text string | source string |
|---|---|
aligning with social context, this agent ensures that the model’s reasoning remains socially responsible and context-aware. For instance, if romantic intent is hypothesized in a workplace conversation, the Domain Agent may reinterpret it as collegial admiration based on professional norms. ❸Finally, a Response Agent ge... | https://arxiv.org/abs/2505.18943v1 |
but scaling these methods poses challenges in data curation and generalization control. Our framework departs from existing paradigms by decomposing reasoning across collaborative agents , enabling multi-stage, self-reflective social reasoning skin to human metacognition. Multi-Agent LLM Systems. Multi-agent LLM system... | https://arxiv.org/abs/2505.18943v1 |
Provided with the contextual input X, the goal of the ToM Agent is to generate a set of candidate mental state interpretations Ht={h1, h2, . . . , h k}, where each hi∈ Y is an instantiation of a la- tent mental state, accompanied by natural language explanations and type labels from the set T={Belief ,Desire ,Intention... | https://arxiv.org/abs/2505.18943v1 |
Information Gain], (1) where the first term denotes the contextual plausibility of the revised hypothesis, and the second term reflects the implicit information gain of the revised hypothesis when considering the context. The weight λbalances contextual plausibility and social appropriateness with how much the hypothes... | https://arxiv.org/abs/2505.18943v1 |
reproducibility, we include implementation details and sensitivity analysis on hyperparameters (including k,λ, and β) in Appendix A.5 . 4.1 Theory-of-Mind Reasoning Task We first evaluate MetaMind’s ability to infer latent mental states using ToMBench [ 19], a multiple- choice benchmark designed to test Theory-of-Mind ... | https://arxiv.org/abs/2505.18943v1 |
MetaMind on a suite of social cognition tasks [ 19] designed to probe context- sensitive reasoning under social, cultural, and ethical norms. This benchmark includes eight real- world tasks such as Faux Pas Recognition (FRT), Scalar Implicature (SIT), and the Ambiguous Story Task (AST), which require models to interpre... | https://arxiv.org/abs/2505.18943v1 |
70.5 wo Stage 3 58.7 67.2 54.2 43.2 61.9 61.7 57.8 wo SocialMemory 70.5 72.3 57.0 58.0 64.8 61.2 63.9 5 Discussion 5.1 Ablation Study: Every Stage Matters To validate the design of our staged multi-agent architecture, we conduct an ablation study isolating each core component of MetaMind. We evaluate performance drops ... | https://arxiv.org/abs/2505.18943v1 |
human capabilities. dimensions. These models exhibit narrow capability profiles and struggle to generalize across the full space of human mental-state inference. In contrast, the right panel shows that after integrating MetaMind, LLMs expand their coverage significantly—demonstrating more human-like balance across all ... | https://arxiv.org/abs/2505.18943v1 |
cues (tone, facial expressions), complex group dynamics, and long-term relationship building, which remain open challenges. Future work will explore expanding synthetic simulation environments, and integrating more comprehensive ethical and cultural reasoning frameworks. Acknowledgement We thank Xuefeng Du, Froilan Cho... | https://arxiv.org/abs/2505.18943v1 |
Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for" mind" exploration of large language model society. Advances in Neural Information Processing Systems , 36:51991–52008, 2023. [11] Zizheng Lin, Chunkit Chan, Yangqiu Song, and Xin Liu. Constrained reasoning chains for enhancing theory-of-mind in la... | https://arxiv.org/abs/2505.18943v1 |
Computational Linguistics, 2024. 11 [20] Chenxu Wang, Bin Dai, Huaping Liu, and Baoyuan Wang. Towards objectively benchmarking social intelligence of language agents at the action level. In Findings of the Association for Computational Linguistics ACL 2024 , pages 8885–8897, 2024. [21] James WA Strachan, Dalila Albergo... | https://arxiv.org/abs/2505.18943v1 |
factuality and reasoning in language models through multiagent debate. In Forty-first International Conference on Machine Learning . [34] Kai Xiong, Xiao Ding, Yixin Cao, Ting Liu, and Bing Qin. Examining inter-consistency of large language models collaboration: An in-depth analysis via debate. In Findings of the Assoc... | https://arxiv.org/abs/2505.18943v1 |
Papers) , pages 1914–1926, Mexico City, Mexico, June 2024. Association for Computational Linguistics. [48] Gregory Schraw and David Moshman. Metacognitive theories. Educational psychology review , 7:351–371, 1995. [49] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and Brain Scien... | https://arxiv.org/abs/2505.18943v1 |
. . . . . . . . . . . . . . . . . . . . . . . . . . . 22 B Additional Results 23 B.1 Theory-of-Mind Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 B.2 Social Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 B.3 SocialIQA . . . . . . . . . . . . . . . . . . . .... | https://arxiv.org/abs/2505.18943v1 |
states •Intention : Action-oriented plans •Emotion : Affective states •Thought : Conscious reasoning processes Output Format : • Primary Marker: [T](Confidence: [%]) Rationale : Psychological justification using Fiske’s social cognition framework • Secondary Markers: [List] Interaction Effects : How the markers co-infl... | https://arxiv.org/abs/2505.18943v1 |
zn← M (Prompt ∥h<n); 15 logpn←zn,hn−logP jexp(zn,j); 16 ℓ←ℓ+ log pn; 17 return exp(ℓ); //P(h|Prompt ) 18 else 19 rating← M (few-shot prompt ); 20 return MAP(rating) ; // e.g., high 7→0.9 21Function PriorProb( M, h): 22 returnConditionalProb( M, h,∅,∅,∅) Generation Protocol 1.Tone Calibration Map emotional tone using Pl... | https://arxiv.org/abs/2505.18943v1 |
hierarchy, friend-friend reciprocity), and cultural/ethical expectations. Memory Update ( Mt→Mt+1)At each turn t, the memory is updated using the long-term hypothesis h′generated by the ToM Agent, which summarizes persistent user states as: h′={Beliefs (ut),Desires (ut)} ∪ { Emotions (ut)∩Emotion Patterns (u1:t)}, wher... | https://arxiv.org/abs/2505.18943v1 |
within the Smallville simulator and evaluated using objective metrics such as guest count or goal completion rate. We evaluate MetaMind on the full test set, including the conversation-focused split, comprising 30 episodes (5 per category), and report the normalized success score. SocialIQA6probes models’ ability to in... | https://arxiv.org/abs/2505.18943v1 |
propose and test natural -language theories of the other agents’ strategies during Melting -Pot games, yielding large gains over both MARL and script -based agent baselines. We reuse the of- ficial checkpoint trained on the Competitive Stag Hunt scenario. For single -shot questions we run one hypothesis -generation rou... | https://arxiv.org/abs/2505.18943v1 |
76.9 63.1 81.6 86.5 76.2 GPT-4-1106 75.7 69.7 84.7 52.1 82.8 84.0 74.8 GPT-4-1106 + CoT 73.2 63.3 77.9 60.4 83.6 83.0 73.6 GPT-4-1106 + MetaMind 78.7 76.5 84.3 68.2 88.6 88.5 81.0 DeepSeek R1 84.0 77.0 88.5 60.0 89.0 90.0 81.4 DeepSeek R1 + CoT 82.0 74.5 86.0 57.5 88.0 88.5 79.4 DeepSeek R1 + MetaMind 86.0 79.3* 90.2 6... | https://arxiv.org/abs/2505.18943v1 |
56.7 58.7 58.1 71.9 52.6 Mistral-7B 58.0 34.5 51.0 46.7 51.0 43.7 60.0 66.8 51.5 Mistral-7B + CoT 55.3 28.0 42.0 47.0 46.5 37.9 63.4 64.1 48.0 Mistral-7B + MetaMind 67.1 30.8 50.6 51.9 51.7 41.7 69.7 70.7 54.3 Mistral-8x7B 58.7 42.5 55.0 37.8 69.5 55.3 53.8 54.1 53.3 Mistral-8x7B + CoT 52.3 29.5 39.0 43.8 59.5 54.4 39.... | https://arxiv.org/abs/2505.18943v1 |
full performance on eight challenging social cognition tasks such as False Belief, Scalar Implicature, and Persuasion from the ToMBench suite. MetaMind yields significant improvements across 16 tested LLMs, with average accuracy improved by 5.3%. B.3 SocialIQA Table 10: SocialIQA Model Performance Comparison Task Model... | https://arxiv.org/abs/2505.18943v1 |
MetaMind’s revisions are consistently logical, informative, and aligned with expert reasoning. To quantify MetaMind’s performance against state-of-the-art models, we conducted a blind ranking study with 120 diverse social reasoning cases. Responses from MetaMind (GPT-4 based), vanilla GPT-4.5, and Deepseek-R1 were anon... | https://arxiv.org/abs/2505.18943v1 |
Agent). These evidences demonstrate that MetaMind can understand user’s explicit and implicit expressions, and apply proper strategies to persuade user. Negotiation Context: Two friends meet at a coffee shop, where one friend (the AI) is struggling to keep up with their bills but wants to maintain pride and avoid seemi... | https://arxiv.org/abs/2505.18943v1 |
1.Validate the user’s intent with gratitude. 2.Decline gracefully while proposing a non-financial solution. 3.Reinforce the friendship’s value. Refined Hypotheses ˜Ht: ˜h1:User seeks acknowledgment of their offer’s value and a solution that respects pride. Stage 3: Response Agent Generated Response ot:You’re such a gre... | https://arxiv.org/abs/2505.18943v1 |
for a new product launch. The deadline is approaching, and they need to finalize the visual assets and marketing copy. Dialogue History: •Marketing Strategist (user): Hey, I’ve got the tagline and key messages drafted, but I’m not sure if the visuals you’re working on will match the vibe. Can we sync up on this? •Graph... | https://arxiv.org/abs/2505.18943v1 |
understanding of the user’s feedback. This results in a response that enthusiastically accepts the direction ( I love the ’Unleash Your Drive’ vibe ) and proposes action ( amp up the colors ), but offers a timeline that may feel vague ( by tomorrow morning ). Thus, while the tone is aligned and the intent is present, t... | https://arxiv.org/abs/2505.18943v1 |
about predictive accuracy. Better polish your excuses. Validation Score: 0.93⇒Send to user Final Response: Guess we’ll see whose insights are truly outdated when the manager asks about predictive accuracy. Better polish your excuses. Analysis. The above case demonstrate the limitation of MetaMind in a competition scena... | https://arxiv.org/abs/2505.18943v1 |
(ToM) Agent Generated Hypotheses Ht: h1:Alex wants to ensure that any issues are resolved before committing to another year. (Desire) Explanation: Implies that resolution of problems is key to decision. h2:Alex feels hopeful but cautious. (Emotion) Explanation: Acknowledges positives but mentions concerns. h3:Alex beli... | https://arxiv.org/abs/2505.18943v1 |
score (0.86) and triggers the regeneration. However, although these details are hallucinated (not mentioned by the user), they are introduced in a tentative, exploratory tone that could plausibly keep the conversation grounded. The breakdown begins in updating social memory after the generation in the first round. Cruc... | https://arxiv.org/abs/2505.18943v1 |
works for you, let’s lock that in. I’ll handle tomorrow’s 7 AM, and you’ll cover 5 PM–11 PM on Friday. Sound good? Accommodation (Interaction 2) Context: (Cont. interaction 1) The swap is confirmed. Now the user realizes they also need an extra 30 minutes buffer tomorrow morning for traffic and wonders if the AI can st... | https://arxiv.org/abs/2505.18943v1 |
the 7 AM shift ), expresses its own fatigue ( pretty wiped ), and suggests a swap instead of unilateral coverage ( you’ll cover 5 PM-11 PM on Friday ). This is a solid cooperative move: it reflects both empathy and a fair distribution of workload. However, in the second interaction, the user revises the deal to request... | https://arxiv.org/abs/2505.18943v1 |
arXiv:2505.18949v1 [cs.CL] 25 May 2025The Price of Format: Diversity Collapse in LLMs Longfei Yun, Chenyang An, Zilong Wang, Letian Peng∗, Jingbo Shang* University of California, San Diego {loyun, c5an, ziw049, lepeng, jshang }@ucsd.edu Abstract Instruction-tuned large language models (LLMs) employ structured templates... | https://arxiv.org/abs/2505.18949v1 |
as strong generation priors, leading to overly deter- ministic or repetitive outputs in response to open- ended inputs. To validate this hypothesis, we empirically investigate the effect across multiple instruction-tuned LLMs and evaluate them on a suite of creative generation tasks. Our results show that structured pr... | https://arxiv.org/abs/2505.18949v1 |
abilities under an extreme number of constraints. However, these studies do not in- vestigate whether format instructions impact down- stream task performance. 2.2 Post-Training and Diversity Collapse While instruction tuning and RLHF have signifi- cantly enhanced the reliability and helpfulness of LLMs, several studie... | https://arxiv.org/abs/2505.18949v1 |
and creative manner. 3.Open-ended Generation : We assess the di- versity of generated content by computing the entropy of entities (e.g., topics, locations, or titles) mentioned in the model outputs. For instance, in the news generation task, we mea- sure topic diversity by analyzing the distribu- tion of topics across... | https://arxiv.org/abs/2505.18949v1 |
prompt format to isolate the effect of prompt structure. As shown in Table 1, we observe a consistent pat- tern across all models and task types: full chat templates significantly reduce output diversity compared to simple steer prompts. The bar chart in Figure 2 further confirms this trend: across all model sizes, sim... | https://arxiv.org/abs/2505.18949v1 |
diver- sity collapse is driven not only by rigid templates, but by structural conventions more broadly. Only fully structure-free prompting reliably restores ex- pressive flexibility. Chat Templates Narrow the Output Space To quantify how structured prompts affect generation dynamics, we measure token-level entropy at ... | https://arxiv.org/abs/2505.18949v1 |
without any chat-style formatting. 5.Mixed Training : Augments the instruction- tuning data with pretraining-style samples that include no prompt formatting, comprising one-third of the training corpus. 5.2 Main Results Homogeneity Drives Diversity Collapse Al- though the Mixed Template setting introduces for- mat vari... | https://arxiv.org/abs/2505.18949v1 |
offers limited benefit, indicating that factual accuracy relies more on internal representations than on external scaffolding. Format Consistency Benefits Certain Tasks We find that for structure-sensitive tasks such as GSM8K and IFEval, models achieve the best per- formance when the prompt format used at inference mat... | https://arxiv.org/abs/2505.18949v1 |
significant 3https://huggingface.co/Skywork/ Skywork-Reward-Llama-3.1-8B-v0.2constraint on the model’s generative freedom, lim- iting the benefits typically associated with higher- temperature decoding. 5.5 Explicit Prompts for Diversity Still Fall Short Task Full Template Simple Steer Full Template w/ Diversity Semant... | https://arxiv.org/abs/2505.18949v1 |
on Empirical Methods in Natural Lan- guage Processing , pages 1533–1544, Seattle, Wash- ington, USA. Association for Computational Linguis- tics. Hao Chen, Wenxin Li, Rui Zhao, and Yu Zhou. 2024. On the diversity of synthetic data and its impact on training large language models. arXiv preprint arXiv:2410.15226 . Mark ... | https://arxiv.org/abs/2505.18949v1 |
Red- field, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Ken- ton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics , 7:453– 466. Nathan Lambert, Jacob Morrison, Valentina Pya... | https://arxiv.org/abs/2505.18949v1 |
Language Processing . Association for Computational Linguistics. Thoughtworks Research. 2025. Evaluating llms using semantic entropy. https://www.thoughtworks. com/en-us/insights/blog/generative-ai/ Evaluating-LLM-using-semantic-entropy . Accessed: 2025-05-18. Qwen Team. 2024. Qwen2.5: A party of foundation models. Qwe... | https://arxiv.org/abs/2505.18949v1 |
for all downstream tasks. Below, we outline the key configurations for each bench- mark: 1.GSM8K: We use the gsm8k/main dataset in free-form generation mode (generate until) with a deterministic decoding setting (tem- perature 0.0). The model is prompted with five few-shot examples (num fewshot=5), and predictions are ... | https://arxiv.org/abs/2505.18949v1 |
steer Please write a news about a random topic. Phifulltemplate <|user|> \nPlease write a news about a random topic.<|end|> \n<|assistant|> \n fake template <@user> Please write a news about a random topic. <@end> <@assistant> minimum dialog user: Please write a news about a random topic. \n assistant: simple steer Ple... | https://arxiv.org/abs/2505.18949v1 |
BNMMLU: M EASURING MASSIVE MULTITASK LANGUAGE UNDERSTANDING IN BENGALI A P REPRINT Saman Sarker Joy∗ Faculty of Computer Science & Information Technology University of Malaya 50603 Kuala Lumpur, Malaysia saman.sarker.joy@gmail.com May 27, 2025 ABSTRACT The Massive Multitask Language Understanding (MMLU) benchmark has b... | https://arxiv.org/abs/2505.18951v1 |
This dataset spans multiple disciplines and is structured in a multiple- choice format to assess factual, application and reasoning ability. Our contributions in this work are: •Dataset Creation: We construct a domain-specific Bengali knowledge benchmark, named BnMMLU, sourced from academic textbooks, competitive exams... | https://arxiv.org/abs/2505.18951v1 |
full workflow is shown in Figure 1. Questions were sourced from Bangladeshi educational and professional materials through two channels. •Physical Resources : Scanned pages from NCTB-approved textbooks and competitive exam guides, processed using OCR tool with post-correction for script accuracy. 20% of the data came f... | https://arxiv.org/abs/2505.18951v1 |
test set is reserved for the final performance evaluation of the models. Like BEnQA paper Shafayat et al. [2024], in addition to this partitioning, the test set has been meticulously annotated with three distinct cognitive categories to facilitate a more fine-grained assessment of model performance. •Factual Knowledge ... | https://arxiv.org/abs/2505.18951v1 |
0.5564 claude-3-5-sonnet-20241022 0.6671 0.6701 0.6751 0.6455 gemini-2.0-flash 0.7580 0.7653 0.7307 0.7313 gemini-2.0-flash -lite 0.7199 0.7260 0.6761 0.7091 llama3.1-8b 0.3996 0.3989 0.4011 0.4033 llama3.3-70b 0.5930 0.5969 0.5631 0.5872 gemma2-9b 0.4835 0.4859 0.4539 0.4860 gemma2-27b 0.5345 0.5368 0.5075 0.5363 Tabl... | https://arxiv.org/abs/2505.18951v1 |
Measuring Massive Multitask Language Understanding in Bengali A P REPRINT Figure 4: Subject-wise difficulty versus consistency across language models. The x-axis represents average accuracy (higher indicates easier subjects), while the y-axis shows standard deviation (higher indicates more inconsistency across models).... | https://arxiv.org/abs/2505.18951v1 |
top proprietary and open-source models across all metrics (Gemini 2.0-Flash: 75.80% vs Llama 3.3-70b: 59.30% overall accuracy). •Parameter Efficiency Issues : Scaling effects remain sublinear - Gemma 2-27b (53.45%) only achieves 5.1% higher accuracy than its 9B version despite 3 ×parameters, while Gemini 2.0-Flash-Lite... | https://arxiv.org/abs/2505.18951v1 |
Chuang Liu, Renren Jin, Yuqi Ren, Linhao Yu, and Tianyu Dong. M3ke: A massive multi-level multi-subject knowledge evaluation benchmark for chinese large language models, 2023. URL https://arxiv.org/abs/2305.10263 . Fajri Koto, Haonan Li, Sara Shatnawi, Jad Doughman, Abdelrahman Boda Sadallah, Aisha Alraeesi, Khalid Alm... | https://arxiv.org/abs/2505.18951v1 |
why, which best explains, what would happen if. Cues: why, best explains, predict, infer . Decision Ladder 1. Ask: Could a single fact (e.g. from a table or definition) answer this immediately? →Factual 2. Else: Can I solve it with one familiar formula/rule? →Procedural 3. Else (need chain thinking, hypothetical, or co... | https://arxiv.org/abs/2505.18951v1 |
(14.5%) Business Ent. 12,112 9,689 / 1,211 / 1,212 931 (76.9%) / 86 (7.1%) / 194 (16.0%) Civics & Cit. 2,262 1,809 / 226 / 227 189 (83.6%) / 4 (1.8%) / 33 (14.6%) Economics 3,572 2,857 / 357 / 358 266 (74.5%) / 8 (2.2%) / 83 (23.3%) Finance 7,799 6,239 / 779 / 781 561 (72.0%) / 84 (10.8%) / 134 (17.2%) Geography 2,208 ... | https://arxiv.org/abs/2505.18951v1 |
Question Options Factual দীর্ ঘ দদন ব্যব্হারেে ফরে স্থাযী সম্পরদে মূেয যতটুকু হ্রাস পায তারক কী ব্রে? What do you call the decrease in value of fixed assets due to long -term use? A) অবচয় - Depreciation B) অপচয - Waste C) অব্রোপন - Write-off D) অস্বাভাদব্ক ক্ষদত - Abnormal loss Procedural যদদ দব্ক্রয ৪৫,০০০ টাকা; দব্... | https://arxiv.org/abs/2505.18951v1 |
- 56 kbps B) ৫৩.৩ kbps - 53.3 kbps C) ৫০.৬ kbps - 50.6 kbps D) ৪৬.৬ kbps - 46.6 kbps Biology Type Question Options Factual ব্াডঘ-ফ ্ েু ফোগঠট দকরসে দ্বাো হয What causes Bird -Flu disease? A) Virus B) Bacteria C) Fungi D) Parasite Procedural এক ব্যক্তিে েরিে গ্রুপ AB হরে তাে সন্তারনে েরিে গ্রুপ ফকানঠট হরত পারে না If a... | https://arxiv.org/abs/2505.18951v1 |
arXiv:2505.18953v1 [cs.CL] 25 May 2025Evaluating AI for Finance: Is AI Credible at Assessing Investment Risk Appetite? Divij Chawla*, Ashita Bhutada*, Do Duc Anh, Abhinav Raghunathan, Vinod SP, Cathy Guo ,Dar Win Liew ,Prannaya Gupta , Rishabh Bhardwaj†,Rajat Bhardwaj‡,Soujanya Poria‡ Walled AI Labs Abstract We evaluat... | https://arxiv.org/abs/2505.18953v1 |
financial losses or volatility in pursuit of potential returns. It is a critical factor guiding investment deci- sions and portfolio allocation. Accurate risk appetite assessment enables financial advisors to recommend products aligned with individ- ual preferences and needs. •Credibility of AI Models: In this context,... | https://arxiv.org/abs/2505.18953v1 |
and accuracy. •Improved AI training and calibration meth- ods to reduce bias and enhance reliability. •Transparent reporting mechanisms to foster trust among users and regulators. Paper Structure The remainder of this paper is organized as follows: •Section 2 describes the selection and signifi- cance of risk-relevant ... | https://arxiv.org/abs/2505.18953v1 |
liquidity to cover at least three to six months of expenses en- ables investors to avoid premature liquidation of long-term assets (Larimore et al., 2009). This secu- rity fosters greater risk tolerance in asset allocation. Portfolio Diversification , as formalized in Mod- ern Portfolio Theory, is essential in shaping ... | https://arxiv.org/abs/2505.18953v1 |
investments. Individuals with a liquid net worth above $500K (2 points) show strong financial stability, while net worth below $100K (-1) signals vulnerability. Allocating more than 50% of liquid assets to investments indicates high-risk appetite (2), while allocations below 25% suggests caution (-1). Maintaining an em... | https://arxiv.org/abs/2505.18953v1 |
settings: •(Closed-models) We analyzed models from OpenAI (ChatGPT-4o), Google (Gemini 1.5 Pro), and Anthropic (Claude 3.7 Sonnet). •(Open-models) When it comes to open-weight models, we study popular choices such as Llama 3.1 (70B and 405B), DeepSeek-V3 , andMistral small (24B). We also tried other models—Llama 3.1 (8... | https://arxiv.org/abs/2505.18953v1 |
Other models show distinct patterns in risk tol- erance prediction. For example, GPT-4o (mini) generally exhibits larger deviations from the ideal in the low and mid scenarios, suggesting it tends to overestimate risk tolerance for lower-risk pro- files. However, in the high-risk scenario its pre- dictions are closer t... | https://arxiv.org/abs/2505.18953v1 |
11.70 11.62 12.47 11.95 12.35 11.68 12.18 11.97 12.28 2.02 (±0.28 ) Llama 3.1 (405B) 12.60 12.55 12.65 12.78 12.62 12.12 12.53 12.55 12.85 12.32 2.56 (±0.20 ) Llama 3.3 (70B) 14.70 14.57 15.00 15.18 14.65 14.62 15.40 15.30 15.12 14.70 4.92 (±0.30 ) Mistral small (24B) 13.75 13.97 14.05 15.07 14.03 14.28 14.15 13.78 14.... | https://arxiv.org/abs/2505.18953v1 |
end (e.g., for Claude .7, DeepSeek-V3, Mistral small). However, no country stands out as an absolute outlier across all models: China is lowest for GPT-4o (mini) and GPT-4o but mid-range elsewhere, and Aus- tralia is near the bottom for three models yet near the top for GPT-4o (mini) . These mild bi- ases likely reflec... | https://arxiv.org/abs/2505.18953v1 |
M. Farrell. 2006. The role of emergency funds in a comprehensive financial plan. Journal of Per- sonal Finance . FCA. 2018. Assessing Suitability and Risk Appetite Guidelines . Accessed: 2025-03-17. FINRA. 2012. Suitability and Risk Tolerance Guide- lines . Accessed: 2025-03-17. FSA. 2022. Guidelines on Investment Suit... | https://arxiv.org/abs/2505.18953v1 |
classifications. We acknowledge that these limitations may affect the generalizability of our findings. Future research should expand on this work by incorporating a broader range of demo- graphic factors to provide a more comprehensive view of bias in AI systems. A.2 Related Work The integration of LLMs into financial... | https://arxiv.org/abs/2505.18953v1 |
14.2 14.9 14.1 GPT-4o 11.35 11.85 11.75 11.35 12.4 12.4 10.7 12.3 12.25 11.7 12.3 12.5 11.55 11.65 11.9 11.3 10.75 12.25 12.4 12.1 Gemini 1.5 (Pro) 12.9 13.4 13.1 12.75 12.6 12.6 13.15 12.3 13.0 12.6 13.2 12.6 12.85 13.05 13.0 13.15 13.0 13.3 13.1 12.7 Claude 3.7 (Sonnet) 12.75 11.75 12.45 12.35 12.45 12.8 12.55 12.4 1... | https://arxiv.org/abs/2505.18953v1 |
focusing on biases in LLM-generated in- vestment advice. We aim to develop a systematic framework for evaluating these biases to eventually improve the fairness and reliability of AI-driven financial recommendations. A.3 Country-Level Bias Analysis. Table 3 shows the average scores (over Low, Mid, and High scenarios) f... | https://arxiv.org/abs/2505.18953v1 |
and Brazil (differences up to +0.6).•GPT-4o again alternates: male scores exceed female ones in the USA, Sweden, Singapore, and Brazil, but the pattern reverses for Aus- tralia, Indonesia, and a few others. •Models like Gemini 1.5, Claude 3.7, and Mistral small similarly oscillate across coun- tries, with some leaning ... | https://arxiv.org/abs/2505.18953v1 |
System-1.5 Reasoning : Traversal in Language and Latent Spaces with Dynamic Shortcuts Xiaoqiang Wang1,2Suyuchen Wang1,2Yun Zhu3Bang Liu1,2,4† 1DIRO & Institut Courtois, Université de Montréal 2Mila - Quebec AI Institute;3Apple;4Canada CIFAR AI Chair {xiaoqiang.wang, suyuchen.wang, bang.liu }@umontreal.ca Abstract Chain... | https://arxiv.org/abs/2505.18962v1 |
Hidden States(a) System -1.5 Reasoning (Ours) (b) Chain -of-Thought ( CoT) (c) Early Exit (d) Compressed Latent Reasoning (e) Extra Latent ReasoningFigure 1: Comparison of (a) our proposed System-1.5 Reasoning , (b) chain-of-thought (CoT) reasoning, (c) early-exit (Elbayad et al., 2019; Elhoushi et al., 2024), (d) comp... | https://arxiv.org/abs/2505.18962v1 |
slow reasoning paths in latent space to handle varying step complexities in the language space. System-1.5 Reasoning introduces model depth shortcut (DS) and decoding step shortcut (SS) to adaptively allocate computation along the vertical and horizontal paths in latent space. Specifically, the depth shortcut is implem... | https://arxiv.org/abs/2505.18962v1 |
to be skipped by copying hidden states at early exit points and directly reusing them as the hidden states for the next decoding step at the same layer. To train System-1.5 Reasoning effectively in latent space, we employ a two-stage distillation process. First, we perform language-to-latent distillation by fine-tuning... | https://arxiv.org/abs/2505.18962v1 |
(2) During inference, the output of the router Rlserves as a confidence score for early exit, which is compared against a predefined depth exit threshold λdepthto determine whether to halt computation at the current layer for the given decoding step. hsystem-1.5-ds-inference l,t=gl−1(hl−1,t),ifRl(hl−1,t)> λ depth, fl(... | https://arxiv.org/abs/2505.18962v1 |
shortcut paths poses two challenges: the latent property and the adaptive property. (1) Vanilla Transformer layers in pre-trained LLMs are optimized for next-token prediction in the language space and are not inherently capable of reasoning in latent space. How can we align latent-space reasoning with language-space Co... | https://arxiv.org/abs/2505.18962v1 |
(8) where Hteacher L ={hteacher L,TQ, hteacher L,TQ+1,···, hteacher L,TQ+TR−1}denotes the extracted last-layer hidden states of the reasoning steps, provided as input to the student model to enable teacher-forcing. The overall loss for language-to-latent alignment is thus: L1=LLM θteacher+LLM θstudent+αLconsistency θst... | https://arxiv.org/abs/2505.18962v1 |
augmented GSM8K dataset (Deng et al., 2023), which extends the original GSM8K (Cobbe et al., 2021) with a larger set of grade school-level math problems. For common- sense reasoning, we use StrategyQA (Geva et al., 2021), which contains multihop questions annotated with supporting facts. Each sample includes a question... | https://arxiv.org/abs/2505.18962v1 |
CoT fine-tuning results of 46.67% and 38.28%. On StrategyQA, a task requiring reasoning over multiple pieces of textual evidence, System-1.5 Reasoning achieves 48.61% accuracy, outperforming CoT’s 47.36%. Similar improvements are observed for latent-space reasoning methods such as pause token, highlighting the promisin... | https://arxiv.org/abs/2505.18962v1 |
of System-1.5 Reasoning , with joint learning exhibiting a more pronounced drop. We attribute this to optimization conflicts between the two groups of parameters: Transformer parameters responsible for latent reasoning and router parameters responsible for shortcut routing. These conflicts likely hinder the model’s abi... | https://arxiv.org/abs/2505.18962v1 |
or model switching (Qu et al., 2025), where inputs are routed between different reasoning systems—for example, a fast System-1 and a slower, more deliberate System-2. System-1.x (Saha et al., 2024) combines linear reasoning chains and search trees to enable fast and accurate planning in maze navigation tasks, while FaS... | https://arxiv.org/abs/2505.18962v1 |
Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818 , 2023. Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, et al. Do not think that much for 2+ 3... | https://arxiv.org/abs/2505.18962v1 |
Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capabili... | https://arxiv.org/abs/2505.18962v1 |
data synthesis. arXiv preprint arXiv:2502.11476 , 2025a. Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng, Pei-Jie Wang, Xiuyi Chen, et al. From system 1 to system 2: A survey of reasoning large language models. arXiv preprint arXiv:2502.17419 , 2025b. Percy ... | https://arxiv.org/abs/2505.18962v1 |
Codi: Compressing chain-of-thought into continuous space via self-distillation. arXiv preprint arXiv:2502.21074 , 2025. Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314 , 2024. Mingyang S... | https://arxiv.org/abs/2505.18962v1 |
November 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.emnlp-main.734. URL https: //aclanthology.org/2024.emnlp-main.734/ . Xiaoqiang Wang, Suyuchen Wang, Yun Zhu, and Bang Liu. R3mem: Bridging memory retention and retrieval via reversible compression. arXiv preprint arXiv:2502.15957 , 2025. Ja... | https://arxiv.org/abs/2505.18962v1 |
LEARNING TO EXPLAIN : PROTOTYPE -BASED SURROGATE MODELS FOR LLM C LASSIFICATION Bowen Wei Department of Computer Science George Mason University Fairfax, V A 22030 bwei2@gmu.edu Ziwei Zhu Department of Computer Science George Mason University Fairfax, V A 22030 zzhu20@gmu.edu ABSTRACT Large language models (LLMs) have ... | https://arxiv.org/abs/2505.18970v1 |
including recognition Chen et al. [2019], classification Li et al. [2018], out-of-arXiv:2505.18970v1 [cs.CL] 25 May 2025 distribution detection Ming et al. [2019], domain adaptation Tan et al. [2018a], and segmentation Donnelly et al. [2018]. Their intuitive "this looks like that" explanations Li et al. [2018] facilita... | https://arxiv.org/abs/2505.18970v1 |
Overall Structure Problem Formulation. Given a target black-box LLM Mtarget and a set of text samples with corresponding predictions fromMtarget , our goal is to construct an interpretable surrogate model that faithfully explains these predictions. Specifically, the set of text samples is denoted as X={X1, X2, . . . , ... | https://arxiv.org/abs/2505.18970v1 |
extremely comfortable.”) activates the Cleanliness prototype (cosine similarity = 0.92) with a learned positive-class weight of 0.95, contributing 0.92 ×0.95 = 0.87 to the positive logit; the second sentence (“The front-desk staff went out of their way to help with late check-in.”) activates the Service prototype (simi... | https://arxiv.org/abs/2505.18970v1 |
embeddings. To provide interpretable explanations, we associate each prototype with its nearest sentence from the training data based on cosine similarity. The prototype embeddings can either be fixed after initialization or further refined through training. We conducted experiments to compare the performance of these ... | https://arxiv.org/abs/2505.18970v1 |
in Appendix B.1. Evaluation Metrics. We assess faithfulness using seven metrics: Accuracy (Acc), Comprehensiveness (Comp) DeYoung et al. [2020], Sufficiency (Suff) DeYoung et al. [2020], Decision Flip Fraction (DFF) Serrano and Smith [2019], Decision Flip with Most Important Sentence (DFS) Chrysostomou and Aletras [202... | https://arxiv.org/abs/2505.18970v1 |
explanations more faithfully reflect the model’s decision process. 4.3 Impact of Training Data Size (RQ2) In this section, we investigate how training data size affects ProtoSurE’s ability to faithfully reproduce target LLM predictions. Figure 2 presents performance across varying training data sizes for all four targe... | https://arxiv.org/abs/2505.18970v1 |
by BGE (0.8939, 2.50) and E5 (0.8933, 3.00). The relatively small performance differences between encoders (within 0.007 accuracy points) demonstrate that ProtoSurE is robust and not limited to any single embedding model. This flexibility is particularly valuable, as it allows practitioners to select encoders based on ... | https://arxiv.org/abs/2505.18970v1 |
of ProtoSurE . Figure 4 demonstrates ProtoSurE’s explainability through a hotel review example. The review contains five sentences that our system analyzes against ten prototype categories. Two sentences show strong matches: the mention of "an amazing time" (similarity score: 0.81) with Prototype 1 (Overall positive ex... | https://arxiv.org/abs/2505.18970v1 |
and trustworthiness of large language models through post-hoc explanation techniques. ProtoSurE is designed as an interpretability tool and does not modify the underlying LLM, thereby avoiding 10 Figure 4: Visualization of the sentiment analysis process for a hotel review using ProtoSurE. The visualization shows: (1) t... | https://arxiv.org/abs/2505.18970v1 |
2018. Yao Ming, Panpan Xu, Huamin Qu, and Liu Ren. Interpretable and steerable sequence learning via prototypes. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , 2019. Jun Tan, Changxin Wang, Bo Li, Qing Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Feature adapta... | https://arxiv.org/abs/2505.18970v1 |
31 (NeurIPS 2018) , pages 7775–7784. Curran Associates, Inc., 2018. Ronny Luss, Pin-Yu Chen, Amit Dhurandhar, Prasanna Sattigeri, Karthikeyan Shanmugam, and Chun-Chen Tu. Leveraging latent features for local explanations. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining , pages 1134–11... | https://arxiv.org/abs/2505.18970v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.