Pith. sign in

REVIEW 3 major objections 4 minor 35 references

SALT: Salience-Aware Lexical Trie for Long-Context Compression

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read SALT reframes long-context prompt compression as budget allocation across lexical themes, using a keyword trie ordered by sentence frequency to keep rare themes from being squeezed out by dominant ones under tight budgets.

desk verdict Solid, well-engineered prompt compressor with a real gap: the SF quantile pruning drops the minority themes the trie allocation is supposed to protect. read the letter →

arxiv 2607.17486 v1 pith:2GZ2IA3K submitted 2026-07-20 cs.PF cs.AIcs.LG

Oteo Mamo , Hyunjin Yi , Joydhriti Choudhury , Shangqian Gao , Weikuan Yu This is my paper
classification cs.PFcs.AIcs.LG
keywords long-contextinferencepromptcompressionthemecollapselexicaltriesentencefrequencybudgetallocationmulti-turndialogueKVcache
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 claims that the usual way to compress long prompts before feeding them to an LLM — score each sentence, keep the top-scoring ones up to a budget — fails silently when the budget is tight: the document's dominant theme consumes the whole budget and rarer, task-relevant themes disappear. SALT instead builds a lightweight trie — a tree whose shared prefixes merge repeated keyword sequences — of the document's recurring keywords, ordered by how many sentences each keyword anchors, and allocates the budget across the trie's theme branches before any sentence is picked. This makes thematic coverage the primary selection objective rather than a post-hoc diversity fix. If the claim holds, then cheap, model-agnostic preprocessing can preserve minority themes, cut prefill computation and memory, and compose with KV-cache methods that optimize decoding — which is what the paper's experiments report at a 20% budget.

What carries the argument

The SF-ordered keyword trie: internal nodes are labeled by salient keywords, leaves store sentence identifiers, and each sentence becomes a root-to-leaf path sorted by decreasing sentence frequency. Shared prefixes capture recurring keyword co-occurrences — the paper's proxy for thematic structure — and selection measures 'uncovered mass' per subtree as the sum of normalized sentence frequencies of keywords not yet covered by chosen sentences. This trie does two jobs: it is the allocation unit (budget is distributed across depth-1 branches before any sentence is scored, with a floor plus a sublinear share of residual mass to protect low-mass branches) and the retrieval index (in query mode,

What would settle it

Build a synthetic document with two themes: a dominant one occupying 90% of the sentences and a rare one occupying the remaining 10%, where each rare-theme keyword appears in only one sentence so all of them fall below the 0.9-quantile salience cutoff. At a 20% budget, ask a question whose answer lies in the rare theme. If SALT fails to retain any rare-theme sentence, the claim that the trie prevents theme collapse is falsified; running the same test with synonym-expressed themes probes the lexical-recurrence assumption.

Watch

Extended reading notes

Core claim

The central claim is that extractive prompt compression is not a one-dimensional ranking problem: under a fixed token budget, selection must first allocate the budget across the document's recurring lexical themes, then score sentences within each theme. SALT operationalizes this by extracting per-sentence keywords via [CLS] attention from a small contrastive encoder, computing a sentence frequency (SF) for each keyword, pruning to a salience set, and inserting each sentence as a root-to-leaf path in a trie sorted by decreasing SF. Selection then distributes the budget across depth-1 branches — a floor plus a sublinear share of residual mass — so low-mass branches keep capacity, and scores s

Load-bearing premise

The entire method rides on the assumption that sentence frequency of [CLS]-extracted keywords is a faithful proxy for a document's thematic structure; if themes are expressed with disjoint vocabularies or code-like syntax where keywords don't recur across sentences, the trie branches won't correspond to the real themes and the budget allocation cannot prevent theme collapse.

Editorial extensions

If this is right

  • Preprocessing-based compression can prevent theme collapse with only lightweight lexical statistics; no expensive semantic clustering or per-query re-encoding is needed.
  • A single per-document trie serves many queries and budgets, so multi-turn dialogue over a long document pays a one-time indexing cost and roughly 11 ms per turn thereafter, a reported ~4x end-to-end speedup over an uncompressed baseline.
  • Because SALT outputs plain text before prefill, it can be composed with KV-cache methods that reduce decoding-time cost, rather than competing with them.
  • At a 20% budget, SALT is the only preprocessing method evaluated that stays within a few points of full-context accuracy on the long-context benchmarks, suggesting scalar ranking is the main bottleneck for existing preprocessing approaches.
  • Trie growth is in width rather than depth: node count tracks sentence count near-linearly, while median path depth stays between 2.1 and 2.4 across 32k–256k contexts, so per-sentence traversal cost does not grow with context length.

Reading between the lines

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

  • If the theme-collapse diagnosis is right, it applies beyond prompt compression: any budgeted selection step that ranks by a scalar relevance score—RAG re-ranking, extractive summarization, even cache eviction—is vulnerable, and the allocate-first-then-select principle could be ported to those settings.
  • The method inherits its weakness from lexical recurrence: documents that express the same theme with synonyms, or with code identifiers that do not recur across lines, may scatter the theme across many low-SF keywords and drop out of the salience set before allocation begins; a synthetic synonym-heavy document would test this directly.
  • The 0.9-quantile salience pruning is a hard gate: if a task-relevant theme's keywords all fall below the threshold, no budget allocation can recover it. Making that threshold adaptive to the number of detected branches, or to the downstream budget, is a natural and testable extension.
  • The heavy lifting is done by the choice of keyword extractor; the paper's own comparison of two small encoders shows the attention distribution of [CLS] changes which keywords survive. This suggests SALT's accuracy could shift substantially with a different lightweight encoder, independent of the trie and allocation logic.
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.

Referee Report

3 major / 4 minor

Summary. The paper introduces SALT, a model-agnostic, extractive prompt-compression framework. SALT first indexes a document by extracting per-sentence keywords with [CLS]-attention from a small BERT encoder, computing sentence frequency (SF) over the keyword sets, and pruning to a salience set above the p-th quantile (default p=0.9). The retained keywords are organized into a trie whose branches correspond to recurring lexical themes. The compression budget is then allocated across trie branches before sentence selection, with sentences scored by their marginal reduction of uncovered keyword mass within a branch. The compressed output is a plain-text sentence subset in original document order. The paper claims that this trie-based allocation prevents 'theme collapse'—the overrepresentation of dominant themes and loss of rarer task-relevant themes—and evaluates SALT on LongBench, QuALITY, RULER NIAH, and PG19-based latency/memory profiling, comparing against preprocessing methods (EXIT, RECOMP, CPC, Sentinel) and KV-cache methods (SnapKV, FastKV, DuoAttention, SentenceKV). It reports that SALT outperforms prior preprocessing methods in accuracy, latency, and memory, matches KV-cache methods in efficiency at a modest accuracy cost, and supports multi-turn reuse without re-encoding the document.

Significance. The core idea—allocating a compression budget over lexical theme branches before scoring individual sentences—is a genuinely useful departure from the standard scalar-ranking formulation of extractive prompt compression. The paper is strong in several respects: the method is lightweight and model-agnostic; the index is reusable across queries and turns; the authors provide code; and the multi-turn evaluation on QuALITY shows a clear end-to-end speedup with near-baseline accuracy. The hardware-portability study across V100-to-B200 GPUs is also valuable. If the central claims hold, SALT would be a practical contribution to pre-prefill long-context compression. However, as detailed below, a load-bearing gap in the pruning mechanism undermines the theme-collapse claim, and the empirical comparisons lack the sensitivity and significance analysis needed to support the headline 'outperforms prior preprocessing methods' statement.

major comments (3)
  1. [§3.1 (SF quantile pruning, Eq. 2) and §3.3 (summary mode)] The central theme-collapse claim is undercut by the p=0.9 salience pruning. Section 3.1 keeps only keywords whose SF is above the p-th quantile and states that pruning 'does not determine which themes are ultimately represented.' But a rare theme expressed in one or two sentences will typically have all of its keywords below the p-th quantile. Such sentences have no trie path—Table 4 reports 8–11% 'no-anchor' sentences—and in summary mode they are eligible only in GlobalFill, after branch quotas are committed. Under tight budgets, no residual budget may remain. Thus the allocation mechanism cannot protect the exact 'less-frequent yet task-relevant' themes it was designed to preserve. Query mode partially mitigates this by reactivating pruned keywords that appear in the query (§3.3), but summary mode does not. The claim should be revised, or the pruning should be made theme-preserving (e.
  2. [§4.1 (Table 1) and §3.3 (α, β_q, p, rmax)] The main empirical claim that SALT 'outperforms prior preprocessing methods' rests on point estimates at a single 20% budget with no error bars or significance tests. In Table 1, the average lead over RECOMP is 3.3 points, but category-wise differences are within 0.3–1 point on Multi-Doc QA (41.42 vs. 41.09) and negative on Code (37.06 vs. 37.37). More importantly, the method has several tunable parameters—p, α, β_q, rmax—whose influence on the central theme-collapse mechanism is not analyzed. There is no ablation varying α or p, no budget sweep for accuracy, and no demonstration that the reported gains are robust to reasonable parameter changes. The conclusion that trie-based allocation is responsible for the gains is therefore not yet established.
  3. [§4.5 (Figure 6)] The NIAH evaluation compares SALT only to the uncompressed baseline, not to other preprocessing methods. The abstract and introduction claim SALT outperforms prior preprocessing methods; the NIAH result is presented as 'SALT preserves NIAH accuracy... matching the uncompressed baseline,' but without comparisons to RECOMP, EXIT, CPC, or Sentinel on the same benchmark, it is impossible to tell whether this is a distinctive property of SALT or common to preprocessing methods. The claim of superiority is thus incomplete for this evaluation target.
minor comments (4)
  1. [Eq. (2)] The definition of SF(w) should use cardinality bars explicitly: SF(w) = |{i : w ∈ K_i}|. The current notation in the text is ambiguous.
  2. [Algorithm 1] The line 'S_eff ← S ∪ (K_q ∩ ∪_i K_i)' is ambiguous; it should be written as S ∪ (K_q ∩ (∪_i K_i)).
  3. [§3.3 / Algorithm 1] The BranchAllocate procedure is referenced but not given a precise formula. The text describes a 'fixed floor plus a residual share proportional to (M_b + ε)^α' but no closed-form expression. Providing the exact equation would improve reproducibility.
  4. [Figure 6 caption] The caption says 'White cells indicate unavailable full-context runs whose final tokenized prompts exceed the model context window.' The sentence is redundant; consider simplifying.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; evaluation is external and parameters are not fit to the target benchmarks.

full rationale

SALT's derivation chain is not circular. The indexing representation (SF, salience set, trie) is built from the input document via a fixed, parameterized pipeline (Section 3.1-3.2); the selection objective (Eq. 4-5) and budget allocation are defined on that structure and are not fitted to the evaluation targets. Accuracy, latency, and memory are measured against external benchmarks (LongBench, QuALITY, RULER, PG19) with baselines under matched budgets (Section 4, Tables 1-2, Figures 3-6), so the claimed improvements are not true by construction. Self-citations are absent; the only cited priors (MMR, submodularity, Kneedle, AttentionRank, BGE) are external and are used as building blocks or comparisons, not as load-bearing justifications of SALT's own result. The p=0.9 SF quantile pruning can exclude rare-theme anchors and thereby weaken the theme-collapse claim (Section 3.1 vs. Table 4's no-anchor sentences), but that is a correctness/scope limitation about whether the proxy achieves the stated goal, not a circular derivation: the paper's equations do not reduce the reported outcomes to the method's inputs. Therefore no circular step is identified.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or mathematical entities are introduced; 'theme collapse' is a descriptive label, and the trie is a data structure, not an invented entity. The central burden is carried by the free parameters (p, rmax, alpha, beta_q, budget, priors) and domain assumptions about lexical theme approximation.

free parameters (6)
  • Salience quantile p = 0.9 (default)
    Retains keywords with SF above the p-th quantile; bounds trie size. No sensitivity analysis; hand-chosen.
  • Per-sentence keyword cap rmax = 0.4
    Caps keyword count at 40% of content words to prevent diffuse sentences; chosen, not swept.
  • Branch allocation exponent alpha = not reported (0<alpha<1)
    Controls sublinear compression of high-mass branches; value not given in paper; likely tuned.
  • Query anchor budget beta_q = not reported
    Controls relevance-coverage trade-off in query mode; not reported; likely tuned.
  • Compression budget B = 20% of context
    Set based on observed task-relevant span of 5-20%; no accuracy evaluation at other budgets.
  • Sentence-level priors L(n_i) and I(s_i) = not specified
    Used in Eq. (5) but never defined; must be guessed by re-implementers.
assumptions (5)
  • domain assumption [CLS] attention of a contrastively trained BERT encoder ranks word importance faithfully.
    BGE-small's [CLS] attention is used to rank keywords (Section B.1); no validation that this generalizes to all domains, especially code.
  • domain assumption Sentence frequency of selected keywords approximates document thematic structure.
    Central proxy (Section 3.1); fails for code and possibly for semantically cohesive but lexically varied text.
  • standard math Kneedle algorithm reliably identifies a good cutoff.
    Used from Satopaa et al.; treated as a black-box heuristic.
  • domain assumption Reconstructing compressed sentences in document order preserves task-relevant coherence.
    SALT always returns sentences in original order (Algorithm 1); this may not be optimal for all tasks.
  • domain assumption Full-context baseline and benchmark datasets are reliable indicators of real-world long-context performance.
    Evaluation relies on LongBench, QuALITY, RULER, PG19; these may not represent all long-context workloads.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SALT: Salience-Aware Lexical Trie for Long-Context Compression." pith.science (2026). https://pith.science/paper/2GZ2IA3K

@misc{pith2026260717486,
  author       = {Pith},
  title        = {Pith review of: SALT: Salience-Aware Lexical Trie for Long-Context Compression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2GZ2IA3K}},
  note         = {Machine review of arXiv:2607.17486}
}
read the original abstract

As large language models (LLMs) process increasingly longer prompts, computation and KV-cache memory costs have emerged as major bottlenecks in inference systems. Existing input-level prompt compression methods address this, but rank each sentence by a scalar relevance score, treating the document as an unstructured pool of words and sentences. Under tight budgets, this causes theme collapse, where the dominant theme(s) of a document consumes the budget, discarding less-frequent yet task-relevant themes. Preserving thematic coverage instead requires allocating the budget across recurring themes rather than scoring sentences in isolation. To this end, we propose SALT, a model-agnostic extractive framework that organizes per-sentence keywords into a trie ordered by sentence frequency (SF), a lightweight, reusable proxy for document thematic structure. This trie-based organization smooths memory allocation and prevents dominant themes from monopolizing the budget. Multi-anchor retrieval activates trie nodes labeled by query keywords at any depth, and the trie persists across dialogue turns, supporting multi-turn use without re-encoding the document. By preserving document themes, SALT reduces the prefill computation and memory cost of long-context prompts while remaining composable with KV-cache methods that target decoding-time latency and memory.

Figures

Figures reproduced from arXiv: 2607.17486 by the authors.

Figure 1
Figure 1. Wall-time for preprocessing methods (EXIT, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the SALT pipeline. (Satopaa et al., 2011) to the monotone envelope c¯t = max τ≤t cτ , t = 1, . . . , ni , (1) where ni is the number of content words in sen￾tence i. The running maximum suppresses local dips and yields a non-decreasing reconstruction curve on which the kneedle criterion is defined. The resulting knee ki determines how many words are retained, but we cap it at 40% of ni to prevent diffuse… view at source ↗
Figure 4
Figure 4. End-to-end latency and TPOT of SALT on Llama-3.1-8B-Instruct across different NVIDIA GPUs and raw-context lengths at a 20% retention budget [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: End-to-end efficiency of SALT against pre [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 5
Figure 5. Figure 5: QuALITY at 20% budget, 50 articles, 972 turns. (a) Per-turn accuracy is bounded for all methods. (b) Cumulative compute over a 19-turn conversation diverges sharply. where the same document is queried repeatedly across a conversation. We evaluate this regime on QuALITY…
Figure 7
Figure 7. Figure 7: (a) Cosine accumulation curve for a sample [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Trie scaling on the PG19 profiling inputs. (a) [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: End-to-end latency and TPOT of SALT on Llama-3.1-8B-Instruct across different NVIDIA GPUs and [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 1 canonical work pages

  1. [1]

    DuoAttention: Efficient Long-Context

    Guangxuan Xiao and Jiaming Tang and Jingwei Zuo and junxian guo and Shang Yang and Haotian Tang and Yao Fu and Song Han , booktitle=. DuoAttention: Efficient Long-Context. 2025 , url=

    2025

  2. [2]

    2026 , url=

    Adamas: Hadamard Sparse Attention for Efficient Long-context Inference , author=. 2026 , url=

    2026

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    H2o: Heavy-hitter oracle for efficient generative inference of large language models , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    Dao, Tri , booktitle=. Flash

  5. [5]

    Jacob Devlin and Ming-Wei Chang and Kenton Lee and Kristina Toutanova , title =. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , month =. 2019 , address =

    2019

  6. [6]

    arXiv preprint arXiv:2502.01068 , year=

    Fastkv: Kv cache compression for fast long-context processing with token-selective propagation , author=. arXiv preprint arXiv:2502.01068 , year=

    arXiv

  7. [7]

    and Hart, Peter E

    Cover, Thomas M. and Hart, Peter E. , title =. IEEE Transactions on Information Theory , volume =

  8. [8]

    arXiv preprint arXiv:2308.03281 , year=

    Towards general text embeddings with multi-stage contrastive learning , author=. arXiv preprint arXiv:2308.03281 , year=

    arXiv

Show all 35 references
  1. [9]

    Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval , pages=

    The use of MMR, diversity-based reranking for reordering documents and producing summaries , author=. Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval , pages=

  2. [10]

    Workshop on Efficient Systems for Foundation Models II @ ICML2024 , year=

    Characterizing Prompt Compression Methods for Long Context Inference , author=. Workshop on Efficient Systems for Foundation Models II @ ICML2024 , year=

  3. [11]

    Yang and Mohammad Mohammadi Amiri , booktitle=

    Yuxuan Zhu and Ali Falahati and David H. Yang and Mohammad Mohammadi Amiri , booktitle=. Sentence. 2025 , url=

    2025

  4. [12]

    2024 , url=

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

    2024

  5. [13]

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

    Hwang, Taeho and Cho, Sukmin and Jeong, Soyeong and Song, Hoyun and Han, SeungYoon and Park, Jong C. EXIT : Context-Aware Extractive Compression for Enhancing Retrieval-Augmented Generation. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.1865...

    2025 doi

  6. [14]

    , title =

    Liskavets, Barys and Ushakov, Maxim and Roy, Shuvendu and Klibanov, Mark and Etemad, Ali and Luke, Shane K. , title =. 2025 , isbn =. doi:10.1609/aaai.v39i23.34639 , articleno =

    2025 doi

  7. [15]

    arXiv preprint arXiv:2505.23277 , year=

    Sentinel: Attention Probing of Proxy Models for LLM Context Compression with an Understanding Perspective , author=. arXiv preprint arXiv:2505.23277 , year=

  8. [16]

    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/v1/2...

    2023 doi

  9. [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 doi

  10. [18]

    2023 , eprint=

    C-Pack: Packaged Resources To Advance General Chinese Embedding , author=. 2023 , eprint=

    2023

  11. [19]

    Finding a "Kneedle" in a Haystack: Detecting Knee Points in System Behavior , year=

    Satopaa, Ville and Albrecht, Jeannie and Irwin, David and Raghavan, Barath , booktitle=. Finding a "Kneedle" in a Haystack: Detecting Knee Points in System Behavior , year=

  12. [20]

    2023 , eprint=

    LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding , author=. 2023 , eprint=

    2023

  13. [21]

    2024 , eprint=

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

    2024

  14. [22]

    and Salakhutdinov, Ruslan and Manning, Christopher D

    Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W. and Salakhutdinov, Ruslan and Manning, Christopher D. , booktitle=

  15. [23]

    A ttention R ank: Unsupervised Keyphrase Extraction using Self and Cross Attention

    Ding, Haoran and Luo, Luo. A ttention R ank: Unsupervised Keyphrase Extraction using Self and Cross Attention. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2021

    2021

  16. [24]

    2020 , publisher=

    spaCy: Industrial-strength natural language processing in Python , author=. 2020 , publisher=

    2020

  17. [25]

    M u S i Q ue: Multihop Questions via Single-hop Question Composition

    Trivedi, Harsh and Balasubramanian, Niranjan and Khot, Tushar and Sabharwal, Ashish. M u S i Q ue: Multihop Questions via Single-hop Question Composition. Transactions of the Association for Computational Linguistics. 2022. doi:10.1162/tacl_a_00475

    2022 doi

  18. [26]

    Advances in Neural Information Processing Systems 32 , pages =

    PyTorch: An Imperative Style, High-Performance Deep Learning Library , author =. Advances in Neural Information Processing Systems 32 , pages =. 2019 , publisher =

    2019

  19. [27]

    The Twelfth International Conference on Learning Representations , year=

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning , author=. The Twelfth International Conference on Learning Representations , year=

  20. [28]

    Advances in Neural Information Processing Systems 37 (NeurIPS 2024) , year=

    SnapKV: LLM Knows What You Are Looking for before Generation , author=. Advances in Neural Information Processing Systems 37 (NeurIPS 2024) , year=

    2024

  21. [29]

    SIGIR '98: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval , year=

    The use of MMR, diversity-based reranking for reordering documents and producing summaries , author=. SIGIR '98: Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval , year=

  22. [30]

    arXiv preprint arXiv:2404.06654 , year=

    RULER: What's the Real Context Size of Your Long-Context Language Models? , author=. arXiv preprint arXiv:2404.06654 , year=

    arXiv

  23. [31]

    Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies , year=

    A Class of Submodular Functions for Document Summarization , author=. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies , year=

  24. [32]

    doi:10.5281/zenodo.12608602 , url =

    Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and...

    doi

  25. [33]

    arXiv preprint arXiv:1911.05507 , year=

    Compressive transformers for long-range sequence modelling , author=. arXiv preprint arXiv:1911.05507 , year=

    1911 arXiv

  26. [34]

    Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    QuALITY: Question answering with long input texts, yes! , author=. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2022

  27. [35]

    2024 , note =

    Ministral-8B-Instruct-2410 , author =. 2024 , note =

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.