Pith. sign in

REVIEW 4 major objections 5 minor 85 references

Deterministic linguistic rules alone can rival neural prompt compressors on longer documents

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:43 UTC pith:VWZQBMKL

load-bearing objection The deployment cost advantage is real, but the paper's own statistics do not support the headline parity claim and the per-benchmark evolution confounds the comparison. the 4 major comments →

arxiv 2607.25335 v2 pith:VWZQBMKL submitted 2026-07-28 cs.CL cs.AI

Every Time I Hire a Linguist, Inference Costs Go Down: On Linguistic Rules as Effective Prompt Compressors

Jianfei Ma , Zhaoxin Feng , Emmanuele Chersoni , Si Chen This is my paper
classification cs.CL cs.AI
keywords prompt compressionlinguistic rulesevolutionary searchLLM inference costdeterministic compressiondual-path evaluationdirect accuracyreconstruction accuracy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks whether prompt compression needs an LLM to score token importance at compression time. It claims that deterministic linguistic rules—such as keeping content words, negation markers, core argument triples, and discourse connectives—can be combined into compressors that match the performance of recent neural prompt-compression methods on several medium-to-long document benchmarks, while using only CPU-side processing and no LM forward pass at deployment. Because the space of rule combinations is too large to design by hand, the paper uses an offline evolutionary search to discover effective programs. If the claim holds, long-context LLM inference could be made substantially cheaper and faster by adding a lightweight, interpretable pre-compression step.

Core claim

The central claim is that linguistic rules alone, evolved offline, can serve as effective prompt compressors. The paper shows that an evolutionary search over lexical, syntactic, semantic, and discourse rules yields deterministic compressors that, on medium-to-long documents, achieve direct and reconstruction QA accuracy within the range of advanced neural baselines, with the lowest cross-model variability on two datasets and near-zero deviation from the target compression ratio. The best evolved compressors fuse signals from multiple linguistic levels and shift from token-level pruning at light compression to sentence-level extraction at aggressive compression.

What carries the argument

The mechanism is an offline evolutionary search over a space of deterministic linguistic rules. Starting from 42 seeds spanning lexical (content verses function words), syntactic (dependency core arguments, subtree pruning), semantic (negation, modals, quantifiers), discourse (sentence position, entity chains), and composite pipelines, an LLM-guided mutator iteratively recombines and refines code. Candidates are filtered through a three-stage cascade and scored by the minimum of Direct and Reconstruction QA accuracies, which forces the compressor to work both when the compressed text is used directly and when a reconstruction model regenerates prose.

Load-bearing premise

The comparison is unequal: the evolutionary search was tuned on 100 training examples from each of the same four benchmark families used for testing, while the neural baselines were not given per-benchmark tuning; if much of the observed parity comes from that benchmark-family tuning rather than from the linguistic rules themselves, the central claim would not transfer beyond the tuned families.

What would settle it

Run the evolved compressors on held-out long-document QA tasks from several benchmark families that were not used during evolution, at the same compression ratios and with the same receiver models. If direct accuracy falls substantially below the neural baselines on multiple such families, the claim that linguistic rules alone are competitive in longer-document settings is undermined. The paper's OOD results on MuSiQue and NarrativeQA are a step in this direction but cover only two families.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Long-context LLM inference can become cheaper and faster because compression runs as a deterministic CPU-side step with no GPU forward pass at deployment.
  • Compression behavior becomes interpretable: which tokens survive is decided by explicit part-of-speech, dependency, and discourse rules rather than opaque salience scores.
  • The rule-based approach keeps the compression ratio on target because an explicit word budget is enforced, avoiding the drift seen in several neural baselines.
  • The evolutionary analysis indicates that effective compression requires fusing multiple linguistic levels, and sentence-level extraction dominates under aggressive budgets—a design principle for future compressors.
  • The dual-path min(D,R) fitness objective could be adopted to improve robustness in compression systems that feed reconstruction or paraphrase models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the parity holds beyond the tuned benchmark families, deterministic linguistic compressors could substitute for neural compressors in deployment settings where compression must run on CPU, such as on-device or edge inference with limited GPU availability.
  • The observation that query-aware matching was eliminated by evolution suggests that indirect preservation of entities, predicates, and numbers can suffice for QA, which could inform retrieval-augmented generation about what content to retain.
  • A testable extension would be to rerun the evolutionary search on a length-balanced training pool and check whether short-document (RACE-like) performance improves; the paper attributes its RACE weakness to the length-biased training signal.
  • Applying the dual-path min(D,R) objective to train neural compressors could make them more robust on reconstruction paths, bridging the gap between hard and soft compression methods.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes replacing LM-based token scoring in prompt compression with deterministic linguistic rules discovered by offline evolutionary search. Forty-two handcrafted lexical, syntactic, semantic, and discourse seeds are evolved with an AlphaEvolve-style framework at four target compression ratios. The resulting Python programs use only spaCy and other lightweight NLP tools at compression time. Evaluation uses a dual-path protocol (Direct QA and Reconstruction-then-QA) on four datasets (RACE, Qasper, Multi-doc QA, LongMemEval), plus OOD transfer to MuSiQue/NarrativeQA, summarization, and classification. The main claim is that the evolved linguistic compressors achieve performance similar to recent neural prompt-compression baselines in many longer-document settings, with far lower deployment cost and no LM forward pass.

Significance. If the claim were established, the result would be practically and scientifically valuable: it would show that explicit, interpretable linguistic rules can substitute for costly LM-based scoring in at least part of the prompt-compression design space. The paper is commendably transparent: it reports bootstrap significance tests, OOD checks, per-dataset cost benchmarks, human validation of the judge, and an explicit limitations section. The main empirical apparatus is careful. However, the headline parity claim is not supported by the paper's own statistics: Appendix F reports 0 significant wins, 9 significant losses, and 47 inconclusive comparisons, and the text explicitly states that inconclusive is not equivalence. Combined with the unequal optimization budgets between the evolved method and the neural baselines, the evidence supports a more modest claim: linguistic-rule compressors produce point estimates in a similar range, but the paper does not yet demonstrate that they are competitive in the statistical sense or that the effect generalizes beyond benchmark-family-specific evolution.

major comments (4)
  1. [§5.1; Appendix F (Tables 20–21)] The central claim of 'performance similar to that of recent advanced prompt-compression strategies' outruns the statistical evidence. Appendix F reports 47 inconclusive, 9 significant losses, and 0 significant wins across 56 comparisons; the text itself notes that inconclusive outcomes establish neither a difference nor equivalence. The 12 direct judge comparisons on the three longer datasets are all inconclusive, and there are significant losses on RACE (2.5×, 4×, 6×) and on the reconstruction path for Qasper (2.5×, 6×). The abstract and contribution list should be revised to a weaker claim ('point estimates are within a few points, but not statistically distinguishable'), or the paper should supply equivalence testing / power analysis showing that the sample size can detect meaningful differences. As written, the headline claim is not established.
  2. [§4.1; Limitations] The comparison rests on unequal optimization budgets. The evolutionary search uses 100 training examples from each of the same four benchmark families used at test time, while the neural baselines are not given per-benchmark tuning. The OOD experiments partially mitigate this, but MuSiQue/NarrativeQA share the LongBench genre with the training families, and summarization/classification are tasks where generic extractive heuristics may suffice. Thus the results cannot distinguish 'linguistic rules are generally effective' from 'evolution specialized to these benchmark families.' Please add per-benchmark-tuned baselines (e.g., fine-tune or tune each baseline on the same 100-example splits) or evaluate on genuinely disjoint domains (e.g., biomedical, legal, or code) without re-evolution, or both.
  3. [Figure 4; Tables 12–15] Accuracy comparisons are made at nominal target ratios even though actual ratios differ substantially across methods. For example, LLMLingua-2 on RACE at the 6× target has an actual ratio of 7.74 while Ours is 6.00 (Table 12); several baselines also drift at other targets. The paper credits Ours with 'near-zero deviation' but does not adjust the accuracy comparisons for the actual compression applied. If a baseline is evaluated under more aggressive actual compression, the parity claim is confounded. Please either match actual ratios post hoc (e.g., analyze only cells with comparable r, or interpolate), or include actual compression ratio as a covariate.
  4. [Appendix F] The statistical significance analysis excludes R2C, even though R2C is included in the main accuracy tables and is often the strongest baseline (e.g., on Multi-doc QA, Table 1). The stated reason—paired per-example outputs were unavailable—is understandable, but excluding a strong competitor can make Ours appear closer to the best baseline than it is. Please provide paired outputs for R2C or report per-cell which baseline is used as the comparator in Tables 20–21 and justify the exclusion more fully. At minimum, the main text should state that the 'best baseline' comparisons do not include R2C.
minor comments (5)
  1. [Title/Abstract] The title and abstract use 'linguistic rules alone' while several evolved programs are surface heuristics (entity preservation, number/digit retention, sentence-position cues) rather than deeply linguistic operations. This is a terminological point, but the paper should clarify which components are genuinely linguistic and which are generic extractive heuristics.
  2. [§5.1, first paragraph] The sentence 'when only one token in six is retained' should say 'word' rather than 'token,' since the compression ratio is defined on word counts. The distinction matters because the ratio axis is word-based.
  3. [Appendix G] Typo: 'evoluntionary search' should be 'evolutionary search.' Also, the cost statement 'totaling about 26 hours and US$50 around' is awkward; recommend 'about 26 hours and US$50.'
  4. [Table 3 and Appendix E] The 'Best baseline' column excludes R2C from transfer experiments and PartPrompt from OOD QA. Please state this explicitly in the Table 3 caption or main text so the reader knows the comparison set is not the full baseline set.
  5. [Appendix A.2] The human validation reports 83% agreement and Fleiss' κ=0.76. This is moderate agreement and should be mentioned in the main text limitations, since the open-ended QA results depend on the judge.

Circularity Check

0 steps flagged

No significant circularity: the central comparison rests on held-out evaluation rather than definitional reduction.

full rationale

The paper's derivation chain is an empirical search-and-evaluate pipeline, not a formal derivation that reduces a prediction to its inputs. Forty-two linguistic seeds are evolved under fitness f = min(D, R) on 100-example training splits per dataset, and the selected compressors are then evaluated on disjoint 200-example test splits, plus OOD settings without re-evolution. No equation makes the held-out test accuracy equal to the fitness function, a fitted constant, or an initial seed; the train/test split is explicit and the test sets are separate. The acknowledged asymmetry—100 training samples from each of the same four benchmark families used at test time, while baselines such as LLMLingua-2 are not given the same per-benchmark optimization budget—is a real fairness and generalization threat, and the paper states it candidly in Limitations. That threat affects external validity, not circularity: the reported numbers are produced by genuine evaluation on disjoint examples rather than by construction. The OOD experiments reduce but do not remove the asymmetry, so the parity claim should be read as benchmark-family-tuned; this is a correctness caveat, not a circular step. There are no self-citations used as load-bearing evidence, and no uniqueness theorem is imported from the authors' prior work. The only by-construction element is the compressor's near-zero compression-ratio deviation, which the paper itself explains as arising from an explicit word budget enforced directly; that is an implementation property, not the central QA-parity claim, and it does not make the main result circular.

Axiom & Free-Parameter Ledger

10 free parameters · 10 axioms · 0 invented entities

The central claim rests on a substantial number of design choices: 42 handcrafted seeds, evolutionary-search hyperparameters, screening thresholds, evolved program weights, and target ratios. These are fitted to training data rather than derived from theory. No new physical or conceptual entities are postulated; the only 'new' artifact is the evolved Python compressor itself. The most consequential unstated assumptions are the reliability of spaCy annotations, the validity of the judge, and the fairness of the per-benchmark optimization budget.

free parameters (10)
  • Handcrafted seed population (42 seeds) = 42 rule programs across 4 linguistic levels + composites
    The choice and content of the initial seeds bias the search space; several seeds are intentionally overlapping (Appendix B), and the search only explores combinations reachable from these seeds.
  • Evolutionary search configuration = 6 islands, 6×8 MAP-Elites grid, 15% migration, 30 generations
    These hyperparameters are chosen by the authors, not derived; the Limitations note that island count, grid, migration rate, and mutator schedule are not ablated.
  • Cascade screening thresholds = Stage 2 threshold = α·(avg_no_compression_accuracy / target_ratio), α<1; Stage 3 = avg_no_compression_accuracy / target_
    The threshold formulas and the n-gram overlap admissibility constraint (from the mutator prompt, Appendix L.1) materially determine which candidates survive evolution and are not derived from theory.
  • Target compression ratios and acceptance windows = 1.5×, 2.5×, 4×, 6× with windows ±0.25, ±0.5, ±0.75, ±1.0
    The specific ratios and window widths are experimental choices that define the evaluation regime and the archive cells.
  • Token scoring weights in the 1.5× compressor = POS: NUM=88, PROPN=72, NOUN=60, VERB=48/40; dependency roles=55; negation=95; sentence bonuses 18–22; entity-chain stren
    These numeric weights appear in the evolved program pseudocode (Appendix H.1) and were produced by evolutionary search on training data.
  • Tiered-fill hybrid truncation parameters (2.5×) = Tier 1 ROOT subtrees; Tier 2 entities/numbers/negation/cues; Tier 3 noun-chunk heads; top-half priority + even spacing
    The tier priorities and hybrid truncation split are evolved choices, not derived from a model of information content.
  • Sentence-scoring weights in the 4× cascade = entities 2.5, numbers 2.0, negation 1.5, connectives/temporal/ROOT 1.0
    Weights in the sentence-then-token cascade (Appendix H.3) were tuned by the evolutionary search.
  • Sentence-selection cues and subtree-deletion rules (6×) = first sentence, WH words, '?' and ':' cues; delete advcl/ccomp/xcomp/relcl/acl/appos subtrees; drop det/aux/cop/mark/cas
    These categorical keep/drop decisions are evolved, hand-discovered program logic rather than derived from theory.
  • Dual-path fitness min(D,R) and judge majority = f = min(D, R); 3 judge calls with majority vote
    The conservative min objective and judge repetition count are design choices that shape which compressors survive.
  • Train/test split sizes and random seed = 100 train / 200 test per dataset; seed 42
    The sample sizes are constrained by LongMemEval availability and the split seed affects all downstream comparisons.
axioms (10)
  • domain assumption spaCy en_core_web_sm annotations (POS, dependencies, NER, noun chunks) are accurate enough for the target English prose.
    The entire compressor pipeline selects content based on these annotations; parser errors would change compression choices and final QA accuracy. The Limitations acknowledge parser dependence.
  • domain assumption Linguistic cues such as content/function word distinctions, argument structure, negation, quantifiers, and discourse connectives are informative for question answering.
    This is the motivating hypothesis of the paper, grounded in cited linguistic literature in §2.2, but it is not independently proven for this task.
  • domain assumption Receiver LLM accuracy on compressed or reconstructed text measures information preservation in a way that generalizes across receivers.
    The dual-path protocol uses three small LLMs and a judge; the results could differ with larger or differently trained receivers, as the Limitations acknowledge.
  • domain assumption GPT-5.4-mini as LLM-as-judge with majority voting is a valid proxy for correctness of open-ended answers.
    Human validation found 83% agreement and Fleiss' κ=0.76, but this does not eliminate judge-model or prompt sensitivity (Appendix A.2).
  • domain assumption The 200-example test subsets are representative of RACE, Qasper, Multi-doc QA, and LongMemEval.
    The paper uses balanced subsets rather than full benchmarks; the Limitations state that generality is not established and that LongMemEval case availability constrained the split.
  • domain assumption LLM mutators can reliably produce valid SEARCH/REPLACE edits and follow the hard constraints in the system prompt.
    The search depends on three proprietary LLM tiers behaving as instructed; syntax errors are filtered at Stage 1, but semantic adherence to constraints is not guaranteed.
  • standard math The paired-bootstrap and Benjamini–Hochberg procedures are appropriate for the significance analysis.
    Standard statistical tools; the paper uses 10,000 resamples and fixed seed 42 in Appendix F.
  • ad hoc to paper Enforcing target compression ratio by truncating or back-filling to ⌊n_src/r⌉ words does not distort the comparison.
    All evolved seeds share a ratio-enforcement wrapper, but baselines use their own budgets and sometimes drift from the target (Figure 4). The comparability of '1.5×' across methods is assumed.
  • ad hoc to paper OOD settings (MuSiQue/NarrativeQA, summarization, classification) are valid transfer tests without re-evolution.
    Appendix E uses these tasks to claim transferability, but they are not part of the fitness search and coverage/exclusions differ across methods.
  • ad hoc to paper The AlphaEvolve/OpenEvolve adaptation remains effective in this multimodal linguistic-rule space.
    The paper adapts AlphaEvolve without ablating island count, migration rate, or mutator schedule; the Limitations explicitly say this trade-off is empirical rather than settled.

pith-pipeline@v1.3.0-alltime-deepseek · 41791 in / 11594 out tokens · 120106 ms · 2026-08-01T02:43:01.271523+00:00 · methodology

0 comments
read the original abstract

Prompt compression shortens LLM input to reduce inference cost, yet existing methods score token importance through LM forward passes. It remains questionable whether such nuanced, costly token selection is necessary. Compression requires identifying informative content, a problem that linguistic research has long addressed through cues that can be operationalized as deterministic rules. We therefore ask: can \textbf{linguistic rules alone} serve as effective prompt compressors, without LM-based scoring at compression time? To address this, we conduct offline evolutionary search over lexical, syntactic, semantic, and discourse seeds to find competitive rule combinations. The resulting linguistic compressor requires no LM forward pass at deployment and uses only CPU-side processing for compression. We evaluate it with a dual-path protocol to balance compression quality and reconstruction fidelity. Across short passages, multi-document reasoning, and dialogue-memory QA datasets, evolved compressors achieve performance similar to that of recent advanced prompt-compression strategies. Performance is strongest under light-to-moderate compression and degrades as compression becomes more aggressive, while the Direct and Reconstruction paths exhibit distinct patterns. Evolutionary analysis reveals that effective compression fuses signals across linguistic levels and, as the compression ratio increases, rules shift from token pruning to sentence extraction.

Figures

Figures reproduced from arXiv: 2607.25335 by Emmanuele Chersoni, Jianfei Ma, Si Chen, Zhaoxin Feng.

Figure 1
Figure 1. Figure 1: Illustration of linguistic rule-based prompt [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our evolutionary framework for discovering a linguistic prompt compressor. An LLM mutator [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cross-model stability measured by the coefficient of variation (CV, %) of Direct accuracy across three [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Deviation of the actual compression ratio [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Survival of linguistic rules in the best evolved [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Memory usage and latency across compres [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

85 extracted references · 8 canonical work pages

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    Prompt Compression for Large Language Models: A Survey

    Li, Zongqian and Liu, Yinhong and Su, Yixuan and Collier, Nigel. Prompt Compression for Large Language Models: A Survey. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.naacl-long.368

    doi:10.18653/v1/2025.naacl-long.368 2025

  9. [9]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Learning to Compress Prompts with Gist Tokens , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  10. [10]

    Adapting Language Models to Compress Contexts

    Chevalier, Alexis and Wettig, Alexander and Ajith, Anirudh and Chen, Danqi. Adapting Language Models to Compress Contexts. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.232

    doi:10.18653/v1/2023.emnlp-main.232 2023

  11. [11]

    The Twelfth International Conference on Learning Representations , year=

    In-context Autoencoder for Context Compression in a Large Language Model , author=. The Twelfth International Conference on Learning Representations , year=

  12. [12]

    500x C ompressor: Generalized Prompt Compression for Large Language Models

    Li, Zongqian and Su, Yixuan and Collier, Nigel. 500x C ompressor: Generalized Prompt Compression for Large Language Models. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.1219

    doi:10.18653/v1/2025.acl-long.1219 2025

  13. [13]

    Xin Cheng and Xun Wang and Xingxing Zhang and Tao Ge and Si-Qing Chen and Furu Wei and Huishuai Zhang and Dongyan Zhao , booktitle=. x. 2024 , url=

    2024

  14. [14]

    Compressing Context to Enhance Inference Efficiency of Large Language Models

    Li, Yucheng and Dong, Bo and Guerin, Frank and Lin, Chenghua. Compressing Context to Enhance Inference Efficiency of Large Language Models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.391

    doi:10.18653/v1/2023.emnlp-main.391 2023

  15. [15]

    LLML ingua: Compressing Prompts for Accelerated Inference of Large Language Models

    Jiang, Huiqiang and Wu, Qianhui and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili. LLML ingua: Compressing Prompts for Accelerated Inference of Large Language Models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.825

    doi:10.18653/v1/2023.emnlp-main.825 2023

  16. [16]

    L ong LLML ingua: Accelerating and Enhancing LLM s in Long Context Scenarios via Prompt Compression

    Jiang, Huiqiang and Wu, Qianhui and Luo, Xufang and Li, Dongsheng and Lin, Chin-Yew and Yang, Yuqing and Qiu, Lili. L ong LLML ingua: Accelerating and Enhancing LLM s in Long Context Scenarios via Prompt Compression. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024....

    doi:10.18653/v1/2024.acl-long.91 2024

  17. [17]

    LLML ingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression

    Pan, Zhuoshi and Wu, Qianhui and Jiang, Huiqiang and Xia, Menglin and Luo, Xufang and Zhang, Jue and Lin, Qingwei and R. LLML ingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.57

    doi:10.18653/v1/2024.findings-acl.57 2024

  18. [18]

    Discrete prompt compression with reinforcement learning

    Jung, Hoyoun and Kim, Kyung-Joong. Discrete prompt compression with reinforcement learning. IEEE Access

  19. [19]

    Learning to Compress Prompt in Natural Language Formats

    Chuang, Yu-Neng and Xing, Tianwei and Chang, Chia-Yuan and Liu, Zirui and Chen, Xun and Hu, Xia. Learning to Compress Prompt in Natural Language Formats. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.naacl-long.429

    doi:10.18653/v1/2024.naacl-long.429 2024

  20. [20]

    2024 , url=

    Fangyuan Xu and Weijia Shi and Eunsol Choi , booktitle=. 2024 , url=

    2024

  21. [21]

    2025 , eprint=

    FrugalPrompt: Reducing Contextual Overhead in Large Language Models via Token Attribution , author=. 2025 , eprint=. doi:10.48550/arXiv.2510.16439 , url=

    work page doi:10.48550/arxiv.2510.16439 2025

  22. [22]

    Parse trees guided LLM prompt compression

    Mao, Wenhao and Hou, Chengbin and Zhang, Tianyu and Lin, Xinyu and Tang, Ke and Lv, Hairong. Parse trees guided LLM prompt compression. IEEE Trans. Pattern Anal. Mach. Intell

  23. [23]

    Selection-p: Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability

    Chung, Tsz Ting and Cui, Leyang and Liu, Lemao and Huang, Xinting and Shi, Shuming and Yeung, Dit-Yan. Selection-p: Self-Supervised Task-Agnostic Prompt Compression for Faithfulness and Transferability. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.646

    doi:10.18653/v1/2024.findings-emnlp.646 2024

  24. [24]

    2002 , publisher=

    Syntactic structures , author=. 2002 , publisher=

    2002

  25. [25]

    2015 , publisher=

    Elements of structural syntax , author=. 2015 , publisher=

    2015

  26. [26]

    and Nivre, Joakim and Zeman, Daniel

    de Marneffe, Marie-Catherine and Manning, Christopher D. and Nivre, Joakim and Zeman, Daniel. U niversal D ependencies. Computational Linguistics. 2021. doi:10.1162/coli_a_00402

    doi:10.1162/coli_a_00402 2021

  27. [27]

    Rhetorical Structure Theory: Toward a functional theory of text organization

    Mann, William C and Thompson, Sandra A. Rhetorical Structure Theory: Toward a functional theory of text organization. Text - Interdiscip. J. Study Discourse

  28. [28]

    IBM Journal of research and development , volume=

    The automatic creation of literature abstracts , author=. IBM Journal of research and development , volume=. 1958 , publisher=

    1958

  29. [29]

    Journal of the ACM (JACM) , volume=

    New methods in automatic extracting , author=. Journal of the ACM (JACM) , volume=. 1969 , publisher=

    1969

  30. [30]

    Summarization beyond sentence extraction: A probabilistic approach to sentence compression , journal =

    Kevin Knight and Daniel Marcu , keywords =. Summarization beyond sentence extraction: A probabilistic approach to sentence compression , journal =. 2002 , issn =. doi:https://doi.org/10.1016/S0004-3702(02)00222-9 , url =

    work page doi:10.1016/s0004-3702(02)00222-9 2002

  31. [31]

    Dependency Tree Based Sentence Compression

    Filippova, Katja and Strube, Michael. Dependency Tree Based Sentence Compression. Proceedings of the Fifth International Natural Language Generation Conference. 2008

    2008

  32. [32]

    The Bell system technical journal , volume=

    A mathematical theory of communication , author=. The Bell system technical journal , volume=. 1948 , publisher=

    1948

  33. [33]

    2016 , publisher=

    Human behavior and the principle of least effort: An introduction to human ecology , author=. 2016 , publisher=

    2016

  34. [34]

    , title =

    Etzioni, Oren and Banko, Michele and Soderland, Stephen and Weld, Daniel S. , title =. Commun. ACM , month = dec, pages =. 2008 , issue_date =. doi:10.1145/1409360.1409378 , abstract =

    arXiv 2008

  35. [35]

    The P enn D iscourse T ree B ank 2.0

    Prasad, Rashmi and Dinesh, Nikhil and Lee, Alan and Miltsakaki, Eleni and Robaldo, Livio and Joshi, Aravind and Webber, Bonnie. The P enn D iscourse T ree B ank 2.0. Proceedings of the Sixth International Conference on Language Resources and Evaluation ( LREC '08). 2008

    2008

  36. [36]

    2014 , publisher=

    Cohesion in english , author=. 2014 , publisher=

    2014

  37. [37]

    The Twelfth International Conference on Learning Representations , year=

    Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers , author=. The Twelfth International Conference on Learning Representations , year=

  38. [38]

    The Twelfth International Conference on Learning Representations , year=

    Large Language Models as Optimizers , author=. The Twelfth International Conference on Learning Representations , year=

  39. [39]

    Promptbreeder: self-referential self-improvement via prompt evolution , year =

    Fernando, Chrisantha and Banarse, Dylan and Michalewski, Henryk and Osindero, Simon and Rockt\". Promptbreeder: self-referential self-improvement via prompt evolution , year =. Proceedings of the 41st International Conference on Machine Learning , articleno =

  40. [40]

    2025 , eprint=

    AlphaEvolve: A coding agent for scientific and algorithmic discovery , author=. 2025 , eprint=

    2025

  41. [41]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume=

  42. [42]

    and Lin, Kevin and Hewitt, John and Paranjape, Ashwin and Bevilacqua, Michele and Petroni, Fabio and Liang, Percy

    Liu, Nelson F. and Lin, Kevin and Hewitt, John and Paranjape, Ashwin and Bevilacqua, Michele and Petroni, Fabio and Liang, Percy. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics. 2024. doi:10.1162/tacl_a_00638

    doi:10.1162/tacl_a_00638 2024

  43. [43]

    Large language models can be easily distracted by irrelevant context , year =

    Shi, Freda and Chen, Xinyun and Misra, Kanishka and Scales, Nathan and Dohan, David and Chi, Ed and Sch\". Large language models can be easily distracted by irrelevant context , year =. Proceedings of the 40th International Conference on Machine Learning , articleno =

  44. [44]

    2023 , url=

    Guohao Li and Hasan Abed Al Kader Hammoud and Hani Itani and Dmitrii Khizbullin and Bernard Ghanem , booktitle=. 2023 , url=

    2023

  45. [45]

    The Thirteenth International Conference on Learning Representations , year=

    LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory , author=. The Thirteenth International Conference on Learning Representations , year=

  46. [46]

    1994 , publisher=

    Information structure and sentence form: Topic, focus, and the mental representations of discourse referents , author=. 1994 , publisher=

    1994

  47. [47]

    2025 , publisher =

    OpenEvolve: an open-source evolutionary coding agent , author =. 2025 , publisher =

    2025

  48. [48]

    2015 , eprint=

    Illuminating search spaces by mapping elites , author=. 2015 , eprint=

    2015

  49. [49]

    Journal of computing and information technology , volume=

    The island model genetic algorithm: On separability, population size and convergence , author=. Journal of computing and information technology , volume=. 1999 , publisher=

    1999

  50. [50]

    1989 , publisher=

    A natural history of negation , author=. 1989 , publisher=

    1989

  51. [51]

    Linguistics and Philosophy , volume=

    Generalized Quantifiers and Natural Language , author=. Linguistics and Philosophy , volume=. 1981 , publisher=. doi:10.1007/BF00350139 , url=

    work page doi:10.1007/bf00350139 1981

  52. [52]

    2020 , doi =

    Honnibal, Matthew and Montani, Ines and Van Landeghem, Sofie and Boyd, Adriane , title =. 2020 , doi =

    2020

  53. [53]

    RACE : Large-scale R e A ding Comprehension Dataset From Examinations

    Lai, Guokun and Xie, Qizhe and Liu, Hanxiao and Yang, Yiming and Hovy, Eduard. RACE : Large-scale R e A ding Comprehension Dataset From Examinations. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. 2017. doi:10.18653/v1/D17-1082

    doi:10.18653/v1/d17-1082 2017

  54. [54]

    and Gardner, Matt

    Dasigi, Pradeep and Lo, Kyle and Beltagy, Iz and Cohan, Arman and Smith, Noah A. and Gardner, Matt. A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021. doi:10.18653/v1/2021.na...

    doi:10.18653/v1/2021.naacl-main.365 2021

  55. [55]

    H otpot QA : A Dataset for Diverse, Explainable Multi-hop Question Answering

    Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William and Salakhutdinov, Ruslan and Manning, Christopher D. H otpot QA : A Dataset for Diverse, Explainable Multi-hop Question Answering. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018. doi:10.18653/v1/D18-1259

    doi:10.18653/v1/d18-1259 2018

  56. [56]

    Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps

    Ho, Xanh and Duong Nguyen, Anh-Khoa and Sugawara, Saku and Aizawa, Akiko. Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps. Proceedings of the 28th International Conference on Computational Linguistics. 2020. doi:10.18653/v1/2020.coling-main.580

    doi:10.18653/v1/2020.coling-main.580 2020

  57. [57]

    L ong B ench: A Bilingual, Multitask Benchmark for Long Context Understanding

    Bai, Yushi and Lv, Xin and Zhang, Jiajie and Lyu, Hongchang and Tang, Jiankai and Huang, Zhidian and Du, Zhengxiao and Liu, Xiao and Zeng, Aohan and Hou, Lei and Dong, Yuxiao and Tang, Jie and Li, Juanzi. L ong B ench: A Bilingual, Multitask Benchmark for Long Context Understanding. Proceedings of the 62nd Annual Meeting of the Association for Computation...

    doi:10.18653/v1/2024.acl-long.172 2024

  58. [58]

    NLTK : The Natural Language Toolkit

    Bird, Steven and Loper, Edward. NLTK : The Natural Language Toolkit. Proceedings of the ACL Interactive Poster and Demonstration Sessions. 2004

    2004

  59. [59]

    W ord N et: A Lexical Database for E nglish

    Miller, George A. W ord N et: A Lexical Database for E nglish. Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, F ebruary 23-26, 1992. 1992

    1992

  60. [60]

    2019 , institution=

    Language Models are Unsupervised Multitask Learners , author=. 2019 , institution=

    2019

  61. [61]

    2023 , eprint=

    Llama 2: Open Foundation and Fine-Tuned Chat Models , author=. 2023 , eprint=

    2023

  62. [62]

    Unsupervised Cross-lingual Representation Learning at Scale

    Conneau, Alexis and Khandelwal, Kartikay and Goyal, Naman and Chaudhary, Vishrav and Wenzek, Guillaume and Guzm \'a n, Francisco and Grave, Edouard and Ott, Myle and Zettlemoyer, Luke and Stoyanov, Veselin. Unsupervised Cross-lingual Representation Learning at Scale. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. ...

    doi:10.18653/v1/2020.acl-main.747 2020

  63. [63]

    2024 , eprint=

    GPT-4 Technical Report , author=. 2024 , eprint=

    2024

  64. [64]

    G lob E nc: Quantifying Global Token Attribution by Incorporating the Whole Encoder Layer in Transformers

    Modarressi, Ali and Fayyaz, Mohsen and Yaghoobzadeh, Yadollah and Pilehvar, Mohammad Taher. G lob E nc: Quantifying Global Token Attribution by Incorporating the Whole Encoder Layer in Transformers. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.1...

    doi:10.18653/v1/2022.naacl-main.19 2022

  65. [65]

    2020 , eprint=

    ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators , author=. 2020 , eprint=

    2020

  66. [66]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  67. [67]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  68. [68]

    2025 , eprint=

    Gemma 3 Technical Report , author=. 2025 , eprint=

    2025

  69. [69]

    2026 , note =

    OpenAI , title =. 2026 , note =

    2026

  70. [70]

    2026 , note =

    DeepSeek , title =. 2026 , note =

    2026

  71. [71]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Fundamental Limits of Prompt Compression: A Rate-Distortion Framework for Black-Box Language Models , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  72. [72]

    Prompt Compression based on Key-Information Density , journal =

    Yuhao Lin and Wenya Guo and Ying Zhang and Chengyi Yang and Zengxiang Li , keywords =. Prompt Compression based on Key-Information Density , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.eswa.2025.127738 , url =

    arXiv 2025

  73. [73]

    Understanding and Improving Information Preservation in Prompt Compression for LLM s

    ajewska, Weronika and Hardalov, Momchil and Aina, Laura and Anna John, Neha and Su, Hang and Marquez, Lluis. Understanding and Improving Information Preservation in Prompt Compression for LLM s. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.949

    work page doi:10.18653/v1/2025.findings-emnlp.949 2025

  74. [74]

    , title =

    Liskavets, Barys and Ushakov, Maxim and Roy, Shuvendu and Klibanov, Mark and Etemad, Ali and Luke, Shane K. , title =. Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence , ar...

    doi:10.1609/aaai.v39i23.34639 2025

  75. [75]

    ICLR 2025 Workshop on Building Trust in Language Models and Applications , year=

    An Empirical Study on Prompt Compression for Large Language Models , author=. ICLR 2025 Workshop on Building Trust in Language Models and Applications , year=

    2025

  76. [76]

    Prompt Compression in the Wild: Measuring Latency, Rate Adherence, and Quality for Faster LLM Inference , year =

    Kummer, Cornelius and Jurkschat, Lena and F\". Prompt Compression in the Wild: Measuring Latency, Rate Adherence, and Quality for Faster LLM Inference , year =. Advances in Information Retrieval: 48th European Conference on Information Retrieval, ECIR 2026, Delft, The Netherlands, March 29--April 2, 2026, Proceedings, Part I , pages =. doi:10.1007/978-3-0...

    work page doi:10.1007/978-3-032-21289-4_17 2026

  77. [77]

    2026 , eprint=

    Telegraph English: Semantic Prompt Compression via Structured Symbolic Rewriting , author=. 2026 , eprint=

    2026

  78. [78]

    Cheng, Huangfei and Wu, Jiawei and Li, Tiantian and Cao, Bin and Fan, Jing , title =. PRICAI 2022: Trends in Artificial Intelligence: 19th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2022, Shanghai, China, November 10--13, 2022, Proceedings, Part II , pages =. 2022 , isbn =. doi:10.1007/978-3-031-20865-2_20 , abstract =

    work page doi:10.1007/978-3-031-20865-2_20 2022

  79. [79]

    Sentence Compression with Semantic Role Constraints

    Yoshikawa, Katsumasa and Iida, Ryu and Hirao, Tsutomu and Okumura, Manabu. Sentence Compression with Semantic Role Constraints. Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2012

    2012

  80. [80]

    From Reading to Compressing: Exploring the Multi-document Reader for Prompt Compression

    Choi, Eunseong and Lee, Sunkyung and Choi, Minjin and Park, Jun and Lee, Jongwuk. From Reading to Compressing: Exploring the Multi-document Reader for Prompt Compression. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.864

    work page doi:10.18653/v1/2024.findings-emnlp.864 2024

Showing first 80 references.