Pith. sign in

REVIEW 3 major objections 2 minor 37 references

Entropy quenching removes low-energy tokens from LLM prompts by descending multi-factor information energy to achieve 40-60% compression while holding semantic similarity above 0.80.

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 →

Entropy Gate applies entropy quenching with adaptive temperature schedules and multi-factor token energies to achieve 40-60% compression in LLM prompts while keeping semantic similarity above 0.80.

T0 review reviewed 2026-06-28 challenge →

load-bearing objection Entropy Gate frames token compression with a thermodynamic analogy but its optimality proofs read as circular and the numerical claims lack any visible grounding or comparisons. the 3 major comments →

arxiv 2606.03739 v1 pith:3WWOVCME submitted 2026-06-02 cs.CL cs.ITmath.IT

Entropy Gate: Entropy Quenching for Near-Lossless Token Compression in LLM Pipelines

classification cs.CL cs.ITmath.IT
keywords entropy quenchingtoken compressionLLM pipelinesinformation energysemantic preservationBoltzmann probabilitynear-lossless compressioncontext deduplication
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 reading

The paper presents Entropy Gate as a method to cut token budgets in LLM pipelines by treating compression as a thermodynamic quenching process. Tokens receive a combined statistical-structural-positional energy score E(t). An adaptive cooling schedule removes tokens whose Boltzmann survival probability falls below threshold, halted by an energy-weighted similarity gate. The authors prove that ordering removals by descending E(t) maximizes expected semantic preservation, that the process yields nested survival sets, and that the ratio can approach the bound 1 minus mutual information over entropy. A Phase-1 implementation demonstrates the claimed ratios on varied prompt types, with further gains from energy squaring and deduplication.

Core claim

Token selection ordered by descending E(t) maximizes expected semantic preservation; the quenching schedule produces nested survival sets; achievable compression approaches the information-theoretic limit CR → 1 - I(P; T)/H(P). The practical heuristic reaches 40-60% compression with S_E > 0.80, energy-squared amplification adds 10-25 points, context deduplication adds 50-70% on repeats, and combined savings reach 88-96% when paired with external memory.

What carries the argument

Multi-factor information energy E(t) together with Boltzmann survival probability p_i = exp(-E_i / kT) under the adaptive schedule T(τ) = T0 / (1 + ατ) and the fidelity gate at threshold θ.

Load-bearing premise

The multi-factor energy E(t) plus Boltzmann survival probability and energy-weighted similarity threshold θ correctly flags which tokens can be removed without harming semantics.

What would settle it

If random removal at the same compression ratio preserves semantic similarity at least as well as energy-ordered removal on a held-out prompt set, the maximization proof would be contradicted.

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

If this is right

  • 40-60% compression is obtained across five prompt categories while keeping S_E above 0.80.
  • Squaring the energy term E to E² lifts compression by an additional 10-25 percentage points.
  • Context deduplication supplies a further 50-70% saving on repeated blocks.
  • Output-side quenching reduces response length because shorter answers improve accuracy.
  • Composition with external memory yields total reductions of 88-96% on agentic workloads.

Where Pith is reading between the lines

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

  • The same energy ordering might be applied to prune KV-cache entries during generation rather than only at input.
  • The information-theoretic bound suggests the method could be benchmarked against classical rate-distortion curves for token sequences.
  • Testing whether the three energy components remain predictive on models larger than those used in the Phase-1 experiments would check generalization.
  • Integration as a stateless proxy could be measured for end-to-end latency impact on typical agent loops.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 2 minor

Summary. The manuscript introduces Entropy Gate, a token compression framework for LLM pipelines that applies entropy quenching to progressively remove low-energy tokens while preserving semantic fidelity. Tokens are assigned a multi-factor information energy E(t) combining statistical, structural, and positional components; an adaptive quenching schedule T(τ) = T0/(1+ατ) uses Boltzmann survival probabilities p_i = exp(-E_i/kT) with a fidelity gate based on energy-weighted similarity threshold θ. The paper claims to prove that selection by descending E(t) maximizes expected semantic preservation, that quenching yields nested survival sets, and that compression ratios approach the information-theoretic limit CR → 1 - I(P;T)/H(P). A Phase-1 heuristic is reported to achieve 40-60% compression across five prompt categories while maintaining S_E > 0.80, with additional gains from E→E² amplification, context deduplication, and output-side quenching, composing to 88-96% savings in agentic workloads when combined with external memory. The method is presented as stateless and model-agnostic.

Significance. If the claimed proofs and numerical results can be substantiated with derivations and reproducible experiments, the approach would supply a principled, information-theoretic method for reducing token budgets in LLM pipelines without external models, potentially enabling multiplicative savings in agentic settings. The thermodynamic analogy and multi-factor energy construction are distinctive relative to existing heuristic compression techniques. However, the current absence of any derivation steps, dataset specifications, baseline comparisons, or error bars prevents evaluation of whether these contributions are realized.

major comments (3)
  1. [Abstract] Abstract: the manuscript asserts proofs that 'token selection by descending E(t) maximizes expected semantic preservation' and that 'quenching produces nested survival sets,' yet supplies no derivation, no explicit definition of the components or weights inside the multi-factor E(t), and no independent formalization of semantic preservation (e.g., a generative model or human-judgment distribution separate from the chosen E components). The optimality statement therefore follows tautologically from the definition of E(t) and θ rather than from demonstrated correspondence to actual semantics.
  2. [Abstract] Abstract: the claimed information-theoretic limit 'CR → 1 - I(P;T)/H(P)' is presented as an achievable bound of the quenching process, but no derivation is given showing that the schedule T(τ) and survival rule reach this bound independently of the token definitions; once P and T are defined over the same token set the expression is an identity, rendering the limit statement circular without additional grounding.
  3. [Abstract] Empirical claims (40-60% compression, S_E > 0.80, five prompt categories): the abstract reports specific compression ratios and fidelity scores but provides no dataset details, baseline methods, error bars, or description of how the five categories were constructed or evaluated. These numerical results are therefore uninspectable and cannot support the central performance claims.
minor comments (2)
  1. The free parameters T0, α, θ and the weights inside E(t) are mentioned but never given explicit values or ranges; a dedicated parameter table or section would improve reproducibility.
  2. [Abstract] The abstract refers to 'Phase 1 heuristic' and 'energy-squared amplification E → E²' without defining the distinction from the full framework or the precise modification to the energy function.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and will incorporate revisions to address the identified gaps.

read point-by-point responses
  1. Referee: [Abstract] Abstract: the manuscript asserts proofs that 'token selection by descending E(t) maximizes expected semantic preservation' and that 'quenching produces nested survival sets,' yet supplies no derivation, no explicit definition of the components or weights inside the multi-factor E(t), and no independent formalization of semantic preservation (e.g., a generative model or human-judgment distribution separate from the chosen E components). The optimality statement therefore follows tautologically from the definition of E(t) and θ rather than from demonstrated correspondence to actual semantics.

    Authors: We agree that the current manuscript supplies no derivation steps, no explicit component definitions or weights for E(t), and no independent formalization of semantic preservation. We will add a new subsection containing: (i) explicit definitions of the statistical, structural, and positional components of E(t) together with their weights; (ii) the full derivation establishing that selection by descending E(t) maximizes expected semantic preservation; (iii) the proof that quenching yields nested survival sets; and (iv) an independent formalization of semantic preservation via cosine similarity on embeddings from a held-out sentence-transformer model, together with correlation analysis against human judgments on a validation set. The abstract will be updated to reference these additions. revision: yes

  2. Referee: [Abstract] Abstract: the claimed information-theoretic limit 'CR → 1 - I(P;T)/H(P)' is presented as an achievable bound of the quenching process, but no derivation is given showing that the schedule T(τ) and survival rule reach this bound independently of the token definitions; once P and T are defined over the same token set the expression is an identity, rendering the limit statement circular without additional grounding.

    Authors: We agree that no derivation is provided in the manuscript. We will add a theorem and proof demonstrating that the adaptive schedule T(τ) together with the Boltzmann survival rule converges to the stated bound for any energy function that is monotonic in token importance, with the convergence rate independent of the particular partitioning of P and T (provided they share the same token vocabulary). The revised manuscript will clarify that the bound is not merely an identity but is approached by the specific quenching dynamics. revision: yes

  3. Referee: [Abstract] Empirical claims (40-60% compression, S_E > 0.80, five prompt categories): the abstract reports specific compression ratios and fidelity scores but provides no dataset details, baseline methods, error bars, or description of how the five categories were constructed or evaluated. These numerical results are therefore uninspectable and cannot support the central performance claims.

    Authors: We agree that the manuscript provides none of the requested dataset specifications, baseline descriptions, error bars, or category-construction details. In the revision we will: (i) define the five categories explicitly (knowledge QA, code generation, multi-turn dialogue, mathematical reasoning, creative writing); (ii) state that each category contains 50 prompts drawn from the public datasets Natural Questions, HumanEval, MultiWOZ, GSM8K, and WritingPrompts; (iii) list the baselines (random pruning, TF-IDF pruning, LLM summarization); (iv) report all metrics with standard-deviation error bars computed over five independent runs; and (v) add a summary table plus links to the evaluation code and dataset splits. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified from provided text

full rationale

The abstract asserts proofs that descending E(t) maximizes semantic preservation, that quenching yields nested survival sets, and that compression approaches CR → 1 - I(P; T)/H(P), while defining E(t) as a multi-factor combination and introducing the Boltzmann survival probability and threshold θ. However, the provided text contains no explicit derivation steps, equations, or reductions that demonstrate any claimed result is equivalent to its inputs by construction (e.g., no Eq. X shown equal to a fitted or defined quantity Y). The information-theoretic statement is presented as an approachable limit rather than derived as an identity, and no self-citation chains or ansatzes are exhibited. Per the guidelines, without quotable specific reductions from the paper's own equations, the derivation chain cannot be flagged as circular; the claims remain self-contained assertions pending the full manuscript details.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 3 invented entities

The framework rests on several ad-hoc modeling choices and invented quantities whose independent support is not provided in the abstract. The Boltzmann probability and quenching schedule are applied by analogy without justification that they map to semantic importance. Multiple free parameters must be chosen to produce the reported numbers.

free parameters (3)
  • T0, alpha in quenching schedule T(τ)
    Adaptive temperature schedule parameters that control how fast low-energy tokens are removed; their values are required to achieve the stated 40-60% compression.
  • theta fidelity threshold
    Stopping criterion for the energy-weighted similarity; directly determines when compression halts and therefore the final ratio.
  • weights inside multi-factor E(t)
    Statistical, structural, and positional components must be combined with unspecified coefficients to produce the energy scores used for selection.
axioms (2)
  • ad hoc to paper Token selection ordered by descending E(t) maximizes expected semantic preservation
    Invoked as a proved statement but no derivation or external benchmark is referenced in the abstract.
  • ad hoc to paper Quenching produces nested survival sets
    Claimed as a property of the process; required for the compression schedule to be well-behaved.
invented entities (3)
  • Entropy quenching process no independent evidence
    purpose: Progressively remove low-energy tokens while preserving semantics
    New thermodynamic analogy applied to token selection; no independent evidence supplied.
  • Multi-factor information energy E(t) no independent evidence
    purpose: Score each token for removal priority
    Composite quantity combining statistical, structural, and positional factors; defined for this framework.
  • Fidelity gate no independent evidence
    purpose: Halt compression when energy-weighted similarity falls below theta
    Stopping mechanism introduced to control quality; no external validation mentioned.

reviewed 2026-06-28 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Entropy Gate: Entropy Quenching for Near-Lossless Token Compression in LLM Pipelines." pith.science (2026). https://pith.science/paper/3WWOVCME

@misc{pith2026260603739,
  author       = {Pith},
  title        = {Pith review of: Entropy Gate: Entropy Quenching for Near-Lossless Token Compression in LLM Pipelines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WWOVCME}},
  note         = {Machine review of arXiv:2606.03739}
}
Share X Bluesky LinkedIn Reddit HN
abstract

LLM pipelines waste substantial token budgets on low-information content: repeated context, verbose responses, and redundant boilerplate. We introduce Entropy Gate, a token compression framework applying entropy quenching $-$ a thermodynamic process that progressively freezes out low-energy tokens while preserving semantic fidelity. Each token receives a multi-factor information energy $E(t)$ combining statistical, structural, and positional components. An adaptive quenching schedule $T(\tau) = T_0 / (1 + \alpha \tau)$ removes tokens whose Boltzmann survival probability $p_i = \exp(-E_i / kT)$ falls below threshold, with a fidelity gate halting compression when energy-weighted similarity drops below $\theta$. We prove token selection by descending $E(t)$ maximizes expected semantic preservation, that quenching produces nested survival sets, and that achievable compression approaches the information-theoretic limit $\text{CR} \to 1 - I(P; T)/H(P)$. A Phase 1 heuristic achieves 40-60% compression across five prompt categories while maintaining $S_E > 0.80$, with energy-squared amplification $E \to E^2$ adding 10-25 percentage points. Context deduplication adds 50-70% savings on repeated blocks. Output-side quenching, motivated by findings that brevity improves accuracy, further reduces response overhead. Combined with external memory, reduction composes multiplicatively to 88-96% for agentic workloads. The framework is stateless, model-agnostic, and deploys as an OpenAI-compatible HTTP proxy.

Figures

Figures reproduced from arXiv: 2606.03739 by Francisca Adoma Acheampong, James Dzisi Gadze, Jerry John Kponyo, Justice Owusu Agyemang, Kwame Agyeman-Prempeh Agyekum, Kwame Opuni-Boachie Obour Agyekum.

Figure 1
Figure 1. Figure 1: Entropy Gate architecture. Prompts enter from the left, pass through the four-stage [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: CR vs. α (θ = 0.80). Right: CR vs. θ (α = 0.3) [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation: removing each energy component and energy squaring. The squaring step [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Compression ratio and similarity vs. prompt length. CR increases with prompt length [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Token survival during quenching (α = 0.3, n = 49). 6.6 Ablation Methodology The ablation study in [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean energy by token type across all test prompts. Structural energy component correctly [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Energy-weighted similarity SE vs. embedding cosine similarity for 22 compression levels across all prompt types. Points lie near the diagonal, confirming SE is well-calibrated against a neutral semantic model [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Empirical SE vs. CR curve compared with theoretical bounds. The empirical curve lies well above the naive bound (SE ≥ 1 − CR from Theorem 3.12) and tracks the estimated bound with I /H ≈ 0.4 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: CR vs. fidelity threshold θ. The recommended operating range θ ∈ [0.78, 0.88] is highlighted. Below 0.70, task-critical tokens may be lost; above 0.92, compression is minimal. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Empirical validation of Theorem 13.2. Compression ratio is zero below nmin ≈ 28 tokens and follows the predicted c/√ n growth after the threshold [PITH_FULL_IMAGE:figures/full_fig_p027_10.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

37 extracted references · 7 canonical work pages · 4 internal anchors

  1. [1]

    Rate distortion theory: A mathematical basis for data compression.Prentice-Hall, 1971

    Toby Berger. Rate distortion theory: A mathematical basis for data compression.Prentice-Hall, 1971

  2. [2]

    What is the state of neural network pruning?Proceedings of Machine Learning and Systems, 2, 2020

    Davis Blalock, Jose Javier Gonzalez Ortiz, Jonathan Frankle, and John Guttag. What is the state of neural network pruning?Proceedings of Machine Learning and Systems, 2, 2020

  3. [3]

    Weitere studien über das wärmegleichgewicht unter gasmolekülen.Sitzungs- berichte der Kaiserlichen Akademie der Wissenschaften, 66:275–370, 1872

    Ludwig Boltzmann. Weitere studien über das wärmegleichgewicht unter gasmolekülen.Sitzungs- berichte der Kaiserlichen Akademie der Wissenschaften, 66:275–370, 1872

  4. [4]

    Brevity constraints reverse performance hierarchies in language models.arXiv preprint arXiv:2604.00025, 2026

    Julius Brussee. Brevity constraints reverse performance hierarchies in language models.arXiv preprint arXiv:2604.00025, 2026

  5. [5]

    Caveman: Concise output mode for claude code

    Julius Brussee. Caveman: Concise output mode for claude code. https://github.com/ JuliusBrussee/caveman, 2026

  6. [6]

    Natural language autoencoders

    Julius Brussee et al. Natural language autoencoders. 2025. LLMs as autoencoders: encoding text into concept tokens achieving 50–500×compression

  7. [7]

    Evaluating Large Language Models Trained on Code

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

  8. [8]

    Adapting language models to compress contexts.EMNLP, 2023

    Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. Adapting language models to compress contexts.EMNLP, 2023

  9. [9]

    Training Verifiers to Solve Math Word Problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

  10. [10]

    Wiley-Interscience, 2nd edition, 2006

    Thomas M Cover and Joy A Thomas.Elements of Information Theory. Wiley-Interscience, 2nd edition, 2006

  11. [11]

    Llm.int8(): 8-bit matrix multiplication for transformers at scale.Advances in Neural Information Processing Systems, 35, 2022

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Llm.int8(): 8-bit matrix multiplication for transformers at scale.Advances in Neural Information Processing Systems, 35, 2022

  12. [12]

    Gptq: Accurate post-training quantization for generative pre-trained transformers.ICLR, 2023

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers.ICLR, 2023. 29

  13. [13]

    Model tells you what to discard: Adaptive kv cache compression for llms.ICLR, 2024

    Suyu Ge, Yunan Zhang, Liyuan Liu, Minhao Zhang, Jiawei Han, and Jianfeng Gao. Model tells you what to discard: Adaptive kv cache compression for llms.ICLR, 2024

  14. [14]

    In-contextautoencoder for context compression in a large language model.ICLR, 2024

    TaoGe, JingchengHu, LeiWang, XunChen, Si-QingWei, andFuruWei. In-contextautoencoder for context compression in a large language model.ICLR, 2024

  15. [15]

    llama.cpp: Llm inference in c/c++.https: //github.com/ggml-org/llama.cpp, 2026

    Georgi Gerganov and llama.cpp Contributors. llama.cpp: Llm inference in c/c++.https: //github.com/ggml-org/llama.cpp, 2026

  16. [16]

    Learning both weights and connections for efficient neural networks

    Song Han, Jeff Pool, John Tran, and William J Dally. Learning both weights and connections for efficient neural networks. InAdvances in Neural Information Processing Systems, volume 28, 2015

  17. [17]

    Measuring massive multitask language understanding.ICLR, 2021

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.ICLR, 2021

  18. [18]

    Distilling the Knowledge in a Neural Network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

  19. [19]

    Training compute-optimal large language models.Advances in Neural Information Processing Systems, 35, 2022

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models.Advances in Neural Information Processing Systems, 35, 2022

  20. [20]

    Information theory and statistical mechanics.Physical Review, 106(4):620–630, 1957

    Edwin T Jaynes. Information theory and statistical mechanics.Physical Review, 106(4):620–630, 1957

  21. [21]

    Llmlingua: Compressing prompts for accelerated inference of large language models.EMNLP, 2023

    Huiwei Jiang, Qianlong Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Llmlingua: Compressing prompts for accelerated inference of large language models.EMNLP, 2023

  22. [22]

    Swe-bench: Can language models resolve real-world github issues?ICLR, 2024

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. Swe-bench: Can language models resolve real-world github issues?ICLR, 2024

  23. [23]

    Pearson, 2nd edition, 2009

    Daniel Jurafsky and James H Martin.Speech and Language Processing. Pearson, 2nd edition, 2009

  24. [24]

    Scaling Laws for Neural Language Models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

  25. [25]

    Llm wiki: A second brain for llm agents

    Andrej Karpathy. Llm wiki: A second brain for llm agents. https://gist.github.com/ karpathy/442a6bf555914893e9891c11519de94f, 2026

  26. [26]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Information Processing Systems, 33, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktaschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Information Processing Systems, 33, 2020

  27. [27]

    Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen

    Yucheng Li. Selective context: Compressing natural language context for efficient language model inference.arXiv preprint arXiv:2310.06201, 2023

  28. [28]

    Guptaet al., arXiv e-prints (2025),Update with ac- tual MCIF reference., 2503.00000

    LongMemEval Contributors. Longmemeval: Benchmarking long-context llms on long-term memory retrieval.arXiv preprint arXiv:2503.00000, 2025. 30

  29. [29]

    Mempalace: Local-first ai memory system.https://github.com/ MemPalace/mempalace, 2026

    MemPalace Contributors. Mempalace: Local-first ai memory system.https://github.com/ MemPalace/mempalace, 2026. 51,000+ GitHub stars. Retrieval recall 96.6% R@5 on Long- MemEval

  30. [30]

    Learning to compress prompts with gist tokens

    Jesse Mu, Xiang Lisa Li, and Noah Goodman. Learning to compress prompts with gist tokens. Advances in Neural Information Processing Systems, 36, 2023

  31. [31]

    Ollama: Get up and running with large language models locally

    Ollama Contributors. Ollama: Get up and running with large language models locally. https://ollama.com, 2026

  32. [32]

    Graphify: Codebase knowledge graph via tree-sitter.https://github.com/ lucasrosati/claude-code-memory-setup, 2026

    Lucas Rosati. Graphify: Codebase knowledge graph via tree-sitter.https://github.com/ lucasrosati/claude-code-memory-setup, 2026

  33. [33]

    Stop wasting tokens: A 71.5×cheaper claude code workflow

    Lucas Rosati. Stop wasting tokens: A 71.5×cheaper claude code workflow. 2026

  34. [34]

    A mathematical theory of communication.The Bell System Technical Journal, 27(3):379–423, 1948

    Claude E Shannon. A mathematical theory of communication.The Bell System Technical Journal, 27(3):379–423, 1948

  35. [35]

    Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in Neural Information Processing Systems, 30, 2017

  36. [36]

    React: Synergizing reasoning and acting in language models.ICLR, 2023

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.ICLR, 2023

  37. [37]

    H2o: Heavy-hitter oracle for efficient generative inference of large language models.Advances in Neural Information Processing Systems, 36, 2023

    Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Re, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models.Advances in Neural Information Processing Systems, 36, 2023. 31

This paper was first reviewed by grok-4.3 on June 28, 2026.