Lost in the Middle: How Language Models Use Long Contexts
Open accessThe DOI matches its Crossref record; 48 of the 59 references it cites were found in a registry or an index; 15 of the 23 sources its claims cite were checked for support.
Abstract
While recent language models have the ability to take long contexts as input, relatively little is known about how well they use longer context. We analyze the performance of language models on two tasks that require identifying relevant information in their input contexts: multi-document question answering and key-value retrieval. We find that performance can degrade significantly when changing the position of relevant information, indicating that current language models do not robustly make use of information in long input contexts. In particular, we observe that performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts, even for explicitly long-context models. Our analysis provides a better understanding of how language models use their input context and provides new evaluation protocols for future long-context language models.
Overview
CoChat Summary
TL;DR
Language models’ performance on tasks requiring access to long input contexts is strongly affected by where relevant information appears, with a U-shaped curve: best at the beginning or end and worst in the middle. Extended-context models are not necessarily better at using context, and reader performance saturates before retriever recall in open-domain QA.
Key findings
- In multi-document QA, performance is highest when the answer-containing document is at the start or end of the context and degrades in the middle; GPT-3.5-Turbo can drop by more than 20%, and in 20- and 30-document settings middle-position performance is lower than its closed-book performance of 56.1% (oracle: 88.3%).
- Extended-context models are not necessarily better at using input context: when the input fits both windows, GPT-3.5-Turbo and GPT-3.5-Turbo(16K) perform nearly identically as a function of relevant-information position.
- In synthetic key-value retrieval, Claude-1.3 and Claude-1.3(100K) do nearly perfectly, while other models struggle especially with 140 or 300 key-value pairs; query-aware contextualization gives GPT-3.5-Turbo(16K) perfect performance at 300 pairs, versus worst-case 45.6% without it.
- In open-domain QA, reader performance saturates long before retriever recall: using 50 instead of 20 retrieved documents only marginally improves performance (∼1.5% for GPT-3.5-Turbo and ∼1% for Claude-1.3).
Methods
- Controlled multi-document QA on NaturalQuestions-Open: 2,655 queries whose annotated long answer is a paragraph; each context has exactly one answer-containing Wikipedia passage plus k−1 Contriever/MS-MARCO distractor passages; the paper varies answer position and context length (10, 20, 30 documents) and measures accuracy.
- Synthetic key-value retrieval: inputs are serialized JSON objects with unique random UUID key-value pairs; models must return the value for a specified key; experiments use 75, 140, and 300 key-value pairs (500 examples each), varying key position and number of pairs, with accuracy as the metric.
- Models include open MPT-30B-Instruct and LongChat-13B(16K), closed GPT-3.5-Turbo, GPT-3.5-Turbo(16K), Claude-1.3, and Claude-1.3(100K), plus additional architecture/instruction experiments with Flan-T5-XXL, Flan-UL2, MPT-30B, Llama-2, and GPT-4; an open-domain QA case study uses a Contriever retriever-reader setup.
Limitations
- Stated: the paper uses greedy decoding and leaves exploration of other decoding methods to future work; the investigations into why models are not robust are described as preliminary; GPT-4 is evaluated only on a subset of 500 multi-document QA examples because full evaluation would cost upwards of $6000.
- Stated: NaturalQuestions-Open has ambiguity, so the authors run an unambiguous subset and also test random distractors and randomized distractor order, finding similar trends.
- Evident: the controlled multi-document QA setting always contains exactly one document that answers the question, unlike open-domain QA; the synthetic key-value task removes natural language semantics; results may not generalize to all tasks, models, or decoding settings.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
7 Valid · 0 Not valid · 5 Not checked · 13 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Language models have become an important and flexible building block in a variety of user-facing language technologies, including conversational interfaces, search and summarization, and collaborative writing (Shuster et al., 2022;Thoppilan et al., 2022;Lee et al., 2022, inter alia).”
§ Lost in the Middle: How Language Models Use Long Contexts > 1 Introduction
What the sources say
- Partly supportedcites Shuster et al., 2022 sourceExperimental — accuracy not yet measured
Section 3.1.1 (p. 4) describes a conversational agent built on a language model that generates internet search queries, supporting conversational interfaces and search as part of dialogue. Section 2 (p. 3) mentions summarization only as related work, and the passages do not establish collaborative writing as an application of this paper’s model. (LLM, conf=0.85)
- Supportedcites Thoppilan et al., 2022 source
The Abstract (p. 1) identifies LaMDA as a family of language models specialized for dialog. In this sentence's list of applications, Thoppilan et al. supports the conversational-interface example; the other cited works need not each cover every example. (LLM, conf=0.92)
- Supportedcites Lee et al., 2022 source
Lee et al. is cited alongside other works, and its contribution to the sentence is the collaborative-writing example. Section 4.2.2 describes writers receiving and selecting model-generated suggestions in a writing interface; Section 2.2.1 also describes language models as applicable to dialogue and summarization. The paper need not establish every example in the jointly cited list. (LLM, conf=0.89)
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“These observations complement prior work, which found that non-instruction fine-tuned language models are biased towards recent tokens (i.e., the end of the input context; Khandelwal et al., 2018;Press et al., 2021).”
§ Lost in the Middle: How Language Models Use Long Contexts > 4.3 Effect of Instruction Fine-Tuning
What the sources say
- Partly supportedcites Khandelwal et al., 2018 sourceExperimental — accuracy not yet measured
The Introduction and Section 5 report that the studied LSTM is much more sensitive to recent word order than to distant word order, supporting a recency effect. The paper studies a standard LSTM on two datasets, however, rather than establishing this tendency across non-instruction-fine-tuned language models generally; it also finds that distant words remain relevant. (LLM, conf=0.82)
- Partly supportedcites Press et al., 2021 sourceExperimental — accuracy not yet measured
Section 3.1 (page not available in our copy) finds that very short contexts hurt performance, while increasing subsequence length from 1,024 to 3,072 tokens gives negligible improvement. That is consistent with limited benefit from more distant context, but the paper does not establish that the models are biased toward recent tokens within an input context. (LLM, conf=0.78)
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“We instantiate this task with data from NaturalQuestions-Open (Lee et al., 2019;Kwiatkowski et al., 2019), which contains historical queries issued to the Google search engine, coupled with human-annotated answers extracted from Wikipedia.”
§ Lost in the Middle: How Language Models Use Long Contexts > 2.1 Experimental Setup
What the sources say
- Supportedcites Lee et al., 2019 source
Section 7.1 says the paper creates an open version of Natural Questions and uses it for training and evaluation. The Introduction describes the task as answering questions by retrieving evidence from an open corpus, supporting the cited use of NaturalQuestions-Open for this task. (LLM, conf=0.91)
- Not checked — no full text of the sourcecites Kwiatkowski et al., 2019 source
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Following Kandpal et al. (2022) and Mallen et al. (2023), we use accuracy as our primary evaluation metric, judging whether any of the cor- rect answers (as taken from the NaturalQuestions annotations) appear in the predicted output.”
§ Lost in the Middle: How Language Models Use Long Contexts > 2.1 Experimental Setup
What the sources say
- Supportedcites Kandpal et al. (2022) source
Section 3 (p. 4) says the paper evaluates answers using exact match against a ground-truth answer set, and Section 3.1 (p. 4) reports QA accuracies as its evaluation results. This supports citing Kandpal et al. for using accuracy to evaluate question answering. (LLM, conf=0.90)
- Supportedcites Mallen et al. (2023) source
Section 3.1 (‘Focus and Task’) explicitly names accuracy as the metric and says a prediction is correct when it contains an exact match to any gold answer. This supports the evaluation approach attributed to Mallen et al.; the NaturalQuestions annotations describe the citing authors’ answer source. (LLM, conf=0.94)
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“or when language models are augmented with external information (relevant documents from a search engine, database query results, etc.; Petroni et al., 2020;Ram et al., 2023;Shi et al., 2023;Mallen et al., 2023;Schick et al., 2023, inter alia).”
§ Lost in the Middle: How Language Models Use Long Contexts > 1 Introduction
What the sources say
- Partly supportedcites Petroni et al., 2020 sourceExperimental — accuracy not yet measured
The Abstract (p. 1) says the paper integrates information from a retrieval system with a pre-trained language model, and §3.4.2 (p. 5) specifies a retrieved paragraph as context. This supports augmentation with retrieved documents, but our copy does not show language models being augmented with database query results. (LLM, conf=0.84)
- Partly supportedcites Ram et al., 2023 sourceExperimental — accuracy not yet measured
The Introduction describes augmenting language models with relevant documents retrieved from an external knowledge source, and Section 5 describes using a BM25 retriever to find those documents. The passages do not establish database query results as another type of external information used by this paper. (LLM, conf=0.82)
- Partly supportedcites Shi et al., 2023 sourceExperimental — accuracy not yet measured
Section 3.1 (page unavailable) describes retrieving relevant documents from a corpus, and the Conclusion says REPLUG uses those documents to augment a language model. The paper supports the document-retrieval example, but our copy does not describe augmenting models with database query results or documents from a search engine specifically. (LLM, conf=0.82)
- Partly supported
20 more claims · 2 with their cited sources checked
Create a free account to see all 25 claims and every source verdict.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“In vanilla multi-document or NIAH-based benchmarks (Kamradt, 2023; Liu et al., 2024), models perform consistently well when the needle (or gold document) appears at the very beginning or end of the context window, with minimal impact from context length.”
From NoLiMa: Long-Context Evaluation Beyond Literal Matching (2025) · cites this paper as Liu et al., 2024
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Even within a single long sequence, attention effectiveness degrades with distance due to attention dilution , positional encoding limitations, and token interference, leading to the well-known 'lost-in-the-middle' and context-decay phenomena (Liu et al., 2024; Press et al., 2021a).”
From MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents (2026) · cites this paper as Liu et al., 2024
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Current large language models (LLMs) exhibit significant deficiencies in episodic memory tasks including encoding, storing, and retrieving specific information from temporally dependent events over a long period of time.”
From ARTEM: Enhancing Large Language Model Agents with Spatial-Temporal Episodic Memory (2026) · cites this paper as Liu et al. 2024
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Recent approaches to handle memory tasks in LLMs, such as in-context learning, retrieval-augmented generation (RAG), and fine-tuning, may resolve the long-term retention issues, but are still inadequate to handle tasks requiring chronological awareness of the stored information.”
From ARTEM: Enhancing Large Language Model Agents with Spatial-Temporal Episodic Memory (2026) · cites this paper as Liu et al. 2024
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“If not ranked at the top, the key information may be neglected by the generative module, resulting in incomplete or inaccurate responses to the user queries24.”
From Leveraging long context in retrieval augmented language models for medical question answering (2025) · cites this paper as [24]
Experimental — accuracy not yet measured
86 more claims about this paper
Create a free account to see every claim other papers make about this one.
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
LLaMA: Open and Efficient Foundation Language Models
Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023) · arXiv · doi:10.48550/arxiv.2302.13971 · Cited by 4,602
Transformer-XL: Attentive language models beyond a fixed-length context
Dai, Z., Yang, Z., Yang, Y., Carbonell, J., Le, Q., & Salakhutdinov, R. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1285 · Cited by 3,169 · cited 2×
“Recent improvements in hardware (e.g., faster GPUs with more memory) and algorithms (Dai et al., 2019; Dao et al., 2022; Poli et al., 2023; Rubin and Berant, 2023, inter alia ) have resulted in language models with larger context windows (e.g., 4096, 32K, and even 100K tokens), but it remains…”
BackgroundLlama 2: Open Foundation and Fine-Tuned Chat Models
Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. (2023) · arXiv · doi:10.48550/arxiv.2307.09288 · Cited by 2,604
Longformer: The Long-Document Transformer
Beltagy, I., Peters, M. E., & Cohan, A. (2020) · arXiv · doi:10.48550/arxiv.2004.05150 · Cited by 2,281
Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering
Izacard, G., & Grave, E. (2020) · Conference of the European Chapter of the Association for Computational Linguistics · doi:10.48550/arxiv.2007.01282 · Cited by 2,263
Natural Questions: A benchmark for question answering research
Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. (2019) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00276 · Cited by 2,069 · cited 1×
“We instantiate this task with data from NaturalQuestions-Open (Lee et al., 2019; Kwiatkowski et al., 2019), which contains historical queries issued to the Google search engine, coupled with human-annotated answers extracted from Wikipedia.”
MethodThe serial position effect of free recall.
Murdock, B. B. (1962) · Journal of Experimental Psychology · doi:10.1037/h0045106 · Cited by 1,717 · cited 1×
“The U-shaped curve we observe in this work has a connection in psychology known as the serial-position effect (Ebbinghaus, 1913; Murdock Jr, 1962), that states that in free-association recall of elements from a list, humans tend to best re-member the first and last elements of the list.”
MethodMemory: A contribution to experimental psychology.
Ebbinghaus, H. (1913) · doi:10.1037/10011-000 · Cited by 1,645 · cited 1×
“The U-shaped curve we observe in this work has a connection in psychology known as the serial-position effect (Ebbinghaus, 1913; Murdock Jr, 1962), that states that in free-association recall of elements from a list, humans tend to best re-member the first and last elements of the list.”
Method
Research works printed but not linked: 1
- Sharan, 2018, Prediction with a short memory, 10.18653/v1/K18-3013 · DOI
Cited by
Items 801–900 of 1,592
Gundidza, F., Kikuchi, M., & Ozono, T. (2026) · Big Data and Cognitive Computing · doi:10.3390/bdcc10080269
“LLM extraction quality degrades with chunk length [36], and chunks that span section boundaries mix project context with appendix boilerplate.”
A process-centric mapping of non-technical debt instances to ISO/IEC 5338 in AI-enabled projects
Kapiyasheva, D., Özcan‐Top, Ö., & Temizel, T. T. (2026) · Software Quality Journal · doi:10.1007/s11219-026-09774-3
“We initially attempted LLM-based verification for P-CMM process areas, but limita - tions in handling extended context and semantically dense texts (Liu et al., 2024) led to inaccurate assessments.”
TCC: Using Topic Chains to Compress Prompts for Long Document Question Answering
Chen, S. F., Liang, B., Yan, J., Liu, L., Wang, H., Yu, Y., Wong, K., & Xu, R. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11463253
Human-AI Co-creation: LLMs, Contextual Hints, Performance
Gratsanis, P., Karydis, I., Sioutas, S., & Vonitsanos, G. (2025) · IFIP advances in information and communication technology · doi:10.1007/978-3-031-97313-0_7
An AI Framework for Literature Review: In Search of Sustainable Fuels
Oliveira, J. P., Mendes, A., & Ferreira, M. C. (2026) · World sustainability series · doi:10.1007/978-3-032-22076-9_40 · Cited by 1
HGFuzzer: Directed Greybox Fuzzing via Large Language Model
Xu, H., Zhao, Y., Yu, J., & Wang, H. (2026) · ACM Transactions on Software Engineering and Methodology · doi:10.1145/3841476 · Cited by 18
From Output Errors to Workflow Harm: A Practitioner-Audit Method for LLM-Mediated ResearchPreprint
Austria, D., McCollister, B., Lindsey, J. E., Arowolo, M. O., & Okon, M. (2026) · medRxiv · doi:10.64898/2026.08.13.26360414 · Cited by 1
Self-Consistency Falls Short! The Adverse Effects of Positional Bias on Long-Context Problems
Byerly, A., & Khashabi, D. (2026) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl.a.625 · Cited by 12
Wang, X., Feng, T., Liu, H., & Zhu, W. (2026) · Springer eBooks · doi:10.1007/978-981-95-7750-7_5 · Cited by 1
Cross-Domain and Multilingual Analysis of Retrieval Augmented Generation on Ontologies
Kühn, R., Mitrović, J., Wendlinger, L., & Granitzer, M. (2025) · IEEE International Conference on Tools with Artificial Intelligence (ICTAI) · doi:10.1109/ictai66417.2025.00202
“With single big chunks, we are afraid to encounter the known phenomenon “lost in the middle” where content stored in the middle of a chunk is more difficult for the LLM to recall [22].”
Learning to Forget: Emotional Salience as a Compression Mechanism for Long-Term AI Memory
Yoo, S., & Im, S. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3818653
“As noted in recent studies on the Lost in the Middle phenomenon [7], simply accumulating vast amounts of data often degrades a model’s ability to retrieve specific, relevant insights.”
Benchmarking International Construction Contract Knowledge of Large Language ModelsPreprint
Xia, H., Chen, Y., Qi, X., & Wang, Y. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5775312
Psy-DMTL: Psychology-guided dynamic multi-task learning for LLM-based personality detection
Wang, Z., Yin, Y., Wei, F., Wu, J., Lv, H., & Yang, W. (2026) · Information Sciences · doi:10.1016/j.ins.2026.124245
Sinha, A., & Chakma, K. (2026) · ACM Transactions on Information Systems · doi:10.1145/3827605 · Cited by 2
Ahmed, T., & Choudhury, S. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7181277
Dubois, R., & Howard, B. (2026) · ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys) · doi:10.1145/3744256.3812555
Moussaoui, Y. E., & Adda, M. (2026) · Procedia Computer Science · doi:10.1016/j.procs.2026.08.392
Wang, Z., Gu, Z., Zhang, Z., Ma, Z., Li, B., Xu, C., Dai, S., Wang, Y., & Liu, H. (2026) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2026.116162
Optimizing Chain of Thought Generation Through Hierarchical Dynamic Reasoning
Chen, Y., Chen, H., Shao, P., & Liu, X. (2025) · International Conference on Robotics, Intelligent Control and Artificial Intelligence (RICAI) · doi:10.1109/ricai68060.2025.11385108
“Furthermore, directly scoring all steps is hampered by input length constraints and ”intermediate step loss” [7], making improved and efficient methods necessary.”
Optimizing medical image analysis with prompt-guided pseudo-labeling
Fan, L., Gong, X., Zheng, C., Li, X., Li, J., & Ou, Y. (2026) · Biomedical Signal Processing and Control · doi:10.1016/j.bspc.2026.109867 · Cited by 1
Ferrigno, E., & Perez-Globant, F. (2026) · SPE Argentina Exploration and Production of Unconventional Resources Symposium · doi:10.2118/231242-ms
Optimizing automated scoring in ILSAs with prompt compression
Jung, J. Y., Bezirhan, U., & von Davier, M. (2026) · Computers and Education Artificial Intelligence · doi:10.1016/j.caeai.2026.100558 · Cited by 1
Evaluating Long-Horizon Memory for Multi-Party Collaborative Dialogues
Hu, C., Li, T., Gao, X., Chen, H., Bai, Y., Xu, D., Lin, T., Li, X., Han, Y., Pei, J., et al. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3817589 · Cited by 45
“In practice, however, failures rarely stem from context length alone [18, 21].”
MedicalBench: Evaluating Large Language Models Towards Improved Medical Concept ExtractionPreprint
Yang, Z., Lyng, G. D., Batra, S. S., & Tillman, R. E. (2026) · SSRN Electronic Journal · doi:10.64898/2026.04.12.26350704 · Cited by 1
Optimizing Low-Parameterised LLMs with RAG using Entropy-based Chunking
Sambare, G. B., Nerpagar, H., Patil, O., & Totlani, K. (2026) · International Conference on Inventive Computing and Informatics (ICICI) · doi:10.1109/icici68773.2026.11580778
End-To-End LLM Systems: Data, Training, Evaluation, and Deployment
Dhanasekar, S. (2026) · International Conference on Machine Learning and Autonomous Systems (ICMLAS) · doi:10.1109/icmlas67792.2026.11483701
Benchmarking Quantized LLMS for Faithfulness in Resource-Constrained RAG
Tamilselvi, P., Ahmed Amin, K., K, V. K. V., & Rohith, S. (2026) · International Conference on Multi-Agent Systems for Collaborative Intelligence (ICMSCI) · doi:10.1109/icmsci67830.2026.11469322
“Furthermore, Liu et al. [9] identified the “Lost in the Middle” phenomenon, where LLMs struggle to access information located in the middle of a long context window.”
Key-Augmented Neural Triggers for Knowledge SharingPreprint
Wolf, A., Palma, M. A., Rani, P., & Gall, H. C. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5376750
Does It Still Make Sense? Organizing and Summarizing Resources in Cross-Session Aggregated Search
Momeni, M., & Hoeber, O. (2026) · Conference on Human Information Interaction and Retrieval · doi:10.1145/3786304.3787873 · Cited by 2
SERank: Sentence-Level evidence selection for LLM-based document reasoning reranking
谢道伟, Luo, J., & Tu, X. (2026) · Expert Systems with Applications · doi:10.1016/j.eswa.2026.134288
Reliable Extraction of Cyber-Physical System Threat Knowledge with Multi-Run LLMs
Habashy, L., Poplawski, I., & Rhéaume, F. (2026) · IEEE International Conference on AI in Cybersecurity (ICAIC) · doi:10.1109/icaic67076.2026.11395857
“This “lost-in-the-middle” phenomenon indicates that models sample the context inconsistently, even when the input is fixed [36].”
DyGADBench: A Comprehensive Benchmark for Anomaly Detection in Dynamic Graphs
Mezhoudi, M. N., Lachaud, G., Diao, Y., Möller, T., Barry, M., & Bifet, A. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3817594
“Mean pooling systematically attenuates it, analogous to the “lost in the middle” phenomenon observed in large language models[37], where relevant information embedded within a long context is diluted by surrounding content.”
Романків, Н., & Ситніков, Д. (2026) · ГРААЛЬ НАУКИ · doi:10.36074/grail-of-science.17.04.2026.069
One Size Does Not Fit All: On the Role of Batch Size in Classifying Requirements with LLMs
van Can, A. T., Aydemir, F. B., & Dalpiaz, F. (2025) · IEEE International Requirements Engineering Conference Workshops (REW) · doi:10.1109/rew66121.2025.00009
Automated Payment Verification of Legal Documents
Carpi, M. D. M., Lima, J. P., Bitelli, B., Ferrari, E., & Finger, M. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-39898-7_10
The Demon at Your Shoulder: The Trade-offs of Partnering with AI for Creative WritingPreprint
Marshall, H. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6278738 · Cited by 1
Integrate-and-Fire Compressor: Learning to Compress Context for LLMs Adaptively
Zhao, Y., Li, X., Wang, Z., Wu, H., Han, M., & Xu, B. (2025) · IEEE International Conference on Multimedia and Expo (ICME) · doi:10.1109/icme59968.2025.11210203
“Following previous work, we conduct experiments on the three datasets NaturalQuestions [20], TriviaQA [21] and HotpotQA [22].”
Scalable Task Planning via Large Language Models and Structured World Representations
Pérez‐Dattari, R., Li, Z., Babuška, R., Kober, J., & Santina, C. D. (2025) · Advanced Robotics Research · doi:10.1002/adrr.202500002 · Cited by 2
Chunk Size and Retrieval Depth Optimization in a Minimal Retrieval-Augmented Generation Pipeline
Ahmad, S., Hussain, J., Najeeb, H., Ahmad, F., Shah, S. H., & Ibad, M. S. (2026) · International Journal of Innovations in Science and Technology · doi:10.33411/ijist/2053
Ribeiro, A., Garrido, R., Ramos, V., Teixeira, S., & Георгиева, П. (2026) · doi:10.1109/icarai70085.2026.11635662
“The number of retrieved chunks ( k ) balances comprehensive context against information overload, documented as the “lost in the middle” effect [8].”
Parallel LLM Reasoning for Bias-Resilient, Robust Conceptual Abstraction
Adeseye, A., Isoaho, J., & Adeseye, A. (2026) · Lecture notes in networks and systems · doi:10.1007/978-3-032-32732-1_6
Lee, G., Kim, S., Byun, Y., Lee, J., & Choi, H. (2026) · International Journal of Precision Engineering and Manufacturing-Green Technology · doi:10.1007/s40684-026-00890-w · Cited by 4
“…present in long prompts, it may not be effectively utilized, and performance can exhibit a U-shaped pattern depending on evidence location, being high at the beginning and end of the context but substantially lower in the middle, and in some cases even falling below closed-book performance [161].”
Retrieval Augmented Tree of Thoughts
Liu, X., & Liu, J. (2024) · Lecture notes in computer science · doi:10.1007/978-981-97-9434-8_36 · Cited by 1
ACoRN: Noise-Robust Abstractive Compression in Retrieval-Augmented Language Models
Kim, S., Jung, G., & Lee, S.-W. (2025) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn64981.2025.11228443 · Cited by 2
“Due to limitations of language models, such as limited relevance discrimination [8] and attention dispersion in long contexts [5], existing abstractive compressors struggle with significant information loss when retrieving multiple documents.”
CARGO: Context-Aware Retrieval-Augmented Generation Compression Optimizer over 6G Links
Shin, Y., Yoo, B., Yang, J., Kim, C. H., Lee, J.-H., Chung, E., Kim, H. W., & Song, H. J. (2025) · International Conference on Information and Communication Technology Convergence (ICTC) · doi:10.1109/ictc66702.2025.11388778
“Long-context studies show that today’s models use context non-uniformly and often under-utilize information located away from the beginning/end of the sequence [2].”
Agentic AI for Chip Design Verification: Failure Modes, Metrics, and Coverage Closure
Marosok, N., Halm, M., Gubbi, K. I., Tarighat, M., Javidnia, N., Shabgahi, S. Z., Huang, K., Rafatirad, S., Sayadi, H., Koushanfar, F., et al. (2026) · International Symposium on Quality Electronic Design (ISQED) · doi:10.1109/isqed69900.2026.11534779
“Current agents often experience ”lost-in-the-middle” failures [15] while parsing large files, only remembering content at the beginning and end of a large input stream.”
iBRAHIMOĞLU, N., Demirezen, M. U., & Yıldız, F. (2026) · Cluster Computing · doi:10.1007/s10586-026-06537-4 · Cited by 3
Tabular Data Wrangling in the Era of Large Language Model: A Survey
He, Y. Q., Yang, Q., Yang, Y., Zhou, Y., Hu, Z., Yuan, S., Xiao, Y. H., & Li, Z. (2026) · IEEE Transactions on Knowledge and Data Engineering · doi:10.1109/tkde.2026.3716977
“Excessive input length not only increases computational cost but also significantly degrades model performance [218], largely due to the inclusion of redundant or irrelevant information.”
Martín, C. A., Torres, J., Aguilar, R. M., Alayón, S., & Bacallado, M. A. (2026) · Applied Sciences · doi:10.3390/app16136793
Edge-AdaptiveKG: Resource-Efficient Knowledge Graph Construction for RAG on Edge Devices
Du, Y., Niu, J., Xue, C. J., & Guan, N. (2026) · IEEE International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA) · doi:10.1109/rtcsa71153.2026.00031
“Existing research [35] has shown that irrelevant content in prompts significantly degrades the capabilities of LLMs, as they are susceptible to hallucinations [28] and the “lost-in-the-middle” problem [19].”
MemAGent: A Cognitive-Hierarchical Multi-Agent System for Long-Form Story Generation
Li, J., & Ding, G. (2025) · International Conference on Computer, Vision and Intelligent Technology (ICCVIT) · doi:10.1109/iccvit67848.2025.11391379
“This results in well-documented failures such as plot decay, character inconsistency, and an inability to maintain a global narrative arc [12], [21].”
Satija, S., Bhatt, A., Jani, P., & Rawal, D. (2026) · ACM Conference on AI and Agentic Systems (CAIS) · doi:10.1145/3786335.3813158 · Cited by 1
“Long contexts degrade LLM performance when relevant information appears in the middle, exhibiting a U-shaped primacy-recency bias [17].”
Lee, J., & Li, W.-S. (2026) · IEEE Access · doi:10.1109/access.2026.3721147 · Cited by 1
“Long-context benchmark studies, including Lost in the Middle , show that simply enlarging the available window does not guarantee faithful use of the right evidence [4], [5], [6], [7].”
A Multimodal Retrieval-Augmented Generation System for Intelligent Question Answering
Liu, B., Ning, Y., He, S., Guo, F., Jia, S., & Zhu, L. (2025) · Lecture notes in computer science · doi:10.1007/978-3-032-06310-6_1 · Cited by 3
Differential Attentive Sequential Recommendation
Wang, H. P., Li, Y., & Chen, J. (2025) · International Journal of High Speed Electronics and Systems · doi:10.1142/s0129156425408204
Aligning Extraction and Generation for Robust Retrieval-Augmented Generation
Song, H., Choi, J., & Kim, M. (2026) · ACM International Conference on Web Search and Data Mining (WSDM) · doi:10.1145/3773966.3777939 · Cited by 5
TAPE: A Temporal Graph-Based Memory System for Personal LLM Agents
Luo, C., Liu, Q., Zhang, W., et al. (2026) · Proceedings - International Conference on Data Engineering · doi:10.1109/icde65706.2026.00318
“Moreover, it suffers from the “lost in the middle” problem [2], where the model struggles to recall information buried within long contexts.”
Knowledge-Guided Reasoning Chain of Pre-trained LLM in Industrial Domain
Lou, J., Ren, H., Ma, G., Liang, X., Yang, C., & Gui, W. (2025) · Communications in computer and information science · doi:10.1007/978-981-96-4509-1_26 · Cited by 1
Wang, Z., Wang, H., Luo, N., Zhang, L., Liu, P., Cao, Y., & Sun, X. (2026) · Lecture notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering · doi:10.1007/978-3-032-32767-3_13 · Cited by 1
Data Access Control in Large Language Models
Oualha, N., & Janneteau, C. (2025) · IEEE International Conference on Trust, Privacy and Security in Intelligent Systems and Applications (TPS-ISA) · doi:10.1109/tps-isa67132.2025.00023
Основные параметры конфигурации системы при проведении эксперимента
Кузнецов, И. (2026) · Cifra LLC Journals · doi:10.60797/irj.2026.171.79.2
A Unifying Scheme for Extractive Content Selection Tasks
Amar, S., Shapira, O., Slobodkin, A., & Dagan, I. (2025) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl.a.53 · Cited by 2
Question Answering on Lengthy Arabic Narrative Books
Ateeq, M. A., Tiun, S. B., Abdelhaq, H., & Kaur, W. (2026) · IEEE Access · doi:10.1109/access.2026.3661853
“A critical challenge identified by [8] is the underperformance of language models when key information is embedded in the middle of long inputs compared to when it appears at the beginning.”
InfluentialZhong, C., Wu, S., Yu, J., Wu, W., Herremans, D., & Zhang, K. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7199726
E³-Guarded Generation: Provably Mitigating Hallucinations in Large Language Models
Wang, G., & Nixon, M. S. (2026) · 2012 International Conference on Computing, Networking and Communications (ICNC) · doi:10.1109/icnc68183.2026.11416906
“…memorization failures, generating plausible but incorrect facts [6]; extrinsic hallucinations contradict or lack support from retrieved evidence [7]; compositional hallucinations arise from incorrect combinations of individually correct facts [8]; and temporal hallucinations mix information…”
STARS: Static Analysis-Guided Assertion Synthesis using Large Language Models
Cao, J., Wang, H., Yan, H., Wen, M., & Pradel, M. (2026) · Proceedings of the ACM on software engineering. · doi:10.1145/3832283
“To manage token limits and avoid overwhelming the LLM with excessively long input, which can lead to performance degradation as noted in prior work [39], we then restrict each generation pass to take no more than 𝑘 𝑎𝑖 assertion components as input components, and generate 𝑘 𝑎𝑜 as output…”
Why Do Vision Language Models Struggle To Recognize Human Emotions?
Agarwal, M., Tsaftaris, S. A., Sevilla-Lara, L., & McDonagh, S. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-37020-4_9 · Cited by 2
GraphRAG-V: Fast Multi-hop Retrieval via Text-Chunk Communities
Yu, T., Srinivasan, V., & Thomo, A. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-14107-1_1
Dialogue Analyzer– Conversation Analysis System
Tom, A. A., & Abraham, A. (2026) · International Conference on Inventive Computing and Informatics (ICICI) · doi:10.1109/icici68773.2026.11580557
“VADER performs well with short text-based social media posts but fails to capture deeper context and often performs badly with conversational structures [7].”
Generative AI and Augmented Reality for Personalized Biophilic Design in Senior Living Environments
Mallepally, A. R., Zhang, R., & Bulbul, T. (2026) · ASCE Construction Research Congress (CRC) · doi:10.1061/9780784486962.090
Foundations of Generative Information Retrieval
Ai, Q., Zhan, J., & Liu, Y. (2024) · The information retrieval series · doi:10.1007/978-3-031-73147-1_2 · Cited by 5
Jun, L., Yan, X., Lin, J.-C., & Zhang, D.-Z. (2026) · Applied Sciences · doi:10.3390/app16125885
CrossBench: Reproducible LLM Memory Evaluation Across Conversational and Coding Workloads
Radu, V., Foley, M., Pour, M. V., Pegoraro, M., Faricy, F., & Girolami, M. (2026) · International Conference on Mobile Systems, Applications, and Services · doi:10.1145/3812836.3814760
Barros, L. A. M., Silva, V. M. A., dos Santos, T. A., et al. (2026) · doi:10.5753/semish.2026.23941
“A primeira caracteriza a queda de qualidade em contexto longo [Liu et al. 2024, Bai et al. 2025].”
Li, H., Huang, J., Ji, M., et al. (2024) · doi:10.2196/preprints.66098
Evaluating the role of provider on safety alignment in large language models
Naser, M. Z. (2026) · Neurocomputing · doi:10.1016/j.neucom.2026.135173
Neighborhood Deviation Attack Against In-Context Learning
Hou, D., Yang, Z., Zheng, L., Jin, B., Xu, H., Li, Y., Xu, B., & Peng, K. (2025) · Applied Sciences · doi:10.3390/app15084177 · Cited by 1
Popovych, Z. B. (2026) · Science and innovation · doi:10.15407/scine22.04.003
Vitale, J. D. (2026) · Machine Learning and Knowledge Extraction · doi:10.3390/make8070207
“RoPE systematically concentrates attention on tokens near the beginning and end of the causal context window, the “lost in the middle” phenomenon documented by Liu et al. (2024) [42], which constrains H 2 ’s recursive discourse from obtaining adequate full thread attention mechanisms.”
InfluentialGraphRAG-Driven Multi-Stage Context Retrieval for Thai Text-to-SQL Using SLM Agents
Sompong, C., Khanthacha, K., Chaukampeng, T., Panyakun, S., Ngokwong, N., & Puntumnunt, V. (2026) · IEEE International Conference on Cybernetics and Innovations (ICCI) · doi:10.1109/icci68752.2026.11506507
“Such long-context inputs often trigger the “Lost in the Middle” phenomenon [19] where the model’s ability to retrieve crucial information degrades significantly when it is buried amidst irrelevant schema noise.”
Myung, J.-H., Chang, S., & Lee, J. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7380663
Deng, W., Liu, Y., & Cui, Z. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-3426-4_8
Die KI-Achse: Allgemeine und modellabhängige Fehlermuster
Elsharkawy, A. E. (2026) · Essentials · doi:10.1007/978-3-662-74123-8_5
Guan, Z., Liang, X., & Zhang, S. (2025) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp49660.2025.10890215
“Despite the widespread utility of RAG , recent experiments highlight a striking deficiency: LLMs struggle to locate relevant documents when they are placed in the middle of their input prompts [9].”
Chen, H., Liu, Y., Li, C., & Hu, W. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-0378-9_20
The recipe for intelligence in natural and artificial systems
Summerfield, C., & Stachenfeld, K. (2026) · Neuron · doi:10.1016/j.neuron.2026.08.013 · Cited by 1
Masethe, M. A., Ojo, S. O., & Masethe, H. D. (2025) · Conference on Information Communications Technology and Society (ICTAS) · doi:10.1109/ictas64866.2025.11155866
“In addition to advances in model architecture, ongoing pre-training strategies have been employed to broaden context windows, with instruction tuning supporting retention of instruction adherence and reasoning skills [24][25].”
Nascimento, V. S. D. O., Tedardi, M. V., Rogério, G. D. S., Pinello, K. C., & Dagli, M. L. (2026) · Cancers · doi:10.3390/cancers18172728
“…contexts are supplied directly to an LLM. Long-context inference increases token consumption, computational cost, and operational complexity, while relevant information may https://doi.org/10.3390/cancers18172728 receive insufficient attention when embedded among numerous competing entries [30].”
InfluentialKnowledge Intensive AgentsPreprint
Liu, Z., Liu, Z., Huang, P., et al. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5459034
From Evaluation to Feedback: A Feature-Based and LLM-Constrained Tool for Korean Writing Assessment
Jang, C., Go, G., Yun, J., Ahn, S., Shin, M. S., Kil, H., Chang, S., Kim, D.-G., & Seo, Y.-D. (2026) · ACM Symposium on Applied Computing · doi:10.1145/3748522.3780021
Reimann, T. E., Haug, V., Verri, F. M., Leinert, C., Fleiner, T., Denkinger, M. D., & Kocar, T. D. (2026) · Journal of Medical Internet Research · doi:10.2196/101882
Beránek, D. (2026) · Figshare · doi:10.6084/m9.figshare.33420025
SciEval: A Benchmark for Automatic Evaluation of K–12 Science Instructional Materials
Li, Z. D., He, P., Chen, Z., Liu, H., Wang, Z., Li, T., & Xiong, J. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-29744-0_36 · Cited by 1
Epistemic Resolve: Governed Uptake and the Architecture of Mind
Crawford, K., & Hammer, P. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-33010-9_12
Operational AI Assistance in Slack-Based Troubleshooting: A Measurement Framework and Field Results
Bommareddy, R. K. R. (2026) · IEEE International Conference on Industry 4.0, Artificial Intelligence, and Communications Technology (IAICT) · doi:10.1109/iaict71158.2026.11620929
“However, more recent evaluations of these methods have revealed weaknesses in aspects of these models, such as their rejection of negative information [4]-[10].”
Enabling Retriever-LLM Connection across the Semantic Gap in Retrieval-Augmented Generation
Ye, F.-D., Li, S.-Y., Zhang, Y.-Q., Chen, L., & Qu, S.-B. (2026) · ACM Transactions on Knowledge Discovery from Data · doi:10.1145/3847108
“For NQ dataset, we utilize the NQ-10 version standardized by [62], which contains 10 candidate documents per question.”
InfluentialUncertainty-Aware Context Compression for Retrieval-Augmented Generation: A Training-Free Framework
Li, Z., Zhao, S., Wang, M., & Zhao, Y. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-2862-1_8
FictionRAG: A Stateful Metacognitive Framework for High-Fidelity Long-Narrative Role-Playing
Deng, Y., Zhang, Y., Yang, J., & Fang, M. (2026) · Algorithms · doi:10.3390/a19050383 · Cited by 2
Metadata
- Authors
- Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P.
- Affiliations
- Stanford University (US)
- University of California, Berkeley (US)
- Samaya AI, UK
- Corresponding author
- Liu, Nelson F.
- Venue
- Transactions of the Association for Computational Linguistics
- Publisher
- MIT Press
- Published
- 2024
- Volume
- 12
- Pages
- 157-173
- ISSN
- 2307-387X (ISSN-L)
- OpenAlex
- W4391876619
- Semantic Scholar
- 1733eb7792f7a43dd21f51f4d1017a1bffd217b5
- Cited by
- 1,612
- References
- 59
- Citations by year
- 2026: 901 · 2025: 516 · 2024: 172
- FWCI
- 233.95
- Type
- journal article
- Language
- en
- Open access
- diamond
- doi.org
- Licence
- Locations
- Transactions of the Association for Computational Linguistics (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Keywords
- long-context language models, multi-document question answering, position bias
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-09-27 (prompt extract-v2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.