text string | source string |
|---|---|
prediction/suppression neu- rons (two of the output-based functional roles of Gurnee et al., 2024), i.e., each of the six neurons is a predic- tion/suppression neuron as well as exemplifying one of our six classes. For ease of interpretability, we choose that prediction/suppression neuron of a particular IO class with ... | https://arxiv.org/abs/2505.17936v1 |
again . On the negative side, the activations do not have any obvious semantic relationship to again . We hypoth- esize that sometimes the residual stream ends up near “minus again ” for semantically unrelated reasons (there are many more possible concepts than dimensions, so the corresponding directions cannot be full... | https://arxiv.org/abs/2505.17936v1 |
(Their analysis is not weight-based, so these may or may not be depletion neurons in our weight-based sense.) This confirms the importance of our work for models with gated activation functions: their internal structure is quite different from older models with GeLU or ReLU. Despite minor differences (especially in the... | https://arxiv.org/abs/2505.17936v1 |
especially with conditional enrichment neurons in early- middle layers. These neurons seem a good fit for the feature engi- neering stage (Lad et al., 2024), corresponding to en- richment as defined by Geva et al. (2023). Indeed, they output a direction similar to the one they detect, which could correspond to related ... | https://arxiv.org/abs/2505.17936v1 |
to investigate the evolution of IO functionalities during model training. Finally, we would like to go beyond the analysis of sin- gle neurons and address the question of how neurons work together within and across IO classes. 8 Limitations This paper focuses on a parameter-based interpretation ofsingle neurons . This ... | https://arxiv.org/abs/2505.17936v1 |
As- sociation for Computational Linguistics. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key- value memories. In Proceedings of the 2021 Confer- ence on Empirical Methods in Natural Language Pro- cessing , pages 5484–5495, Online and Punta Cana, Dominican Republic.... | https://arxiv.org/abs/2505.17936v1 |
Ananya Jha, Sachin Ku- mar, Li Lucy, Xinxi Lyu, Nathan Lambert, Ian Magnusson, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew Peters, Ab- hilasha Ravichander, Kyle Richardson, Zejiang Shen, Emma Strubell, Nishant Subramani, Oyvind Tafjord, Evan Walsh, Luke Zettlemoyer, Noah Smith, Han- naneh H... | https://arxiv.org/abs/2505.17936v1 |
software at this Github URL. See the readme file for detailed documentation. The repository also contains the visualizations of max/min activations for the neuron case studies in Sec- tion 6. Everything else can be quickly reproduced, and the plots are included in this paper. The repository is under the Apache 2.0 lice... | https://arxiv.org/abs/2505.17936v1 |
texts and visualize them. We use TransformerLens (Nanda and Bloom, 2022). A colleague kindly provided us with a version that also supports OLMo. E More on SwiGLU Figure 6 visualizes a SwiGLU neuron (described in Section 3). F IO classes vs. functional roles We compare our results with those of another classifica- tion ... | https://arxiv.org/abs/2505.17936v1 |
prediction/suppression class is disjoint from par- tition . This gives a (very high) excess kurtosis of 230.9736. •Entropy : Following Stolfo et al. (2024), we focus on the last layer, and we define the null space of WUas the subspace of model space spanned by its last 40 singular vectors. We find that two neurons have... | https://arxiv.org/abs/2505.17936v1 |
enrichment 6 0 0 00 18 24 at. enrichment 15 0 1 00 220 236 total 1,319 410 1,000 24 21 349,521 352,256 Table 2: Contingency table of IO classes (rows) vs Gurnee et al. (2024)’s functional roles (columns) for OLMo-7B- 0424. c = conditional. at = atypical. Cutoffs for prediction/suppression and partition were chosen as d... | https://arxiv.org/abs/2505.17936v1 |
to the Arabo-Islamic world. The same goes for −wout(as it is similar to win). Activates on Muhammad . 24.4880 : For all three weight vectors the first four to- kens (but not more) are Philippine-related (even though the gate vector is actually not very similar to the oth- ers). The gate vector also reacts to other geog... | https://arxiv.org/abs/2505.17936v1 |
three vectors correspond to tokens re- lated to cities. Moreover, −woutseems to correspond to non-city places, such as national governments or vil- lages. winis actually not that similar to wgate, wout(in terms of cosine similarities), but all three correspond to city-related tokens. When using WE, in all three weights... | https://arxiv.org/abs/2505.17936v1 |
andwinare also similar and involve iTunes . Activates oniTunes . 29.9734 :wgatereacts to the East in a broad sense as opposed to the West ( Iran, Kaz-akhstan, Kash-mir, Ukraine ...),winmostly to male first names without pre- ceding space. woutseems to produce word pieces that could begin a foreign name. Activates on Mu... | https://arxiv.org/abs/2505.17936v1 |
con- ditional enrichment neurons each. •In some models, especially the OLMo ones, there is a non-negligible number of conditional depletion neurons. They tend to appear in middle-to-late lay- ers, shortly after the conditional enrichment wave. The clearest example is OLMo-1B, with a peakof 1418 conditional depletion ne... | https://arxiv.org/abs/2505.17936v1 |
3698 3403 4349 4358 3417 3897 4923 4789 4265 3766 4413 3639 3702 3387 2881 2975 3778 2997 2976 1738 735 271 147 83 93 89 54 47 15 77 10174 22 20 78 91 137 239 176 350 363 660 613 836 985 1565 1318 1447 1820 1249 1397 1583 1085 979 759 747 477 423 305 203 217 228 271 200 185 163 157 173 209 346 501 600 88314061 14247 14... | https://arxiv.org/abs/2505.17936v1 |
atypical conditional depletion 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27176 597 812 1495 3516 3495 3860 3778 3801 3420 3386 3644 3282 3361 3447 3057 1591 1004 627 497 165 78 37 44 39 19 35 3415 71 52 67 77 102 148 312 366 421 542 474 297 173 89 103 87 73 65 71 62 54 67 134 249 507 753 89... | https://arxiv.org/abs/2505.17936v1 |
18729 18743 18510 1660836 2 4 0 3 1 10 35 38 12 22 13 17 37 27 12 17 18 23 35 26 27 16 15 18 28 43 819Qwen/Qwen2.5-7B enrichment atypical enrichment conditional enrichment atypical conditional enrichment proportional change atypical proportional change orthogonal output depletion atypical depletion conditional depletio... | https://arxiv.org/abs/2505.17936v1 |
Layer 9 Layer 10 Layer 11 1 01Layer 12 Layer 13 Layer 14 Layer 15 1 01Layer 16 Layer 17 Layer 18 Layer 19 1 01Layer 20 Layer 21 Layer 22 Layer 23 1 0 11 01Layer 24 1 0 1Layer 25 1 0 1 1 0 1 1.00 0.75 0.50 0.25 0.000.250.500.751.00 cos(wgate,win) cos(wgate,wout)cos(win,wout)gemma-2-2bFigure 13 23 1 01Layer 0 Layer 1 Lay... | https://arxiv.org/abs/2505.17936v1 |
Layer 13 Layer 14 Layer 15 1 01Layer 16 Layer 17 Layer 18 Layer 19 1 01Layer 20 Layer 21 Layer 22 Layer 23 1 01Layer 24 Layer 25 Layer 26 Layer 27 1 0 11 01Layer 28 1 0 1Layer 29 1 0 1Layer 30 1 0 1Layer 31 1.00 0.75 0.50 0.25 0.000.250.500.751.00 cos(wgate,win) cos(wgate,wout)cos(win,wout)mistral-7bFigure 19 28 1 01La... | https://arxiv.org/abs/2505.17936v1 |
arXiv:2505.17950v1 [cs.CL] 23 May 2025Volume x(x), x—xx. http://dx.doi.org/xxx-xxx-xxx Handling Symbolic Language in Student Texts: A Comparative Study of NLP Embedding Models Tom Bleckmann1, Paul Tschisgale2* Abstract Recent advancements in Natural Language Processing (NLP) have facilitated the analysis of student-gen... | https://arxiv.org/abs/2505.17950v1 |
1 NLP-based LA research between 2021 and 2023, generally outperforming other text analysis methods. The general principle behind NLP embedding models is to transform words, sentences, or even entire documents into high-dimensional numerical vectors—known as embeddings —that capture their semantic content (Qiu et al., 2... | https://arxiv.org/abs/2505.17950v1 |
of the symbolic expressions. To mitigate such issues, some studies have opted to exclude symbolic expressions from textual data (Shang, Huang, Zeng, Zhang, & Wang, 2022). However, both approaches—omitting symbolic expressions or neglecting the information they convey—risk losing potentially valuable information. Since ... | https://arxiv.org/abs/2505.17950v1 |
embedding models is specifically optimized for science education. 2.1 Evaluation of Embedding Model Performance via Similarities 2.1.1 Data Sources and Preparation In the first approach, we examined textual responses from German students to two physics tasks (Bleckmann, 2024; Tschisgale et al., 2023). In one task, stud... | https://arxiv.org/abs/2505.17950v1 |
recipe” “apple pie recipe” 4.Category IRC: Pairs (SEi,IRC i)including incorrect related concepts IRC ithat are somewhat related to the correct related concept RC ibut not directly associated with the symbolic expression SE iitself. 5.Category OT: Pairs (SEi,OTi)where OTi=“apple pie recipe” , serving as an off-topic (OT... | https://arxiv.org/abs/2505.17950v1 |
pie recipe“ γα β A cosine similarity of 1indicates that two embeddings point in the same direction within the embedding space, representing maximum similarity (e.g., synonymous terms), while a value of −1indicates that they point in opposite directions, suggesting strong dissimilarity or antonymy. A value of 0implies t... | https://arxiv.org/abs/2505.17950v1 |
is a paired difference test, meaning it evaluates the distribution of differences by subtracting the similarity values of two paired categories associated with the same symbolic expression (see Eq. (2a) and Eq. (2b)). The test then assesses whether the distribution of these differences is symmetric about zero (null hyp... | https://arxiv.org/abs/2505.17950v1 |
or related concepts, respectively. This may be regarded as a first indicator that the embedding models are, to some extent, capable of interpreting symbolic expressions meaningfully. Interestingly, only the three models German_Semantics_STS_V2 , paraphrase-multilingual-mpnet-base-v2 , and GPT-text-embedding-3-large ass... | https://arxiv.org/abs/2505.17950v1 |
SEiand the correct literal translation LTi(or correct related concept RCi) exceeds the expression’s similarity with the corresponding incorrect literal translation ILT i(or incorrect related concept IRC i), as formally represented in Equations (1a) and (1b). To assess statistical significance, we conducted a paired one... | https://arxiv.org/abs/2505.17950v1 |
to the originally used model ( German_Semantic_STS_V2 ) for both types of student responses. However, the performance of the G-SciEdBERT model in classifying both types of student responses was significantly lower, consistent with the results presented in Section 3.1. Overall, the GPT-text-embedding-ada-002 model provi... | https://arxiv.org/abs/2505.17950v1 |
into account. One such consideration is cost: while a one-time payment for generating embeddings for a fixed text corpus in a research scenario may be feasible, continuous usage in large-scale LA applications could result in substantial financial costs. Another critical consideration is regulatory compliance, including... | https://arxiv.org/abs/2505.17950v1 |
and practitioners of carefully selecting NLP embedding models for applications involving science-related language products, particularly those containing symbolic expressions such as formulas and equations. Our results reveal meaningful differences in the capabilities of current embedding models to handle physics-speci... | https://arxiv.org/abs/2505.17950v1 |
De Lozano, S. R., & Cardenas, M. (2002). Some learning problems concerning the use of symbolic language in physics. Science and Education ,11(6), 589–599. doi: 10.1023/A:1019643420896 Ethayarajh, K. (2019). How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embedding... | https://arxiv.org/abs/2505.17950v1 |
learning. In 2022 IEEE 25th International Conference on Computer Supported Cooperative Work in Design (CSCWD) (pp. 1053–1058). Hangzhou, China: IEEE. doi: 10.1109/CSCWD54268.2022.9776230 Treagust, D., Chittleborough, G., & Mamiala, T. (2003, November). The role of submicroscopic and symbolic rep- resentations in chemic... | https://arxiv.org/abs/2505.17950v1 |
arXiv:2505.17952v1 [cs.CL] 23 May 2025Beyond Distillation: Pushing the Limits of Medical LLM Reasoning with Minimalist Rule-Based RL Che Liu1∗, Haozhe Wang2∗, Jiazhen Pan3∗, Zhongwei Wan4,Yong Dai5,Fangzhen Lin2, Wenjia Bai1,Daniel Rueckert1,3,Rossella Arcucci1 1Imperial College London,2HKUST,3Technical University of M... | https://arxiv.org/abs/2505.17952v1 |
AlphaMed , the first work designed to incentivize reasoning capability solely through minimalist rule-based RL, going beyond conventional approaches that rely on SFT with CoT data. Instead of depending on distilled CoT data supervision, AlphaMed is trained directly via simple rule-based rewards derived from multiple-ch... | https://arxiv.org/abs/2505.17952v1 |
as rule-based supervision signals [16, 18, 22, 23]. Open-Source Medical LLMs. Open-source medical LLMs have emerged as promising tools for domain-specific clinical reasoning, yet most remain heavily dependent on supervised data or hand- crafted feedback. HuatuoGPT [24] was instruction-tuned on ChatGPT-distilled medical... | https://arxiv.org/abs/2505.17952v1 |
adopt Llama3.1-8B-Instruct andLlama3.1-70B-Instruct as back- bone models. All experiments are conducted under full parameter tuning with a batch size of 512, meaning each batch contains 64 QA pairs and each question generates 8 candidate answers, trained for 300 steps. We use verl2[31], a framework designed for rule-ba... | https://arxiv.org/abs/2505.17952v1 |
all three datasets. 4For PubMedQA [34], only questions with definitive answer labels (i.e., A/B/C) are retained. 4 Figure 1: Performance comparison on six medical QA benchmarks. Our models are initialized withLlama3.1-8B-Instruct [45] and trained using minimalist rule-based RL on one of three balanced subsets: MedQA-Su... | https://arxiv.org/abs/2505.17952v1 |
to con- sistent performance gains, highlighting the value of informative data. In contrast, PubMedQA-Sub shows no improvement, reflecting the limitations of low-informative data sources. Figure 4: Effect of data diversity. Aver- age accuracy across six medical QA bench- marks when models are trained individu- ally on s... | https://arxiv.org/abs/2505.17952v1 |
data are incorporated only through separate training runs, not incrementally during training. While performance on MedXpert [37] increases consistently, trends on other benchmarks vary. Final models trained on the full set (L1–L6) generally achieve comparable or superior performance to HuatuoGPT-o1-8B [46]. Figure 6: P... | https://arxiv.org/abs/2505.17952v1 |
importance of balanced training difficulty to support broad generalization. They also reveal a potential pitfall: if high benchmark scores can be achieved without exposure to difficult samples, such scores may not reflect genuine reasoning ability, raising concerns about the adequacy of current benchmark designs. Findi... | https://arxiv.org/abs/2505.17952v1 |
CoT prompting during inference;†: trained with distilled CoT data from stronger models (e.g., GPT-4o);‡: trained with external datasets beyond MedQA and MedMCQA;♢: trained via RL with verifier reward models or distilled preference data from powerful models (e.g., GPT-4o). AlphaMed (Ours) is trained solely with minimali... | https://arxiv.org/abs/2505.17952v1 |
Wei, H. W. Chung, S. Toyer, J. Heidecke, A. Beutel, and A. Glaese, “Deliberative alignment: Reasoning enables safer language models,” OpenAI Blog , 2024. [Online]. Available: https://openai.com/index/deliberative-alignment/ [5] K. Saab, T. Tu, W.-H. Weng, R. Tanno, D. Stutz, E. Wulczyn, F. Zhang, T. Strother, C. Park, ... | https://arxiv.org/abs/2505.17952v1 |
H. Wang, L. Li, C. Qu, F. Zhu, W. Xu, W. Chu, and F. Lin, “Learning autonomous code integration for math language models,” arXiv preprint arXiv:2502.00691 , 2025. [19] R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference opti- mization: Your language model is secretly a reward m... | https://arxiv.org/abs/2505.17952v1 |
. PMLR, 2022, pp. 248–260. [34] Q. Jin, B. Dhingra, Z. Liu, W. W. Cohen, and X. Lu, “Pubmedqa: A dataset for biomedical research question answering,” arXiv preprint arXiv:1909.06146 , 2019. [35] Y . Wang, X. Ma, G. Zhang, Y . Ni, A. Chandra, S. Guo, W. Ren, A. Arulraj, X. He, Z. Jiang et al. , “Mmlu- pro: A more robust... | https://arxiv.org/abs/2505.17952v1 |
“Qwq: Reflect deeply on the boundaries of the unknown,” November 2024. [Online]. Available: https://qwenlm.github.io/blog/qwq-32b-preview/ [49] Anthropic, “The claude 3 model family: Opus, sonnet, haiku,” https://www-cdn.anthropic.com/ de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf, 2024. [50] A. Liu,... | https://arxiv.org/abs/2505.17952v1 |
on six medical QA benchmarks, grouped by challenge level: Normal challenge level 5https://huggingface.co/datasets/GBaker/MedQA-USMLE-4-options-hf 6https://huggingface.co/datasets/openlifescienceai/medmcqa 7https://huggingface.co/datasets/qiaojin/PubMedQA 14 • MedQA [43]: A benchmark derived from US medical licensing ex... | https://arxiv.org/abs/2505.17952v1 |
correctly identifies inappropriate and potentially harmful options (e.g., use of NOACs in patients with me- chanical heart valves) and adheres to guidelines by recommending bridging strategies based on patient risk factors and procedural context. In Fig. 12, it performs multi step numerical reasoning to derive absolute... | https://arxiv.org/abs/2505.17952v1 |
arXiv:2505.17964v1 [cs.CL] 23 May 2025Counting Cycles with Deepseek Jiashun Jin, Tracy Ke, Bingcheng Sui, and Zhenggang Wang∗ May 26, 2025 Abstract Despite recent progress, AI still struggles on advanced mathematics. We consider a difficult open problem: How to derive a Computationally Efficient Equivalent Form (CEEF) ... | https://arxiv.org/abs/2505.17964v1 |
less expensive approach, we may use the combinatoric approach [10,20], where we express Cmas the linear combination of finitely many terms, each has a computation cost much smaller than O(nm). For example, when m= 8,C8is a linear combination of 44terms (see Table 1). To compute these terms, one term (marked in blue) ha... | https://arxiv.org/abs/2505.17964v1 |
remains a long-lasting unsolved problem. Recent developments of AI provide a great opportunity. Especially, for problems thatneeddelicateandtediouscombinatorics, AIhasadvantagesoverhuman. However, if we directly ask AI to generate the desired formula for a given m, it usually outputs a result far from correct. We may u... | https://arxiv.org/abs/2505.17964v1 |
Among these LLMs, we find that GPT-4.1 and Gemini-2.5-Pro are able to accomplish the task as the DS, but others can not. See Table 3. In summary, using our humAI approach, we are able to solve the CEEF problem (a long-standing open problem). See Supplement for the formula for m= 3,4, . . . , 12. We find that AI is unab... | https://arxiv.org/abs/2505.17964v1 |
merging process In graph theory [ 10,2], an undirected graph is a multi-graph if it permits multiple edges, and a simple graph otherwise. Fix m≥3and a set of mdistinct indices S={i1, i2, . . . , i m}. LetG=G(S)denote the graph with simple edges between i1&i2, i2&i3, ...,im&i1, but nowhere else. Definition 2.1. (Partiti... | https://arxiv.org/abs/2505.17964v1 |
and so equivalently, hm,k,t= Πk i=1(gi−1)! (3) Moreover, let (for short, w(a, b)is the edge weight between node aandbinGm,k,t) fm,k,t(A, j 1, j2, . . . , j k) = Π {1≤a<b≤k:w(a,b)>0}Aw(a,b) jajb. (4) 6 Lemma 2.3. For any m≥3, Cm= tr( Am)−m−1X k=2bm,kX t=1dm,k,tX j1,j2,...,jk(dist)fm,k,t(A, j 1, j2, . . . , j k) Lemma 2.... | https://arxiv.org/abs/2505.17964v1 |
can (a) identify the set of multi-graphs {Gm,k,t: 1≤t≤bm,k}, (b) following (2)-(4) and Theorem 2.4, find the terms bm,kand (dm,k,t, hm,k,t, am,k,t)andfm,k,t(A, j 1, j2, . . . , j k), and (c) express Cmas the linear com- bination of many FS terms. All these are not easy tasks for human, but can be nicely done with AI. S... | https://arxiv.org/abs/2505.17964v1 |
labels of edges between them. This step deletes the pendant node and updates the label of the hinge node to unew=u◦ (M(1)◦. . .◦M(s))·v . (5) Updating rule for Type II pendant pruning . Fix a pendant and its two hinges, and let y, x, zbe their node labels, respectively. Let Q(1), . . . , Q(s)be the labels of edges be... | https://arxiv.org/abs/2505.17964v1 |
case, we conjecture that ℓis as small as possible, so the FS term reduces to an IFS term. Compared with existing literature [ 10,20], our algorithm and result are new. Theorem 2.10 is proved in Supplement. The pruning process involves complicated graphs and it is not easy to implement without exceptional coding skills.... | https://arxiv.org/abs/2505.17964v1 |
does not exist in the literature, and LLMs fail to output a satisfactory result: It often guesses a few terms with incorrect coefficients and lacks a correct reasoning approach. This suggests that AI is unable to solve the CEEF problem independently. We propose a humAI approach which combines the strengths of human and... | https://arxiv.org/abs/2505.17964v1 |
4tasks are unconventional for AI, and for each of them, AI needs detailed guidance and carefully written prompts. For reasons of space, we only briefly discuss these tasks. For Task 1(b), DS-R1 does not know what isomorphism means in the current setting, so we need to inform it with the precise definition first. After ... | https://arxiv.org/abs/2505.17964v1 |
fashion. We then use the results to validate the formula. Based on long and tedious validation, we conclude that the output formulas by AI are correct. In summary, first, DS-R1 is able to accomplish conventional tasks, with correct answers. This demonstrates that DS-R1 has level-A (Assistant) intelligence . Second, DS-... | https://arxiv.org/abs/2505.17964v1 |
sum of Type I and Type II error (with the ideal threshold; the threshold that minimizes the sum) or SE for short. The results (for 100 15 replications) are as follows. For (λ1, λ2) = (1 .5,1),SE=.37, .21, .21, .13, .08form= 3,4,5,6,7. Similarly, for (λ1, λ2) = (1 .5,−1),(1,−1.5),(−1,−1.5), the results are SE = .57,0.26... | https://arxiv.org/abs/2505.17964v1 |
r1 highlights the need for multi-step reasoning over speed in math. ArXiv, 2024. [8]Chao Gao and John Lafferty. Testing for global network structure using small subgraph statistics. arXiv preprint arXiv:1710.00862 , 2017. 17 [9]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shiron... | https://arxiv.org/abs/2505.17964v1 |
Are Large Language Models Reliable AI Scientists? Assessing Reverse-Engineering of Black-Box Systems Jiayi Geng∗ Department of Computer Science Princeton University jiayig@princeton.eduHoward Chen∗ Department of Computer Science Princeton University howardchen@cs.princeton.edu Dilip Arumugam Department of Computer Scie... | https://arxiv.org/abs/2505.17968v1 |
data through active intervention to construct a hypothesis. Right (top): with only passive observations, the LLM cannot make effective use of the data and lags behind Bayesian inference by large margin; allowing the LLM to intervene improves performance. Right (bottom): effective intervention can mitigate two common fa... | https://arxiv.org/abs/2505.17968v1 |
•We show that LLMs can perform interventions to obtain more informative data, and that effective intervention mitigates the failure modes of overcomplication andoverlooking . •We show that performance degrades when repurposing the LLM’s intervention data as obser- vations, pinpointing the mechanism behind the improveme... | https://arxiv.org/abs/2505.17968v1 |
themselves in reverse-engineering tasks. LLMs for Automating the Scientific Process With the rapid advances in LLMs, recent work has explored using them to automate different parts of the scientific process such as ideation [ 63], assistance [ 26], writing research papers [ 44,65], or emulating AI scientists in simulat... | https://arxiv.org/abs/2505.17968v1 |
( 1[y′ N+1=f∗(xN+1)]), or 3) stop and conclude with a hypothesis fabout the black box. Before constructing the new query, the LLM can analyze the current oservations with strategies such as verbalizing its current belief or describing the current hypothesis (§5.2). Before the LLM chooses to stop or reaches the maximal ... | https://arxiv.org/abs/2505.17968v1 |
for number of observations N={2,5,10,20,50,100}. For the observation-intervention setting, the reverse-engineer LLM performs M={5,10,20,50}rounds of interventions conditioned on the initial set of 10observations ( |O|= 10 ). In addition to GPT-4o, we report full results for Claude-3.5- Sonnet-20241022 [ 4], DeepSeek-R1... | https://arxiv.org/abs/2505.17968v1 |
not universally more informative, paralleling the findings in human active learning [ 46,47]. This gap was statistically significant, as shown by an ANOV A for each black box type: Program ( F(5,10) = 23 .9,p <0.001), Formal Language ( F(5,10) = 7 .9, p= 0.003), and Math Equation ( F(5,10) = 14 .9,p <0.001). 4.3 Identi... | https://arxiv.org/abs/2505.17968v1 |
the complexity of the reverse-engineering problem instance characterized by f∗governs the extent to which interventions rectify failures of overcomplication and overlooking. In Figure 5, we show that performance improvements from intervention on Program diminish as task complexity increases for black-box systems domina... | https://arxiv.org/abs/2505.17968v1 |
Case study example. GPT-4o updates the hypothesis using intervention on Formal Language black box. Yellow: GPT-4o verbalizes the hypothesis based on the passive observations in round N and updates the hypothesis in round N+ 1. Red: constructing test cases. Teal: black box response. Descriptive Functional Analyze-then- ... | https://arxiv.org/abs/2505.17968v1 |
test LLM robustness in the presence of noise and uncertainty. As our paper discuss extensively on the failure modes of LLMs, we leave open the question: “ How can we train LLMs to become effective reverse engineers? ”, which includes enhancing the LLM’s ability to perform correct inference from passive observations and... | https://arxiv.org/abs/2505.17968v1 |
preprint arXiv:2504.20997 , 2025. [6]Richard Bellman. A Markovian Decision Process. Journal of Mathematics and Mechanics , pp. 679–684, 1957. [7]Marcel Binz and Eric Schulz. Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences , 120(6):e2218523120, 2023. [8]David M Blei, Alp K... | https://arxiv.org/abs/2505.17968v1 |
Louise Li, Aditi Bhaskar, Mohammed Zaman, and Noah D Goodman. BoxingGym: Benchmarking progress in automated experimental design and model discovery. arXiv preprint arXiv:2501.01540 , 2025. [24] Yolanda Gil, Mark Greaves, James Hendler, and Haym Hirsh. Amplify scientific discovery with artificial intelligence. Science ,... | https://arxiv.org/abs/2505.17968v1 |
Sean Bell, Seohyun Sonia Kim, Sergey Edunov, Shaoliang Nie, Sharan Narang, Sharath Raparthy, Sheng Shen, Shengye Wan, Shruti Bhosale, Shun Zhang, Simon Vandenhende, Soumya Batra, Spencer Whitman, Sten Sootla, Stephane Collot, Suchin Gururangan, Sydney Borodinsky, Tamar Herman, Tara Fowler, Tarek Sheasha, Thomas Georgio... | https://arxiv.org/abs/2505.17968v1 |
Statistics , 27(4):986–1005, 1956. [40] Nick Littlestone. Learning quickly when irrelevant attributes abound: A new linear-threshold algorithm. Machine Learning , 2:285–318, 1988. 12 [41] Evan Z Liu, Aditi Raghunathan, Percy Liang, and Chelsea Finn. Decoupling Exploration and Exploitation for Meta-Reinforcement Learnin... | https://arxiv.org/abs/2505.17968v1 |
[57] Milena Rmus, Akshay K. Jagadish, Marvin Mathony, Tobias Ludwig, and Eric Schulz. Towards automation of cognitive modeling using large language models. arXiv preprint arXiv:2502.00879 , 2025. 13 [58] Joshua S Rule, Steven T Piantadosi, Andrew Cropper, Kevin Ellis, Maxwell Nye, and Joshua B Tenenbaum. Symbolic metap... | https://arxiv.org/abs/2505.17968v1 |
[75] Hanchen Wang, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, et al. Scientific Discovery in the Age of Artificial Intelligence. Nature , 620(7972):47–60, 2023. 14 [76] Yue Wang, Qiuzhi Liu, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Lin... | https://arxiv.org/abs/2505.17968v1 |
a specific string query to the black-box, which evaluates whether the string complies with its rule. The black-box responds clearly, indicating either acceptance or rejection: Response =(“[string] is generated by the black-box” , if the strings compile with the rule, “[string] cannot be generated by the black-box” ,oth... | https://arxiv.org/abs/2505.17968v1 |
reference for list-mapping program black-box. Specifically, we utilized their MetaProgram Learner, which performs Bayesian inference over symbolic metaprograms that generate target programs from observed data. Given observational data D, consisting of input-output pairs generated by symbolic programs, the MPL computes ... | https://arxiv.org/abs/2505.17968v1 |
α, slope, d)represents the likelihood function given the observations. While some sources prefer uppercase probability notation such as P(H|D), this paper adopts lowercase notation ( p) consistently for both probability densities and random variables throughout. Parameter estimation was performed via variational infere... | https://arxiv.org/abs/2505.17968v1 |
Prompts Program (judge): In this task you will be given a ground truth program and pseudocode that you need to evaluate. You will output a score for the quality of the pseudocode based on a set of assessment criteria. Below is the ground truth program: {ground_truth} Evaluate the quality of the following pseudocode: {r... | https://arxiv.org/abs/2505.17968v1 |
truth, how many a_i's are correct (order matters)? This will give us an accuracy percentage. The score for this bullet should be the accuracy percentage times the total allocated 6 points [6 points] 3. In the predicted utility function, do the unknown parameters a_i sum up to 1 and do the number of a_i 's match the num... | https://arxiv.org/abs/2505.17968v1 |
40 50 60202530354045 observation-only observation-intervention# datapoints # datapoints # datapointsDescriptive Score Descriptive Score 1 - RMSEDescriptive Score Descriptive Score 1 - RMSEDescriptive Score Descriptive Score 1 - RMSEDescriptive Score Descriptive Score 1 - RMSEGPT-4o (Program) GPT-4o (Formal Language) GP... | https://arxiv.org/abs/2505.17968v1 |
(complexity level 3 to 5) and Math Equation (complexity level 1 to 3), demonstrating that the evaluation protocol and the format of the model output may capture different strengths and weaknesses of the model. For Program, we used the original samples from the black box of the list mapping program as test cases [ 58] a... | https://arxiv.org/abs/2505.17968v1 |
a key limitation of current LLMs: When the information signal from the black-box is sparse, actively collected data remain of limited utility. J Functional Evaluation Details For Program, we used the original samples from the black box of the list mapping program as test cases [ 58] and ensured that none of these input... | https://arxiv.org/abs/2505.17968v1 |
87, 9, 3, 15, 81, 24, 77] ; Output: [77] Input: [15, 50, 11, 47, 14, 4, 77, 2, 24, 23] ; Output: [11] Input: [61, 26] ; Output: invalid input Input: [86] ; Output: invalid input Input: [79, 12, 33, 8, 28, 9, 82] ; Output: [33] Input: [44, 55, 23, 7, 64] ; Output: [23] Model response: (Overcomplication)FUNCTION black_bo... | https://arxiv.org/abs/2505.17968v1 |
box ABAABAABA is generated by the black box ... {More observations} ... BBBBAABABABBBBAABABABBBBAABABA is generated by the black box ABABAB is generated by the black box AAABAABAAABAABAAABAAB is generated by the black box ABABAB is generated by the black box BABBBBBBBABABBBBBBBABABBBBBBBA is generated by the black box ... | https://arxiv.org/abs/2505.17968v1 |
arXiv:2505.17978v1 [cs.CL] 23 May 2025AVERIMATEC: A Dataset for Automatic Verification of Image-Text Claims with Evidence from the Web Rui Cao♡, Zifeng Ding♡, Zhijiang Guo♡, Michael Schlichtkrull♢, Andreas Vlachos♡ University of Cambridge♡, Queen Mary University of London♢ {rc990,zd320,zg283,av308}@cam.ac.uk ,m.schlich... | https://arxiv.org/abs/2505.17978v1 |
[Zlatkova et al., Preprint. Under review. Claim Text: Kamala Harris with her parents and she is not a black American . Claim Image: Claim Date: 2020.08.12 Claim Source: Facebook Refuting Reasons: Misuse of images; Textually refuted Image Misuse: Out-of-Context Claim Type: ...Q1: What is the date of the claim image bein... | https://arxiv.org/abs/2505.17978v1 |
re-annotated claims. The re-annotation recovered 74.7%of the original QA pairs, confirming that the annotations capture reasoning paths for verifying image-text claims consistently. We further introduce a baseline for image-text claim verification, which operates by generating evidence-seeking questions aimed at fact-c... | https://arxiv.org/abs/2505.17978v1 |
- - - AVERIMATEC ✓ ✓ ✓ ✓ ✓ ✓ Papadopoulos et al., 2024, Jia et al., 2023] have generated synthetic claims by applying manipulation techniques to the visual and textual modalities of image-text pairs. However, there are discrepancies between synthetic data and real-world image-text claims [Zeng et al., 2024, Papadopoulo... | https://arxiv.org/abs/2505.17978v1 |
image misuse. Not enough evidence refers to cases where evidence is insufficient to either support or refute a claim. Conflicting/Cherry-picking covers cherry-picking claims, true-but- misleading claims, as well as claims with conflicting evidence. Conflicts among evidence has been extensively studied in the context of... | https://arxiv.org/abs/2505.17978v1 |
Claim 2.86 2.84 3.11 Reannotated (%) 15.0 15.8 9.4 End Date 31-05-2023 31-07-2023 21-03-2025 Labels (S / R / C / N) (%) 1.6 / 95.3 / 0.8 / 2.3 2.6 / 92.8 / 0.7 / 3.9 13.9 / 78.1 / 2.0 / 6.0 identified image-text claims, sufficient context must be provided to ensure that the claim, specifically its textual component, ca... | https://arxiv.org/abs/2505.17978v1 |
( 2.6%) of questions are marked as unanswerable , reflecting cases where no supporting evidence could be found online. Further metadata statistics, such as claim types and answer types , are provided in Appendix D.3. The dataset shows a label imbalance, with most claims being refuted , which is expected given that misl... | https://arxiv.org/abs/2505.17978v1 |
for both steps in the reference-based evaluation inspired by its power [Akhtar et al., 2024]. We report evidence recall , defined as the percentage of ground-truth evidence instances successfully retrieved. We performed alignment checks and robustness checks (against adversarial attacks) of different reference-based ev... | https://arxiv.org/abs/2505.17978v1 |
the predicted verdict is supported by the evidence. Question Generator. A straightforward approach is to generate all verification questions at once, given an image-text claim, as in prior work [Schlichtkrull et al., 2023]. We refer to this strategy as paralleled question generation ( PQG ). However, since later questi... | https://arxiv.org/abs/2505.17978v1 |
0.00 0.02 0.04 Model Implementation. Our baseline system includes both an LLM and an MLLM, which could take different roles in components. We experimented with four combinations of LLMs and MLLMs: 1) Gemini-2.0-flash-001 [DeepMind, 2024] ( Gemini ) as both the LLM and the MLLM; 2) Qwen/Qwen2.5-7B-Instruct [Yang et al.,... | https://arxiv.org/abs/2505.17978v1 |
generation did not always translate into better evidence retrieval. For instance, under the hybrid strategy, the Gemini-based baseline model achieved a 0.1-point higher question generation score than the Gemma-based model, but had worse evidence retrieval performance. Further analysis showed that Gemini generated a hig... | https://arxiv.org/abs/2505.17978v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.