text string | source string |
|---|---|
Rewriter Agent Given a piece of content containing queries to search, your task is to rewrite the queries in order to obtain more comprehensive search results. Please provide at least three rewritten queries. The output format should be the following JSON structure: {"queries": ["query 1", "query 2", "query 3"]} Here a... | https://arxiv.org/abs/2505.20285v2 |
the director of the movie "Sense and Sensibility", I will take the following steps: 1. First, determine who directed the movie "Sense and Sensibility". 2. Then, search for this director’s educational background, focusing on undergraduate education. I will now begin with step one by searching online to identify the dire... | https://arxiv.org/abs/2505.20285v2 |
the suspense drama that premiered on MBC? Thought :<think> To solve this problem, I need to follow a few steps in my search: 1. First, I need to identify which suspense drama premiered on MBC. 2. Then, I will look up who directed that suspense drama. First, I will search for recent suspense dramas that premiered on MBC... | https://arxiv.org/abs/2505.20285v2 |
Question: {question} Correct Answer: {answer} Model Response: {model_response} E Reinforcement Learning E.1 Training Template Table 7: Training template. The question is appended at the end during RL training and inference. Answer the given question. You must conduct reasoning inside <think> and </think> first every ti... | https://arxiv.org/abs/2505.20285v2 |
the standard answer is included and there is no incorrect or contradictory information. Incorrect Responses Here are examples of Incorrect responses: Question: What are Barack Obama’s children’s names? Standard Answer: Malia Obama and Sasha Obama Model Prediction 1: Malia Model Prediction 2: Malia, Sasha, Susan, and Sa... | https://arxiv.org/abs/2505.20285v2 |
following: Question: {question} Standard Answer: {target} Predicted Answer: {predicted answer} Only return the option represented by Correct orIncorrect , that is, only return A or B, without adding any other text. F RAMP Case Study We illustrate how the model acquires the ability to retrieve and reason using the Retri... | https://arxiv.org/abs/2505.20285v2 |
in the National Football League for the Washington Redskins. He played college football at Eastern Kentucky University. </information> <think> After finding out that David Hoelscher is indeed associated with the National Football League (NFL) and a specific team, I will now confirm the name of the team he played for in... | https://arxiv.org/abs/2505.20285v2 |
1998. He played college football at Eastern Kentucky University. </answer> G Broader Impacts As an exploration in the field of Search Agents, M ASKSEARCH has the following impacts: •Provides a general pre-training method for Search Agent: MASKSEARCH introduces the Retrieval Augmented Mask Prediction Pre-training Task, ... | https://arxiv.org/abs/2505.20285v2 |
Visualized Text-to-Image Retrieval Di Wu*, Yixin Wan*, Kai-Wei Chang University of California, Los Angeles {diwu,elaine1wan,kwchang}@cs.ucla.edu Abstract We propose Visualize-then-Retrieve ( VisRet ), a new paradigm for Text-to-Image (T2I) retrieval that mitigates the limitations of cross-modal similarity alignment of ... | https://arxiv.org/abs/2505.20291v1 |
the desired entity, posture, and camera angle at the same time. Second, by operating entirely within the image modality during retrieval, VisRet avoids the weaknesses of cross-modal retrievers and instead 1arXiv:2505.20291v1 [cs.CV] 26 May 2025 T2I ModelLVLM Reader Text Embedding Can’t see the underside of wings! Imag ... | https://arxiv.org/abs/2505.20291v1 |
as textual query expansion (Levy et al., 2023; Lee et al., 2024) and reranking (Liu et al., 2024; Feng et al., 2025). Finally, a recent line of work introduces generative image retrieval (Li et al., 2024; Qu et al., 2025), which trains a generative model to directly memorize an index of the image corpus. Different from... | https://arxiv.org/abs/2505.20291v1 |
To aggregate the mseparate retrieval results, we apply Reciprocal Rank Fusion (RRF) (Cormack et al., 2009). RRF assigns a fusion score to eachcandidate image rbased on its rank across mlists: score RRF(r) =mX i=11 λ+rank i(r), where rank i(r)is the rank position of image rin list R(vi,I), and λis a hyperparameter that ... | https://arxiv.org/abs/2505.20291v1 |
model and the rephrase LLM in Appendix B.1 and Appendix B.2. 4.3 Downstream QA Performance To assess the utility of VisRet in real-world applica- tions, we evaluate its downstream VQA accuracy in a RAG pipeline. We compare three settings: (1) using only the model’s internal knowledge, (2) RAG with original text query-b... | https://arxiv.org/abs/2505.20291v1 |
explore solutions to similar trustworthy issue with multi- modal generative models. Artifact Release Our Visual-RAG-ME annota- tion is based on Visual-RAG, which is under CC BY-NC 4.0 license and the images shared by the iNaturalist 2021 dataset, which are under one of CC BY 4.0, CC BYNC 4.0, CC BY-NC-ND 4.0, CC BY-NC-... | https://arxiv.org/abs/2505.20291v1 |
et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 . Xintong Han, Zuxuan Wu, Phoenix X Huang, Xiao Zhang, Menglong Zhu, Yuan Li, Yang Zhao, and Larry S Davis. 2017. Automatic spatially-aware fashion concept discovery. In Proceedings of the IEEE international conference on computer vision , pages 1... | https://arxiv.org/abs/2505.20291v1 |
Wenjie Wang, Yongqi Li, Liqiang Nie, and Tat-Seng Chua. 2025. Tiger: Unifying text-to-image generation and retrieval with large multimodal models. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ily... | https://arxiv.org/abs/2505.20291v1 |
Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. 2022. Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res. , 2022. Mert Yüksekgönül, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. 2023. When and why vision-language models behave like bags- of-words, and what to d... | https://arxiv.org/abs/2505.20291v1 |
questions can cover overly similar topics. After the filtering step, we ended up with 50 high quality multi-entity queries. Retrieval Label Annotation and Balancing Fi- nally, for each question, we collect images of the second entity from iNaturalist and annotate their retrieval label. A positive label is assigned only... | https://arxiv.org/abs/2505.20291v1 |
investigate the effectiveness of VisRet from more perspectives, including the choices of T2I generation Model, T2I Instruction LLM, the downstream VQA LVLM reader. Finally, inspired by the generative retrieval literature, we conduct a pilot study of whether the generated images could be directly used as the knowledge c... | https://arxiv.org/abs/2505.20291v1 |
VQA experiments with two additional LVLMs: GPT-4o-mini (version gpt-4o-mini-2024-07-18 ) and GPT-4.1 (version gpt-4.1-2025-04-14 ). Overall, we observe sim- ilar trends as those presented in Figure 2. Both direct T2I retrieval and VisRet outperform only relying on the model’s knowledge, with VisRet substantially outper... | https://arxiv.org/abs/2505.20291v1 |
diversity. Therefore, we followed this setting in this paper and save further perturbing the instruction as future work. VisRet : Retrieval After obtaining the generated visualizations, we encode both the visualized images and the image corpus via an off-the-shelf CLIP3or E5-V4encoder and perform a similarity search. C... | https://arxiv.org/abs/2505.20291v1 |
into a short descriptive phrase that highlights the key part of the entity where the queried feature could be found. DO NOT include the asked feature (shape, color, etc.) but instead include the part of the entity where the feature could be found. Output only the rephrased query. Examples: Original query: What shape ar... | https://arxiv.org/abs/2505.20291v1 |
the reasoning. Please always follow the answer format without bolding texts: "### Reasoning: {reasoning}\n### Answer: {your_answer}" Figure 6: Prompt for VQA on Visual-RAG. You are a model that rigorously answers a question that compares a visual feature of two organisms (animal, plant, etc.) using systematic reasoning... | https://arxiv.org/abs/2505.20291v1 |
the score is needed, do not start a new line and append after remark with brackets, e.g. "Score: {score} | Redundant | (Explanation: abc)". Following are few examples: Question: Is there any specific color marking around the eyes of a semipalmated plover (scientific name: Charadrius semipalmatus)? Reference Answer: bla... | https://arxiv.org/abs/2505.20291v1 |
arXiv:2505.20293v1 [cs.CL] 26 May 2025 Enhancing the Comprehensibility of Text Explanations via Unsupervised Concept Discovery Yifan Sun1,2, Danding Wang1, Qiang Sheng1, Juan Cao1,2, and Jintao Li1,2 1Media Synthesis and Forensics Lab, Institute of Computing Technology, Chinese Academy of Sciences 2University of Chines... | https://arxiv.org/abs/2505.20293v1 |
LayerConcept EvaluatorBeer Review Linear LayerPredefined Concepts Appearance Aroma PalateTaste ①SUP. Baseline ②UNSUP. Baseline ③ECO -Concept Discover new concepts Easy to comprehend Prediction① Prediction③ ② EnhanceConcept ExtractorBeer Review Linear Layer Prediction Figure 1: Comparison of explanations between our pro... | https://arxiv.org/abs/2505.20293v1 |
that models truly comprehend or employ the adopted concepts, as the models were not exposed to these concepts during training (Rudin, 2019, Poeta et al., 2023). Additionally, some unsupervised post-hoc methods may extract concepts lacking semantic meaning. Concept-based Self-explaining Methods aim to provide a built-in... | https://arxiv.org/abs/2505.20293v1 |
Loss: 𝓛𝒄𝒐𝒏Concept Prototypes Optimization (b) Concept Extractor Figure 2: (a) Illustration of the proposed framework ECO-Concept. ECO-Concept consists of a concept extractor, a classifier, and a concept evaluator. (b) The concept extractor takes the encoded text Xas input and interacts with the concept prototypes C... | https://arxiv.org/abs/2505.20293v1 |
Aby applying a sparse softmax function ϕacross the concept slots (along the Maxis). This normalization introduces competition among slots to attend to each input token. The sparsity normalization ensures that each input token is primarily associated with a limited number of concepts, facilitating a more focused and int... | https://arxiv.org/abs/2505.20293v1 |
tmin a mini-batch ℬ, the concept importance score βmis: βm=1 ∣ℬ∣∑ t∈ℬtmΩ ∑ ω=1∣Wm,ω∣. (8) Concept Comprehensibility. To measure the comprehensibility of concepts, we refer to the automated interpretability evaluation methods used for explaining neurons in LLMs (Bills et al., 2023, Templeton et al., 2024). In summary, t... | https://arxiv.org/abs/2505.20293v1 |
The selected baselines include black-box, supervised concept-based, and unsupervised concept- based methods. The first group, black-box methods, directly tackles text classification tasks without inter- pretatability, including a BERT-based classifier (Devlin et al., 2019) and a RoBERTa-based classifier (Liu et al., 20... | https://arxiv.org/abs/2505.20293v1 |
post-hoc methods, and ProtoTEx is a self-explaining method. They extract concepts as representative training samples or text segments. For a fair comparison, we summarize the concept interpretation of these methods using the top 5 representative samples per concept, applying the same summary prompt as ECO-Concept. Note... | https://arxiv.org/abs/2505.20293v1 |
in identifying the intruder indicates that the concepts are more intuitive and easier for humans to understand. Table 3 summarizes the results. Our method achieves the highest intruder detection accuracy across all tasks, demonstrating that the concepts extracted by our approach are more comprehensible. Table 3: Accura... | https://arxiv.org/abs/2505.20293v1 |
variant models: w/o ℒconand w/oℒdist, which respectively remove the concept consistency regularizer and the concept distinctiveness regularizer. Experimental results are shown in Table 5. The results show that the best task performance and conceptinterpretabilityareachievedwhenbothregularizersareappliedtogether. Bycomp... | https://arxiv.org/abs/2505.20293v1 |
Base model), ECO-Concept, and ECO-Concept (w/o ℒcom). 5. Conclusion To automatically extract human-understandable concept explanations with no predefined concept anno- tations, we proposed ECO-Concept, an intrinsically interpretable framework. ECO-Concept employs an object-centric architecture based on the slot attenti... | https://arxiv.org/abs/2505.20293v1 |
Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, an... | https://arxiv.org/abs/2505.20293v1 |
of the International Conference on Neural Information Processing Systems, volume 29, pages 2288–2296. Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. 2018. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In Internati... | https://arxiv.org/abs/2505.20293v1 |
of Text Explanations via Unsupervised Concept Discovery Zhen Tan, Tianlong Chen, Zhenyu Zhang, and Huan Liu. 2024a. Sparsity-guided holistic explanation for llms with interpretable inference-time intervention. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 21619–21627. Zhen Tan, Lu ... | https://arxiv.org/abs/2505.20293v1 |
as concept labels for supervised concept-based methods. Hotel(Wang et al., 2010) is a multi-aspect hotel reviews dataset, comprising review texts annotated with seven concept labels: value, rooms, location, cleanliness, check-in/front desk, service, and business service. Since the original labels are on a scale of 0 to... | https://arxiv.org/abs/2505.20293v1 |
demonstrates superior perfor- mance in interpretability compared to the best baseline, achieving enhanced comprehensibility (semantics, distinctiveness, consistency) of 25%. We believe the limited additional computational cost during training is a worthwhile trade-off for these significant benefits, especially as it do... | https://arxiv.org/abs/2505.20293v1 |
each concept along with its summary and ask participants to rate these examples based on the following criteria using a 1-5 scale: Consistency : Do you think this concept formed by all these highlighted text parts has consistent semantic meanings? Clarity: Do you think the semantic meaning of this concept is clear and ... | https://arxiv.org/abs/2505.20293v1 |
review case 20 Enhancing the Comprehensibility of Text Explanations via Unsupervised Concept Discovery E. Parameter Sensitivity Analysis E.1. Impact of the Number of Concepts To evaluate the impact of the number of concepts on task performance and concept-related metrics, we conducted experiments with 10, 20, 30, 40, a... | https://arxiv.org/abs/2505.20293v1 |
we randomly applied one of these adversarial strategies: synonym replacement or spelling error. For synonym replacement, we selected a synonym from WordNet (Miller, 1995), prioritizing the one with the lowest frequency in the training set. For spelling errors, we randomly altered a single character in the word. To eval... | https://arxiv.org/abs/2505.20293v1 |
evaluation Highest education level completed Avg. Age High School Undergraduate Graduate Doctorate 3 10 7 1 29.1 Ethnicity Female:Male White Black Asian Mixed 8 6 5 2 1:1.33 Table 13: Demographic information of participants in forward simulatability experiments Highest education level completed Avg. Age High School Und... | https://arxiv.org/abs/2505.20293v1 |
arXiv:2505.20295v1 [cs.CL] 26 May 2025Self-reflective Uncertainties: Do LLMs Know Their Internal Answer Distribution? Michael Kirchhof AppleLuca Füger Independent ResearcherAdam Goli ´nski Apple Eeshan Gunesh Dhekane AppleArno Blaas AppleSinead Williamson Apple Abstract To reveal when a large language model (LLM) is un... | https://arxiv.org/abs/2505.20295v1 |
government and many commercial hubs.Marseilles is one of France's most popular and vibrant cities.It's Paris.It's Toulouse.The capital of France is Paris.Its main city is Paris.Normal (greedy) answer: 'The capital of France is Paris.'Figure 1: LLMs have internal answer distributions about user queries. Rather than just... | https://arxiv.org/abs/2505.20295v1 |
This considers the full distribution over strings, but it still reduces the uncertainty to a single number and presents this number alongside a single string from the distribution. Moving towards richer uncertainty explications, Xu et al. [2024] generate multiple samples from an LLM, use GPT-4 to summarize the distribu... | https://arxiv.org/abs/2505.20295v1 |
theoretical foundation for sufficient statistics in string spaces in Section 3.1 and develop the SelfReflect metric in Section 3.2. 3.1 Summaries as predictive sufficient statistics BΘQ A(1:N) S Figure 2: Graphical model for the suf- ficiency that SelfReflect quantifies.Suppose we have an LLM (which we denote LLM θ), p... | https://arxiv.org/abs/2505.20295v1 |
Bifrom the rest of the words B−iwith the extra context of either the sampled answers A(1:N)or their summary S. Identical behavior in this masked-token prediction task turns out to be equivalent to predictive sufficiency (and hence, Definition 3.1): Proposition 3.2 (Informal; towards the SelfReflect metric) .For answers... | https://arxiv.org/abs/2505.20295v1 |
Australian Prime Minister Edmund _ was elected in 1901. Please provide only the missing word "_", not the whole sentence. <|im_end|> <|im_start|>assistant candidate summary si.i.d. samples from pasted into contextpθ(A|q) Predicted token vector: pJ(Bi|q,s,b−i)=(0.70,0.28,0.01,...)Predicted token vector: pJ(Bi|q,a(1:N),b... | https://arxiv.org/abs/2505.20295v1 |
al., 2019], we let LLM Jsplitsinto a “distribution” over atomic statements and likelihoods, compute a pairwise entailment matrix and return the Earth Mover’s distance to pθ(A|q). Ablations. We also ablate key characteristics of SelfReflect. SR-PMI forgoes the masked-out task and directly compares the log likelihoods of... | https://arxiv.org/abs/2505.20295v1 |
in literature, can no longer distinguish these fine-grained quality differences. To investigate this further, we test the metrics’ ability to score the existence of details and the relative likelihoods of possible answers in isolation. First, we subsample questions where all answers in the answer distribution are the s... | https://arxiv.org/abs/2505.20295v1 |
Judge 0.80±0.02 1.00±0.00 Opt. Transport 0.63 ±0.02 0.80±0.00 Embedding 0.29 ±0.03 0.18±0.02 SR-PMI -0.03 ±0.03 -0.20 ±0.00 SR-sampling-free 0.57 ±0.03 0.83±0.00 SR-P(True) 0.66 ±0.03 1.00±0.00 SelfReflect 0.66 ±0.03 1.00±0.00Table 2 shows that most metrics have a positive rank correlation with the reference metric. Th... | https://arxiv.org/abs/2505.20295v1 |
±0.125 Human vs human 0.723 ±0.027 0.988 ±0.013 0.906 ±0.035 0.871 ±0.048 0.441 ±0.075 0.636 ±0.064 0.452 ±0.069 As we see from Table 3, SelfReflect has the highest overall alignment with the majority human judgement ( α= 0.690). This is close to the inter-human alignment ( α= 0.723) and significantly higher than any o... | https://arxiv.org/abs/2505.20295v1 |
: methods which utilize only one decoding, without explicitly elicitating intermediate samples. Of particular interest is whether any of the single- decoding methods are able to match the performance of the multi-sample methods. We consider three single-decoding methods: a) Greedy : a baseline simply using a greedy-dec... | https://arxiv.org/abs/2505.20295v1 |
to samples from the answer distribution, focusing on a single large-scale model, Qwen2.5 72B Instruct. To explore whether CoT correctly captures the spread of the answer distribution, i.e., whether it focuses on a single answer when the true distribution is unimodal and includes multiple options when the true distribut... | https://arxiv.org/abs/2505.20295v1 |
is intended to guide the field towards developing methods to make LLMs honestly describe all possible answers to a question. We have seen in our benchmark that this is a 9 hard task, but a solution to this problem would be a fundamental building block in many applications: Describing internal LLM distributions in a str... | https://arxiv.org/abs/2505.20295v1 |
L. Xu, L. Xia, M. Zhang, M. Zhang, M. Tang, M. Li, M. Wang, M. Li, N. Tian, P. Huang, P. Zhang, Q. Wang, Q. Chen, Q. Du, R. Ge, R. Zhang, R. Pan, R. Wang, R. J. Chen, R. L. Jin, R. Chen, S. Lu, S. Zhou, S. Chen, S. Ye, S. Wang, S. Yu, S. Zhou, S. Pan, S. S. Li, S. Zhou, S. Wu, S. Ye, T. Yun, T. Pei, T. Sun, T. Wang, W.... | https://arxiv.org/abs/2505.20295v1 |
Massive multilingual text embedding benchmark. arXiv preprint arXiv:2502.13595 , 2025. E. Fadeeva, R. Vashurin, A. Tsvigun, A. Vazhentsev, S. Petrakov, K. Fedyanin, D. Vasilev, E. Gon- charova, A. Panchenko, M. Panov, T. Baldwin, and A. Shelmanov. LM-polygraph: Uncertainty es- timation for language models. In Y . Feng ... | https://arxiv.org/abs/2505.20295v1 |
M. Janiewicz, M. Felardos, M. Darrin, M. Hoff, M. Seznec, M. J. Kenyon, N. Derwiche, N. C. Zaragoza, N. Faurie, N. Moreau, N. Schuhl, N. Raghuraman, N. Muhs, O. de Garrigues, P. Rozé, P. Wang, P. von Platen, P. Jacob, P. Buche, P. R. Muddireddy, P. Savas, P. Stock, P. Agrawal, R. de Peretti, R. Sauvestre, R. Sinthe, R.... | https://arxiv.org/abs/2505.20295v1 |
58th Annual Meeting of the Association for Computational Linguistics , July 2020. A. Santilli, A. Golinski, M. Kirchhof, F. Danieli, A. Blaas, M. Xiong, L. Zappella, and S. Williamson. Revisiting uncertainty quantification evaluation in language models: Spurious interactions with response length bias results. arXiv pre... | https://arxiv.org/abs/2505.20295v1 |
proofs. We begin with the definition of predictive sufficiency and provide a proof of its two equivalent characterizations in the context of the SelfReflect metric. We then prove an equivalence between solving the masked- token prediction task of the SelfReflect metric and the desired predictive sufficiency of the summ... | https://arxiv.org/abs/2505.20295v1 |
assumed to be X, the set of all finite-length sentences generated from vocabulary V. 3.The summarization mechanism that inputs the sampled answers A(1:N)and generates their summary Sis assumed to be a function ψ. Formally, ψ:XN−→ X inputs any Nsampled answers A(1:N)from the LLM and generates their summary SasS:=ψ A(1:... | https://arxiv.org/abs/2505.20295v1 |
related to answering the given question Q. For this, requiring the summary to be predictive sufficient serves the purpose precisely. Now, in the context of the Definition A.1 of predictive sufficiency, Definition 3.1 of ideal summary, and the graphical model of Figure 5, we prove Proposition 3.1, which asserts the equi... | https://arxiv.org/abs/2505.20295v1 |
represents the k−th sampled answer from the LLM. For each Bi, its support is going to be the vocabulary Vand the supports of B−iandB areVL−1andVL≡ XLrespectively. With this setup, we can prove Proposition 3.2, which asserts that under assumptions from subsection A.1, SelfReflect metric provides an equivalent formulatio... | https://arxiv.org/abs/2505.20295v1 |
dividing by the same terms and step (2)follows from rearranging the terms and recognizing x(L)=b∗by definition. Now, we consider the ℓ−th term from the Equation 11 and simplify it as follows: p B=x(ℓ−1)|A(1:N)=a(1:N) p B=x(ℓ)|A(1:N)=a(1:N) =(1)p B1=b∗ 1,···, Bℓ−1=b∗ ℓ−1, Bℓ=bℓ, Bℓ+1=bℓ+1,···, BL=bL|A(1:N)=a(1:N) ... | https://arxiv.org/abs/2505.20295v1 |
sponding S=s:=ψ a(1:N) ), we get the desired results involving all underlying random variables: p B|A(1:N) =p(B|S). This proves the reverse direction of the equivalence. A.4 Modeling with LLM: From derivation to implementation Now, having proved the equivalence of the basis of the SelfReflect metric and the desired... | https://arxiv.org/abs/2505.20295v1 |
where one or more of the assumptions are loosened. Also, further generalizing the SelfReflect metric in cases where the assumptions are loosened or proving that the current formulation holds in those scenarios remains an interesting direction for future theoretical work. 20 B Convergence of the SelfReflect metric In th... | https://arxiv.org/abs/2505.20295v1 |
(n=50)Figure 9: Convergence of the SelfReflect score with N=M= 5 and an increasing number of queries we evaluate on. Answer Distributions of Qwen 2.5 72B Instruct on Natural Questions. 0 1000 2000 3000 4000 5000 Number of Questions/Answer Distributions Averaged Over0.0700.0750.0800.0850.0900.095Average SelfReflect Scor... | https://arxiv.org/abs/2505.20295v1 |
3.1 8B Instruct, Phi 4 14B, Qwen 2.5 7B Instruct (which we ultimately use in the paper), and Qwen 2.5 72B Instruct. We generate answer distributions on Natural Questions for different LLM θ(Llama 3.1 8B Instruct, Phi 4 14B, and Qwen 2.5 7B Instruct), then use Gemini 2.0 to generate summaries like in Section 4.1, and ca... | https://arxiv.org/abs/2505.20295v1 |
question posed to the LLM is “Who received the first Nobel Prize in physics?” . As can be seen below, the LLM’s answer distribution includes Wilhelm Conrad Röntgen as most likely answer, as well as Hendrik Antoon Lorentz and Pieter Zeeman or Henri Becquerel as additional possibilities, and details on their work. Let us... | https://arxiv.org/abs/2505.20295v1 |
on the effect of magnetic fields on the spectrum of light emitted by atoms, known as the Zeeman effect.mid summary: It's most likely that Wilhelm Conrad Röntgen received the first Nobel Prize in Physics. But the laureates could also have been Hendrik Antoon Lorentz and Pieter Zeeman or Henri Becquerel. Figure 12: SelfR... | https://arxiv.org/abs/2505.20295v1 |
second summary which previously had a score of 0.084(or 84) and state that Henri Becquerel was the most likely first Nobel laureate, which is in conflict with the LLM’s internal answer distribution: “It’s most likely that Henri Becquerel received the first Nobel Prize in Physics. But the laureates could also have been ... | https://arxiv.org/abs/2505.20295v1 |
over the vocabulary size for the current token of the fill-in word. If a fill-in word consists of multiple tokens, where we add the tokens of the true fill-in word one after another into the autoregressive context of the assistant answer. Given the two fill-in token vectors conditioned either on the summary or on the c... | https://arxiv.org/abs/2505.20295v1 |
distribution, normalize them and take the inner product to form cosine distances. We average over all samples. The reason why we select this particular embedding model is that at the time of submission it was the best-performing open-source model on the MTEB benchmark [Enevoldsen et al., 2025]. 27 <|im_start|>user Who ... | https://arxiv.org/abs/2505.20295v1 |
well-organized. It should not just be a heap of related information, but should build from sentence to sentence to form a coherent body of information about the topic. Here are some examples: <4 few-shot examples> Now here is the summary whose coherence you are supposed to rate: Summary: {summary} Coherence: Prompt for... | https://arxiv.org/abs/2505.20295v1 |
case: {question} Individual answers: x_1 = ’{answer}’ ... x_{n_answers} = ’{answer}’ Summarized answer: ’{summary}’ Please provide the reason and the score of how good the summarized answer matches the distribution of individual answers. E.8 Optimal Transport score The optimal transport metric consists of two steps. Fi... | https://arxiv.org/abs/2505.20295v1 |
summaries of answer distributions. In this version, the summaries are written by humans rather than by Gemini, on a disjoint set of questions. Mean ±95% interval. Confidence intervals are larger than in Table 1 because we have less manually written summaries of answer distributions than the automated ones in Table 1. M... | https://arxiv.org/abs/2505.20295v1 |
’{answer}’ ... x_{n_answers} = ’{answer}’ Please provide the summarized answer. Gemini 2.0 Flash prompt to generate an ’almost-good’ summary from a ’good’ summary in Table 1. Also used to generate ’truncated’ from ’detailed’ summaries. Below, you are given an answer to the question ’{question}’. Your goal is to shorten... | https://arxiv.org/abs/2505.20295v1 |
list of [x_1, x_26, ...]. Here are the {n_answers} answers: x_1 = ’{answer}’ ... x_{n_answers} = ’{answer}’ Please output the two json files, one after another. Each json file should start with “‘json We then count how many member each cluster has (manually in code as opposed to asking the LLM since this increases accu... | https://arxiv.org/abs/2505.20295v1 |
golden answer in 80% of tasks to proceed. The remaining ten questions were periodically included as verification checks. A total of 215 raters passed the entrance exam and contributed ratings. Confidence intervals were calculated using 100 bootstrapped samples. 35 H Automatic summary generation H.1 Experimental details... | https://arxiv.org/abs/2505.20295v1 |
and "could also be". •The format of the summarized answer should be the same as a typical answer and be stand-alone. •If there is only clear answer to the question, just provide that answer, without hedging across possibilities. The output should be in the following format: Reasoning: [REASONING ABOUT WHICH POSSIBLITIE... | https://arxiv.org/abs/2505.20295v1 |
89 88 −1 92+3 83−6 82−7 Ministral 8B Instruct 2410 [Jiang et al., 2024] 17% 101 99 −1 99−2 95−6 94−7 Llama 3.3 70B Instruct [Meta AI, 2024b] 68% 91 96 +5 104 +13 86−5 85−6 Llama 4 Scout 17B 16e Instruct [Meta AI, 2025] 59% 90 96 +6 104 +14 88−2 87−4 Gemma 3 1B Instruct [Gemma Team et al., 2025] 22% 113 126 +13 127 +14 ... | https://arxiv.org/abs/2505.20295v1 |
Qwen2.5 1.5B Instruct [Yang et al., 2024a] 0.94 0.91 ×0.96 1.29×1.37 1.89×2.01 Qwen2.5 3B Instruct [Yang et al., 2024a] 0.84 0.85 ×1.02 1.00×1.20 1.82×2.18 Qwen2.5 7B Instruct [Yang et al., 2024a] 0.80 0.84 ×1.05 1.09×1.36 1.91×2.38 Qwen2.5 14B Instruct [Yang et al., 2024a] 0.96 1.01 ×1.05 1.33×1.38 2.44×2.53 Qwen2.5 3... | https://arxiv.org/abs/2505.20295v1 |
matrices between certainty of CoT summaries vs. actual answer distributions for Qwen2.5 72B Instruct. Judged by Gemini 2.0 Flash for 500random questions per dataset. We show the results per dataset in Fig. 17. By just looking at the marginals of the answer distribution, we can infer that the question difficulty increas... | https://arxiv.org/abs/2505.20295v1 |
goal is to classify which type of distribution this is: A. The answers all do not contradict each other, up to one or two that differ from the majority answer. B. The answers give multiple mutually exclusive answer options, but there is one answer option that is given in the majority of cases. C. The answers give multi... | https://arxiv.org/abs/2505.20295v1 |
arXiv:2505.20296v1 [cs.CL] 26 May 2025Reasoning LLMs are Wandering Solution Explorers Jiahao Lu∗Ziwei Xu∗Mohan Kankanhalli NUS AI Institute National University of Singapore jiahao.lu@u.nus.edu, ziwei.xu@u.nus.edu, mohan@comp.nus.edu.sg Abstract Large Language Models (LLMs) have demonstrated impressive reasoning abiliti... | https://arxiv.org/abs/2505.20296v1 |
be an acceptable way of reasoning for easy problems with small solution spaces. More importantly, such deterioration could appear minor or negligible for small to moderately complex problems and cause illusions of achieving perfect performances in limited benchmarks. However, the AI model’s performance could suddenly s... | https://arxiv.org/abs/2505.20296v1 |
2 solutions), while qwcaptures the RLLM’s ability to explore systematically – higher values correspond to more consistent search behaviour. Eq. (1) reveals that success probability drops exponentially withdfor wandering RLLMs. As shown in Fig. 2, RLLMs may exhibit a performance “plateau” at lowd, particularly when mult... | https://arxiv.org/abs/2505.20296v1 |
of a Wandering Exploration A wandering exploration violates at least one of the properties of a systematic exploration outlined in Definition 2. This section identifies possible failures modes of a wandering exploration, which can be generally categorised into the following three classes below •Invalid Exploration : wh... | https://arxiv.org/abs/2505.20296v1 |
including (a)controllable problem size : the required number of atomic reasoning steps can be controlled by changing problem specifications; (b) verifiable trace : the solution is decomposable to atomic steps in a common symbolic system, which enables tracking and comparison of solutions; and (c) standard solving proce... | https://arxiv.org/abs/2505.20296v1 |
maintain awareness of global constraints. 4.1.2 Procedure Omission Definition . Aprocedure omission refers to any exploratory trajectory that terminates prematurely or skips essential sub-regions of the search space that are required to reach a correct solution. Examples . In problems with multiple goal states ( e.g.,p... | https://arxiv.org/abs/2505.20296v1 |
permutations of multiset is 9!/(8!1!) which is 9. . . . Wait, but how does the algorithm place the 2 in different positions between the first and last elements?. . . The final path before backtracking would be [1,1,1,1,1,1,1,1,2], which is done. Then backtrack all the way up to the first element, then choose the next o... | https://arxiv.org/abs/2505.20296v1 |
. Hmm. Wait, maybe I need to try a different combination. . . . For example, ( ( (13 - (8/2)) * (12 / something) ).. . . No, already used. Hmm. . . . (Author’s note: Repeating this loop for 90times and exhausting all token budgets!) ♂lightbulbFinal Answer (Author’s note: all token budgets are consumed, no answer blocks... | https://arxiv.org/abs/2505.20296v1 |
actions. UnfaithfulConclusionS0S1S2S3 Thinking:Answer:S0S3♂question-circleTask Using four integers [12, 13, 2, 8] to play the 24 game. (Author’s note: a possible solution is8∗(2 + 13 −12)). /simplybuiltModel Deepseek-R1 ☁Thinking . . . Wait, maybe 13 - (12/(8/2))? Let’s see. . . . 13 - (12/(8/2))? Let’s see. . . . 12 *... | https://arxiv.org/abs/2505.20296v1 |
all its trial-and-error history, it recalls only a small subset. This phenomenon highlights a failure in faithful conclusion despite valid prior explorations. Potential causes . The RLLM is not explicitly trained to faithfully summarize its entire reasoning history; instead, it is primarily optimized to generate what a... | https://arxiv.org/abs/2505.20296v1 |
small benchmarks yet degrade rapidly on deeper, more complex tasks. This calls for new evaluation tools that go beyond final-answer accuracy and assess the process of problem solving. For example, solution trace validity, search completeness, or coverage metrics could be the crucial components of such a benchmark. Addi... | https://arxiv.org/abs/2505.20296v1 |
Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv:2408.03314 , 2024. [17] Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Hanjie Chen, et al. Stop overthinking: A survey on efficient reasoning ... | https://arxiv.org/abs/2505.20296v1 |
preprint arXiv:2501.07301 , 2025. [34] Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559 , 2024. [35] Zhi Zheng, Zhuoliang Xie, Zhenkun Wang, and Brya... | https://arxiv.org/abs/2505.20296v1 |
solvers or other search algorithms. PuzzleBench [13] collects NP-hard combinatorial puzzles to reveal how current chain-of-thought and tool-augemented strategies break down on deeper search tasks. ProcessBench [34] targets Olympiad-level mathematics and provides step-by-step gold chains so that models need to not only ... | https://arxiv.org/abs/2505.20296v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.