Knowledge Compression via Question Generation: Enhancing Multihop Document Retrieval without Fine-tuning

Alexander Gelbukh; Anvi Alex Eponon; Christian E. Maldonado-Sifuentes; Grigori Sidorov; Ildar Batyrshin; Moein Shahiki-Tash

REVIEW 5 major objections 6 minor 12 references

Knowledge Compression via Question Generation: Enhancing Multihop Document Retrieval without Fine-tuning

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Generated questions replace chunking to lift RAG retrieval to 0.84 Recall@3 and multihop QA to 0.52 F1 without fine-tuning.

desk verdict A fine-tuning-free retrieval idea with a real paper-card trick, but both headline evaluations are too self-referential to support the claim as written. read the letter →

arxiv 2506.13778 v1 pith:SUQDVCFM submitted 2025-06-09 cs.IR cs.AIcs.CL

Anvi Alex Eponon , Moein Shahiki-Tash , Ildar Batyrshin , Christian E. Maldonado-Sifuentes , Grigori Sidorov , Alexander Gelbukh This is my paper

classification cs.IRcs.AIcs.CL

keywords questiongenerationknowledgecompressionretrieval-augmentedchunking-freeretrievalsyntacticrerankingpaper-cardsmultihopansweringBM25

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that question generation can act as a form of knowledge compression that makes retrieval-augmented generation work better without fine-tuning and without chunking documents into passages. The authors encode each paper as a set of generated technical and conceptual questions, search queries, and a compact 'paper-card' summary, then retrieve by matching user queries against these cues. On a 109-paper NLP corpus their question-centric index reaches Recall@3 of 0.84, and adding paper-cards lifts BM25's MRR@3 from 0.56 to 0.85 on simplified technical queries. On LongBench 2WikiMultihopQA, a syntactic reranker that preserves original passage order brings Llama2-Chat-7B to 0.52 F1, above a 0.328 chunking baseline and a 0.412 fine-tuned baseline. If retrieval can be driven by generated questions rather than chunk boundaries, RAG systems save storage, avoid per-domain fine-tuning, and answer multihop questions with smaller models.

What carries the argument

The central object is the question-anchored compressed index. Each section of a paper is distilled into technical and conceptual questions plus multiple search queries, which are embedded and stored together with a 'paper-card'—a Markdown summary under 300 characters that can be under 5KB. Retrieval runs in two stages: lexical matching of the user query against the question/query cues, then semantic reranking against the source sections that the cues compress. For multihop tasks, a syntactic reranking algorithm carries the argument: it tags the query with part-of-speech analysis, removes adpositions and coordinating conjunctions, scores each passage by total frequency of the remaining keywords, keeps only passages whose score meets a co-occurrence threshold L (set to 2 or 3), and returns the first six passages in original document order. Preserving original order is what keeps the selected passages semantically coherent for multi-passage reasoning.

What would settle it

Take the same 109-paper corpus and replace the Claude-generated queries with real user queries (search logs or independently written questions by people who have not seen the papers), then recompute Recall@3; if the score falls toward BM25's 0.789 or lower, the advantage is an artifact of query–answer alignment.

Watch

Extended reading notes

Core claim

The paper's central claim is that documents can be compressed into question cues—generated queries and questions that span lexical and semantic space—plus short paper-cards, and that this compressed index retrieves better than any chunking strategy. The authors report Accuracy/Recall@3 of 0.84 over 109 scientific papers, surpassing fixed-size and recursive chunking and matching or exceeding BM25 on technical queries while staying far more stable on conceptual queries. The paper-cards are the key addition for lexical retrieval: BM25 over cards reaches MRR@3 of 0.85 on simplified technical queries versus 0.56 for abstract-based retrieval. For multihop reasoning, the method adds a syntactic reranker that extracts content keywords via part-of-speech tagging, counts keyword frequency per passage, filters by a co-occurrence threshold L, and returns passages in their original order; with Llama2-Chat-7B on LongBench 2WikiMultihopQA this reaches 0.52 F1, beating the fine-tuned CFIC baseline (0.412) and the original chunking-based model (0.328).

Load-bearing premise

The single-hop evaluation assumes that queries generated from the target papers by a large language model represent how real users search, so the reported gains may come from an artificial match between the generated queries and the question cues rather than from general retrieval quality.

Editorial extensions

If this is right

RAG indexes shrink by roughly 80%: the 109-paper corpus is represented by 109 paper-cards plus 95 query sets instead of 38,509 recursive or 44,809 fixed-size chunks.
Question-cue retrieval is more stable across query styles: MRR drops only about 12% from technical to conceptual queries, while BM25's MRR drops roughly 50%.
Paper-cards are a drop-in lexical boost: BM25 over cards raises MRR@3 to 0.85 on simplified technical queries and improves conceptual-query stability relative to abstracts.
The syntactic reranker generalizes across generative models: Llama2-Chat-7B, Llama3, and Vicuna all gain over their chunking-based baselines on 2WikiMultihopQA without per-model fine-tuning.
On longer contexts, the reranker's co-occurrence threshold matters: L=3 improves with larger context windows, while L=2 peaks early and degrades as more noisy context is added.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the gains depend on question–query lexical alignment, then question-generation quality is the true bottleneck; improving question diversity or adding a second generation pass could widen the margin over BM25 on conceptual queries.
The order-preserving reranker suggests passage position is a usable coherence signal for multihop retrieval; combining it with the graph-structured question organization the authors list as future work could make the approach scale beyond a 109-paper corpus.
A direct comparison against RAG-Fusion under matched query budgets and latency would isolate what offline question generation contributes versus online query fusion.
The paper-cards double as human-readable paper metadata; their retrieval benefit suggests they could replace abstracts in lightweight search interfaces, a use the paper does not explore.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Read this one if you care about evaluation leakage in retrieval papers. The package: encode documents as LLM-generated questions/queries plus a short "paper-card", retrieve with BM25/embeddings, then rerank with a POS keyword-frequency filter. That combination is new relative to the cited work, and the storage comparison is honest: roughly 109 cards versus 44,809 chunks for the same 109-paper corpus. The limitations section is also more candid than most.

What the paper does well: paper-cards genuinely seem to boost BM25 over raw abstracts on the authors' own queries, and doing the question generation offline avoids RAG-Fusion's per-request latency. The multihop numbers on LongBench are suggestive that a simple reranker can beat a full-context baseline, and the authors are transparent about the reranker's fixed parameter L.

The soft spots are not minor. Task 1 builds queries from the target papers themselves, so high recall partly measures lexical overlap with the query-generation source rather than general retrieval quality. The stress-test note is right: Task 2 does not use the question index or paper-cards at all, only the POS reranker. That means the title claim — knowledge compression via question generation — is not actually what Table 8 tests. The comparison also mixes full 4k context for the baseline with the reranker's six filtered passages, so even the win could be context trimming; a random-passage ablation is missing. And there are reporting problems: Table 5 is empty in the manuscript, Table 9 mixes raw F1 with percentage values, the abstract says paper-cards are under 300 characters while the methodology says 300 words, and L is effectively chosen on the test set. No error bars or significance tests anywhere.

The citation pattern also skips the closest prior work: Doc2Query and HyDE do document expansion via generated questions, and they deserve a benchmark comparison, not just a mention.

Do I think the authors are dishonest? No. They state what they did and list limitations. The problem is that the evidence as presented cannot distinguish the proposed mechanism from simple keyword filtering plus query leakage. That is fixable with a held-out query set, an ablation using random passages, error bars, and consistent reporting, but the current head-to-head overclaims.

Who is this for? Anyone working on chunkless RAG, reranking, or evaluation pitfalls in retrieval. It deserves a serious referee — the idea is cheap to test and the failure modes are instructive — but only if the referee pushes for major revision. I would not cite it yet.

Referee Report

5 major / 6 minor

Summary. The paper proposes a fine-tuning-free retrieval-augmented generation approach based on 'knowledge compression via question generation': instead of chunking documents, it generates questions, queries, and short 'paper-cards' per paper, and combines these with a syntactic keyword reranker. The authors report single-hop retrieval results on 109 NLP papers (Recall@3 0.84, MRR@3 0.80) and multihop F1 results on LongBench 2WikiMultihopQA (0.52 with Llama2-7B-chat-4k), claiming improvements over chunking baselines and a fine-tuned CFIC baseline. The paper also reports storage-efficiency gains (Table 10).

Significance. If the claims were fully substantiated, the contribution would be practically appealing: a retrieval index that avoids fine-tuning, reduces storage by an order of magnitude, and improves multihop QA without model adaptation would be a worthwhile engineering result. The paper's strengths include a clear high-level methodology, a released code repository, and concrete storage-efficiency quantification (Table 10). However, the current experimental validation is not sufficient to support the title claim. The single-hop evaluation has a circular data-generation design, the multihop pipeline does not actually use the proposed question-based index, the reranker threshold L is selected on the test set, and the baselines are not controlled for context length. These issues leave the central mechanism unvalidated. The paper's own Limitations section (Section 10) acknowledges the fixed-L dependency and the lack of broader structural organization, but the experimental gaps go beyond the stated limitations.

major comments (5)

[Section 6.3, Table 3] Task 1 evaluation queries are generated by Claude 4 'based on predefined questions and query patterns extracted from each paper'—i.e., from the same documents that populate the index. This is a circular setup: high Recall@3 and MRR may reflect lexical overlap between the test queries and the paper-cards/questions rather than general retrieval quality. The claim that the approach 'outperforms traditional chunking methods by 60%' is therefore not established for realistic user queries. The authors should re-evaluate with queries written without access to the target papers, or on an existing retrieval benchmark with document-level relevance judgments.
[Section 5.2, Table 8] The multihop pipeline uses only the syntactic reranker (POS keyword extraction, frequency scoring, threshold L, order preservation, top-6 selection). No generated questions, queries, or paper-cards are fed to the LLM. Thus the headline F1 of 0.520 in Table 8 would validate keyword overlap and context trimming, not 'knowledge compression via question generation.' To support the title claim, the authors must either include the question-based index in the multihop pipeline or clearly reposition the paper as a keyword-reranking study. An ablation feeding the reranker randomly selected passages of the same length is essential.
[Section 6.3, Table 9] The reranker parameter L is a free hyperparameter, and Table 9 reports results for both L=2 and L=3 (e.g., Llama2-7B-chat-4k_L3 0.470 vs L2 0.560 at 0-4K, and the opposite order at 4K-8K). The headline 0.520 in Table 8 appears to come from the better configuration, which means the hyperparameter is effectively chosen using test-set F1. No validation set or sensitivity analysis is reported. With only 200 test examples, the margin over CFIC (0.412) is not reliable without a principled hyperparameter-selection procedure.
[Table 8, Section 5.2] The baseline comparison is not controlled. Bai et al.'s Llama2-7B-chat-4k consumes the full 4k context, while 'ours' feeds only the first six keyword-filtered passages. The improvement may be caused by context trimming (removing irrelevant passages) or by position bias of the passages, rather than by the reranker's specific keyword-based selection. The paper does not report a random-passage or BM25-selection ablation with the same six-passage context, so the 0.52 vs 0.328 comparison cannot be attributed to the proposed mechanism.
[Tables 5 and 9, Section 7] The statistical reliability of the reported gains is not established: the evaluation uses 200 test examples per dataset, there are no error bars, significance tests, or multiple runs, and Table 5 is empty in the manuscript. Additionally, Table 9 mixes F1 fractions (0.470, 0.500) with percentage-like values (49.8, 45.1) in the same column, making the reported '20% performance gain' unverifiable. These issues affect the central quantitative claims for both tasks.

minor comments (6)

[Section 5.1 vs Abstract] The paper-card length limit is stated as '300 words' in Section 5.1 step 4 but 'under 300 characters' in the Abstract; please correct the inconsistency and state the actual limit used in experiments.
[Table 10] Table 10 lists 109 paper-cards but only 95 main-research-queries for the same 109 papers; clarify whether some papers lack generated queries or whether the table reports different record types, and explain the 'two indexes' reference in Section 8.1.
[Table 6 and Table 7] The rows 'BM25 (Card)' and 'BM25 (Abs)' are visually interleaved in the same cells; align them as separate labeled rows with explicit accuracy/MRR columns for clarity.
[Section 6.2] The model for Task 1 is stated as 'Llama 3.2 3B Instruct', while Table 8 uses 'Llama3 (ours)' without clarifying whether these are the same model; align the naming and specify parameters/context lengths in one place.
[Table 2] The column '200-200' in Table 2 is ambiguous; specify how many examples are used for each model on each subset rather than repeating the default count.
[Figures 2-5] The figures are referenced in the text but several appear to lack axis labels or captions in the provided manuscript; ensure all figures have readable labels and that Figure 1's caption is complete.

Circularity Check

1 steps flagged · score 6.0 of 10

The single-hop evaluation queries are generated from the same per-paper questions and query patterns that form the proposed index, so the headline Recall@3 gain is partly self-confirming; the multihop benchmark is external but does not test the question-compression mechanism.

self definitional [Section 6.3 (Experimental Setup, Task 1) and Appendix (prompts); cf. Section 5.1 (Queries Generation, Lexical Matching)]
"The queries were generated using a zero-shot prompting strategy with Claude 4, based on predefined questions and query patterns extracted from each paper (prompt details available in Appendix 10). ... Appendix prompt: 'generate csv documents of a query per item based on the queries and questions that result in a simulated simple user query that combines the main technical approach or method and the broader goal or alternative perspective.'"

In the proposed method, the retrieval index is built from per-paper generated questions and queries: Section 5.1 states 'Queries Generation: From generated questions and identified keywords, we construct multiple search queries associated with each paper,' and retrieval's first stage is 'Lexical Matching: The query is passed through a filtering approach to match with predefined questions and queries at the lexical level.' The Task 1 evaluation queries are generated from exactly the same 'predefined questions and query patterns extracted from each paper,' as the Appendix prompt confirms by creating a 'simulated simple user query' from the paper's own queries and questions.

full rationale

The central single-hop claim is partially circular by construction: the evaluation queries are not independent samples of real user queries but are generated from the same 'predefined questions and query patterns extracted from each paper' that constitute the question-based index, and the retrieval pipeline explicitly matches against those predefined questions and queries at the lexical level. This makes the Task 1 comparison against chunking baselines unfair, because the baselines do not have access to those per-paper cues. The multihop evaluation on LongBench 2WikiMultihopQA is an external benchmark and is not circular; however, it tests only the syntactic reranker, not the question-generation or paper-card encoding claimed in the title, and the reranker parameter L is swept between two values on the test set, so the multihop result does not independently validate the proposed knowledge-compression mechanism. The paper's own Limitations section notes the fixed-L dependency and the absence of adaptive parameter selection, which further supports treating the multihop result as a reranker benchmark rather than a validation of question-based encoding. No load-bearing self-citation chain or imported uniqueness theorem is present. Overall, the partial circularity is localized to the single-hop prediction, giving a score of 6.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical or theoretical entities are introduced; 'paper-cards' are text summaries, not postulated entities. The main free parameter is the reranker threshold L, with a secondary design choice on card length.

free parameters (2)

L (reranker threshold) = 2 and 3
Minimum number of query keywords that must co-occur in a passage; Section 6.3 says experiments use L=2 and L=3, and Table 9 shows performance varies with L (e.g., Llama2 0.470 vs 0.560 in 0-4K). The paper's Limitations section notes the reranker cannot adaptively set L.
Paper-card length limit = 300 characters (abstract) / 300 words (Section 5.1)
The card size is a design constraint that affects how much information is preserved; the paper is inconsistent about the limit, which adds ambiguity to the method.

assumptions (4)

domain assumption Generated questions and queries faithfully span the lexical and semantic space of the source documents.
Section 5.1 assumes that question generation 'captures the core contributions' and creates targeted retrieval cues; if questions miss important content, retrieval fails.
domain assumption Part-of-speech filtering (exclude ADP, CCONJ, punctuation) yields keywords that are sufficient for passage relevance.
Section 5.2 uses POS-based keyword extraction as the sole relevance signal in the reranker.
domain assumption LongBench 2WikiMultihopQA is a valid proxy for multihop retrieval performance.
The paper uses only this dataset for Task 2 and generalizes from it.
domain assumption Reported baseline numbers (LongBench, CFIC) are accurately transcribed and the comparisons are fair.
The paper does not reproduce the baselines and Table 9 shows unit inconsistencies, so the trustworthiness of baseline comparisons is assumed rather than verified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Knowledge Compression via Question Generation: Enhancing Multihop Document Retrieval without Fine-tuning." pith.science (2026). https://pith.science/paper/SUQDVCFM

@misc{pith2026250613778,
  author       = {Pith},
  title        = {Pith review of: Knowledge Compression via Question Generation: Enhancing Multihop Document Retrieval without Fine-tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SUQDVCFM}},
  note         = {Machine review of arXiv:2506.13778}
}

read the original abstract

This study presents a question-based knowledge encoding approach that improves retrieval-augmented generation (RAG) systems without requiring fine-tuning or traditional chunking. We encode textual content using generated questions that span the lexical and semantic space, creating targeted retrieval cues combined with a custom syntactic reranking method. In single-hop retrieval over 109 scientific papers, our approach achieves a Recall@3 of 0.84, outperforming traditional chunking methods by 60 percent. We also introduce "paper-cards", concise paper summaries under 300 characters, which enhance BM25 retrieval, increasing MRR@3 from 0.56 to 0.85 on simplified technical queries. For multihop tasks, our reranking method reaches an F1 score of 0.52 with LLaMA2-Chat-7B on the LongBench 2WikiMultihopQA dataset, surpassing chunking and fine-tuned baselines which score 0.328 and 0.412 respectively. This method eliminates fine-tuning requirements, reduces retrieval latency, enables intuitive question-driven knowledge access, and decreases vector storage demands by 80%, positioning it as a scalable and efficient RAG alternative.

Figures

Figures reproduced from arXiv: 2506.13778 by the authors.

**Figure 4.** Perfomances on the 2WikiMultihopQA_e. LLama3 and LLama2 correspond to the models using our approach with some specificities on the parameter L which can be either 3 or 2. 8 Discussions Our investigation began with the hypothesis that queries and question anchors enhance information retrieval effectiveness in document-based RAG systems 8.1 Task 1 The first evaluation conclusively demonstrates that our approach outpe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 2 canonical work pages

[3]

ArXiv, abs/2407.21783

The llama 3 herd of models. ArXiv, abs/2407.21783. Hongyu Gong, Yelong Shen, Dian Yu, Jianshu Chen, and Dong Yu

arXiv
[4]

ArXiv, abs/2402.05131

Financial report chunking for effective retrieval augmented genera - tion. ArXiv, abs/2402.05131. Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Kuttler, Mike Lewis, Wen tau Yih, Tim Rock - täschel, Sebastian Riedel, and Douwe Kiela

arXiv
[7]

ArXiv, abs/2402.03367

Rag-fusion: a new take on retrieval -augmented generation . ArXiv, abs/2402.03367. Hugo Touvron, Louis Martin, Kevin R. Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M. Bikel, Lukas Blecher, Cris tian Cantón Ferrer, Moya Chen, Guillem Cucu rull, David Esiobu, Jude Fernande...

arXiv
[9]

ArXiv, abs/2503.09600

Moc: Mixtures of text chunking learners for retrieval -augmented generation system . ArXiv, abs/2503.09600. Jihao Zhao, Zhiyuan Ji, Pengnian Qi, Simin Niu, Bo Tang, Feiyu Xiong, and Zhiyu Li

arXiv
[10]

arXiv preprint arXiv:2410.12788

Meta- chunking: Learning efficient text segmentation via logical perception. arXiv preprint arXiv:2410.12788. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dache ng Li, Eric P. Xing, Haotong Zhang, Joseph E. Gonzalez, and Ion Stoica

arXiv
[11]

ArXiv, abs/2306.05685

Judging llm -as-a-judge with mt -bench and chatbot arena. ArXiv, abs/2306.05685. Zhiying Zhu, Zhiqing Sun, and Yiming Yang

arXiv
[12]

ArXiv, abs/2403.04307

Halueval-wild: Evaluating hallucinations of language models in the wild. ArXiv, abs/2403.04307. Appendix Prompts used to generate queries from prede- fined questions and queries: Generate only one query for each, which com- bines the main technical approach or method and the application domain or specific tech- nique. generate csv documents of a query per...

arXiv
[2019]

BERT: Pre -training of deep bidirectional transformers for language under - standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Andr...

work page 2019

Show all 12 references

[2020]

ArXiv, abs/2005.11401

Retrieval-augmented generation for knowledge - intensive nlp tasks. ArXiv, abs/2005.11401. Carlo Merola and Jaspinder Singh

2005 arXiv
[2023]

ArXiv, abs/2307.09288

Llama 2: Open foundation and fine-tuned chat models. ArXiv, abs/2307.09288. Shicheng Xu, Liang Pang, Huawei Shen, and Xueqi Cheng

arXiv
[2024]

ArXiv, abs/2406.17526

Lumberchunker: Long-form narra - tive document segmentation. ArXiv, abs/2406.17526. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony S. Hartshorn, Aobo Yan...

arXiv
[2025]

Association for Computa- tional Linguistics

Is semantic chunking worth the computational cost? In Findings of the Association for Computational Linguistics: NAACL 2025 , pages 2155 –2177, Al - buquerque, New Mexico. Association for Computa- tional Linguistics. Zackary Rackauckas

2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

ArXiv, abs/2407.21783

The llama 3 herd of models. ArXiv, abs/2407.21783. Hongyu Gong, Yelong Shen, Dian Yu, Jianshu Chen, and Dong Yu

arXiv

[2] [4]

ArXiv, abs/2402.05131

Financial report chunking for effective retrieval augmented genera - tion. ArXiv, abs/2402.05131. Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Kuttler, Mike Lewis, Wen tau Yih, Tim Rock - täschel, Sebastian Riedel, and Douwe Kiela

arXiv

[3] [7]

ArXiv, abs/2402.03367

Rag-fusion: a new take on retrieval -augmented generation . ArXiv, abs/2402.03367. Hugo Touvron, Louis Martin, Kevin R. Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M. Bikel, Lukas Blecher, Cris tian Cantón Ferrer, Moya Chen, Guillem Cucu rull, David Esiobu, Jude Fernande...

arXiv

[4] [9]

ArXiv, abs/2503.09600

Moc: Mixtures of text chunking learners for retrieval -augmented generation system . ArXiv, abs/2503.09600. Jihao Zhao, Zhiyuan Ji, Pengnian Qi, Simin Niu, Bo Tang, Feiyu Xiong, and Zhiyu Li

arXiv

[5] [10]

arXiv preprint arXiv:2410.12788

Meta- chunking: Learning efficient text segmentation via logical perception. arXiv preprint arXiv:2410.12788. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dache ng Li, Eric P. Xing, Haotong Zhang, Joseph E. Gonzalez, and Ion Stoica

arXiv

[6] [11]

ArXiv, abs/2306.05685

Judging llm -as-a-judge with mt -bench and chatbot arena. ArXiv, abs/2306.05685. Zhiying Zhu, Zhiqing Sun, and Yiming Yang

arXiv

[7] [12]

ArXiv, abs/2403.04307

Halueval-wild: Evaluating hallucinations of language models in the wild. ArXiv, abs/2403.04307. Appendix Prompts used to generate queries from prede- fined questions and queries: Generate only one query for each, which com- bines the main technical approach or method and the application domain or specific tech- nique. generate csv documents of a query per...

arXiv

[8] [2019]

BERT: Pre -training of deep bidirectional transformers for language under - standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Andr...

work page 2019