Cognitive Load-Aware Inference: A Neuro-Symbolic Framework for Optimizing the Token Economy of Large Language Models

Yilun Zhang

REVIEW 4 major objections 5 minor 1 cited by

Cognitive Load-Aware Inference: A Neuro-Symbolic Framework for Optimizing the Token Economy of Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that framing LLM inference as cognitive load management—measuring intrinsic difficulty, removing extraneous computation, and spending saved tokens on productive reasoning—cuts token consumption by up to 45% without…

desk verdict Token savings look real but the cognitive-load attribution is untested; the paper is a useful prompt recipe that overclaims its theoretical basis. read the letter →

arxiv 2507.00653 v1 pith:SQOT3UAK submitted 2025-07-01 cs.LG cs.AI

Yilun Zhang This is my paper

classification cs.LGcs.AI

keywords cognitiveloadtheoryLLMinferenceoptimizationtokeneconomymeta-promptinginstructionfine-tuningcontextcompressionchain-of-thoughtemergentproblemdecomposition

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

The paper's central claim is that LLM inference can be managed like a human cognitive process: measure each problem's intrinsic complexity, strip away extraneous computation, and spend the saved tokens on productive reasoning. It formalizes these three loads as $ICL_{LLM}$, $ECL_{LLM}$, and $GCL_{LLM}$, and solves the resulting token-allocation problem with two mechanisms: a zero-shot meta-prompt (CLAI-Prompt) and a fine-tuned model (CLAI-Tune). Across GSM8K, MATH, LongBench QA, and HumanEval, the methods report up to 45% token reduction with equal or better accuracy. If true, this gives a principled, theory-driven route to cheaper inference and a possible explanation for why a fine-tuned model spontaneously decomposes hard problems.

What carries the argument

The load decomposition: $ICL_{LLM}$ (intrinsic problem complexity, estimated by the LLM's own decomposition and complexity score), $ECL_{LLM}$ (extraneous load from irrelevant context or redundant reasoning, identified via attention or explicit filtering), and $GCL_{LLM}$ (germane load spent on a structured reasoning chain). The CLAI principle is to minimize $ECL_{LLM}$, then allocate the saved token budget to $GCL_{LLM}$ according to $ICL_{LLM}$. The meta-prompt pipeline in CLAI-Prompt stages these three steps; CLAI-Tune distills the staged behavior into a single forward pass by supervised fine-tuning on teacher-generated trajectories.

What would settle it

Run a control prompt instructing the same models to 'be concise, structured, and ignore irrelevant context' without any cognitive-load framing; if the token reduction matches CLAI's, the mechanism is generic instruction-following, not cognitive load management. Alternatively, check whether $ICL_{LLM}$ scores correlate with human-labeled difficulty or the number of reasoning steps actually required—a lack of correlation would invalidate the budgeting step.

Watch

Extended reading notes

Core claim

The paper claims that framing inference as cognitive economics works: the model itself, guided by a meta-prompt or fine-tuning, can estimate intrinsic load, budget tokens, and filter irrelevant context, producing 31–45% token savings without accuracy loss. It further claims that fine-tuning internalizes this control so the model spontaneously plans instead of solving directly when complexity is high, an emergent behavior not explicitly optimized for.

Load-bearing premise

The central claim rests on the assumption that an LLM's self-assessed complexity score and self-chosen token budget reliably track true problem difficulty, and that attention-based measures faithfully capture extraneous versus germane load; the paper offers no independent validation of these metrics.

Editorial extensions

If this is right

Token consumption drops by roughly 30–45% across reasoning, long-context QA, and code benchmarks, with accuracy staying at or above the standard chain-of-thought baseline.
CLAI-Prompt works zero-shot on existing LLMs, meaning a structured meta-prompt can deliver immediate savings without retraining.
CLAI-Tune reduces end-to-end latency by replacing the multi-stage prompt pipeline with a single forward pass, while also improving accuracy slightly.
On high-complexity problems, CLAI-Tune emits a decomposed multi-step plan rather than a direct solution attempt, showing a form of meta-cognitive control that standard baselines lack.
In RAG settings, prompt-based extraction of query-relevant facts achieves better compression ratios and F1 than statistical compressors like LLMLingua and RECOMP.

Reading between the lines

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

If the load metrics are valid, the same framework could be applied inside a single forward pass using hidden-state probes to allocate computation dynamically, for instance by routing tokens to more layers or experts when estimated intrinsic load is high.
The paper does not test whether the same token savings would occur with an instruction to 'be concise and structured' that omits the cognitive-load framing; a control experiment would isolate the framework's specific contribution.
The emergent decomposition behavior suggests that training on cognitive-load-labeled trajectories could teach models to recognize when to stop and plan, which may generalize to open-ended agentic tasks beyond the four benchmarks.
The framework unifies existing optimization techniques—context compression becomes ECL reduction, chain-of-thought becomes GCL application—so combining these mechanisms with the budgeting step could yield further gains.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the short version: the token savings are probably real, but the paper never shows they come from the cognitive-load machinery. The control is 'Standard CoT' with no instruction to be concise; a trivial 'think step-by-step but be concise' prompt could plausibly match the 31-36% reductions. So the central attribution claim is untested.

What is actually new: the CLT vocabulary is mapped onto LLM inference in a clean way, and the three-stage prompt is a reasonable engineering recipe. The fine-tuning pipeline (distilling a teacher's prompted behavior into a student) is standard. The paper is clearly written, and the appendix gives the full meta-prompt and dataset samples, which is good for reproducibility.

Soft spots, in order of severity. First, no ablations. Stage 1 budget, Stage 2 parsing, and Stage 3 self-correction are added together, and none is removed or tested separately. Second, no error bars, seeds, or significance tests; Tables 1-3 report single numbers. Third, the ICL/ECL/GCL metrics are never actually computed; the prompt just asks the model to self-assess and set a budget. Eq. (1) is never solved or its terms quantified. Fourth, the 'emergent decomposition' claim is contradicted by the training data: the high-ICL outputs are explicitly decomposition plans (Appendix B.2), so the model was trained on exactly what the paper calls emergent. The paper itself says the ability 'was present in the training data.'

The limitations section is honest about the heuristic nature of ICL estimation and teacher-model dependence, which I credit. But the experimental section doesn't support the theoretical story.

Who this is for: practitioners who want a drop-in prompt for token reduction might find the template useful. As a scientific contribution, it needs substantial revision: ablation against a plain-conciseness baseline, multiple seeds, and a measured ICL metric. I'd send it to review because the empirical claim, if properly tested, matters, and the reviewers can demand those experiments.

Referee Report

4 major / 5 minor

Summary. The paper introduces a Cognitive Load-Aware Inference (CLAI) framework that maps Cognitive Load Theory constructs (ICL, ECL, GCL) to LLM inference. Two implementations are presented: CLAI-Prompt, a zero-shot multi-stage meta-prompt that decomposes the query, estimates complexity, prunes context, and generates a budgeted chain-of-thought, and CLAI-Tune, a fine-tuned model trained on synthetic data generated by the teacher executing CLAI-Prompt. Experiments on GSM8K, MATH, LongBench QA, and HumanEval report token reductions up to 45% with similar or better accuracy, and CLAI-Tune is claimed to exhibit emergent autonomous decomposition.

Significance. If the causal claims were established, this would be an interesting conceptual contribution connecting cognitive science to token-efficiency optimization. The paper is clearly written and provides a complete meta-prompt template and hyperparameters in appendices, which supports reproducibility of the method as described. However, the experimental evidence does not support the causal role of the cognitive-load components, the 'emergent' ability is directly trained, and the formal metrics are not operationalized as defined. As presented, the contribution reduces to a generic 'structure and be concise' prompt heuristic plus distillation from that prompt, with no statistical validation.

major comments (4)

[5.2 (Tables 1-3)] All reported results are single numbers without error bars, multiple seeds, or significance tests. For example, Table 1 reports GSM8K accuracy 92.5 vs. 92.3 and states the drop is 'within the margin of statistical error,' but no margin is given. Since the core claim is that token reductions do not sacrifice accuracy, the absence of variance estimates makes this claim unverifiable. The authors should run multiple seeds (or use bootstrapping) and report confidence intervals and significance tests for both accuracy and token counts.
[3.2-3.4 and 5.2] The causal attribution of token savings to the cognitive-load framework is untested. CLAI-Prompt adds, at once, sub-question decomposition, an explicit token budget, context filtering, and a self-correction pass. No ablation or control prompt (e.g., 'think step-by-step but be concise; avoid repetition') is compared. Without such a control, the 31-45% token reductions in Tables 1-3 could come entirely from generic instruction to be concise and structured, and the ICL_LLM estimation (3.2), ECL_LLM filtering (3.3), and GCL_LLM budgeting (3.4) would contribute nothing. This is the central empirical claim, so the missing ablation is a load-bearing gap.
[4.2, Appendix B.2, 5.3] The claimed 'emergent' autonomous decomposition is directly taught. Appendix B.2 shows high-complexity training samples whose target output is a decomposed plan, and Section 4.2 states that for high ICL_LLM the output is 'a decomposed multi-step solution plan.' Section 5.3's statement that the behavior 'was present in the training data but not directly optimized for' is internally inconsistent: the decomposition plan is the explicit supervised target. The paper should either remove the emergence claim or reframe it as a learned behavior.
[2.3 (Eq. 1), 3.3, 5.1] The formal optimization problem in Eq. (1) is never operationalized. The metrics ICL_LLM, ECL_LLM, and GCL_LLM are defined via heuristics and proposed attention-based proxies, but no experiment measures or quantifies them, and neither implementation actually uses attention-based ECL/GCL computation: Stage 2 (3.3) is a prompt-based extraction, while Stage 1 (3.2) is LLM self-assessment. There is thus no evidence that the framework's theoretical constructs drive the observed results, and the 'cognitive economics optimization' claim is not supported by quantitative data.

minor comments (5)

[References] Several entries are non-archival or informal (e.g., 'LADDER' cited from Reddit, 'Confident AI' blog, Kim (2024) Medium), and Wei et al. 2022a and 2022b are the same title; please revise to a consistent, peer-reviewed reference list.
[5.1 and Appendix C] Section 5.1 says decoding uses 'greedy decoding or nucleus sampling,' but Appendix C states all experiments used greedy decoding with temperature=0.0; please clarify which was used per experiment.
[Table 2] Table 2 reports only input-token compression ratios for the RAG benchmark, while the paper's token-economy claims concern generated tokens; please report output token counts and total token reduction for the RAG setting as well.
[Algorithm 1 and Table 3] In Algorithm 1, the self-correction step is marked optional; the latency figures in Table 3 for CLAI-Prompt presumably include multiple LLM calls, but the token counts may count only the final answer. Please specify how tokens and latency are accounted across stages.
[1.6] The claim of being the 'first' systematic mapping of CLT to LLM inference is too strong given the existing techniques cited in Section 1.2 and the related prompt decomposition literature (Self-Ask, Least-to-Most); please temper the novelty claim.

Circularity Check

2 steps flagged · score 7.0 of 10

Token savings and 'emergent' decomposition are hard-wired into the method: the prompt makes the model set and obey its own token budget, and CLAI-Tune is fine-tuned on teacher outputs that already contain the decomposed plans.

self definitional [Section 3.2 (Stage 1) and Appendix A (Stage 3); results in Tables 1–3]
"Subsequently, the LLM must propose a 'token budget' for the subsequent reasoning stage (i.e., the application of GCL LLM) based on this complexity score."

The reported token reductions are produced by instructing the model to set its own budget in Stage 1 and then to adhere to it in Stage 3 ('Adhere to the token budget'). The measured Avg. Tokens/Problem is therefore a direct consequence of a self-imposed constraint, not an independent empirical finding that validates the ICL/ECL/GCL metrics. Any prompt that imposes a length limit could produce similar savings, and the paper never ablates the cognitive-load components against a generic 'be concise and structured' instruction. The efficiency claim thus reduces to the method's own construction.
fitted input called prediction [Section 4.2 (high-ICL output format), Appendix B.2, and Section 5.3 (Qualitative Analysis)]
"When the teacher model assesses a very high complexity in Stage 1, the 'output' is not a direct problem-solving process but a decomposed multi-step solution plan. ... Beyond quantitative metrics, the most exciting result from CLAI-Tune is its emergent ability, which was present in the training data but not directly optimized for."

The SFT training objective (Section 4.3) is to reproduce the 'output' field for each training instruction, and for high-ICL samples that output is a decomposed plan (Appendix B.2). The 'emergent' decomposition is therefore a supervised target inserted by construction, and the paper's own statement that the behavior 'was present in the training data' contradicts the claim that it was 'not directly optimized for.' The baseline comparison only shows the model learned its training labels; it does not demonstrate an independently emergent meta-cognitive skill.

full rationale

The accuracy results against external benchmarks (GSM8K, MATH, LongBench, HumanEval) are not circular, because those labels are fixed and the reported scores are comparable to baselines. However, the two headline claims reduce by construction. For CLAI-Prompt, token savings come from a self-imposed token budget that the meta-prompt requires the model to set and then obey; the paper does not ablate this against a simple conciseness instruction, so the cognitive-load attribution is unsupported. For CLAI-Tune, the claimed 'emergent' problem decomposition is explicitly present in the training data as the desired output for high-ICL examples, and the fine-tuning loss directly optimizes reproduction of that output. The self-citation to Zhang et al. (2024) for meta-prompting is not load-bearing. Overall, the framework's efficiency and emergence claims are partially circular, while the accuracy comparisons retain independent content.

Assumptions & free parameters 0 free parameters · 4 assumptions · 3 invented entities

The framework introduces new cognitive-load metrics that are neither formally defined nor measured. The methods depend on the unvalidated assumption that cognitive-load constructs map cleanly onto token usage and that the LLM's self-assessment is reliable. No free parameters are fitted because the metrics are not computed at all; they are only narrative labels.

assumptions (4)

domain assumption Cognitive Load Theory applies to LLM inference, so human working memory constructs map onto token computation.
Section 2.3 maps ICL, ECL, and GCL to LLM metrics without empirical validation of the mapping.
domain assumption The Transformer attention mechanism is a valid proxy for the model's cognitive state and can quantify extraneous and germane load.
Section 2.2 draws an analogy between attention weights and attention in the brain, then relies on it for defining ECL and GCL.
domain assumption An LLM's self-reported complexity score and token budget are accurate enough to guide inference.
Stage 1 of CLAI-Prompt (Section 3.2) uses the LLM's own decomposition and scoring without external calibration.
domain assumption Token count is an adequate proxy for computational cost and cognitive load.
The optimization objective in Eq. (1) and all efficiency metrics treat tokens as the resource to minimize.

invented entities (3)

ICL_LLM (Intrinsic Cognitive Load for LLM)
purpose: Quantify the inherent complexity of a query to set a token budget.
Defined via heuristics such as structural analysis and classifier pre-estimation, but never measured against an objective standard.
ECL_LLM (Extraneous Cognitive Load for LLM)
purpose: Quantify wasted computation on irrelevant or redundant tokens.
Proposed as attention-based, but no attention scores are computed or validated in the experiments.
GCL_LLM (Germane Cognitive Load for LLM)
purpose: Quantify productive reasoning tokens to be allocated according to ICL.
Equated with chain-of-thought tokens, but never independently measured.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cognitive Load-Aware Inference: A Neuro-Symbolic Framework for Optimizing the Token Economy of Large Language Models." pith.science (2026). https://pith.science/paper/SQOT3UAK

@misc{pith2026250700653,
  author       = {Pith},
  title        = {Pith review of: Cognitive Load-Aware Inference: A Neuro-Symbolic Framework for Optimizing the Token Economy of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQOT3UAK}},
  note         = {Machine review of arXiv:2507.00653}
}

abstract

The escalating computational costs of Large Language Model (LLM) inference have become a critical barrier to their widespread and sustainable deployment. While existing optimization strategies are effective, they are predominantly based on statistical heuristics or architectural modifications, lacking a guiding cognitive theory to manage the inference process itself. This paper aims to bridge this gap by introducing a novel paradigm: the Cognitive Load-Aware Inference (CLAI) framework, which operationalizes principles from Cognitive Load Theory (CLT) and neuroscience for LLM inference. We formalize the concepts of Intrinsic Cognitive Load, Extraneous Cognitive Load, and Germane Cognitive Load into quantifiable LLM metrics ($ICL_{LLM}$, $ECL_{LLM}$, and $GCL_{LLM}$), thereby reframing the inference process as a cognitive economics optimization problem: based on the intrinsic complexity of a problem ($ICL_{LLM}$), minimize wasteful computation ($ECL_{LLM}$), and strategically allocate the token budget to productive reasoning ($GCL_{LLM}$). We propose two implementation paths: CLAI-Prompt, a zero-shot method that guides a base LLM through cognitive control steps via a structured meta-prompt, and CLAI-Tune, a fine-tuned model that internalizes these principles for spontaneous cognitive economy. Across a range of benchmarks in complex reasoning, long-context question answering, and code generation, our methods achieve significant reductions in token consumption (up to 45\%) without sacrificing accuracy. Furthermore, CLAI-Tune exhibits an emergent ability to autonomously decompose difficult problems, a key characteristic of human expert cognition. This work demonstrates that by emulating the brain's resource management strategies, we can build more efficient, robust, and capable artificial intelligence systems.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

When Iterative RAG Beats Ideal Evidence: A Diagnostic Study in Scientific Multi-hop Question Answering
cs.CL 2026-01 conditional novelty 7.0 of 10

On ChemKGMultiHopQA, iterative retrieval-reasoning outperformed oracle gold-context static RAG for all 11 LLMs tested, with gains up to 25.6 percentage points.

Reference graph

Works this paper leans on

34 extracted references · 27 canonical work pages · cited by 1 Pith paper

[1]

Anonymous. (2024). LADDER: Learning through Autonomous Difficulty-Driven Example Recursion. Reddit

work page 2024
[2]

Chaouachi, M., & Jraidi, I. (2025). Challenging Cognitive Load Theory: The Role of Educational Neuroscience and Artificial Intelligence in Redefining Learning Efficacy. PubMed Central

work page 2025
[3]

Confident AI. (2024). The Definitive Guide to Synthetic Data Generation Using LLMs. Confident AI Blog

work page 2024
[4]

Forn, C., et al. (2021). The Effect of Cognitive Load on the Retrieval of Long-Term Memory: An fMRI Study. Frontiers in Behavioral Neuroscience

work page 2021
[5]

Gong, L., & Zhang, Q. (2024). Context-Aware Systems with LLMs: Semantic Decomposition and Selective Context Filtering. arXiv preprint arXiv:2502.11444

arXiv 2024
[6]

Huang, J., et al. (2024). Decomposing, Evaluating, and Analyzing the Self-Correction Behaviors of Large Language Models. arXiv preprint arXiv:2412.19513

arXiv 2024
[7]

Jiang, H., et al. (2024). LongLLMLingua: Compressing and Reorganizing Prompts for Long Context Large Language Models. In Proceedings of ACL 2024

work page 2024
[8]

Kim, D. (2024). A Comprehensive Review: Model Compression for Large Language Models (LLMs). Medium

work page 2024

Show all 34 references

[9]

Leppink, J., & van den Heuvel, A. (2015). The evolution of cognitive load theory and its application to medical education. Perspectives on medical education

2015
[10]

Li, X., et al. (2025). AttentionRAG: An Attention-Guided Context Pruning Method for RAG Systems. arXiv preprint arXiv:2503.10720

2025
[11]

Ma, X., et al. (2023). LLM-Pruner: On the Structural Pruning of Large Language Models. Advances in Neural Information Processing Systems

2023
[12]

Madaan, A., et al. (2023). Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems

2023
[13]

Paas, F., Renkl, A., & Sweller, J. (2003). Cognitive load theory and instructional design: Recent developments. Educational psychologist

2003
[14]

Paas, F., & van Merriënboer, J. J. G. (2020). Cognitive-load theory: Methods to manage working memory load in the learning of complex tasks. Current Directions in Psychological Science, 29(4), 394-398

2020
[15]

Pan, J., & Li, G. (2025). A Survey of LLM Inference Systems. arXiv preprint arXiv:2506.21901

2025 arXiv
[16]

A., & Ungerleider, L

Pessoa, L., Gutierrez, E., Bandettini, P. A., & Ungerleider, L. G. (2002). Neural correlates of visual working memory: fMRI amplitude predicts task performance. Neuron

2002
[17]

Press, O., et al. (2022). Measuring and narrowing the compositionality gap in language models. arXiv preprint arXiv:2210.03350

2022 arXiv
[18]

Prompting Guide. (2024). Retrieval Augmented Generation (RAG) for LLMs. promptingguide.ai

2024
[19]

Shestyuk, A., Kashefi, S., & Knight, R. T. (2019). EEG-Based Prediction of Cognitive Load in Intelligence Tests. Frontiers in Human Neuroscience

2019
[20]

Sörqvist, P., & Marsh, J. E. (2015). How concentration shields against distraction. Frontiers in psychology

2015
[21]

Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive science

1988
[22]

Sweller, J., Ayres, P., & Kalyuga, S. (2011). Cognitive load theory. Springer

2011
[23]

Sweller, J., van Merriënboer, J. J. G., & Paas, F. (2019). Cognitive Architecture and Instructional Design: 20 Years Later. Educational Psychology Review

2019
[24]

Vaswani, A., et al. (2017). Attention is all you need. Advances in neural information processing systems

2017
[25]

Wei, J., et al. (2022a). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems

2022
[26]

Wei, J., et al. (2022b). Chain-of-thought prompting elicits reasoning in large language models. Google AI Blog

2022
[27]

Xia, H., Du, C., Li, Y., Liu, Q., & Li, W. (2025). Tutorial Proposal: Speculative Decoding for Efficient LLM Inference. COLING 2025 Tutorial

2025
[28]

Xu, F., Shi, W., & Choi, E. (2023). RECOMP: Improving Retrieval-Augmented LMs with Compression and Selective Augmentation. OpenReview

2023
[29]

Yan, Y., et al. (2025). Spontaneous Step-level Self-correction for Mathematical reasoning. AAAI Conference on Artificial Intelligence

2025
[30]

Zhang, Y., et al. (2024). Meta Prompting for AGI. arXiv preprint arXiv:2311.11482

2024 arXiv
[31]

X., et al

Zhao, W. X., et al. (2024). Efficient Large Language Models: A Survey. Transactions on Machine Learning Research

2024
[32]

Zhou, D., et al. (2022). Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625

2022 arXiv
[33]

Zhou, P., Wang, X., & Wang, Y. (2024). EEG-based cognitive load recognition in simulated flight missions: A temporal dynamics study. Frontiers in Human Neuroscience

2024
[34]

Zysset, S., Müller, K., Lohmann, G., & von Cramon, D. Y. (2001). Color-word matching stroop task: separating interference and response conflict. NeuroImage

2001

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Anonymous. (2024). LADDER: Learning through Autonomous Difficulty-Driven Example Recursion. Reddit

work page 2024

[2] [2]

Chaouachi, M., & Jraidi, I. (2025). Challenging Cognitive Load Theory: The Role of Educational Neuroscience and Artificial Intelligence in Redefining Learning Efficacy. PubMed Central

work page 2025

[3] [3]

Confident AI. (2024). The Definitive Guide to Synthetic Data Generation Using LLMs. Confident AI Blog

work page 2024

[4] [4]

Forn, C., et al. (2021). The Effect of Cognitive Load on the Retrieval of Long-Term Memory: An fMRI Study. Frontiers in Behavioral Neuroscience

work page 2021

[5] [5]

Gong, L., & Zhang, Q. (2024). Context-Aware Systems with LLMs: Semantic Decomposition and Selective Context Filtering. arXiv preprint arXiv:2502.11444

arXiv 2024

[6] [6]

Huang, J., et al. (2024). Decomposing, Evaluating, and Analyzing the Self-Correction Behaviors of Large Language Models. arXiv preprint arXiv:2412.19513

arXiv 2024

[7] [7]

Jiang, H., et al. (2024). LongLLMLingua: Compressing and Reorganizing Prompts for Long Context Large Language Models. In Proceedings of ACL 2024

work page 2024

[8] [8]

Kim, D. (2024). A Comprehensive Review: Model Compression for Large Language Models (LLMs). Medium

work page 2024