text string | source string |
|---|---|
Specifically, we verbalize the refined successful trajectory to an abstract plan using GPT-4o-mini1to replace the original chain-of-thoughts {ht}T t=1. At a cur- rent time step tc, a plan is generated by refining the trajectory history τt<tc, current observation otc, current action atc, and future trajectory τt>tc 1We ... | https://arxiv.org/abs/2505.20013v1 |
core idea behind rollback is to equip M with the ability to dynamically evaluate the valid- ity of its actions and their consequences at each step of the trajectory. When the outcome of an ac- tion deviates from M’s plans or expectations, the model identifies the erroneous action atand the corresponding state otwhere t... | https://arxiv.org/abs/2505.20013v1 |
30.23 28.26 24.68 5.7 25.0 + R EJ. SAMPLING 34.88 6.98 24.39 38.10 22.73 34.15 48.78 19.05 36.96 29.50 7.5 33.0 + Q WQ D ISTILL 32.56 20.93 23.81 40.48 29.55 34.15 31.71 25.58 36.96 30.65 18.9 52.0 +WEBCOT 39.53 27.91 30.95 59.52 38.64 43.90 46.34 27.91 54.35 41.04 20.8 56.0 Table 1: Performance comparison across WebV ... | https://arxiv.org/abs/2505.20013v1 |
data generation approaches for fine-tuning Llama-3.3-70b-Instruct: (i) Rejection Sampling (Rej. Sampling), (ii) successful trajectories sam- pled by QWQ-32B (QwQ Distill), and (iii) the WEBCOT data proposed in our work for training. 5.3 Main Results Table 1 presents the performance comparison be- tween our method, WEBC... | https://arxiv.org/abs/2505.20013v1 |
newly verbalized CoT will lead to more improvements in general. 5.4.2 Effects of Rationale Verbalization In Reflection & Lookahead and Branching, newly successful queries are executed and their corre- sponding trajectories are added to the training set. InWEBCOT, we further verbalize lookahead plan- ning and action sel... | https://arxiv.org/abs/2505.20013v1 |
only the flawed ones, and find that this leads to degraded perfor- mance. 0 2000 4000 6000 8000050100150Countwebdreamer branching Figure 4: Distribution of generated token number per query for WebDreamer (branching) and WEBCOT. A significantly smaller amount of reasoning tokens are required for our method. 5.4.4 Effect... | https://arxiv.org/abs/2505.20013v1 |
Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 techni- cal report. arXiv preprint arXiv:2303.08774 . Renat Aksitov, Sobhan Miryoosefi, Zonglin Li, Daliang Li, Sheila Babayan, Kavya Kopparapu, ZacharyFisher, Ruiqi Guo, Sushant Prakash, Pranesh Srini- vasan, Manzil Zahee... | https://arxiv.org/abs/2505.20013v1 |
CoRR , abs/2410.19609. Minda Hu, Licheng Zong, Hongru Wang, Jingyan Zhou, Jingjing Li, Yichen Gao, Kam-Fai Wong, Yu Li, and Irwin King. 2024. SeRTS: Self-rewarding tree search for biomedical retrieval-augmented generation. In Findings of the Association for Computational Lin- guistics: EMNLP 2024 , pages 1321–1335, Mia... | https://arxiv.org/abs/2505.20013v1 |
A compact reasoning model with reinforcement learning scaling. https:// huggingface.co/Qwen/QwQ-32B . Apache 2.0 Li- cense. Model available at https://huggingface. co/Qwen/QwQ-32B . Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Re- flexion: language agents with verbal reinforc... | https://arxiv.org/abs/2505.20013v1 |
Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net. Hongming Zhang, Ruixin Hong, and Dong Yu. 2025a. Streaming looking ahead ... | https://arxiv.org/abs/2505.20013v1 |
Details of WebDreamer The prompt designs for proposing possible actions, (h(i) t, a(i) t)k i=1, simulating the outcomes of these ac- tions, and evaluating them based on the simulations are provided in Table 15, Table 16, and Table 17, respectively.C Details of Finetuning We use Megatron-LM2framework to finetune all the... | https://arxiv.org/abs/2505.20013v1 |
objective: This is the task you’re trying to complete. The current observation (web page’s accessibility tree): This is a simplified representation of the webpage, providing key information. Optionally, you may be provided with a screenshot of the webpage. You should pay close attention to the screenshot to make decisi... | https://arxiv.org/abs/2505.20013v1 |
next observation or result will be after performing the action.}} Critic Evaluation: {{Your rationale on the effectiveness of the action as well as a score from 0 (poor performance) to 1 (excellent performance), judging the corresponding action’s s effectiveness.}} 2. ... (continue with subsequent steps as needed in th... | https://arxiv.org/abs/2505.20013v1 |
Prompt for generating alternative thoughts and actions. {Environment Description } {Agent Hints } Table 13: System prompt for web agent. {Environment Description } {Agent Hints } Previously, the action " {a′ j}" has been attempted, and this action will not lead to the task completion. Please provide an action for going... | https://arxiv.org/abs/2505.20013v1 |
link ’Accessories’[37] link ’Apple Card’[38] link ’iOS 18’[39] link ’Shop iPhone’[40] StaticText ’Get $170–’[41] StaticText ’$630 in credit toward iPhone 16 or iPhone 16 Pro when you trade in iPhone 12 or higher.’[42] link ’Footnote * symbol’[43] link ’Shop iPhone’[44] StaticText ’Designed to be loved.’[45] button ’Pla... | https://arxiv.org/abs/2505.20013v1 |
refining the query or seeking additional information. However, given the context, it is likely that the results will provide the necessary information to complete the task. ACTION: “‘click [11]“‘ Table 18: Examples of hallucination in ˆDc L. Parts with hallucination are marked in red. EXCLUDED_WEBSITES_MIND2WEB = { ’ex... | https://arxiv.org/abs/2505.20013v1 |
arXiv:2505.20014v1 [cs.CL] 26 May 2025Does Rationale Quality Matter? Enhancing Mental Disorder Detection via Selective Reasoning Distillation Hoyun Song Huije Lee Jisu Shin Sukmin Cho Changgeon Ko Jong C. Park∗ School of Computing Korea Advanced Institute of Science and Technology (KAIST) {hysong,huijelee,jisu.shin,nel... | https://arxiv.org/abs/2505.20014v1 |
ele v anceFigure 1: Illustration of varying rationale quality. R1 effec- tively connects the social media post to specific symptoms in the DSM-5 criteria for major depressive disorder, demonstrat- ing high relevance. R2 lacks these connections, showing low relevance. These examples were generated by GPT-3.5. Uban et al... | https://arxiv.org/abs/2505.20014v1 |
Specifically, this criterion refers to the extent to which a rationale is explained based on a sufficient understanding of domain knowl- edge, as shown in Figure 1. Our emphasis on do- main relevance stems from its close alignment with the reasoning and diagnostic processes of mental health experts, who utilize establi... | https://arxiv.org/abs/2505.20014v1 |
models, emphasizing the potential for these models to ad- versely affect students’ learning processes (Zhou and Ai, 2024). While these studies demonstrate the general importance of data quality, there remains a need to explicitly investigate the impact of select- ing high-quality and domain-relevant data in the reasoni... | https://arxiv.org/abs/2505.20014v1 |
et al., 2023, 2024). Suppose that we have a mental health detection dataset D={(xi, yi)}N i=1, where xrepresents social media posts and ytheir detection labels. When a post xis given, we use a teacher model Tto predict the authors’ mental condition y′and generate rationales rusing CoT prompting p, explaining why each x... | https://arxiv.org/abs/2505.20014v1 |
symptom-relevant information (i.e., DSM- 5 diagnostic criteria); (2) Symptom Recognition: accurate identification of relevant symptoms from the post; and (3) Symptom Relevancy: alignment of the rationale with identified symptom informa- tion. Details of the prompt for the LLM-evaluation are in Appendix A.2. 3.4 Quality... | https://arxiv.org/abs/2505.20014v1 |
mental disorder-related subreddits and random subreddits (clean text). Detailed statistics of this dataset are provided in Appendix B.1. For evaluating mental disorder detection, we use accu- racy and F1 score as the primary metrics. Details of evaluation metrics are described in Appendix B.2 Models For the teacher mod... | https://arxiv.org/abs/2505.20014v1 |
our approach across these various settings, we aimed to demonstrate its robustness and broad ap- plicability. Table 1 presents the depression detec- tion performance across various combinations of these models and strategies. The experimental results show that distilling rationales from the teacher models generally im-... | https://arxiv.org/abs/2505.20014v1 |
methods, assessing their correlation with human judgments ofConsistency ,Reliability , and Professionality . A high cor- relation indicates a strong alignment with human judgment. (Spearman correlation, *: p<.05, **:p<.01, ***: p<.001) Symptoms (Reference) Relevance Corr.w/Human Corr.w/LLM V ocal Nodule X 0.172 * 0.307... | https://arxiv.org/abs/2505.20014v1 |
a student model trained with rationales generated by a teacher model. The lines connect points repre- senting the same student model trained with and without our selective distillation method. Figure 3 shows a clear positive trend, indicating that higher rationale quality scores are generally associated with greater ac... | https://arxiv.org/abs/2505.20014v1 |
TheWorst rationales often result in the lowest accu- racy, emphasizing the adverse effect of incorporat- ing low-quality rationales. These findings demon- strate the importance of prioritizing high-quality rationales for effective knowledge transfer and im- proved mental health detection. 7 Conclusion This paper empiri... | https://arxiv.org/abs/2505.20014v1 |
to understanding their po- tential benefits and limitations (Won et al., 2025). Future research could investigate more deeply the human-AI interaction regarding how these ratio- nales can assist mental health professionals in di- agnosis, treatment planning, and patient communi- cation by conducting user studies with c... | https://arxiv.org/abs/2505.20014v1 |
diagnoses. For medical diagnosis, the 3Approval number: KH2023-166model’s output should only serve as supplemen- tary indicators, and consultation with professional psychiatrists or clinical practitioners is essential. Acknowledgements This work was supported by the National Re- search Foundation of Korea (NRF) grant f... | https://arxiv.org/abs/2505.20014v1 |
Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adap- tation of large language models. arXiv preprint arXiv:2106.09685 . Hyolim Jeon, Dongje Yoo, Daeun Lee, Sejung Son, Seungbae Kim, and Jinyoung Han. 2024. A dual- prompting for interpretable mental health language models. In Proceedin... | https://arxiv.org/abs/2505.20014v1 |
media using RoBERTa. In Proceed- ings of the 12th International Workshop on Health Text Mining and Information Analysis , pages 59–68, online. Association for Computational Linguistics. Thong Nguyen, Andrew Yates, Ayah Zirikly, Bart Desmet, and Arman Cohan. 2022. Improving the generalizability of depression detection b... | https://arxiv.org/abs/2505.20014v1 |
abilities of large language models. arXiv preprint arXiv:2206.07682 . Hyunseon Won, Migyeong Kang, Minji Kim, Daeun Lee, Hyein Choi, Yonghoon Kim, Daejin Choi, Min- sam Ko, and Jinyoung Han. 2025. " show your mind": Unveiling user experience on an ai-based mental health assessment system with symptom-based evi- dences.... | https://arxiv.org/abs/2505.20014v1 |
for align- ment. Advances in Neural Information Processing Systems , 36. Yuhang Zhou and Wei Ai. 2024. Teaching-assistant- in-the-loop: Improving knowledge distillation from imperfect teacher models in low-budget scenarios. arXiv preprint arXiv:2406.05322 . A Details of Instruction Prompts A.1 Prompts for Rationale Gen... | https://arxiv.org/abs/2505.20014v1 |
(indicating non- depression). B.2 Evaluation Metrics To evaluate the performance of the models for men- tal health problem detection tasks, we use standard classification metrics, including Accuracy (Acc.) and F1-Score (F1). We frame the mental health problem detection task as an explainable mental health analysis, sim... | https://arxiv.org/abs/2505.20014v1 |
such as reading the newspaper or watching television. Moving or speaking so slowly that other people could have noticed. Thoughts that you would be better off dead, or of hurting yourself. Table 10: PHQ-9 items assessing depressive symptoms. DSM-5 diagnostic criteria for GAD Excessive anxiety and worry, occurring more ... | https://arxiv.org/abs/2505.20014v1 |
and the explanation match, but the explanation is difficult to read and contains serious errors. •2: The detection result and the explanationmatch. The explanation is mostly consistent and readable, with a few minor errors. •3: The detection result and the explanation match perfectly. The explanation is natural, consis... | https://arxiv.org/abs/2505.20014v1 |
about their emotions and feelings, which can be a sign of depression. 5. X Seeking Help and Uncertainty : The poster is seeking help and acknowledges that they may be experiencing something wrong, indicating a level of self-awareness and willingness to seek help, which are positive signs. +Selective Yes. 1.✓Expression ... | https://arxiv.org/abs/2505.20014v1 |
our selective distillation framework, applied to student models (Llama-2- 7B, Llama-3.1-8B, Mistral-7B), against other com- petitive SOTA models. Among these, we particu- larly highlight the MentalLLaMA models (Yang et al., 2024), which are specifically fine-tuned for diagnosing mental illness. 6We include these sectio... | https://arxiv.org/abs/2505.20014v1 |
the poster is likely to suffer from very severe Major Depressive Disorder (MDD) for several reasons: 1.△Intensity of negative emotions : The poster describes feeling “low”, “beside myself”, and experiencing a “fog” of depression that makes them feel like a “helpless child”. These phrases suggest a high level of emotion... | https://arxiv.org/abs/2505.20014v1 |
arXiv:2505.20015v1 [cs.CL] 26 May 2025On the class of coding optimality of human languages and the origins of Zipf’s law Ramon Ferrer-i-Cancho(a)1 1Quantitative, Mathematical and Computational Linguistics Research Group, Department of Computer Science, Uni- versitat Polit` ecnica de Catalunya, 08034 Barcelona, Cataloni... | https://arxiv.org/abs/2505.20015v1 |
l(i))<0. (2) It has been demonstrated that, if ⟨l⟩is minimum, then the Kendall τcorrelation between p(i) and l(i) cannot be positive, i.e. [10] τ(p(i), l(i))≤0, which sheds light on the origins of the law of abbreviation. In standard information theory, one assumes that l(i) is the length of the code assigned to unit i... | https://arxiv.org/abs/2505.20015v1 |
not non- singular. Unit Code 1 aa 2 aa 3 a 4 b 5 ba 6 bb by Eq. 5 asN−1 Nconverges quickly to 1 as Nincreases. Second, the solution of the constrained minimization of ⟨l⟩ by means of Lagrange multipliers produces directly Eq. 5 and Eq. 4 results a posteriori by imposing discreteness on l(i) [11]. Here we present a clas... | https://arxiv.org/abs/2505.20015v1 |
of the two coding schemes, namely the code for unit isatisfies the size-rank law [16], i.e. l(i) = anslns min(i) +bns =−anslogNp(i) +bns (13) and the size-probability law, i.e. l(i) = audlud min(i) +bud =audlogNi+bud, (14) where ans,aud,bnsandbudare constants.In that class, ansandbnsare the parameters (slope and interc... | https://arxiv.org/abs/2505.20015v1 |
class. Indeed, an exponential rank distribution has been found in “key signs” in rhesus monkeys [22], “tonemes” in dolphins [23] and calls in warbling vireos [24]. In the past, various authors concluded that “tonemes” in dolphins and calls in warbling vireos exhibit Zipf’s law based on a straight line in logarithmic sc... | https://arxiv.org/abs/2505.20015v1 |
2). Languages (at least Catalan, Spanish and English) ex- hibit such a linear proximity, which is consistent with a strong form of compression. The failure to find Zipf’s law in certain species but an exponential distribution instead [22–24] suggests that such a pressure may not be strong enough for these species or th... | https://arxiv.org/abs/2505.20015v1 |
regarded as a proof that Zipf’s law can be retrieved without involving any optimization [20, 21, 40, 41]. That argument is flawed because random typing is an optimal non-singular coding system. How- ever, one can still retain that random typing is a proof that Zipf’s law is easy to retrieve by virtue of some simple sto... | https://arxiv.org/abs/2505.20015v1 |
be worth for them or the lack of design for unique segmentation may prevent them for getting to longer sequences. To sum up, other species are telling us that reproducing Zipf’s law for word fre- quencies is not as easy as commonly believed [21,41] and optimal coding yields a unified explanation to Zipf’s law of p-5 R.... | https://arxiv.org/abs/2505.20015v1 |
194. [13]Mandelbrot B. ,Information theory and psycholinguis- tics: a theory of word frequencies inReadings in mathe- matical social sciences , edited by Lazarsfield P. F. and Henry N. W. , (MIT Press, Cambridge) 1966 pp. 151– 168. [14]Shannon C. E. ,Bell Systems Technical Journal ,27 (1948) 379. [15]Hern ´andez-Fern ´... | https://arxiv.org/abs/2505.20015v1 |
,5 (2009) e9411. [43]Ferrer-i-Cancho R. andGavald `a R. ,Journal of the American Association for Information Science and Tech- nology ,60(2009) 837. [44]Ferrer-i-Cancho R. ,Complexity ,21(2016) 409. [45]Kapur J. N. andKesavan H. K. ,Entropy optimization principles and their applications inEntropy and Energy Dissipation... | https://arxiv.org/abs/2505.20015v1 |
arXiv:2505.20016v1 [cs.CL] 26 May 2025TTPA: Token-level Tool-use Preference Alignment Training Framework with Fine-grained Evaluation Chengrui Huang1, Shen Gao1, Zhengliang Shi2, Dongsheng Wang1, Shuo Shang1* 1University of Electronic Science and Technology of China, 2Shandong University ry.cr.huang@gmail.com , shengao... | https://arxiv.org/abs/2505.20016v1 |
token-level differences can determine the success or failure of the call. In highly struc- tured outputs like tool calls, even a single token error can lead to complete failure, highlighting the necessity for more precise preference alignment. (2) Furthermore, existing preference data sampling methods typically rely on... | https://arxiv.org/abs/2505.20016v1 |
sampling approach to construct fine-grained preference data. •We propose the Error-oriented Scoring Mech- anism , which captures fine-grained differences between answers, enabling precise alignment of LLM. •Experimental results demonstrate that TTPA sig- nificantly improves tool-use capabilities on three diverse benchm... | https://arxiv.org/abs/2505.20016v1 |
reveal information about the tools or parameters involved (Qin et al., 2023b). In contrast, real-world user queries typically do not ex- plicitly specify the tools to be called or the input pa- rameters. This discrepancy creates a gap between the dataset and real-world applications, ultimately affecting the model’s per... | https://arxiv.org/abs/2505.20016v1 |
from the outputs of L, which predicts a probability distribution Ppredover possible tool calls for the i-th step, given the in- 3 ===? In or not Json.loads Name* Score’Error -oriented Scoring Preference Oriented Tool -use Dataset Construction (1) Reversed Dataset Construction (2) Token -level Preference SamplingTool 1 ... | https://arxiv.org/abs/2505.20016v1 |
will be introduced in § 3.4. 3.4 Error-oriented Scoring Mechanism Existing tool learning methods usually employ LLM-based evaluation or human evaluation to as- sess the quality of generated tool calls, and then use this signal to optimize the model parameters. In this paper, we design an error-oriented scoring mechanis... | https://arxiv.org/abs/2505.20016v1 |
Hammer2.0-7B is fine-tuned from Qwen-2.5-7B, while ToolACE-8B and xLAM-7B-R are based on LLaMA-3.1-8B. Our experiments include models fine-tuned from both the same and different base models, enabling a broader evaluation of TTPA’s effectiveness. 4.3 Dataset & Metric We evaluate the tool learning model fine-tuned with T... | https://arxiv.org/abs/2505.20016v1 |
diverse datasets. The results are shown in Table 1, Table 3, and Table 4 for Toolbench, BFCL, and Our testset, respectively. ToolBench The findings in ToolBench validate the effectiveness of training on tool-use datasets, re- vealing that models with merely 7-8 billion param- eters can achieve comparable or even superi... | https://arxiv.org/abs/2505.20016v1 |
included for a specific case is fewer than the actual possible solutions. This limitation could lead to two main issues: (1) correct tool calls being misclassified as false, thereby reducing accuracy metrics, and (2) potential favoritism toward models trained on specific datasets that the datas’ distri- bution is simil... | https://arxiv.org/abs/2505.20016v1 |
enhances the tool-use capabilities of LLMs. Specifically, we observed substantial improvements across all three benchmark datasets, with performance gains reach- ing up to 39.7%. These findings suggest that con- structing token-level preference datasets for model fine-tuning enables more granular alignment with correct... | https://arxiv.org/abs/2505.20016v1 |
as well as tools used in our experiment were obtained from existing benchmarks and public open re- sources, thus ensuring a high level of transparency and reproducibility in our experimental procedure. To minimize potential bias and promote fair- ness, we use the prompts following existing works, which are publicly acc... | https://arxiv.org/abs/2505.20016v1 |
Wu, Xinzhi Wang, Yong Liu, Yasheng Wang, Duyu Tang, Dandan Tu, Lifeng Shang, Xin Jiang, Ruim- ing Tang, Defu Lian, Qun Liu, and Enhong Chen. 2024a. Toolace: Winning the points of llm function calling. In International Conference on Learning Representations: ICLR . Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, ... | https://arxiv.org/abs/2505.20016v1 |
is secretly a reward model. Advances in Neu- ral Information Processing Systems . Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. Neural Information Processing Systems... | https://arxiv.org/abs/2505.20016v1 |
. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. 2024b. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. arXiv . Qinzhuo... | https://arxiv.org/abs/2505.20016v1 |
on the tool call results and the scenario. Finally, we gen- erate a query corresponding to all the information. Tool-Learning: Then we employ a tool-learning model to solve the generated query. During this process, we sample multiple tool-calling samples from the generated tokens’ distribution. By scoring the samples, ... | https://arxiv.org/abs/2505.20016v1 |
related to the tools provided. IMPORTANT: The scenario you simulate CAN NOT contain any explicit questions. You SHOULD only state the scenario. The scenario you simulate CAN NOT contain any tool name in the tools above. You SHOULD keep the scenario as realistic as possible. YOUR OUTPUT CONTAINS: scenario: str, the scen... | https://arxiv.org/abs/2505.20016v1 |
sequence of tool calls repre- sents the steps to derive this answer. 2. Ensure the question is intricate and closely related to the tool calls and the final answer. 3. Write the question from a first-person perspective, mak- ing it sound natural and human-like. 4. The question should include the necessary information 1... | https://arxiv.org/abs/2505.20016v1 |
"name": "get_news_report", "description": "Retrieves the latest news for a specified location formatted as 'City, State'.", "parameters": { "type": "dict", "required": ["location"], "properties": { "location": { "type": "string", "description": "The location for which to retrieve the news, in the format of 'City, State... | https://arxiv.org/abs/2505.20016v1 |
arXiv:2505.20023v1 [cs.CL] 26 May 2025Training LLM-Based Agents with Synthetic Self-Reflected Trajectories and Partial Masking Yihan Chen1, Benfeng Xu1, Xiaorui Wang2, Yongdong Zhang1, Zhendong Mao1, 1University of Science and Technology of China,2Metastone Technology {chenyihan, benfeng}@mail.ustc.edu.cn Abstract Auto... | https://arxiv.org/abs/2505.20023v1 |
slow down and may eventually plateau. This indicates such methods have limitations in en- 1 Instruction: I'm looking for a bath brush in beigewith a longhandle, and lower than $40.00 Thought 1: I need to find ... Action 1: search[ ... ...] Observation 1: Products: 1.B09CF4X81Y ... Thought 2: I'll check the first one ..... | https://arxiv.org/abs/2505.20023v1 |
self-reflected trajectories achieves comprehensive improvements compared to the model trained with the full dataset of expert trajectories. Additionally, we conduct ablation experiments to verify the ne- cessity of Self-Reflected Trajectories andPartial Masking , as well as test the performance of existing models on th... | https://arxiv.org/abs/2505.20023v1 |
count may leave insufficient con- text to complete the task. Moreover, heuristic de- tection methods can only address a limited range of error types and are unable to detect and correct ar- bitrary errors. Compared to directly utilizing prior Reflection-related works such as Reflexion (Shinn et al., 2023) for trajector... | https://arxiv.org/abs/2505.20023v1 |
STeP utilizes golden trajectories and corresponding instructions to train a Self-reflected LLM-based agent through three stages. Stage 1: Agent Initialization; Stage 2: Self-Reflected Trajectories Synthesizing; Stage 3: SFT with Partial Masking . 3.2 Agent Initialization Following complex agent task instructions is qui... | https://arxiv.org/abs/2505.20023v1 |
4 based on the task requirements, the interaction his- tory, and the feedback ojprovided by eat that step, πteacher will, in the first-person narrative, offer the content, reason, reflection on this error, and the correct action that πθshould take to correct it. To ensure consistency in the training data, we convert th... | https://arxiv.org/abs/2505.20023v1 |
trajectories not fully completing the tasks, we filter out trajectories with reward = 1to con- stitute our training set. We randomly sample 50% of the data from each task and combine it to D1, which is used to train the base LLM agent during the Agent Initialization stage. 4.2 Experiments Setup Our experiments are prim... | https://arxiv.org/abs/2505.20023v1 |
Results Our agent outperforms most baselines in average re- ward across all tasks, demonstrating the efficacy of STeP. Compared to agents trained solely on golden trajectories, STeP achieves better results across all test sets. LLaMA2-7B-chat + STeP outperforms LLaMA2-7B-chat + Golden Trajs by an average of 9.2% on all... | https://arxiv.org/abs/2505.20023v1 |
Our training set combines Self-Reflected Trajecto- riesDrwith a portion of golden trajectories D1. To evaluate the impact of the self-reflected trajecto- ries, we trained LLaMA2-7B-chat using only D1, resulting in the model LLaMA2-7B-chat ( w/o self- reflected trajectories ). As illustrated in Table 3, not utilizing an... | https://arxiv.org/abs/2505.20023v1 |
LLMs. We use Self-Reflected Trajectories Dralongside the half of all golden trajectories D1to train Mistral-7B- Instruct-V0.3 (Jiang et al., 2023) and LLaMA3- 8B-Instruct according to our proposed method. In addition, we compared two other methods based on LLaMA3-8B: WKM (Qiao et al., 2024a), which introduced a paramet... | https://arxiv.org/abs/2505.20023v1 |
in real-time need further improvement. Reflections and corrections of erroneous steps in self-reflected trajectories may not always be appro- priate. We aim to address these issues in future works. Secondly, due to small-scale LLMs have relatively limited capabilities and are unable to gen- erate sufficient high-qualit... | https://arxiv.org/abs/2505.20023v1 |
pages 23813–23825. Curran Associates, Inc. Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Ao- han Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie Huang, Yuxiao Dong, and Jie Tang. 2024. Age... | https://arxiv.org/abs/2505.20023v1 |
. Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2021. Alfworld: Aligning text and embodied environments for interactive learning. Preprint , arXiv:2010.03768. Yifan Song, Da Yin, Xiang Yue, Jie Huang, Sujian Li, and Bill Yuchen Lin. 2024. Trial and error: Explor... | https://arxiv.org/abs/2505.20023v1 |
Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. 2024a. Travelplanner: A benchmark for real-world planning with language agents. In Forty-first Interna- tional Conference on Machine Learning . Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhou- jun Cheng, Dongchan Shi... | https://arxiv.org/abs/2505.20023v1 |
ALFWorld contains text- simulated environments that parallel embodied worlds in the ALFRED dataset (Shridhar et al., 2020). The agent must provide textual actions to solve simulated em- bodied tasks, such as placing a vase in a safe. Specific task formats and requirements are detailed in Figure 6. These task requiremen... | https://arxiv.org/abs/2505.20023v1 |
environment ecorresponding to the agent task. The main evaluation metric is average re- ward , which is the mean of the rewards the agent obtains across all instructions in a certain agent task. Reward calculation varies by task. In ALF- World, reward 1 corresponding to task completion and 0 to failure. For the other t... | https://arxiv.org/abs/2505.20023v1 |
to assist the LLM teacher in making better decisions, we created several corresponding con- siderations based on the characteristics of each task. The LLM teacher will combine the current task in- struction, interaction history, the actions taken by the base LLM agent, and the corresponding feed- back from the environm... | https://arxiv.org/abs/2505.20023v1 |
clickables. An action should be of the following structure: search[keywords] click[value] If the action is not valid, perform nothing. Keywords in search are up to you, but the value in click must be a value in the list of available actions. Remember that your keywords in search should be carefully designed. Your respo... | https://arxiv.org/abs/2505.20023v1 |
arXiv:2505.20027v1 [q-bio.NC] 26 May 2025Published as a conference paper at ICLR 2025 MULTI -MODAL BRAIN ENCODING MODELS FOR MULTI - MODAL STIMULI Subba Reddy Oota1∗, Khushbu Pahwa2∗, Mounika Marreddy3, Maneesh Singh4 Manish Gupta5, Bapi S. Raju6 1Technische Universität Berlin, Germany,2Rice Univ, USA,3Univ of Bonn, Ge... | https://arxiv.org/abs/2505.20027v1 |
representations extracted using Transformer-based (Vaswani et al., 2017) multi-modal models. Our analysis focuses on brain alignment—the degree of similarity when predicting brain activity using both uni-modal and multi-modal models. 1https://github.com/subbareddy248/multi-modal-brain-stimuli 1 Published as a conferenc... | https://arxiv.org/abs/2505.20027v1 |
included with audio (St- Laurent et al., 2023), we investigate several research questions. First, we investigate the effectiveness of stimulus representations obtained using multi-modal models versus unimodal models for brain encoding. Multi-modal models are of two broad types: (i) cross-modal pretrained models, where ... | https://arxiv.org/abs/2505.20027v1 |
(1) To the best of our knowledge, this study is the first to leverage both cross-modal and jointly pretrained multi-modal models to perform brain alignment while subjects are engaged with multi-modal naturalistic stimuli. (2) We evaluate the performance of several unimodal Transformer models (three video and two audio)... | https://arxiv.org/abs/2505.20027v1 |
same mulitmodal model for residual analysis. It should be noted that multi-modal models can be trained broadly using two strategies – cross-modally (dual) or jointly (Frank et al., 2021). It is conceivable that the representations arising from these two kinds of training might have different brain alignment (Oota et al... | https://arxiv.org/abs/2505.20027v1 |
training on the combined brain data from The Bourne supremacy andThe wolf of wall street and testing on the brain data from the movie Life. We present the average cross-subject prediction accuracy across voxels for theMovie10 fMRI dataset across subjects in Appendix B. 4 M ETHODOLOGY Multi-modal Models : To analyse how... | https://arxiv.org/abs/2505.20027v1 |
individual modalities, we use the following methods to obtain embeddings for individual modalities. Video-based models. To extract representations of the video stimulus, we use three popular pretrained Transformer video-based models from Huggingface (Wolf et al., 2020): (1) Vision Transformer Base (ViT-B) (Dosovitskiy ... | https://arxiv.org/abs/2505.20027v1 |
and The wolf of wall street : 6898 TRs). The test set consisted only data from the “Life” movie (2028 TRs). Thus there is no possibility of any information leakage during inference on the test set. Model details and hyper-parameter settings are in Appendix E. Removal of a single modality features from multi-modal repre... | https://arxiv.org/abs/2505.20027v1 |
because fMRI data is considered to have positive dependence (Genovese, 2000)). In all cases, we denote significant differences (p ≤0.05) with a∗or∧. 6 R ESULTS 6.1 H OW EFFECTIVE ARE MULTI -MODAL REPRESENTATIONS ? In Fig. 2, we present the average normalized brain alignment scores for both multi-modal and individual mo... | https://arxiv.org/abs/2505.20027v1 |
Figs. 2 ( Right ). The Wilcoxon signed-rank test shows that the differences in embeddings from the IB Concat and TVLT models are not statistically significant when averaged across language and visual regions. Similar to whole brain performance, in the language regions, cross-modal embeddings are significantly better th... | https://arxiv.org/abs/2505.20027v1 |
TVLT, Unimodal VM, and Unimodal SM. Fig. 2 ( Left) displays the comparison of average normalized brain alignment of randomly generated vectors, pretrained and randomly initialized models. From Fig. 2 ( Left), we observe that randomly initialized models show significantly better alignment than random vectors. However, t... | https://arxiv.org/abs/2505.20027v1 |
398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431Under review as a conference paper at ICLR 2025 *^^^ Whole brain0.10.20.30.40.50.6 Normalized brain alignment *^ ^ ^^*^ Language: AG0.10.20.30.40.50.6 Normalized brain alignment *^ *^*^... | https://arxiv.org/abs/2505.20027v1 |
353 video and audio embeddings partially impacts the brain alignment. We observe similar findings for 354 language ROIs such as PTL, MFG, ATL, PCC and visual regions EVC, OV and FV , as shown in 355 Figs. 8 and 9 in Appendix. These results suggest that there is additional information beyond the 356 unimodal embeddings ... | https://arxiv.org/abs/2505.20027v1 |
the PTL is known for auditory processing, it also contributes to the integration of visual and auditory inputs. The improved alignment of video embeddings here indicates that visual information enhances the processing capabilities in this region. (3) Consistent with the cross-modality models, in jointly pretrained TVLT... | https://arxiv.org/abs/2505.20027v1 |
video embeddings here indicates that visual information enhances the processing capabilities in this region. (3) Consistent with the cross-modality models, in jointly pretrained TVLT models, TVLT video embeddings significantly outperform TVLT audio embeddings, except in PTL region. These observations indicate that vide... | https://arxiv.org/abs/2505.20027v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.