Pith. sign in

REVIEW 3 major objections 4 minor 3 cited by

EXIT: Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper shows that a small, parallel sentence classifier can compress retrieved documents for RAG both more accurately and much faster than abstractive summarization, while also beating the uncompressed baseline.

desk verdict EXIT is a solid, thoroughly evaluated extractive compression method for RAG, but the abstract's 'consistently surpasses' claim is contradicted by the paper's own appendix. read the letter →

arxiv 2412.12559 v3 pith:THXFXXCM submitted 2024-12-17 cs.CL cs.AIcs.IR

Taeho Hwang , Sukmin Cho , Soyeong Jeong , Hoyun Song , SeungYoon Han , Jong C. Park This is my paper · ORCID
classification cs.CLcs.AIcs.IR
keywords retrieval-augmentedgenerationcontextcompressionextractivesentenceclassificationquestionansweringmulti-hopreasoninglatencyreductionadaptiveselection
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

EXIT is an attempt to settle a practical question: when a RAG system retrieves more documents than it can afford to read, what is the cheapest way to keep only the sentences that matter? The paper's answer is to treat compression as a binary classification problem—each sentence is judged 'Yes' or 'No' for relevance, conditioned on the query and its full source document, by a 2-billion-parameter model that reads all sentences in parallel. The selected sentences are then stitched back in their original order. Across single-hop (NQ, TriviaQA) and multi-hop (HotpotQA, 2WikiMultihopQA) benchmarks, EXIT reports higher Exact Match and F1 than every tested compressor and the uncompressed baseline, with an average total latency near one second. If those numbers hold, then the dominant cost in RAG compression—autoregressive summarization—can be replaced by a faster, more accurate filtering step that also removes distracting content that otherwise misleads the reader.

What carries the argument

The load-bearing mechanism is the context-aware binary relevance classifier with single-token probability scoring. Instead of generating compressed text token by token, EXIT defines relevance as the normalized probability of 'Yes' versus 'No' for each sentence given the query and its whole source document (Equation 4). Because the 'Yes'/'No' decision consumes a single next-token probability, the model processes all sentences of the retrieved set in parallel, which turns extra compute into lower wall-clock latency. Around this classifier sit three design choices that carry the argument: sentence-level decomposition (rather than token-level pruning, which breaks entities), conditioning on the full document when judging each sentence (removing this costs 1.2 EM in ablation), and adaptive, threshold-based selection rather than fixed-size selection (fixed four-sentence selection costs 1.4 EM). The classifier is trained with balanced positive, hard-negative, and random-negative samples derived from HotpotQA's sentence-level annotations.

What would settle it

Construct a set of multi-hop questions whose gold evidence sentences are each individually irrelevant to the query taken alone (for example, no word overlap between the query and any single evidence sentence, but a chain through two or more sentences yields the answer), run EXIT at its default threshold, and check whether any required evidence sentence is pruned. If pruning occurs and the reader's accuracy on that set falls below the uncompressed baseline, the per-sentence decision rule is refuted. A complementary check: take EXIT's wrong answers and re-add the pruned sentences to the prompt; any question that flips from wrong to correct is direct evidence that sentence-independent pruning discarded answer-critical content.

Watch

Extended reading notes

Core claim

The core discovery is that the quality of compressed context is governed less by how the retained text is paraphrased than by whether the retained sentences are the right ones, judged in context. EXIT computes, for each sentence $s_{ij}$ in a retrieved document $d_i$, the relevance score $r_{ij} = P(\text{Yes}|q,d_i,s_{ij})/(P(\text{Yes}|q,d_i,s_{ij}) + P(\text{No}|q,d_i,s_{ij}))$ using a fine-tuned Gemma-2B model, keeps every sentence whose score exceeds a threshold (set to 0.5), and concatenates survivors in original document order. Trained only on HotpotQA, this classifier reports consistent EM gains over RECOMP, CompAct, Refiner, and LongLLMLingua on all four datasets—for example, +2.5 EM on HotpotQA and +8.1 on 2WikiMultihopQA with an 8B reader—while running at 0.8 seconds average end-to-end latency, faster than the uncompressed documents. With a 70B reader, the average EM gain over the uncompressed baseline is 3.7 points, at 3.5 seconds. The paper also reports that the method scales to 30 retrieved documents without degradation, cutting tokens by 86.8% while improving accuracy, and transfers to BM25 retrieval, GPT-4o readers, and specialized domains such as BioASQ and COVID-QA.

Load-bearing premise

The approach assumes that a sentence's usefulness can be decided by looking at that sentence alone, with its document as context, and that keeping every sentence above a fixed threshold cannot destroy evidence that only emerges when several individually weak sentences are read together.

Editorial extensions

If this is right

  • Abstractive compressors' latency penalty can be removed: EXIT compresses in roughly 0.36s versus 1.5–8s for RECOMP-Abst and CompAct, while improving EM on every dataset tested.
  • Retrieval sets can grow aggressively: increasing k from 1 to 30 documents raises EXIT's EM from 28.2 to 33.1 on HotpotQA while cutting tokens by 86.8%, so the usual 'lost in the middle' degradation does not appear.
  • The compressor transfers across readers and retrievers: trained only on HotpotQA, it improves accuracy with Llama-3.1-8B, Llama-3.1-70B, and GPT-4o readers, and with both dense (Contriever) and sparse (BM25) retrieval.
  • For API-based pipelines, the compression ratio (retaining roughly 15–31% of tokens) translates directly into lower input cost and lower latency without sacrificing answer quality.
  • The method is plug-and-play: it requires no retraining of the retriever or reader and no architectural changes, so existing RAG systems can adopt it by inserting the classifier between retrieval and reading.

Reading between the lines

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

  • The accuracy gains may come partly from removing misleading distractors rather than from efficient evidence retention: the paper's case studies show the reader answering correctly from its internal knowledge when EXIT prunes all context, so gains could shrink on questions where the model has no prior knowledge to fall back on.
  • A natural extension is to generalize the per-sentence decision to short spans or sentence pairs, allowing joint evidence to pass the threshold; this could close the multi-sentence evidence gap without losing the parallelism.
  • Because the paper shows GPT-4o can produce pseudo-labels with high precision on 'No' sentences, a fully automated data-generation loop is plausible for new domains, removing the manually annotated training data requirement.
  • The same compression module could be applied to any long-context task beyond QA (e.g., summarization or document-grounded dialogue), since the classifier only needs query–sentence relevance labels rather than QA-specific annotations.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces EXIT, an extractive context compression framework for RAG. Retrieved documents are split into sentences; a fine-tuned Gemma-2B classifier scores each sentence's relevance to the query conditioned on its containing document (Equation 4); sentences above a threshold are reassembled in original order. The method is evaluated on NQ, TQA, HQA, and 2WikiMultiHopQA with Contriever and BM25 retrievers and Llama-3.1-8B/70B and GPT-4o readers, reporting EM, F1, token counts, and end-to-end latency. The main Table 1 shows that for Llama-3.1 readers at Top-5, EXIT achieves the highest EM on all four datasets while maintaining sub-second (8B) latency, with additional ablations and analyses on threshold sensitivity, compressor model size, training data composition, and pseudo-annotation.

Significance. If the claims were fully supported, EXIT would be a practical contribution: a lightweight, parallelizable extractive compressor that outperforms both abstractive compressors and uncompressed context in QA accuracy while reducing latency. The paper includes several strengths: evaluation across multiple readers, retrievers, and datasets; latency measured as means and standard deviations over five runs (Table 12); ablations on training data, adaptive selection, and passage context; threshold sensitivity analysis (Figure 7); and a public code release. However, the abstract's universal claim of 'consistently surpasses ... even uncompressed baselines' is contradicted by the paper's own appendix, and no statistical significance is reported for accuracy differences. The central contribution is defensible for the main Llama-3.1 configuration, but the presentation overstates its scope.

major comments (3)
  1. [Abstract and Section 5] The abstract states that EXIT 'consistently surpasses existing compression methods and even uncompressed baselines in QA accuracy,' and Section 5 reports that 'compared to all baseline methods, EXIT consistently improves EM scores.' These universal claims are contradicted by the paper's own data: Appendix C.2, Table 9 (GPT-4o reader, 2WIKI, Top-20) shows Original Docs achieving 40.0 EM versus EXIT's 35.6, and Appendix C.1, Table 8 (BM25 retriever, 2WIKI, Top-5) shows RECOMP-Abst achieving 25.0 EM versus EXIT's 24.4. The paper must either restrict the claim to the configurations in Table 1 (Llama-3.1 readers, Contriever, Top-5) and explicitly reconcile the Appendix counterexamples, or provide a principled reason for excluding those settings. As written, the central claim is falsified by the manuscript's own reported results.
  2. [Section 5 / Table 1] No statistical significance or variance is reported for any EM or F1 difference. The headline gains in Table 1 range from about 1 to 8 EM points, but several 8B-reader differences are only 1–2 points (e.g., NQ 35.9 vs. 34.6, TQA 60.8 vs. 59.2), and the paper does not demonstrate that these exceed noise. Given the counterexamples in the appendix, the paper should report paired significance tests or bootstrap confidence intervals for the main comparisons, and temper the 'consistently improves' wording accordingly.
  3. [Appendix C.14, Table 19] The case study in Table 19 shows EXIT pruning all retrieved context and the reader still answering correctly, which the authors attribute to 'hallucination or model bias.' This is not a harmless anecdote: it implies that a correct answer can be produced from the reader's parametric knowledge when EXIT has removed all evidence. To support the causal claim that compression improves QA accuracy by preserving key evidence, the paper should quantify how often EXIT produces an empty or evidence-free compressed context and report reader accuracy on those cases. Without such analysis, some of the reported gains may reflect the reader's ability to answer without context rather than effective evidence selection.
minor comments (4)
  1. [Section 3.3, Eq. (5)] The loss expression is malformed: the indicator functions are written as '1 l=“Yes”' without proper notation, and the equation omits the explicit indicator brackets. Please rewrite as L = −\mathbb{1}[l=\text{Yes}] \log P(\text{Yes}) − \mathbb{1}[l=\text{No}] \log P(\text{No}).
  2. [Section 6.4, Table 3] The row labels 'w/o Adaptive Sentence Selection (2 sents)' and '(4 sents)' are inconsistent with the text, which describes a fixed-length strategy that 'limits selection to four sentences.' Clarify which fixed sizes were tested and how the numbers in Table 3 correspond to them.
  3. [Section 5, first paragraph] The sentence 'for instance, by 1.3 and 2.0 points on NQ and TQA, and by even larger margins of 2.5 and 8.1 points on HQA and 2WIKI' appears to compare against the uncompressed baseline in Table 1, but the preceding clause compares against 'all baseline methods.' Please specify that the cited margins are relative to Original Docs, or recompute them against the best baseline.
  4. [Section 3.2, Eq. (4)] The method classifies each sentence independently given its document, and selected sentences are concatenated without joint reasoning over the selected set. The paper would benefit from a discussion or failure analysis of multi-hop questions that require combining sentences that are individually irrelevant but jointly necessary; the positive empirical results on HQA and 2WIKI are encouraging, but the paper does not examine cases where EXIT prunes one of the two necessary supporting sentences.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: EXIT's QA accuracy claims are measured against external benchmarks, and the method does not reduce by construction to its training data or to self-citations.

full rationale

The paper's central claims are empirical benchmark results, not derivations. EM/F1 are computed by running a fixed reader on compressed contexts from held-out QA datasets (NQ, TQA, HQA, 2WIKI), and the reported gains are not constructed from the classifier's training objective. The relevance classifier is trained on HQA train and evaluated on HQA dev / other datasets; QA accuracy is a downstream downstream measured outcome, not the training label itself, so this is standard supervised evaluation rather than a fitted-input-called-prediction circularity. The threshold tau=0.5 is a tuned hyperparameter, but Appendix C.6 shows EXIT is robust across tau=0.3-0.9, so the main result does not reduce to that single fitted value. The two same-group citations (Jeong et al. 2024 for adaptive selection, Hwang et al. 2024 for order preservation) are used only as motivational support for design choices that are independently ablated in Tables 3 and 16; they are not load-bearing uniqueness theorems and do not forbid alternatives. The abstract's 'consistently surpasses' wording is contradicted by the paper's own Appendix Tables 8 and 9 in specific settings (e.g., RECOMP-Abst beats EXIT on BM25/2WIKI, and uncompressed docs beat EXIT on GPT-4o/2WIKI), but that is an internal consistency / correctness concern, not a circular derivation. No equation in the paper is equivalent to its input by construction, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on empirical transfer assumptions rather than mathematical derivation. The only fitted scalar is the relevance threshold tau=0.5. There are no invented particles, forces, or new theoretical entities; the system components are standard sentence tokenization, a fine-tuned classifier, and a reader LLM.

free parameters (1)
  • relevance threshold tau = 0.5
    Set based on empirical tuning (Section 4 and Appendix C.6). The paper reports sensitivity across tau values but does not describe a separate held-out selection procedure, so this is a fitted scalar that affects compression ratio and accuracy.
assumptions (4)
  • domain assumption A sentence's relevance can be decided independently per sentence given the query and its containing document.
    Section 3.2, Equation 4, scores each sentence separately and then applies a threshold. If multi-hop evidence requires jointly deciding which combinations of sentences to keep, this assumption can fail.
  • domain assumption HotpotQA sentence-level evidence annotations transfer to other QA distributions, including single-hop and out-of-domain multi-hop datasets.
    Section 3.3 and Section 4: the classifier is trained only on HotpotQA train sentences and evaluated on NQ, TQA, 2WikiMultihopQA, MuSiQue, BioASQ, and COVID-QA. The zero-shot generalization claims depend on this transferability.
  • domain assumption Single-token probabilities from Gemma-2B provide a calibrated enough relevance signal to separate useful from useless sentences.
    Sections 3.2 and 6.5: the method relies on P(Yes) versus P(No) from an instruction-tuned 2B model. The paper does not calibrate these probabilities or compare against a calibrated probabilistic classifier.
  • domain assumption SpaCy sentence segmentation preserves the semantic units needed for question answering.
    Section 4 and Appendix B.3: all documents are split with a rule-based tokenizer. The limitation section notes that sentences that are overly long, noisy, or entity-heavy may not be well handled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EXIT: Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation." pith.science (2026). https://pith.science/paper/THXFXXCM

@misc{pith2026241212559,
  author       = {Pith},
  title        = {Pith review of: EXIT: Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/THXFXXCM}},
  note         = {Machine review of arXiv:2412.12559}
}
read the original abstract

We introduce EXIT, an extractive context compression framework that enhances both the effectiveness and efficiency of retrieval-augmented generation (RAG) in question answering (QA). Current RAG systems often struggle when retrieval models fail to rank the most relevant documents, leading to the inclusion of more context at the expense of latency and accuracy. While abstractive compression methods can drastically reduce token counts, their token-by-token generation process significantly increases end-to-end latency. Conversely, existing extractive methods reduce latency but rely on independent, non-adaptive sentence selection, failing to fully utilize contextual information. EXIT addresses these limitations by classifying sentences from retrieved documents - while preserving their contextual dependencies - enabling parallelizable, context-aware extraction that adapts to query complexity and retrieval quality. Our evaluations on both single-hop and multi-hop QA tasks show that EXIT consistently surpasses existing compression methods and even uncompressed baselines in QA accuracy, while also delivering substantial reductions in inference time and token count. By improving both effectiveness and efficiency, EXIT provides a promising direction for developing scalable, high-quality QA solutions in RAG pipelines. Our code is available at https://github.com/ThisIsHwang/EXIT

Figures

Figures reproduced from arXiv: 2412.12559 by the authors.

Figure 1
Figure 1. Average QA accuracy (EM) and efficiency (To [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our framework. First, the retrieved document is split into sentences. Next, each sentence [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance analysis on HQA across different Top- [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison of compression and reading la [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Ablation on HQA comparing EM scores and latency for different model configurations within EXIT (red dot). CompAct and Original Docs are included as indicators. Experiments on a single A100-80GB GPU. ability to filter out spurious retrieval noise. Adaptive Sentence Sele…
Figure 6
Figure 6. Figure 6: Confusion matrix of GPT-4o zero-shot clas [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Changes in EM, F1 score, and token count as the threshold [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Confusion matrices (row-normalized) for context-aware relevance classification on HQA (in￾domain) and 2WIKI (out-of-domain). narrowing this discrepancy as a valuable future re￾search direction. C.11 Classification Performance under Ablation Settings C.11.1 Analysis of …
Figure 9
Figure 9. Figure 9: Row-normalized confusion matrices for clas [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Row-normalized confusion matrices compar [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MHA-RAG: Improving Efficiency, Accuracy, and Consistency by Encoding Exemplars as Soft Prompts

    cs.AI 2025-10 conditional novelty 6.0 of 10

    MHA-RAG encodes retrieved exemplars into order-invariant soft prompts via multi-head attention, claiming ~20-point effective-accuracy gains over RAG at ~10x lower inference FLOPs.

  2. Beyond Hard and Soft: Hybrid Context Compression for Balancing Local and Global Information Retention

    cs.CL 2025-05 conditional novelty 6.0 of 10

    HyCo2 combines soft global compression with hard local token selection, reporting QA performance near uncompressed retrieval while cutting context tokens by about 88.8%.

  3. SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression

    cs.CL 2025-07 conditional novelty 5.0 of 10

    SARA combines short natural-language snippets with vector-compressed summaries of the remaining retrieved documents, improving RAG answer quality under 512/1024-token context budgets.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages · cited by 3 Pith papers

  1. [1]

    Yes”) and negative (“No

    BioASQ (Tsatsaronis et al., 2015), a biomedi- cal QA dataset using the PubMed corpus (14.9M documents), 2) COVID-QA (Möller et al., 2020), a COVID-specific QA benchmark based on the TREC-COVID collection (Roberts et al., 2021) (171K documents). Although our classifier was trained solely on HQA, a general-domain dataset, EXIT performs competitively or bett...

    work page 2015

  2. [2]

    "Custard

    Creme anglaise However, the ice cream base is much thicker and has various flavourings. The American South it is known as ""Custard."" It can be served like Eggnog during the Christ- mas season. Other names include the French terms ""crème à l’anglaise"" (""English-style cream"") and ""crème française"" (""French cream""). Crème anglaise Crème anglaise (F...

    work page

  3. [3]

    "Île flottante

    Creme anglaise Alternatively, it can be drunk as a dessert on its own, for example in ""Île flottante"" (""floating island""): the cream is poured into a bowl with a piece of meringue (""blancs en neige"") floated on top along with praline. It can also be used as a base for desserts such as ice cream or crème brûlée

    work page

  4. [4]

    CoRR, abs/2403.08295

    Gemma: Open models based on gemini re- search and technology. CoRR, abs/2403.08295. Timo Möller, Anthony Reina, Raghavan Jayakumar, and Malte Pietsch. 2020. COVID-QA: A question an- swering dataset for COVID-19. In Proceedings of the 1st Workshop on NLP for COVID-19 at ACL 2020, Online. Association for Computational Linguistics. Jesse Mu, Xiang Li, and No...

    arXiv 2020

  5. [5]

    A similar recipe with only meringue filling is called Šampita

    Cremeschnitte usually pure thick Custard, less commonly combined with meringue (whipped egg whites and sugar) creme. A similar recipe with only meringue filling is called Šampita. In Australia, the dish is more commonly known as a ’vanilla slice’. Cremeschnitte A cremeschnitte („„„„ ) is a chantilly and Cus- tard cream cake dessert popular in several Cent...

    work page

  6. [11]

    Yes” or “No

    employ dense retrieval and multi-stage rank- ing, while UPR (Sachan et al., 2022) and rank- ing prompting methods (Liang et al., 2023; Qin et al., 2024) refine retrieval through generation- based approaches. However, these methods focus on reordering passages rather than compressing retrieved content. Research on retrieval granularity (Seo et al., 2019; L...

    work page 2022

  7. [14]

    "Crème pâtissière

    Custard lemon. ""Crème pâtissière"" is a key ingredi- ent in many French desserts including mille- feuille (or Napoleons) and filled tarts. It is also used in Italian pastry and sometimes in Boston cream pie. The thickening of the Cus- tard is caused by the combination of egg and cornstarch. Corn flour or flour thicken at 100 °C and as such many recipes i...

    work page

  8. [16]

    "English cream

    Creme anglaise Crème anglaise Crème anglaise (French for ""English cream"") is a light pouring Cus- tard used as a dessert cream or sauce

    work page

Show all 29 references
  1. [17]

    "crème à l’anglaise

    Creme anglaise Other names include the French terms ""crème à l’anglaise"" (""English-style cream"") and ""crème française"" (""French cream""). Crème anglaise Crème anglaise (French for ""English cream"") is a light pour- ing Custard used as a dessert cream or sauce

  2. [19]

    "Crème pâtissière

    Custard ""Crème pâtissière"" is a key ingredient in many French desserts including mille-feuille (or Napoleons) and filled tarts. Answer Custard Predict Eggnog Custard 24 Table 19: Case study comparing compressed contexts and answers between Original Docs and Ours. Despite con...

    1988

  3. [20]

    Wagner Steven Bochco and several of his projects

    Michael I. Wagner Steven Bochco and several of his projects. Wagner was asked by ABC in 1987 to help de- velop a new science fiction series, " Probe ", a light-hearted series about a scientific crime fighter named Austin James. ... Parker Stevenson, who played the lead char- a...

    1987

  4. [21]

    Aliens",

    John Wagner the mid-1990s Wagner worked on a number of licensed properties for Dark Horse Comics in the US, including "Aliens", "Star Wars" – notably solo stories starring Boba Fett and the comics strand of the multimedia project "" – and "". ... It was nominated for the Angou...

  5. [22]

    Martin Wagner (artist) Martin Wagner (artist) Martin Wagner (born April 27, 1966) is an American artist, car- toonist, and filmmaker. ... . His production schedule became increas- ingly protracted and he ceased publishing the series altogether following issue No. 12 in

    1966

  6. [23]

    L.A. Law

    Wired (film) "L.A. Law", "Murphy Brown", and "Sein- feld"), Chiklis gained fame for portraying the lead roles of Commissioner Tony Scali on the ABC police drama "The Commish" (1991- 1996), and LAPD Detective Vic Mackey on the FX police drama "The Shield" (2002-

    1991

  7. [25]

    Phillips disaster of 1989 On October 23, 1989, at 1400 Jefferson Road, Pasadena, Texas , a powerful explosion killed 23 and injured 314. A powerful and devas- tating explosion and fire ripped through the HCC, killing 23 persons—all working at the facility—and injuring 314 othe...

    1989

  8. [26]

    The phrase is: There are only three sports: bullfighting, motor racing, and mountaineering; all the rest are merely games

    Extreme sport usually, but wrongly, attributed to Ernest Hemingway . The phrase is: There are only three sports: bullfighting, motor racing, and mountaineering; all the rest are merely games. Answer Ernest Hemingway Predict There is no information provided about who said that....

    1984

  9. [27]

    1984 Summer Olympics boycott The boycott involved 14 Eastern Bloc coun- tries and allies, led by the Soviet Union, which initiated the boycott on May 8, 1984

    1984

  10. [28]

    Although a boycott led by the Soviet Union depleted the field, 140 NOCs took part

    Summer Olympic Games Eastern Bloc that did attend the 1984 Olympics. Although a boycott led by the Soviet Union depleted the field, 140 NOCs took part. Without Eastern European coun- tries, the 1984 Games were dominated by the host. The Games were also the first time mainland ...

    1984

  11. [29]

    1984 Summer Olympics boycott However, no threat to Eastern Bloc athletes was discovered, and the athletes from the Eastern Bloc country that did attend the 1984 games— Romania —encountered no prob- lems. Answer Romania Predict China Romania 26 Table 21: Zero-shot QA prompt eva...

    1984

  12. [981]

    Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and En- rico Shippole

    Association for Computational Linguistics. Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and En- rico Shippole. 2024. Yarn: Efficient context window extension of large language models. In The Twelfth International Conference on Learning Representa- tions, ICLR 2024, Vienna, Austr...

    2024 arXiv

  13. [1793]

    Junyi Li, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen

    Association for Computational Linguistics. Junyi Li, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023c. Halueval: A large-scale hal- lucination evaluation benchmark for large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Lan...

    2023 arXiv

  14. [2008]

    The Thing

    and as well as Marvel superhero Ben "The Thing" Grimm in the films "Fantastic Four" (2005) and "" (2007). ... Wired (film) Wired is a 1989 biographical film of comedian and actor John Belushi, di- rected by Larry Peerce, and adapted from the 1984 book of the same name by "Wash...

    2005

  15. [2016]

    MS MARCO: A human generated machine reading comprehension dataset. In Proceedings of the Workshop on Cognitive Computation: Integrat- ing neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Infor- mation Processing Systems (NIPS 2016), Barc...

    2016 arXiv

  16. [2017]

    Get to the point: Summarization with pointer- generator networks. In Proceedings of the 55th An- nual Meeting of the Association for Computational Linguistics, ACL 2017, V ancouver , Canada, July 30 - August 4, V olume 1: Long Papers, pages 1073–1083. Association for Computati...

    2017

  17. [2018]

    Ranking sentences for extractive summariza- tion with reinforcement learning. In Proceedings of the 2018 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June ...

    2018

  18. [2021]

    Searching for scientific evidence in a pan- demic: An overview of TREC-COVID. J. Biomed. Informatics, 121:103865. Stephen E. Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and be- yond. F ound. Trends Inf. Retr ., 3(4):333–389. Devendra Singh Sa...

    2009

  19. [2024]

    Michael Günther, Isabelle Mohr, Bo Wang, and Han Xiao

    OpenReview.net. Michael Günther, Isabelle Mohr, Bo Wang, and Han Xiao. 2024. Late chunking: Contextual chunk embed- dings using long-context embedding models. CoRR, abs/2409.04701. Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing A multi-hop ...

    2024 arXiv

  20. [2425]

    ACM. Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aakanksha Chowdh- ery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone...

  21. [4441]

    Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen

    Association for Computational Linguistics. Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen. 2023. En- hancing retrieval-augmented large language models with iterative retrieval-generation synergy. In Find- ings of the Association for Computationa...

    2023 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.