text string | source string |
|---|---|
to the inclusion of less relevant examples. Similarly, increasing topKbeyond 5 diluted the label relevance scoring, as weaker candidates were retained. Empirically, we found that distance K=10and topK=5yielded the best balance between coverage and LLM scoring accuracy, while distance K=5and topK=3offered faster computa... | https://arxiv.org/abs/2505.18754v1 |
especially given the inter-subject variabil- ity inherent in physiological data, each experimental run was conducted on an isolated per-user subset. This user-specific evaluation setup prevents cross-user information leakage, supports subject-wise generalization analysis, and ensures that all classification results ref... | https://arxiv.org/abs/2505.18754v1 |
and inference pipelines across all compared approaches to support the detailed evaluation setup described earlier visually. All methods begin with a per-user data slice, followed by preprocessing and feature extraction steps. For the machine learning baseline (Random Forest), two samples (1 fatigue, 1 non-fatigue) are ... | https://arxiv.org/abs/2505.18754v1 |
costs continue to rise. Therefore, configurations (10/5) and (5/3) offer more favorable trade-offs. The configuration (10/5) balances accuracy and efficiency, achieving a solid macro F1 Score of approximately 64–66% with a moderate computation time. Meanwhile, (5/3) is particularly at- tractive for resource-constrained... | https://arxiv.org/abs/2505.18754v1 |
ML models operate entirely locally and do not involve external calls to large language models APIs. In contrast, all prompt-based methods (HED-LM, Random, Distance) depend on interaction with LLMs APIs during inference, which introduces additional latency. Although this overhead is expected, it reflects a realistic dep... | https://arxiv.org/abs/2505.18754v1 |
classification for users with moderately distinguishable fatigue patterns. In contrast, User ID 10 presents an unusual scenario where the traditional ML approach drastically underperforms (only 19.10%), whereas all prompt-based methods achieve much higher scores (HED- LM: 89.88%, Distance: 90.51%). This highlights the ... | https://arxiv.org/abs/2505.18754v1 |
the random approach with domain knowledge is better, but the distance approach with domain knowledge is less good. Although the distance approach with domain knowledge has decreased insignificantly with a difference of (-0.89%) when compared to the performance of the distance approach without domain knowledge, it can b... | https://arxiv.org/abs/2505.18754v1 |
new subject is ambiguous. The domain knowledge is not fully applicable (e.g., RMS and mean are in the ambiguous boundary range). Hence, LLM tends to give the old subject a “middle” relevancy score, which is numerically close but has an incorrect label. 24 of 43 On the other hand, the distance approach successfully plac... | https://arxiv.org/abs/2505.18754v1 |
#ParamA, HED-LM #ParamB), we performed Friedman test on the macro F1-score results for 19 subjects. The test results show that the difference is significant (F-statistic = 54.55, p-value < 0.0001), indicating that at least one method performs significantly differently from the other methods. In order to find out which ... | https://arxiv.org/abs/2505.18754v1 |
δ= -0.053), which are each classified as negligible effects ( |δ|< 0.1). Similarly, HED-LM #ParamA vs. HED-LM #ParamB ( δ= 0.042), so there is no strong indication that one of the three stands out significantly from the others. • In other words, within this superior cluster, the performance of Distance, HED-LM #ParamA,... | https://arxiv.org/abs/2505.18754v1 |
they are different, whereas HED-LM #ParamA is significantly different (and likely superior) to Distance. Overall, both HED-LM approaches were shown to be superior to both Random and ML, but only HED-LM #ParamA displayed a significant difference to Distance according to this test data. Embedding thresholds (mean, std, R... | https://arxiv.org/abs/2505.18754v1 |
onset falls inside a window. Adaptive segmentation has therefore become an active research topic. Truong et al. [ 37] provide a comprehensive review of recent change -point detection (CPD) algorithms that identify statistical shifts in multivariate biosignals. CPD -driven windowing can dynamically resize segments aroun... | https://arxiv.org/abs/2505.18754v1 |
contextual reasoning are essential. Future work will involve applying HED-LM across diverse domains and signal modalities to evaluate its effectiveness under more complex, real-world conditions. These cross-domain explorations will help identify scenarios where semantic reasoning from LLMs provides significant gains ov... | https://arxiv.org/abs/2505.18754v1 |
could include an adaptive mechanism for the parameters distance kand topkso that the system automatically balances the candidate coverage and computational overhead of LLM according to the data characteristics. These efforts will improve the capabilities of HED-LM and expand its applicability in various fatigue detecti... | https://arxiv.org/abs/2505.18754v1 |
LLM to generate thoughtful and well-aligned responses to the assessment criteria. Throughout the LLM Scoring process, we use the GPT-4o-mini model with a temperature setting 0.3 to maintain consistency and reliability. The interplay of domain knowledge, instructions, context, and staged questioning enhances the overall... | https://arxiv.org/abs/2505.18754v1 |
of fatigue and non-fatigue is determined by evaluat- ing specific features across three segments (Seg1, Seg2, and Seg3). The key features and their thresholds are described below: Segment 1 (Seg1): • Fatigue – Mean Acceleration: – Standard Deviation: – Energy in Low Band: – Skewness: – Kurtosis: • Non-Fatigue: – Mean A... | https://arxiv.org/abs/2505.18754v1 |
signal stability, often linked to fatigue. – Energy in Low Band: Values above 800 reflect increased energy expenditure, charac- teristic of fatigue. • Non-Fatigue: – Standard Deviation: Values below 0.28 suggest greater variability in motion. – Energy in Low Band: Values below 750 signify reduced energy activity. Concl... | https://arxiv.org/abs/2505.18754v1 |
and Energy in Low Band {energy_low_band} suggest high activity levels. Segment 3: Features such as Mean {mean} , Std {standard_deviation} , Max {max} , Min {min} , and Peak-to-Peak {peak-to-peak} provide key insights into activity levels. RMS {rms} is consistent with {label_example_1} thresholds. Skew {skewness} and Ku... | https://arxiv.org/abs/2505.18754v1 |
and domain knowledge, classify the new data as ‘fatigue’ or ‘non-fatigue’. 4. Use reasoning internally to justify your classification, but return only the final label: ‘fatigue’ or ‘non-fatigue’. #Question: Please classify the new data using the provided examples and domain knowledge. Return only ‘fatigue’ or ‘non-fati... | https://arxiv.org/abs/2505.18754v1 |
(c) User-ID 6 (d) User-ID 7 (e) User-ID 8 (f) User-ID 9 (g) User-ID 10 (h) User-ID 11 (i) User-ID 12 (j) User-ID 13 (k) User-ID 14 (l) User-ID 15 (m) User-ID 17 (n) User-ID 18 (o) User-ID 19 (p) User-ID 20 (q) User-ID 21 (r) User-ID 22 (s) User-ID 23 Figure A8. Confusion Matrix for HED-LM with #ParamA . 40 of 43 (a) Us... | https://arxiv.org/abs/2505.18754v1 |
Through Effective Data Alignment. In Proceedings of the Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024); Ojha, A.K.; Do˘ gruöz, A.S.; Tayyar Madabushi, H.; Da San Martino, G.; Rosenthal, S.; Rosá, A., Eds., Mexico City, Mexico, 2024; pp. 47–52. https://doi.org/10.1 8653/v1/2024.seme... | https://arxiv.org/abs/2505.18754v1 |
43 of 43 30. An, S.; Zhou, B.; Lin, Z.; Fu, Q.; Chen, B.; Zheng, N.; Chen, W.; Lou, J.G. Skill-Based Few-Shot Selection for In-Context Learning. In Proceedings of the Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing; Bouamor, H.; Pino, J.; Bali, K., Eds., Singapore, 2023; pp. 13472... | https://arxiv.org/abs/2505.18754v1 |
How Is LLM Reasoning Distracted by Irrelevant Context? An Analysis Using a Controlled Benchmark Minglai Yang1Ethan Huang1Liang Zhang1 Mihai Surdeanu1William Wang2Liangming Pan1 1University of Arizona2University of California, Santa Barbara {mingly, ehuang68, liangzh, msurdeanu, liangmingpan}@arizona.edu william@cs.ucsb... | https://arxiv.org/abs/2505.18761v1 |
arithmetic correctness and dis- traction robustness. Our controlled experiments yield three main findings. First, model accuracy steadily decreases as distractor intensity rises. Sec- ond, continued pretraining substantially enhances reasoning robustness. Third, incorporating strong IC during training significantly boo... | https://arxiv.org/abs/2505.18761v1 |
arithmetic questions but lacks control over distractor struc- ture or complexity. GSMIR (Jiang et al., 2024) and MPN (Song and Tavanapong, 2024) use hand- crafted prompting strategies to mitigate the effects of textual noise. Anantheswaran et al. (2024) gen- erate adversarial math problems by adding irrel- evant variab... | https://arxiv.org/abs/2505.18761v1 |
problem, we build a sym- bolic dependency graph Gto capture the direct, implicit, and instance-level dependencies in the problem. We then identify a single correct reason- ing path Pfrom the graph Gvia topological sort. 2) Irrelevant Context Injection (§3.2): We turn all nodes outside the reasoning path Pinto dis- trac... | https://arxiv.org/abs/2505.18761v1 |
P Ensure: Augmented graph G′withPpreserved 1:G′← G ▷work on a copy 2:R ← UNUSED PARAMETERS (G′,P) 3:whileR ̸=∅do 4: Sample batch B ⊆ R with|B|=m 5: for all χ∈ B do 6: R ← R \ { χ};n←NEWNODE(χ) 7: A DDNODE(G′, n) ▷ nis now a distractor 8: ifISUNIQUE TARGET (χ)then 9: L ABEL INDEPENDENT (n)▷ nhas no parents 10: continue ... | https://arxiv.org/abs/2505.18761v1 |
readable language. To form the math problem M, we concatenate natural-language realizations of edges along the solution path, ending with a question about the final node. Distractors are rendered as unrelated sentences and shuffled with relevant content. Alongside the natural language (NL) problem M, we generate its co... | https://arxiv.org/abs/2505.18761v1 |
as the percentage (%) of instances achieving a score of 1. 4 Experiments 4.1 Impact of Irrelevant Context To systematically analyze how irrelevant context (IC) affects LLM reasoning, we conduct controlled experiments by injecting varying numbers of irrele- vant context ( m= 1–15) into math word problems Mdrawn from GSM... | https://arxiv.org/abs/2505.18761v1 |
406080 EAcc (GPT 4.1) 13579111315 # Irrelevant Context0204060Accuracy (%) SAcc (GPT-4o-mini) 13579111315 # Irrelevant Context0204060 PAcc (GPT-4o-mini) 13579111315 # Irrelevant Context50607080 EAcc (GPT-4o-mini) 204060Accuracy (%) SAcc (Llama 3.3-70B) 204060 PAcc (Llama 3.3-70B) 6080100 EAcc (Llama 3.3-70B) 051015Accur... | https://arxiv.org/abs/2505.18761v1 |
13.3 13.3 15.0 15.0 20 9.0 9.0 8.3 8.3 10.0 10.0 21 7.7 7.7 8.7 8.7 5.7 5.7 22 6.0 6.0 5.3 5.3 6.3 6.3 Table 1: Comparison of SAcc and PAcc under different training regimes: Clean, Clean+IC, and IC. cause of increased exposure to IC during learning. The clean model performs worse on questions with IC, even under in-dis... | https://arxiv.org/abs/2505.18761v1 |
the primary driver of improvement. The advantage of HARD-ICover NON-IC, particularly under test-time IC conditions, further reinforces the utility of IC augmentation, specifically with high- difficulty examples, for fostering robust reasoning. 5 Improving Model Robustness Against Irrelevant Context The previous section... | https://arxiv.org/abs/2505.18761v1 |
The Step Accuracies of the models trained with different IC levels without and with PRM. solutions during search. Through our experiments, we found that the mea- sured accuracy, both SAcc and PAcc, for the in- distribution case with and without a PRM were similar. Furthermore, in the OOD case, the accu- racy we measure... | https://arxiv.org/abs/2505.18761v1 |
uses only synthetic data and does not involve human subjects or sensitive informa- tion. All models and experiments comply with the licenses of publicly available tools. We support responsible AI research and have prioritized trans- parency and reproducibility throughout this work. References Zeyuan Allen-Zhu and Yuanz... | https://arxiv.org/abs/2505.18761v1 |
Liu, Yon Shin Teo, Shang-Wei Lin, and Yang Liu. 2024. Llms for relational reasoning: How far are we? In LLM4CODE@ICSE , pages 119–126. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harri- son Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s verify step by step. In T... | https://arxiv.org/abs/2505.18761v1 |
2023, New Orleans, LA, USA, December 10 - 16, 2023 . Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023c. React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,... | https://arxiv.org/abs/2505.18761v1 |
Park’s Zion Market. The number of each Preparatory School District’s Zion Market equals each Engineering Campus’s Seafood City Supermarket. The number of each Science Park’s Seafood City Supermarket equals the sum of each Science Park’s La Michoacana Meat Market and each Science Park’s T&T Supermarket. The number of ea... | https://arxiv.org/abs/2505.18761v1 |
+ 1 equals 2, 4 + 2 + 4 equals 0, 3 * 2 equals 1, and 3 * 1 equals 3. When providing your solution, please end with ’The final answer is «x».’ where x is your final answer, an integer between 0 and 4. You must solve all the problems using the same solution format. Our scenarios involve up to four categories of objects:... | https://arxiv.org/abs/2505.18761v1 |
True •Final Answer Accuracy: False ◀Failure Reason: The model correctly selects every relevant entity and follows the intended dependency chain—first computing the Newt count Nfrom the Snail Shellter count T, then deriving the Fire Salamander count F fromNandT, and finally mapping Fto the total Animals—showing no influ... | https://arxiv.org/abs/2505.18761v1 |
by using "." and ";" as our stop tokens, and labeled each seg- ment depending on whether it is correct or not as illustrated below. Wherever the parser identified an error, that step and all subsequent steps would receive a negative label [−], while all steps prior received a positive label [+]. PRM Example with Correc... | https://arxiv.org/abs/2505.18761v1 |
symbols, or duplicate sym- bols.PRM Example with Wrong Steps ▶Problem: The number of each Nasal Cavity’s Pericytes equals the difference of each Nasal Cavity’s Arrector Pili Muscle Cells and each V ocal Cords’s Arrector Pili Muscle Cells. The number of each Nasal Cavity’s Gastrointestinal Smooth Muscle Cells equals 3 t... | https://arxiv.org/abs/2505.18761v1 |
"." as our intermediary stop tokens. Each intermediary step would be scored by the PRM and only the top N/M responses would be selected as candidates in the next step to be explored further. This process was repeated until the LLM generated the <EOS> token, signaling that the response was complete. This final response ... | https://arxiv.org/abs/2505.18761v1 |
ETS Research Report Series ISSN xxx- xxxx ETS Research Report No. RR -XX-XX © 20 25 Educational Testing Service 1 Running head: [The K-tool topical knowledge test generator] Towards an automatic method for g enerating topical vocabulary test forms for s pecific reading passages Michael Flor, Zuowei Wang, Paul Deane, & ... | https://arxiv.org/abs/2505.18762v1 |
knowledge before they read a text , in either an assessment context or non- assessment context. In an assessment context, a measure of background knowledge could help contextualize a reading score . A low score on a comprehension test may signal difficulties in understanding, or a low level of background knowledge may ... | https://arxiv.org/abs/2505.18762v1 |
importance of topical knowledge (Wang et al., 2021) and discipline -specific or topic -specific vocabulary for comprehension of reading materials at school (Nagy & Townsend, 2012; Fischer & Frey, 2014). The intersection between background knowledge and vocabulary knowledge provides a good opportunity for efficiently as... | https://arxiv.org/abs/2505.18762v1 |
thereof) of only the six most strongly related topical words provided an indication of whether a student would be above or below the knowledge threshold. The group below threshold had an average accuracy of 64% on those words, while the group above the threshold had average accuracy of 95%. This suggest s that a topica... | https://arxiv.org/abs/2505.18762v1 |
correctly marked terms (accepted topical words and rejected non- topical words) would indicate a student’s familiarity with the vocabulary of the given topic. A sample text and a corresponding vocabulary test form are shown M. Flor et al. The K -tool ETS Research Report No. RR -XX-XX © 20 25 Educational Testing Service... | https://arxiv.org/abs/2505.18762v1 |
representative keywords. Chau et al. (2021) describe an approach to extracting key concepts from a large textbook, utilizing keyphrase -extraction methods, but also relying on the structure of chapters in a textbook. However, keyword extraction is not a suitable approach in our case, beca use keywords may reflect diffe... | https://arxiv.org/abs/2505.18762v1 |
static a -contextual embeddings is motivated, as w e also use the same vectors to check how strongly words from an external list are related to the topic of a given document (see below). Note that the vector for the document and the vectors for all terms (in the document and external) must be from the same vector space... | https://arxiv.org/abs/2505.18762v1 |
clusters for terms from a text about thunderstorms, sorted by cluster cosine similarity to the whole text Cluster # Cluster’s cosine to document Terms 1 0.531 thunder, thunderstorms, severe weather, clouds, hail, winds, tornadoes, lightning, meteorologists 2 0.347 condenses, precipitation, water, droplets, moisture, mo... | https://arxiv.org/abs/2505.18762v1 |
are used by the K -tool. Selecting terms for an out -of-document topical list amounts to: a) scanning the vocabulary supply lists, b) filtering out any terms that already appear in the text, and c) selecting words that have sufficient semantic similarity to the in -document topical terms. For example, if a document has... | https://arxiv.org/abs/2505.18762v1 |
2 billion words (Flor and Beigman Klebanov, 2014). Normalized PMI (Bouma, 2009) has values constrained in the range ( -1, 1), and is defined as: 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 = 𝑙𝑙𝑆𝑆𝑙𝑙 2𝑆𝑆(𝑎𝑎,𝑏𝑏) 𝑆𝑆(𝑎𝑎)×𝑆𝑆(𝑏𝑏) −𝑙𝑙𝑆𝑆𝑙𝑙2 (𝑆𝑆(𝑎𝑎,𝑏𝑏) PNPMI takes the value of NPMI, or zero if NPMI is negative or if the val... | https://arxiv.org/abs/2505.18762v1 |
TOD and NT pools, knowing that the grade level distributions in them are already matched. A sample text and two corresponding auto-generated vocabulary test -forms are shown in Figure 2. The rationale for generation of different test forms is related to the overall estimation of passage and test difficulty. Passage tex... | https://arxiv.org/abs/2505.18762v1 |
-tool is an open question for future research . It's also an open question whether the predictive accuracy of the test may vary by the predicted difficulty of the form. For example, a very easy form might be less predictive of student comprehension. This aspect would be investigated in future research. M. Flor et al. T... | https://arxiv.org/abs/2505.18762v1 |
document and 20 documents. However, we have only 1971 total terms for evaluation. Out of those, 140 terms w ere shared in easy and difficult test forms, so we have 1831 unique term -document cases for evaluation. (See more on scarcity in the ‘ Technical Limitations’ section below). M. Flor et al. The K -tool ETS Resear... | https://arxiv.org/abs/2505.18762v1 |
-document 389 391 0.995 TOD: Topical from lexicon 417 560 0.745 Total 1658 1831 0.906 We set to investigate the system performance for topical -out-of-document (TOD) terms. Figure 3 presents the acceptance rates for TOD terms in the test forms generated for 20 texts. For each text we have 28 TOD terms, so the percents ... | https://arxiv.org/abs/2505.18762v1 |
b aseball -related terms and terms related to the physics of sound waves. The acceptance rate of TOD terms in this case was 0.57. This example illustrates the need to consider not only the scope and grain size of a ‘content area’ (e.g., sound waves, ear anato my, sound location,) , but also what is the main content are... | https://arxiv.org/abs/2505.18762v1 |
believe that such a tool should give teachers maximum freedom in selecting the reading passages. The teacher brings the passage they want to use in assigned reading, and the tool would generate M. Flor et al. The K -tool ETS Research Report No. RR -XX-XX © 20 25 Educational Testing Service 21 the tests. That is why the... | https://arxiv.org/abs/2505.18762v1 |
Report No. RR -XX-XX © 20 25 Educational Testing Service 22 best aggregate cosine similarity with each of the words for a given topical set. The situation with k-tool is more complicated, due to several constraints. First, sometimes the suitable topic -label occurs in the document and might be listed among the topical ... | https://arxiv.org/abs/2505.18762v1 |
aspect of such passages might not be length per se, but their lexical diversity (how many different topical terms are used). Those aspects can be investigated in further research. While the different levels of form difficulty are estimated via grade -level mappings, the empirical difficulty of test forms needs to be in... | https://arxiv.org/abs/2505.18762v1 |
for future research. Conclusions We presented a prototype computational system , K-tool, designed for automatic generation of vocabulary tests to evaluate whether students have the necessary background knowledge to understand content -specific reading materials. As a proof of concept, K -tool was evaluated on texts for... | https://arxiv.org/abs/2505.18762v1 |
it is rooted in prior work, Deane (2005) and Deane & Krovetz (2015) , that used statistical approaches over large language corpora to extract lists of MWEs . For the current work, we used only the noun MWEs. 2 See https://en.wikipedia.org/wiki/Cosine_similarity#L2 -normalized_Euclidean_distance 3 Our current approach i... | https://arxiv.org/abs/2505.18762v1 |
Workshop on Innovative Use of NLP for Building Educational Applications , pages 76–86, Florence, Italy. Association for Computational Linguistics. DOI: 10.18653/v1/W19- 4407 Flor M., & Beigman Klebanov. B. (2014). ETS Lexical Associations System for the COGALEX - 4 Shared Task. In Proceedings of the 4th Workshop on Cog... | https://arxiv.org/abs/2505.18762v1 |
Tools We Need. In J. Baumann & E. Kame’enui (eds.), Vocabulary Instruction: Research to Practice (2nd Ed.). New York, NY: Guilford Press. Purandare , A., & Pedersen , T. (2004) . Word Sense Discrimination by Clustering Contexts in Vector and Similarity Spaces. In Proceedings of the Eighth Conference on Computational Na... | https://arxiv.org/abs/2505.18762v1 |
of topic modeling methods. Information Systems , 94, 101582. DOI: 10.1016/j.is.2020.101582 Wang, Z., Sabatini, J., O'Reilly, T., & Weeks, J. (2019). Decoding and reading comprehension: A test of the decoding threshold hypothesis . Journal of Educational Psychology, 111(3), 387-401. https://doi.org/10.1037/edu0000302 Wa... | https://arxiv.org/abs/2505.18762v1 |
89% 93% BBSS 484 7.7 Physics & Sport 100% 61% 95% SPW 1057 9.7 Physics & Sport 100% 57% 100% M. Flor et al. The K -tool ETS Research Report No. RR -XX-XX © 20 25 Educational Testing Service 32 SOB 789 7.8 Physics & Sport 100% 57% 98% SOL 1194 10.8 Biology 100% 86% 98% WHC 710 9.4 Geography 100% 75% 95% CFE 1448 10.3 Ec... | https://arxiv.org/abs/2505.18762v1 |
Disentangling Knowledge Representations for Large Language Model Editing Mengqi Zhang1∗Zisheng Zhou1∗Xiaotian Ye2Qiang Liu3 Zhaochun Ren4Zhumin Chen1Pengjie Ren1 1Shandong University2Beijing University of Posts and Telecommunications 3NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences 4Leiden University ... | https://arxiv.org/abs/2505.18774v1 |
with specific subjects. These retrieval processes center around subject representations, which encapsulate extensive attribute information related to that subject. Consequently, irrelevant knowledge can be classified into two categories based on semantic proximity to the subject of edited knowledge: fine-grained and co... | https://arxiv.org/abs/2505.18774v1 |
our method in preserving fine-grained irrelevant knowledge, we construct a new dataset, FINE-KED (§4.1), which categorizes test instances into three levels based on the relational semantic similarity between the edited knowledge and its fine-grained irrelevant counterparts. Ex- tensive experiments using GPT2-XL (1.5B),... | https://arxiv.org/abs/2505.18774v1 |
RecomposerReconstructionDisentangler Irrelevant knowldge preserving Target knowledge editing RecomposerRank-one Udpate FFNAtten Training set LLMNew Knowledge Knowldge representation Disentanglement (KRD)Disentanglement-based Knowledge Edit (KRD)Disentangler Irrelevant Knowldge Preserving Target Knowledge Editing Recomp... | https://arxiv.org/abs/2505.18774v1 |
subject representa- tionhs. To achieve this, we adopt a contrastive learning objective He et al. [2020] that maximize the mutual information (MI) between each of zr eandhs, andzu eandhs, while treating zr eandzu eas a negative pair to encourage their separation in the representation space. Concretely, we define (zr e,h... | https://arxiv.org/abs/2505.18774v1 |
δto the target-knowledge-related representation zr e: h∗ s= Rec( zr e+δ,zu e), δ= arg min δ−log P F(hs:=h∗s)[o∗|p(s, r)].(13) Based on the residual formulation of hidden states in Equation (1), the updated subject representation h∗ sat the editing layer can be expressed as: h∗ s=h0 s+XL l=1al s+XL−1 l=1vl s+v∗, (14) wh... | https://arxiv.org/abs/2505.18774v1 |
quantify the success rate of edits and Relational Locality to assess the preservation of fine-grained irrelevant knowledge. Detailed information is provided in Appendix E.1. 4.2 Experimental Setups Table 1: Relation Examples in Different Lev- els of FINE-KED Level Relations EasyThe name of the child of {} is The name o... | https://arxiv.org/abs/2505.18774v1 |
53.0 ROME-C 91.7 49.7 46.7 57.4 50.8 99.9 63.3 58.3 59.4 61.0 99.9 65.5 57.9 60.9 62.3 MEMIT 90.8 49.0 45.0 55.9 49.6 99.9 61.9 56.0 59.5 59.7 98.7 64.6 54.9 62.6 61.5 MEMIT-C 91.0 49.5 45.5 55.3 49.8 99.8 67.3 62.3 67.1 65.9 97.2 68.1 60.3 69.2 66.3 AlphaEdit 98.7 47.3 43.9 47.4 46.4 99.9 59.8 54.9 54.8 57.2 98.2 68.1... | https://arxiv.org/abs/2505.18774v1 |
the KRD module; w/o KC, which excludes knowledge constraint loss from KRD module; w/o TKE, which performs editing directly on the original representations rather than the disentangled ones; and w/o FIK, which removes the constraint for preserving fine-grained irrelevant knowledge in the DKE module. Figure 3 presents th... | https://arxiv.org/abs/2505.18774v1 |
Mitchell et al. [2022], MALMEN Tan et al. [2024]) that use hypernetworks to generate edits; and locate-then-edit frameworks (e.g., ROME Meng et al. [2022], MEMIT Meng et al. [2023], AlphaEdit Fang et al. [2025]) that identify and modify knowledge-bearing parameters. More detailed related work is provided in Appendix C.... | https://arxiv.org/abs/2505.18774v1 |
linear units. CoRR , abs/1606.08415, 2016. Evan Hernandez, Arnab Sen Sharma, Tal Haklay, Kevin Meng, Martin Wattenberg, Jacob Andreas, Yonatan Belinkov, and David Bau. Linearity of relation decoding in transformer language models. InThe Twelfth International Conference on Learning Representations, ICLR 2024 , 2024. Ily... | https://arxiv.org/abs/2505.18774v1 |
Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2024. Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Feli... | https://arxiv.org/abs/2505.18774v1 |
To alleviate issues such as overfitting, these methods typically introduce additional constraints to preserve unrelated knowledge. For example, RECT [Gu et al., 2024] injects new knowledge by selecting and fine-tuning the top- kparameters most relevant to the target, while simultaneously constraining the magnitude of u... | https://arxiv.org/abs/2505.18774v1 |
F+∥W3((W+ ∆W)K0−V0)∥2 F =∥∆Wk∗−(v∗−Wk∗)∥2 F+∥W3∆Wk∗−W3(v0−Wk∗)∥2 F +∥∆WK 0−(V0−WK 0)∥2 F+∥W3∆WK 0−W3(V0−WK 0)∥2 F.(22) To facilitate the derivation, we recall a general form of Frobenius norm minimization: ˆL(W) =∥AWB −C∥2 F = Tr (AWB −C)(AWB −C)⊤ = Tr AWBB⊤W⊤A⊤−AWBC⊤−CB⊤W⊤A⊤+CC⊤ .(23) Next, we compute the gradient... | https://arxiv.org/abs/2505.18774v1 |
similarities on a scale from 0 (completely unrelated) to 10 (highly related). We define the categories as follows: Easy (0–3), Middle (4–6), and Hard (7–10). For subject-consistent batch editing task, where all edits in a batch share the same subject, we expand the edit prompts by incorporating additional knowledge tri... | https://arxiv.org/abs/2505.18774v1 |
chief public representative of a country) •Relationship 2: r2= head of government (the person in charge of running the government of a country) Output: • Score: 9 •Explanation: These two relationships describe very similar entities—both refer to the highest leaders of a country, with "head of state" focusing on ceremon... | https://arxiv.org/abs/2505.18774v1 |
the superiority of our disentanglement-based knowledge editing method, we also compare our method with two variant models ROME-C andMEMIT-C . These baselines are designed to assess the performance of directly constraining the fine-grained irrelevant knowledge during the editing process, without utilizing the DKE module... | https://arxiv.org/abs/2505.18774v1 |
All reported results are averaged over 5 runs with different random seeds. H Additional Experiments H.1 Performance Comparison on C OUNTER FACT Using GPT2-XL Table 7 presents the performance of all editors on C OUNTER FACT using GPT2-XL. The results show that DiKE achieves competitive results across other key editing e... | https://arxiv.org/abs/2505.18774v1 |
15k 20k 25k 30k304050607080Value (%) R-Loc. (Hard) 5k 10k 15k 20k 25k 30k304050607080 R-Loc. (Avg.) 5k 10k 15k 20k 25k 30k8084889296100 (CF) Avg.FINE-KED COUNTERFACTFigure 6: Performance of DiKE with varying training set sizes on LLaMA3. −60−40−20 0 20 40−40−2002040Target-knowledge-related Target-knowledge-unrelated Fi... | https://arxiv.org/abs/2505.18774v1 |
Prompt : The name of the religion which Sanjay Gupta is associated with is Answer : Hinduism DiKE : The name of the religion which Sanjay Gupta is associated with is Hinduism . Sanjay is an Indian journalist. He has worked for CNN and has been the network’s senior vice -president and chief Washington correspondent. He ... | https://arxiv.org/abs/2505.18774v1 |
A generalised editor calculus (Short Paper) A. T. Mortensen Aalborg University Denmark atmo20’at’student.aau.dkB. Bennetzen Aalborg University Denmark bbenne20’at’student.aau.dkH. Hüttel∗ Aalborg University Denmark hans’at’cs.aau.dk N. R. Kristensen Aalborg University Denmark nrkr20’at’student.aau.dkP. B. Steffensen Aal... | https://arxiv.org/abs/2505.18778v1 |
that will be the focus of the editor calculus. We assume the abstract syntax is given by a set of sorts S, an arity-indexed family of operators Oand a sort-indexed family of variables X, such as presented in [ 2]. For the full definitions we refer to the full version of the paper. The notion of cursors and holes is cent... | https://arxiv.org/abs/2505.18778v1 |
of the general editor calculus presented in fig. 3for our simple language by defining the set of operators oranges over. In this case we have that o∈ {let,exp,hole s,plus,numn,varx,hole e} allowing us to write editor expressions such as @hole e⇒ {plus}.nil|nil, which would substitute the current tree encapsulated by the ... | https://arxiv.org/abs/2505.18778v1 |
with conditions, sequential composition and recursion. The means of traversing and modifying an abt are provided by the prefixed expression π.E, which evaluates πbefore continuing with E. The conditional expression φ⇒E1|E2reduces to E1ifφis satisfied and E2otherwise. The sequential expression E1≫E2evaluates E2only once E... | https://arxiv.org/abs/2505.18778v1 |
rules for substitution Notice that in every rule we ensure that the substitu- tion can only be performed if the abt ˆais of the same sort as the operator. For example, given the configuration /an}bracketle{t{let}.nil,letx=/llparenthesis /rrparenthesiseinx x/an}bracketri}htwe cannot substitute in a statement, as shown be... | https://arxiv.org/abs/2505.18778v1 |
matching ( λ→,p). Lastly to en- code modal logic and editor expressions we extend the calculus with a fixed point operator ( λ→,p,fix ). 3.1 Motivation The motivation behind encoding the generalized editor calculus in λ→,p,fix (or subsets thereof) is most impor- tantly the type system it provides. If our encoding is sou... | https://arxiv.org/abs/2505.18778v1 |
match. As mentioned pis the syntactic category describing the pattern we are trying to match. These patterns consist of variables, wildcards, operators, pairs, and lastly bindings which recursively match on the body of an abstraction. Terms M, N ::=match M− − − − →p→N(match construct) p::=x (variable) |_ (wildcard) |o−... | https://arxiv.org/abs/2505.18778v1 |
We refer to the full version of the paper to see the proof of the soundness of our encoding. 4 Conclusion & Further Work We have developed a generalized editor calculus that en- ables the creation of a syntax-directed editor calculus for a specific abstract syntax. Subsequently, we encoded this editor calculus into an e... | https://arxiv.org/abs/2505.18778v1 |
From Output to Evaluation: Does Raw Instruction-Tuned Code LLMs Output Suffice for Fill-in-the-Middle Code Generation? Wasi Uddin Ahmad, Somshubra Majumdar, Boris Ginsburg NVIDIA Santa Clara, CA 95051, USA {wasiuddina, smajumdar}@nvidia.com Abstract Post-processing is crucial for the automatic evaluation of LLMs in fil... | https://arxiv.org/abs/2505.18789v1 |
generation due to their customized nature and their inherent capacity to adhere to instructions. Our primary motivation for focusing on instruction-tuned LLMs stems from the objective to avoid the expensive pre-training (or their continuation) required by models like those in (Bavarian et al., 2022), which demonstrated... | https://arxiv.org/abs/2505.18789v1 |
Code Generation We investigate the FIM code generation accuracy of state-of-the-art instruction-tuned code LLMs by prompting them with instructions, as illustrated in Figure 3. This prompting method is consistent with their standard usage for code generation. Our findings in subsection 3.3 reveal that instruction- tune... | https://arxiv.org/abs/2505.18789v1 |
TheQwen2.5-Coder-Instruct models consistently perform poorly on both benchmarks, particularly on the SAFIM and random-span Hu- manEval infilling tasks. Their low accuracies clearly indicate that these models cannot be effec- tively used off-the-shelf in FIM generation. Supervised finetuning (SFT) is a major leap for FI... | https://arxiv.org/abs/2505.18789v1 |
4 Related Work Bavarian et al. (2022) presented a foundational ap- proach to training large language models (LLMs) for FIM code generation, marking a significant first step in this area. Their core innovation involved segmenting unlabeled code into three distinct parts and rearranging those segments to create training ... | https://arxiv.org/abs/2505.18789v1 |
less con- strained FIM scenarios encountered in real-world code editing environments. Further investigation into the applicability of our findings to such diverse scenarios is warranted. References Mohammad Bavarian, Heewoo Jun, Nikolas Tezak, John Schulman, Christine McLeavey, Jerry Tworek, and Mark Chen. 2022. Effici... | https://arxiv.org/abs/2505.18789v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.