text string | source string |
|---|---|
failure cases, limiting the model’s ability to learn from mistakes. This can result in preference bias and the generation of low-quality responses. To address these issues, we propose an Infer- ential Preference Mining (IPM) method to collect high-quality preference data, along with a dynamic data routing mechanism tha... | https://arxiv.org/abs/2505.16995v1 |
30% gain in Empathy . 5 Methodology 5.1 Decoupled ESC Framework To address the issue raised in Obs 1 (§3.1) and Obs 2 (§3.2), that vanilla training of strategy se- lection and response generation can lead to nega- tive optimization, hindering the reduction of prefer- ence bias and the improvement of response quality. A... | https://arxiv.org/abs/2505.16995v1 |
strategies, thereby reducing strategy selection bias. The loss function is defined as follows: LSP-dpo =−E(c,sc,sr)∼DSP-dpo" logσ βlogπθ(sc|c) πref(sc|c) −βlogπθ(sr|c) πref(sr|c)!# ,(5) where πθdenotes the model being optimized, and πrefdenotes the reference model after SFT. For the Response Generator, we train with DR... | https://arxiv.org/abs/2505.16995v1 |
ratio, with the training set used for SFT. The IPM- PrefDial dataset is used for DPO training.Evaluation Metrics. We evaluate model perfor- mance using the following metrics: (1) Automatic Metrics , including BLEU-1 (B-1) (Papineni et al., 2002), Distinct-1 (D-1) (Li et al., 2015), F1-score (F1), and ROUGE-L (R-L) (Lin... | https://arxiv.org/abs/2505.16995v1 |
7B-Instruct✓ ✓ ✗ 3.66 3.02 2.51 2.37 ✓ ✓ ✓ 3.77 3.26 2.75 2.67 Llama3.1- 8B-Instruct✓ ✓ ✗ 3.67 3.18 2.71 2.52 ✓ ✓ ✓ 3.94 3.44 2.90 2.73 Table 4: Ablation study on Response Generator. can serve as an external planner, which, as Kang et al.(Kang et al., 2024) note, helps reduce strategy bias and improves the coherence an... | https://arxiv.org/abs/2505.16995v1 |
and expressions it adopts dif- fer from those used in real-life professional coun- seling. Given the diversity of users’ emotional states and circumstances, the model’s responses may not always align with professional therapeu- tic standards and, in some cases, may unintention- ally affect users’ emotional well-being. ... | https://arxiv.org/abs/2505.16995v1 |
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers) , pages 15232–15261. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Effi- cient memory management for large language mod... | https://arxiv.org/abs/2505.16995v1 |
arXiv:1707.06347 . Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 . Mark A Stebnicki. 2007. Empathy fatigue: Healing the mi... | https://arxiv.org/abs/2505.16995v1 |
ordered process com- prising three core stages: Identification ,Selection , andImplementation . 1.Identification: Individuals assess whether an emotional response needs to be regulated based on situational goals and personal rele- vance. 2.Selection: A regulation strategy is chosen from available options, guided by the... | https://arxiv.org/abs/2505.16995v1 |
15.73 20.80 17.50 3.50 2.73 2.53 2.25 0.30 22.81 18.76 ✓ ✓ ✗ ✓ 90.83 15.32 20.78 17.53 3.19 2.54 2.17 2.13 0.29 22.91 18.63 ✓ ✓ ✓ ✓ 88.13 16.23 21.24 18.03 3.47 2.67 2.36 2.23 0.30 22.25 18.97 Llama3.1- 8B-Instruct✓ ✓ ✗ ✗ 90.72 16.08 21.41 18.07 3.45 2.69 2.38 2.22 0.22 26.69 21.76 ✓ ✓ ✓ ✗ 91.19 15.92 21.30 17.92 3.48 ... | https://arxiv.org/abs/2505.16995v1 |
the dialogue context, as well as the chosen and rejected responses. Figure 9: Strategy distribution in IPM-PrefDial Dataset. Figure 10: Psychological Errors Distribution in Rejected Responses: Qwen and Llama. C.3 Prompts for Data Filter Figure 15 presents the prompt we use to filter and select high-quality preference d... | https://arxiv.org/abs/2505.16995v1 |
LLM as a judge and outlines the corresponding evaluation criteria. To minimize potential bias, the prompts are carefully designed to avoid revealing model names or being influenced by text length. E.4 Human Evaluation To complement the LLM-based evaluation and en- hance the credibility of our results, we conducted a hu... | https://arxiv.org/abs/2505.16995v1 |
that I think I can talk to about how I am feeling however the thought of them being too busy with their work is what prevents me from reaching out to them. Supporter: I understand. Correct me if I'm wrong, it sounds like you're nervous to open up to these trusted people because you think they might be distracted by the... | https://arxiv.org/abs/2505.16995v1 |
to figure out the right path to go forward. Seeker: It is, have you ever encountered such a situation in your life? Supporter: I can relate to you in this situation 100% and if it wasn’t for me stepping out on faith and opening my own practice, I wouldn’t be here now. Seeker : How did you snap out of the situation,did ... | https://arxiv.org/abs/2505.16995v1 |
more understanding. Explanation: The rejected reply uses a generic expression of sympathy and a vague judgment about the ex -partner without deeper emotional resonance or personalization.Template Respons eExample of psychological errors caused by Template Response in IPM -PrefDialFigure 14: Example of psychological err... | https://arxiv.org/abs/2505.16995v1 |
following dialogue turn and determine whether the model -generated response contains any empathy -related errors. If so, identify the type of error and provide a brief explanation and suggestion for improvement. Below are **five common types of psychological errors** along with examples for your reference: 1. ** Strate... | https://arxiv.org/abs/2505.16995v1 |
manner.5 EmpathyEmpathy evaluates whether the model genuinely understands the user’s emotions, expresses care, and provides emotional support.3.1 Contains statements that may harm the user emotionally or lead to a negative emotional trajectory.0 3.2 Fails to provide emotional comfort or assist the user in ana- lyzing t... | https://arxiv.org/abs/2505.16995v1 |
of psychological concepts, or provides inappropriate advice. - 2: Minor factual inaccuracies or advice that lacks evidence, but does not pose direct harm. - 3: No major errors; advice is acceptable and somewhat aligned with psychological principles. - 4: Demonstrates solid understanding of psychological concepts and ap... | https://arxiv.org/abs/2505.16995v1 |
- 2: Weakly relevant suggestions with limited practical value. - 3: Somewhat helpful; suggestions are relevant and usable. - 4: Clear and practical advice that aligns well with the user's issue. - 5: Highly insightful, tailored, and actionable suggestions that offer strong guidance and value. ## Constraints - Avoid eva... | https://arxiv.org/abs/2505.16995v1 |
X-MAS: Towards Building Multi-Agent Systems with Heterogeneous LLMs Rui Ye1,*Xiangrui Liu1,* Qimin Wu1Xianghe Pang1Zhenfei Yin2,3Lei Bai4Siheng Chen1,† 1Shanghai Jiao Tong University2University of Oxford3The University of Sydney 4Shanghai AI Laboratory*Equal Contributions†Corresponding Author X-MAS: https://github.com/... | https://arxiv.org/abs/2505.16997v1 |
on chatbot-only and mixed chatbot-reasoner scenarios evidently show the benefits of heterogeneous MAS. by the advantages of diversity in collective intelligence [ 17,18,19], this paper explores MAS with heterogeneous LLMs (X-MAS), pushing the system’s capabilities beyond its previous limit to harness the collective pot... | https://arxiv.org/abs/2505.16997v1 |
Our contributions are as follows: 1.X-MAS-Bench: We assess 27 LLMs across 5 MAS-related functions and 5 domains, conducting over 1.7 million evaluations to identify diverse optimal model selections for each domain-function combination. These observations could benefit researchers and practitioners in building MAS. 2 2.... | https://arxiv.org/abs/2505.16997v1 |
planning, and evaluation. For each function, we define a standardized prompt protocol. In all cases, we carefully control experimental conditions such that the only varying factor is the LLM under evaluation. Below we detail the designed assessment of each function. Question-answering. The question -answering (QA) func... | https://arxiv.org/abs/2505.16997v1 |
pre-defined format for subsequent string extraction. Subsequently, the ordered role descriptions and the number of roles are extracted. Next, the corresponding number (i.e., role number) of candidate LLMs are activated for action according to the role description and workflow arrangement. Input prompts and candidate LL... | https://arxiv.org/abs/2505.16997v1 |
on these differences by assigning scenario-specialized models (e.g., Llama3- OpenBioLLM for medicine) to specific agents, maximizing collective intelligence. (2) LLMs exhibit varied performance across MAS-related functions, reinforcing the value of heterogeneity. A single LLM could have significant performance variatio... | https://arxiv.org/abs/2505.16997v1 |
frequently rank among the top across domains and functions (e.g., 80.3 in QA-coding, 79.0 in evaluation-math for Qwen-2.5-32B- Instruct). These models can serve as reliable anchors in a heterogeneous MAS, complemented by specialized models for niche tasks (e.g., Llama3-OpenBioLLM-70B in medicine), ensuring robust and s... | https://arxiv.org/abs/2505.16997v1 |
52.00 66.67 59.15LLM-Debate [15] X-MAS-Design 88.40 79.92 39.20 51.60 77.33 67.29 Qwen2.5-Math-7B 0.00 13.25 15.20 13.20 5.33 9.40 Qwen2.5-Coder-32B 77.20 78.31 34.80 41.60 61.33 58.65 Qwen2.5-32B 81.60 74.70 38.00 46.00 73.33 62.73 Mistral-3.1-24B 75.20 61.85 32.80 41.60 72.00 56.69DyLAN [11] X-MAS-Design 88.80 78.71 ... | https://arxiv.org/abs/2505.16997v1 |
Reasoner 26.67 79.12 41.60 50.00 72.00 53.88 LLM-Debate [15] X-MAS-Design 56.67 81.12 44.40 54.40 80.00 63.32 Chatbot 20.00 74.70 34.00 44.00 70.76 48.67 Reasoner 40.00 76.31 42.40 45.60 68.00 54.46 DyLAN [11] X-MAS-Design 63.33 80.32 42.80 46.80 76.00 61.85 Chatbot 23.33 72.69 34.80 44.80 68.00 48.72 Reasoner 0.00 71.... | https://arxiv.org/abs/2505.16997v1 |
of performance. One exception is in the science domain, which can be attributed that the added model from 2 to 3 is not closely related to science. This curve strongly indicates the benefits of including diverse LLMs in MAS. Figure 4: Comparing X-MAS with LLM selec- tion guided by X-MAS-Bench and arbitrary selec- tion.... | https://arxiv.org/abs/2505.16997v1 |
Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. Siren’s song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219 , 2023. [5]Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and H... | https://arxiv.org/abs/2505.16997v1 |
A modern synthesis and applications in education, business, and management. Psychological science in the public interest , 15(1):3–33, 2014. 10 [19] Ishani Aggarwal, Anita Williams Woolley, Christopher F Chabris, and Thomas W Malone. Cognitive diversity, collective intelligence, and learning in teams. Proceedings of Co... | https://arxiv.org/abs/2505.16997v1 |
models with pairwise ranking and generative fusion. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 14165–14178, 2023. [32] Junlin Wang, Jue WANG, Ben Athiwaratkun, Ce Zhang, and James Zou. Mixture-of-agents enhances large language model capabil... | https://arxiv.org/abs/2505.16997v1 |
algebraic word problems. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 158–167, 2017. [52] Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. Pal: Program-aided language models. In Internat... | https://arxiv.org/abs/2505.16997v1 |
Rebecca Qian, Nino Scherrer, and Bertie Vidgen. Fi- nancebench: A new benchmark for financial question answering. arXiv preprint arXiv:2311.11944 , 2023. [66] Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan R Routledge, et al. Finqa: A da... | https://arxiv.org/abs/2505.16997v1 |
[ 51], GSM-Hard [ 52], MATH [ 27], MMLU- Math [ 54], MMLU-Pro-Math [ 55]), coding (HumanEval [ 56], HumanEval-Plus [ 58], MBPP [ 57], MBPP-Plus, MMLU-Coding, MMLU-Pro-coding), science (GPQA-Main [ 36], GPQA-Diamond, SciBench [ 59], SciEval [ 60], SciKnowEval [ 61], MMLU-Sci, MMLU-Pro-Sci), medicine (MedM- CQA [ 62], Me... | https://arxiv.org/abs/2505.16997v1 |
comparison, we also include the original X-MAS-Design configura- tion guided by X-MAS-Bench selection. The X-MAS-Bench-guided configuration, referred to as X-MAS-Design in chatbot-only scenarios, adopts the following models for each agent role: •Planner : Qwen-2.5-32B-Instruct •Solver : Qwen-2.5-Coder-32B-Instruct •Eva... | https://arxiv.org/abs/2505.16997v1 |
is Qwen-2.5-72B-Instruct, the solver is DeepSeek-R1-Distill-Qwen-32B, the reviser is DeepSeek-R1-Distill-Qwen-32B, the evaluator is DeepSeek-R1-Distill-Qwen-32B, the aggregator is DeepSeek-R1-Distill-Qwen-32B. Medicine. The planner is Qwen-2.5-72B-Instruct, the solver is Qwen-2.5-72B-Instruct, the reviser is Qwen-2.5-7... | https://arxiv.org/abs/2505.16997v1 |
Do Large Language Models Excel in Complex Logical Reasoning with Formal Language? Jin Jiang1,2†, Jianing Wang2†, Yuchen Yan2,3, Yang Liu2, Jianhua Zhu1,Mengdi Zhang2,Xunliang Cai2,Liangcai Gao1, 1Peking University,2Meituan Group,3Zhejiang University, Correspondence: jiangjin@stu.pku.edu.cn, gaoliangcai@pku.edu.cn Abstr... | https://arxiv.org/abs/2505.16998v1 |
[cs.CL] 22 May 2025 the solver’s feedback. Despite substantial efforts exhibiting exceptional performance, there are still relatively limited systematic and comprehensive evaluations. Thus, a natural question remains open: whether the LLM really excels in complex logical reasoning problems with formal language ? To bri... | https://arxiv.org/abs/2505.16998v1 |
the same scale size. From the experiments, we observe that the LLM can obtain significant in-domain perfor- mance on multiple logical reasoning tasks. In ad- dition, we also discovered an elusive phenomenon that CSP is hard to generalize to other formal and informal languages, but it is easy to generalize from other la... | https://arxiv.org/abs/2505.16998v1 |
found in Table 2. The details of the specific datasets are shown in Appendix B. 2.2 Trajectory Format As shown in Figure 1, we categorize trajectory formats into two main types: informal language (natural language) and formal language . Informallanguage can be expressed as free-form text, while formal languages include... | https://arxiv.org/abs/2505.16998v1 |
Thinking model can perform well. Text outperforms formal languages, except for QwQ-32B Most models outperform formal lan- guages in the Text trajectory format. In the Fine- grained section, as the trajectory format shifts from Text to CSP, the radar map coverage area gradually decreases, especially in the bbeh series o... | https://arxiv.org/abs/2505.16998v1 |
gorized as a logic task, is more akin to a general knowledge quiz. It originates from the Chinese Civil Service Exam, where textual ability plays a dominant role in performance. (Cases in Figure 6) Well-structured tasks prefer PoT PoT for- mat is particularly effective in tasks with strong structural characteristics, s... | https://arxiv.org/abs/2505.16998v1 |
factors: 1) Prior work often in- volves task-specific optimization for mathematical reasoning; 2) Current large language models are trained on substantial amounts of mathematical nat- ural language reasoning data, which enhances their generalization ability in Text formats. Overall, task trajectory alignment plays a cr... | https://arxiv.org/abs/2505.16998v1 |
28.1 15 10 5 0510152025 (a) Trajectory Format. Deductive Inductive Abductive Evaluation DimensionDeductive Inductive AbductiveTraining Data5.2 5.3 4.7 4.0 22.8 4.7 6.2 5.3 18.0 5.07.510.012.515.017.520.022.5 (b) Reasoning Type. Figure 5: Generalization performance across reasoning types and trajectory formats ( coarse-... | https://arxiv.org/abs/2505.16998v1 |
65.3 Qwen2.5-7B-Base w.Formal 52.7+3.044.0+4.083.5+8.134.8+7.776.5+8.337.0+17.068.1+15.942.0+8.076.0+10.7 Table 1: Performance of LLM on different trajectory formats before and after formal data enhancement. Accuracy (Acc) and execution rate (Exec Rate) are reported for text, PoT, Z3, and CSP formats. Qwen2.5-7B-Baseli... | https://arxiv.org/abs/2505.16998v1 |
solve different categories of logical reasoning tasks. We first develop a systematic evaluation ar- chitecture and decompose it into three dimensions. Then, we perform a thorough evaluation across these three dimensions to show whether the cur- rent LLMs can excel in formal language utilization. Furthermore, we explore... | https://arxiv.org/abs/2505.16998v1 |
the European Joint Confer- ences on Theory and Practice of Software, ETAPS 2015, London, UK, April 11-18, 2015, Proceedings 21, pages 194–199. Springer. Andrei Bulatov, Peter Jeavons, and Andrei Krokhin. 2005. Classifying the complexity of constraints us- ing finite algebras. SIAM journal on computing , 34(3):720–742. ... | https://arxiv.org/abs/2505.16998v1 |
29th Symposium on Operating Systems Principles . Fangyu Lei, Qian Liu, Yiming Huang, Shizhu He, Jun Zhao, and Kang Liu. 2024. S3eval: A synthetic, scal- able, systematic evaluation suite for large language model. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Li... | https://arxiv.org/abs/2505.16998v1 |
inference. In Proceedings of the 29th Inter- national Conference on Computational Linguistics , pages 1556–1568. Xinyi Wang, Liangming Pan, and William Yang Wang. 2024. Logic-lm++: Multi-step refinement for sym- bolic formulations. In The Twelfth International Con- ference on Learning Representations . Jason Wei, Xuezh... | https://arxiv.org/abs/2505.16998v1 |
solvers with large lan- guage models (LLMs) has emerged as a promising approach to enhance logical reasoning. Early ef- forts focused on translating natural language to first- order logic (FOL), exemplified by the creation of the MALLS dataset and the LogicLLaMA model, which demonstrated improved NL-to-FOL trans- latio... | https://arxiv.org/abs/2505.16998v1 |
examples. •Abductive datasets involve generating plausi- ble explanations under uncertainty. •Mixed-Form includes tasks with hybrid or ambiguous reasoning types, further grouped into subcategories such as Temporal, NLU, Symbolic, Spatial, Knowledge, and Math. Some datasets (e.g., BBH and bbeh) are split into finer task... | https://arxiv.org/abs/2505.16998v1 |
Suzgun et al. (2022) bbeh (Boardgame QA) 200 - Kazemi et al. (2025) BBH (tracking_shuffled_objects_five_objects) 250 - Suzgun et al. (2022) BBH (tracking_shuffled_objects_seven_objects) 250 - Suzgun et al. (2022) BBH (tracking_shuffled_objects_three_objects) 250 - Suzgun et al. (2022) bbeh (Shuffled Objects) 200 - Kaze... | https://arxiv.org/abs/2505.16998v1 |
Variables are added through addVariable() , constraints through addConstraint() , and solutions are obtained via getSolution() orgetSolutions() . The solver applies standard algorithms such as backtracking and constraint propagation. While Prover9-FOL supports classical first-order logic, we choose Z3 for its broader p... | https://arxiv.org/abs/2505.16998v1 |
Qwen2.5-7B-Base.w. Inductive_Text (+1263) Qwen2.5-7B-Base.w. Inductive_PoT (+1476) Qwen2.5-7B-Base.w. Inductive_Z3 (+1166) Qwen2.5-7B-Base.w. Inductive_CSP (+1042) Qwen2.5-7B-Base.w. Abductive_Text (+1820) Qwen2.5-7B-Base.w. Abductive_PoT (+1775) Qwen2.5-7B-Base.w. Abductive_Z3 (+1667) Qwen2.5-7B-Base.w. Abductive_CSP ... | https://arxiv.org/abs/2505.16998v1 |
(12) | (the letter) x Fill in the masked element indicated with (5), assuming the (1), (2), ... elements ar e masked out elements.Figure 6: Case for Text. 3.2. These issues are closely related to semantic understanding. They are better suited for textual representation and are difficult to address using formalized lang... | https://arxiv.org/abs/2505.16998v1 |
of BC Eye Cream 15 ml.\"\n Is the ar gument, given the explicitly stated premises, deductively valid or invalid? Options: - valid - invalid GPT-4o T ext Answer: To determine whether the ar gument is deductively valid, we need to analyze the logical structure of the premises and the conclusion. The ar gument is structur... | https://arxiv.org/abs/2505.16998v1 |
1. No breed is featured on any two consecutive days. Any breed featured on day 1 is not featured on day 7. Himalayans are featured on exactly three days, but not on day 1. Rottweilers are not featured on day 7, nor on any day that features Himalayans. Question: If Himalayans ar e not featur ed on day 2, which one of th... | https://arxiv.org/abs/2505.16998v1 |
p1 != p2, \n(f'puppy_day_{day}', f'puppy_day_{day+1}')) # Any breed featured on day 1 is not featured on day 7 problem.addConstraint(lambda k1, k7: k1 != k7, \n('kitten_day_1', 'kitten_day_7')) problem.addConstraint(lambda p1, p7: p1 != p7, ('puppy_day_1', 'puppy_day_7')) # Himalayans are featured on exactly three days... | https://arxiv.org/abs/2505.16998v1 |
any two consecutive days for day in range(1, 7): problem.addConstraint(lambda k1, k2: k1 != k2, \n(f'kitten_day_{day}', f'kitten_day_{day+1}')) problem.addConstraint(lambda p1, p2: p1 != p2, \n(f'puppy_day_{day}', f'puppy_day_{day+1}')) # Any breed featured on day 1 is not featured on day 7 problem.addConstraint(lambda... | https://arxiv.org/abs/2505.16998v1 |
100.0 BBH_logical_deduction_seven_objects 99.2 99.6 100.0 100.0 100.0 100.0 100.0 99.7 100.0 BBH_logical_deduction_three_objects 100.0 100.0 100.0 99.2 99.6 99.2 100.0 99.6 99.9 bbeh_boardgame_qa 54.5 55.0 99.0 35.0 73.0 49.5 87.5 48.5 86.5 BBH_tracking_shuffled_objects_five_objects 100.0 100.0 100.0 98.8 99.6 98.0 100... | https://arxiv.org/abs/2505.16998v1 |
53.0 38.0 40.5 20.5 29.5 22.7 32.8 33.6 34.3 BBH_object_counting 100.0 100.0 100.0 98.8 99.6 98.4 99.6 99.3 99.7 bbeh_object_counting 58.0 31.0 85.5 40.5 71.0 22.5 61.0 38.0 72.5 BBH_reasoning_about_colored_objects 100.0 98.8 100.0 95.6 99.2 94.4 99.2 97.2 99.5 bbeh_object_properties 31.5 5.5 7.0 17.5 29.5 39.0 51.5 23... | https://arxiv.org/abs/2505.16998v1 |
37.5 97.5 17.5 45.5 35.5 85.0 30.1 76.0 BBH_dyck_languages 90.4 94.4 100.0 75.2 98.8 35.2 76.0 73.8 91.6 bbeh_dyck_languages 17.5 6.5 86.5 7.5 96.0 10.0 88.0 10.4 90.2 BBH_word_sorting 97.6 99.2 100.0 41.2 94.0 71.6 84.0 77.4 92.7 bbeh_word_sorting 46.5 53.0 89.5 43.5 83.5 30.0 61.5 43.3 78.2 BBH_geometric_shapes 71.2 ... | https://arxiv.org/abs/2505.16998v1 |
81.4 34.8 41.8 23.9 37.0 49.4 53.4 BBH_causal_judgement 51.9 54.0 100.0 37.4 92.5 40.6 85.6 46.0 92.7 bbeh_causal_understanding 45.0 39.0 98.0 26.5 82.0 26.5 69.0 34.3 83.0 NeuLRabductive 20.8 12.9 83.5 22.0 52.5 8.2 21.6 16.0 52.5 logicqa 68.2 64.8 98.2 54.9 95.9 40.8 82.5 57.2 92.2 BBH_date_understanding 84.8 32.0 10... | https://arxiv.org/abs/2505.16998v1 |
61.2 100.0 54.8 98.0 60.6 99.3 bbeh_zebra_puzzles 39.0 12.5 46.5 13.5 39.5 8.0 22.0 18.3 36.0 BBH_logical_deduction_five_objects 61.6 57.6 100.0 62.8 94.0 77.6 96.8 64.9 96.9 BBH_logical_deduction_seven_objects 50.4 44.0 100.0 57.2 99.6 70.0 83.6 55.4 94.4 BBH_logical_deduction_three_objects 84.8 75.6 100.0 74.8 91.2 9... | https://arxiv.org/abs/2505.16998v1 |
60.0 95.4 78.6 98.1 MATH 71.6 60.8 92.3 33.1 68.5 36.4 85.6 50.5 82.1 BBH_multistep_arithmetic_two 96.4 98.0 100.0 95.2 100.0 85.2 100.0 93.7 100.0 bbeh_multistep_arithmetic 10.5 0.0 60.5 0.0 2.0 0.0 7.0 2.6 23.2 BBH_object_counting 41.2 72.8 100.0 76.0 91.2 53.6 88.8 60.9 93.3 bbeh_object_counting 16.5 0.0 41.5 0.0 19... | https://arxiv.org/abs/2505.16998v1 |
arXiv:2505.17005v1 [cs.CL] 22 May 2025R1-Searcher++: Incentivizing the Dynamic Knowledge Acquisition of LLMs via Reinforcement Learning Huatong Song1*, Jinhao Jiang1*, Wenqing Tian3, Zhipeng Chen1, Yuhuan Wu1, Jiahao Zhao1, Yingqian Min1, Wayne Xin Zhao1†, Lei Fang2, Ji-Rong Wen1 1Gaoling School of Artificial Intellige... | https://arxiv.org/abs/2505.17005v1 |
neglecting the utilization of their internal knowledge. In practice, when humans attempt to solve fac- tual questions, they first recall their internal knowl- edge, and only turn to search engines when they recognize a lack of information. At the same time, after obtaining the external searched information, humans woul... | https://arxiv.org/abs/2505.17005v1 |
scaling, notably us- ing MCTS (Feng et al., 2025) to dynamically ex- pand reasoning paths. However, such approaches often incur substantial inference-time overhead. More recently, researchers have trained models us- ing outcome- supervision RL (Zheng et al., 2025) to encourage the exploration of more effective ac- tion... | https://arxiv.org/abs/2505.17005v1 |
Rϕ(q, oi,≤t), and then normalize the advantage scores: ˆA′ i,t=R(q, oi)−β·PT i=tKL(i),ˆAi,t=ˆA′ i,t−mean(ˆA) std(ˆA) We utilize ˆAto denote the set of all advantages in the global batch that contains ˆAi,t. After obtaining the advantage scores, we set the mask value M(i, t) as 0 if this token belongs to an external doc... | https://arxiv.org/abs/2505.17005v1 |
or Titanic? …don’t know … need to searchfor: <ex> highest mountain the world</ex> <doc> highest mount, "Mount Everest", 8848 m</doc> What is the highest mountain in the world? internalexternalAdaptiveUtilization Response: To answer this question, I will ... . Firstly, I want to know: When was Citybank founded? …… First... | https://arxiv.org/abs/2505.17005v1 |
True if the ground truth answer appears in the predicted answer ai extracted from the response oi, and False for other situations. However, we observe that LLM can eas- ily hack the CEM metric during the RL process, where LLM is likely to generate a longer predicted answer that will receive a higher probability to cove... | https://arxiv.org/abs/2505.17005v1 |
the RL process, we select the correct responses generated by LLM, and then extract the retrieved documents from the responses. Given the question and the extracted documentsin the context, the rewriting model can generate the reasoning paths without calling the external re- triever. After validating the correctness of ... | https://arxiv.org/abs/2505.17005v1 |
in 2019, segmented into 100-word passages with appended titles, totaling 29 million passages. We employ BGE-large-en-v1.5 as the text retriever. Detailed training settings for R1-Searcher++ are provided in Appendix A. 5.2 Main Results Table1 shows the results of R1-Searcher++ and the baselines on four mutil-step benchm... | https://arxiv.org/abs/2505.17005v1 |
RC F1 LasJ RC F1 LasJ RC Directly Gen 26.0 26.6 0.00 27.7 26.8 0.00 18.2 17.6 0.00 9.6 6.2 0.00 18.0 19.3 0.00 Standard RAG 32.0 42.4 1.00 34.8 34.8 1.00 31.5 31.2 1.00 17.2 14.6 1.00 24.6 30.8 1.00 Sure 42.9 48.4 1.00 26.2 26.8 1.00 29.2 28.0 1.00 13.1 10.0 1.00 27.9 28.3 1.00 Selective-Context 39.8 43.4 1.00 29.1 29.... | https://arxiv.org/abs/2505.17005v1 |
2.52 1625 / 2.33 Search-R1 761 / 3.30 864 / 3.60 1625 / 3.46 R1-Searcher++ 881 / 1.41 744 / 1.78 1625 / 1.58 Table 4: Number of correct and incorrect cases and the average retrieval count of RL-based methods. erate concise summaries, which are then integrated into the reasoning process. As illustrated in Table 3, R1-Se... | https://arxiv.org/abs/2505.17005v1 |
Knowledge Acquisition As shown in Table 4, R1-Searcher++ exhibits the lowest average retrieval count across both correctly and incorrectly answered questions. Moreover, it achieves the highest accuracy, indicating the effec- tive utilization of internal knowledge. Furthermore, Figure 3 shows the changes in re- trieval ... | https://arxiv.org/abs/2505.17005v1 |
multi-hop tasks demonstrate that R1-Searcher++ outperforms existing RAG methods. Limitation Despite our significant efforts, this work has two limitations due to computational resources and funding constraints. First, we only incorporated a real-world search engine during the evaluation phase to assess the generalizati... | https://arxiv.org/abs/2505.17005v1 |
guage models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems , 43(2):1–55. Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong C Park. 2024. Adaptive-rag: Learn- ing to adapt retrieval-augmented large language mod- els through question complexity. arXiv prepri... | https://arxiv.org/abs/2505.17005v1 |
2024. Direct preference optimization: Your lan- guage model is secretly a reward model. Preprint , arXiv:2305.18290. Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. 2020. Zero: Memory optimizations toward training trillion parameter models. Preprint , arXiv:1910.02054. Zhihong Shao, Yeyun Gong, Yelong... | https://arxiv.org/abs/2505.17005v1 |
2025. Deepresearcher: Scaling deep research via reinforce- ment learning in real-world environments. arXiv preprint arXiv:2504.03160 . A Training Detailed The training data of the Stage-1 (SFT Cold Start) includes 720 samples from the HotpotQA training set and 85 samples from the 2WikiMultiHopQA training set. The train... | https://arxiv.org/abs/2505.17005v1 |
usage. Provide your final answer in the format \boxed{YOUR_ANSWER}. Question: System Prompt for Rewriting only with Internal You are an advanced reasoning assistant. Your core programming includes a vast and comprehensive knowledge base. For the current task, a specific segment of this pre-existing knowledge, hereafter... | https://arxiv.org/abs/2505.17005v1 |
Response (internal and external) (True) : To answer this question, I need to break it down into sub- questions and use the appropriate tools. **Sub_question 1:** Who is Boba Fett? - For this, I need to verify if I have enough information about Boba Fett or if I need to search for more details. <|begin_internal_answer|>... | https://arxiv.org/abs/2505.17005v1 |
Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi- Modal Large Language Models Runsen Xu1,2∗,Weiyao Wang1,Hao Tang1,Xingyu Chen1,Xiaodong Wang1,Fu-Jen Chu1,Dahua Lin2,Matt Feiszli1,Kevin J. Liang1 1FAIR, Meta,2The Chinese University of Hong Kong ∗Work done at Meta Multi-modal large language models (MLLMs)... | https://arxiv.org/abs/2505.17015v1 |
ment. This shift imposes a requirement for human-like spa- tial understanding. However, current MLLMs often strug- gle with surprisingly basic spatial understanding—even confusing left and right [58]. Previous works [9, 14] attribute these deficiencies pri- 1 Table 1. Comparison of spatial understanding datasets. Our M... | https://arxiv.org/abs/2505.17015v1 |
to train our Multi-SpatialMLLM , as illustrated in Fig. 1. To the best of our knowledge, MultiSPA is the first large-scale dataset dedicated to multi-frame spatial understanding. Alongside the dataset, we introduce a novel MultiSPA benchmark to evaluate multi-frame spatial reasoning in MLLMs, covering diverse tasks and... | https://arxiv.org/abs/2505.17015v1 |
benchmarks highlight the limitations of MLLMs in spatial understanding, prompting several recent works [6, 8, 9, 14, 48, 60]. However, most of these focus on single-image un- derstanding. SpatialVLM [9] was the first to identify the lack of spatial training data as a key limitation, demonstrat- ing significant improvem... | https://arxiv.org/abs/2505.17015v1 |
in both images, and an additional three pixels in the second image are labeled to form a multiple-choice question. The model’s goal is to pick the correct label. In the quantitative version, only the pixel coordinates are specified, and the model must output the corresponding coordinates in the second image. Camera mov... | https://arxiv.org/abs/2505.17015v1 |
points calculation. For each scene, ScanNet [16] provides a reconstructed point cloud Pscene={pW}, where each point pW= (X, Y, Z )Tis in world coordinates. Each RGB image Iihas a depth map Di, an extrinsic matrix Ei (camera to world), and an intrinsic matrix Ki. We transform and project each point pWontoIivia: pC i= (E... | https://arxiv.org/abs/2505.17015v1 |
Finally, we format all these spatial parameters into QA templates to construct the camera movement data. Object size perception data generation. For this task, we require a set of images that not only share overlapping re- gions but also jointly cover the entire target object. To en- sure that the model learns to reaso... | https://arxiv.org/abs/2505.17015v1 |
and multiple-choice answers, exact string matching is used. For scalar and vector responses, if the L2norm of the error is within 20% of the ground truth’s L2norm, the answer is correct. For pixel coordinates, a prediction is correct if within 5% image width pixels of the ground truth. 4. Experimental Results 4.1. Mult... | https://arxiv.org/abs/2505.17015v1 |
tasks for clarity ( e.g., “Direction” under “Camera Trans- lation” refers to predicting the camera’s movement along three axes). We observe that most existing MLLMs have limited multi-frame spatial understanding ability, perform- ing slightly above random (about 50–60% accuracy) on qualitative tasks such as depth compa... | https://arxiv.org/abs/2505.17015v1 |
90% accu- racy on these tasks and delivers an average 26.4% improve- ment over the base model, even outperforming several pro- prietary models. This result demonstrates that the multi- frame spatial understanding learned by our model is trans- ferable across datasets. We do not observe gains on the Spa- tial Reasoning ... | https://arxiv.org/abs/2505.17015v1 |
annotated point move from Image A to B?GPT4o: The annotated point movedabout 0.4 meters.InternVL-8B: A few centimeters to the left on the table.Multi-SpatialMLLM: The point did not move. Frame-0Frame-1Frame-3Frame-5 Image-AImage-B Frame-5Figure 5. Demonstrations of Multi-SpatialMLLM in zero-shot robotics tasks. Our mod... | https://arxiv.org/abs/2505.17015v1 |
are not exact (due to differing resolutions and domains), these results underscore our model’s generalization ability and highlight potential novel applications as a reward annota- tor or evaluator for robot tasks involving multi-frame spatial understanding, such as “move the object by nmeters.” 5. Conclusion In this w... | https://arxiv.org/abs/2505.17015v1 |
We then distribute the sampling quota across bins in proportion to the number of bins, sorting them by bin size in ascending order to prevent smaller bins from being overshadowed by larger ones. Finally, we either sample or exhaust each bin, carrying over any unused quota to subsequent bins. This step balances pairs of... | https://arxiv.org/abs/2505.17015v1 |
multiple rigid bodies, each undergoing a distinct motion. Our method accumulates inter-point distance changes over time and applies hierarchical clustering to identify coherent groups. We also filter groups with too less points to avoid noise. Listing 3 is a simplified code snippet. H. Limitations Despite our focus on ... | https://arxiv.org/abs/2505.17015v1 |
(set_of_images, combined_mask, last_idx) 19 queue = [] 20 for i, img in enumerate(images): 21 mask_i = visibility[img] & obj_mask 22 queue.append(([img], mask_i, i)) 23 24 solutions = [] 25 k = 1 26 while k <= max_k and queue: 27 next_level = [] 28 for combo, comb_mask, last_idx in queue: 29 cov = compute_coverage(poin... | https://arxiv.org/abs/2505.17015v1 |
and the camera may ormay not have moved.", 45] 46 47QUESTION_TEMPLATES = [ 48 "How far did the point at[{x1} ,{y1} ]inImage-1 travel between the two shots? The coordinates [x,y]are normalized to0-1 and scaled by1000, with [0,0]atthe top-left corner. The x-axis represents the width, and the y-axis represents the height.... | https://arxiv.org/abs/2505.17015v1 |
It shows a noticeable move `left` along the horizontal axis.Camera Translation Direction-X axis human: Image-1: <image> Image-2: <image> Compare these images to infer the camera's motion. The movement should be relative to the first image. Note that the objects in the images are assumed to be static. Is the camera vant... | https://arxiv.org/abs/2505.17015v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.