(�... | https://arxiv.org/abs/2505.20245v1 |
training them on their own generations that ultimately lead to the correct answers. Nevertheless, despite its effectiveness for one-time generation tasks, this process is inherently flawed when applied to recent RAG sys- tems in the MHQA scenario: for a complex multi-hop question, even when the final prediction is corr... | https://arxiv.org/abs/2505.20245v1 |
inCompletion #1 the ... of England James Watt Priestley Riotsis written by refers to ......Exploration #2 Completion #2 was educated at... James Watt ......... ... ... ...University of Glasgow was educated at Exploration #3 Backtracing the rioting ... Birmingham University of Glasgow was educated at...... ... ... ...Un... | https://arxiv.org/abs/2505.20245v1 |
.304 .425 4.1 Experimental Setup Datasets. We evaluate our KnowTrace over three standard MHQA benchmarks under the open-domain setting: HotpotQA [ 50], 2Wiki- MultihopQA (2Wiki) [ 10], and MuSiQue [ 41]. We use the same data splits as previous works [ 20,42] for evaluation. In order to create the open-domain setting, w... | https://arxiv.org/abs/2505.20245v1 |
generations and construct a finetuning dataset. The statistical characteristics are analyzed in Section 4.4. On top of the base LLM, we train two distinct LoRA adapters [ 12] to specialize the capabilities of knowledge exploration and knowledge comple- tion, respectively. We tune the training epoch in {1,2,3}, batch si... | https://arxiv.org/abs/2505.20245v1 |
positive trajectories should be filtered, we then calculate the ratio of the tokens that should be filtered to all output tokens. We refer to this ratio as FA (Filtered-to-All). A larger FA means that the collected positive trajectories contain more useless generations, indicating an inferior quality of the finetuning ... | https://arxiv.org/abs/2505.20245v1 |
aligns with standard parameter-efficient finetuning (approximately 2–3 hours on an NVIDIA A100 GPU). Retrieval Models. We validate the compatibility of our framework across different retrievers. Specifically, in addition to BM25 used in Table 1, we further conduct experiments with two other retrievers: DPR and Contriev... | https://arxiv.org/abs/2505.20245v1 |
decision-making tasks, has yet to be explored. Second, although our KnowTrace can retrospectively distill high-quality data for bootstrapping, how to proactively correct erroneous tra- jectories without finetuning remains an open challenge. Conclusion. This work introduces KnowTrace, an elegant iterative RAG framework ... | https://arxiv.org/abs/2505.20245v1 |
2020. Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reason- ing Steps. In Proceedings of the 28th International Conference on Computational Linguistics . 6609–6625. [11] Arian Hosseini, Xingdi Yuan, Nikolay Malkin, Aaron Courville, Alessandro Sor- doni, and Rishabh Agarwal. 2024. V-STaR: Training ... | https://arxiv.org/abs/2505.20245v1 |
and Siamak Shakeri. 2023. Trigger- ing Multi-Hop Reasoning for Question Answering in Language Models using Soft Prompts and Random Walks. arXiv preprint arXiv:2306.04009 (2023). [25] Kamal Nigam and Rayid Ghani. 2000. Analyzing the Effectiveness and Applicabil- ity of Co-training. In Proceedings of the 2000 ACM CIKM In... | https://arxiv.org/abs/2505.20245v1 |
and Noah Fiedel. 2024. Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models. Transactions on Machine Learning Research (2024). [39] Xin Su, Tiep Le, Steven Bethard, and Phillip Howard. 2023. Semi-Structured Chain-of-Thought: Integrating Multiple Sources of Knowledge for Improved Language Mo... | https://arxiv.org/abs/2505.20245v1 |
Multiple Chains of Thought. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing . 5942–5966. [53] Zheng Yuan, Hongyi Yuan, Chengpeng Li, Guanting Dong, Keming Lu, Chuanqi Tan, Chang Zhou, and Jingren Zhou. 2023. Scaling relationship on learning math- ematical reasoning with large l... | https://arxiv.org/abs/2505.20245v1 |
hop questions that require either compositional reasoning or com- parative analysis. Both structured and unstructured information from Wikipedia and Wikidata are combined for data construction. MuSiQue. The multi-hop questions in this dataset is constructed by carefully selecting and composing single-hop questions obta... | https://arxiv.org/abs/2505.20245v1 |
retrieve more relevant knowledge for the subsequent LLM inference.SG-Prompt [ 20].This approach first constructs a semantic graph structures through information extraction from all retrieved text, and then leverages this symbolic information (including entities and semantic relations) to enhance the inference quality. ... | https://arxiv.org/abs/2505.20245v1 |
arXiv:2505.20246v1 [cs.AI] 26 May 2025ONPATH TO MULTIMODAL HISTORICAL REASONING : HISTBENCH AND HISTAGENT Organizing Team Jiahao Qiu∗1, Fulian Xiao∗2, Yimin Wang∗3, Yuchen Mao∗4, Yijia Chen∗5, Xinzhe Juan3, Siran Wang2, Xuan Qi6, Tongcheng Zhang4, Zixin Yao7, Jiacheng Guo1, Yifu Lu1, Charles Argon8, Jundi Cui2, Daixin ... | https://arxiv.org/abs/2505.20246v1 |
on real-world general tasks. Code:https://github.com/CharlesQ9/HistAgent Dataset:https://huggingface.co/datasets/jiahaoq/HistBench ∗Equal contribution. †Correspondence to: gaoxi@fudan.edu.cn ,mengdiw@princeton.edu . HistAgent (gpt-4o)pass@1pass@2 ODR-smolagents (gpt-4o)pass@1pass@2 gpt-4o: online Grok 3: online DeepSee... | https://arxiv.org/abs/2505.20246v1 |
task and is annotated with difficulty level, question type, and reasoning dimension. The tasks span a wide range of historical problems—from 2 factual retrieval based on primary sources to interpretive analysis of manuscripts and images, to interdisciplinary challenges involving archaeology, linguistics, or cultural hi... | https://arxiv.org/abs/2505.20246v1 |
browsing, and a suite of modular tools—including OCR for manuscripts, image provenance retrieval, multilingual translation, literature parsing, and document analysis—into a cohesive framework for historical research. It is designed to mirror the interpretive workflow of human historians, enabling document-level reasoni... | https://arxiv.org/abs/2505.20246v1 |
integration. ChemCrow [ 3], for instance, leverages a suite of 18 expert-designed chemistry tools to perform complex tasks in organic synthesis, drug discovery, and materials design, showcasing enhanced performance in chemistry-related problem-solving. Other advanced approaches include Case-Based Reasoning (CBR), enabl... | https://arxiv.org/abs/2505.20246v1 |
of the historical questions in HLE, although the capabilities of LLMs in historical research is indeed evaluated (as opposed to the benchmarks mainly focused on historical knowledge), the total number of questions is only 56. [ 5] For existing benchmarks, there are limitations in both the scope of the questions and the... | https://arxiv.org/abs/2505.20246v1 |
LLM-based pre-evaluation and professional review (history of rigor and validity of reasoning) to ensure the quality of these questions in terms of both form and content,especially their academic value and challenges to current LLMs. In terms of coverage, HistBench spans 29 ancient and modern languages, over 20 historic... | https://arxiv.org/abs/2505.20246v1 |
historical research and embody key aspects of the research processes that characterize professional historical scholarship. 3.3 Difficulty Stratification 3.3.1 Difficulty Stratification Criteria To support layered evaluation and targeted analysis, all questions in HistBench are categorized into three difficulty levels—... | https://arxiv.org/abs/2505.20246v1 |
of historical research appropriate to its level of difficulty. This multi-tiered design allows the benchmark to evaluate the capabilities of LLMs across a spectrum of historical tasks aligned with real-world scholarly practices. A selection of representative questions from each difficulty level is provided in Appendix ... | https://arxiv.org/abs/2505.20246v1 |
linguistic diversity presents substantive challenges for AI systems. It enables evaluation of multilingual retrieval, cross-lingual historical reasoning, and OCR robustness—addressing a core limitation of existing benchmarks that predominantly rely on English-language content. A complete list of languages and their fre... | https://arxiv.org/abs/2505.20246v1 |
in region-specific expertise (e.g., Dunhuang studies, Slavic paleography) as well as underrepresented areas such as papyrology and Siberian ethnography, ensuring broad civilizational and epistemic representation. In addition to thematic and regional diversity, HistBench evaluates the capabilities of historical research... | https://arxiv.org/abs/2505.20246v1 |
SSIM filtering). Literature Search Scholarly text Peer-review retrieval and PDF parsing (Scholar websites, SpringerDownloadAndParseTool ). File Processing Documents Typed file parsing: PDFTool ,DOCXTool ,XLSXTool ,PPTXTool . OCR Image text Manuscript transcription (Transkribus, Asian-script OCR). Speech Recognition Aud... | https://arxiv.org/abs/2505.20246v1 |
web searches, and Springer Nature’s API, to locate exact phrases, download accessible PDFs, and extract verbatim quotes along with full citation metadata. Functionality. The agent issues a multi-stage search: it first queries Google Books (with random domain rotation and robots.txt compliance) via BookMatchExtractorToo... | https://arxiv.org/abs/2505.20246v1 |
o.w PDF- formatted Output with Metadata Figure 7: Overview of Literature Search Agent. It is specifically designed for academic searching, including a priority search module and web search tools. Functionality. When given an audio file path, the agent verifies its existence and measures its size; if the file exceeds 25... | https://arxiv.org/abs/2505.20246v1 |
4.2.1 Priority Search Protocol A core challenge in automating academic research is ensuring that retrieved information aligns with scholarly standards. To address this, our system introduces a priority-based retrieval mechanism that favors academically reputable sources over general-purpose content. This design choice ... | https://arxiv.org/abs/2505.20246v1 |
is posed to the agent in a fresh session. We limit the total number of tool invocations (agent calls) per question to a fixed budget to prevent infinite loops and to enforce a fair comparison. Both systems output a final answer for each question, which we compare to the ground-truth answer. Additionally, we compare our... | https://arxiv.org/abs/2505.20246v1 |
summary (as described for the HLE benchmark before). Subsequently, we employ LLM-based Judging, where an LLM assesses the response for semantic equivalence to the ground truth, factual completeness, and logical coherence, issuing a binary judgment (“Correct”/“Incorrect”) as the HLE official scoring function released on... | https://arxiv.org/abs/2505.20246v1 |
comparison, both HistAgent and ODR-smolagents are based on GPT-4o. 5.3.3 Performance on GAIA Validation We also test HistAgent on GAIA benchmark, which achieves 60.00% pass@1 accuracy, showing that our domain- specific adaptations don’t hinder HistAgent’s competitive performance on real-world general tasks. We evaluate... | https://arxiv.org/abs/2505.20246v1 |
Vertical Format Sample Entries from HistBench 18 A.2 Sample Questions Across Difficulty Levels To illustrate the scope and structure of questions in HistBench, we present three annotated examples—one for each diffi- culty level. These samples highlight variation in required source processing, historical reasoning, and ... | https://arxiv.org/abs/2505.20246v1 |
previous owner of part of the land . D. Johan Pattukqua was the receiver of the land E. Johan Pattukqua was a friend of the sachem . 20 A.3 Language Distribution in HistBench Table 9 provides the full list of languages represented in the HistBench dataset, along with their frequencies. These include both modern languag... | https://arxiv.org/abs/2505.20246v1 |
contribution of the work. E Details in Experiment E.1 Experiment Compute Resources M1 chips, 512 MB, 20 minutes for 1 question. E.2 Prompt Template JUDGE_PROMPT Template JUDGE_PROMPT = """You are a fair evaluator. Judge whether the following [response] to [question] is semantically consistent with the [correct_answer] ... | https://arxiv.org/abs/2505.20246v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.