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 901–1,000 of 1,597
Enabling Retriever-LLM Connection across the Semantic Gap in Retrieval-Augmented Generation
Ye, F., Li, S., Zhang, Y., Chen, L., & Qu, S. (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
Chen, Y., Zeng, Z., & Hassan, M. S. (2026) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2026.1784200
Fan, Y.-C., Yang, Z., Qin, H.-T., & Wu, S. (2026) · International Journal of Multimedia Information Retrieval · doi:10.1007/s13735-026-00413-y
Graph Representation of Tables+Text and Compact Subgraph Retrieval for QA Tasks
Kumar, V., Sen, J., Chelani, B., & Chakrabarti, S. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-88708-6_11 · Cited by 1
Gueddes, A., Louhichi, B., & Mahjoub, M. A. (2026) · Cluster Computing · doi:10.1007/s10586-026-06436-8
“Empirical studies have demonstrated that language models exhibit a pronounced positional bias, whereby information located in the middle of a long context is systematically under-attended—the so-called “Lost in the Middle” phenomenon [10 ].”
Explainable Fake News Detection via Multi-Aspect Semantic Discovery
Nguyen, H.-D., & Le, T. (2025) · Procedia Computer Science · doi:10.1016/j.procs.2025.09.446 · Cited by 1
Chatbot-Assisted Explainability of AI-Powered Models and Collection of User Feedback
Razuvayevskaya, O., Foster, M., & Bontcheva, K. (2026) · Springer eBooks · doi:10.1007/978-3-032-11782-3_13
Retrieval-Augmented Generation for Smart Classroom Applications
Zhengyuan, F., Yan-li, L., & Zhang, H. (2025) · International Conference on Computing, Networks and Internet of Things (CNIOT) · doi:10.1109/cniot65435.2025.11070736 · Cited by 1
“Then, the classroom recording is fed into the large model, and fifty pieces of data related to the content of the lesson are refined through prompt words [12].”
Document Haystack: A Long Context Multimodal Image/Document Understanding Vision LLM Benchmark
Huybrechts, G., Ronanki, S., Jayanthi, S. M., FitzGerald, J., & Veeravanallur, S. (2025) · IEEE International Conference on Computer Vision Workshops · doi:10.1109/iccvw69036.2025.00428 · Cited by 5
“Nova Lite exhibits a mild manifestation of the “lost in the middle” phenomenon described in [36], with slightly decreased accuracy in middle-depth ranges.”
Covert Knowledge Poisoning Attacks in Retrieval-Augmented Code Generation
Zhang, Y., Chen, M., He, X., Cong, T., Xu, K., & Li, Q. (2026) · IEEE Transactions on Dependable and Secure Computing · doi:10.1109/tdsc.2026.3686845
“In RAG, this enables information exchange across retrieved document boundaries [39], [40].”
FACTUM: Mechanistic Detection of Citation Hallucination in Long-Form RAG
Dassen, M., Kotula, R., Murray, K., Yates, A., Lawrie, D., Kayi, E. S., Mayfield, J., & Duh, K. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-21289-4_18 · Cited by 9
Generative Optimization for Incentivized Advertising with Global Level Constraints
Chen, G., Luo, N., Jiang, H., Li, D., Shu, W., Sha, T., Zeng, Y. J., Tai, W., Zhou, F., & Liu, X. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3818423
[Withdrawn] AI Agents Need Memory Control Over More ContextPreprint
Bousetouane, F. (2026) · Qeios · doi:10.32388/mzqb3t
Wu, J., Li, J., Ning, Z., & Wang, L. (2026) · Machine Learning · doi:10.1007/s10994-026-07109-8
“…often contain redundancy, noise, or partially conflicting information, and the most decisive supporting evidence may be buried beneath passages that are topically related but not truly helpful for answering the ques - tion (Yoran et al., 2024; Liu et al., 2024; Niu et al., 2024 ), as shown in Fig.”
Rohde, W. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6577818
Jia, H., He, M. X., Yin, Z., Wu, L. J., Fan, J., & Sang, J. (2026) · International Conference on Automation Control, Algorithm and Intelligent Bionics (ACAIB) · doi:10.1145/3828836.3828924
Fianko, S. K., Addo, A., Agyemang, K., Gbadago, F. Y., Adjaye-Gyamfi, O., Quartey, O., Agropah, F., Larttey, R., Amoah, N., & Nooni, I. K. (2026) · Discover Education · doi:10.1007/s44217-026-01837-8
“Context-length effects may compound this, since appending full marking schemes substantially increases input length and research has shown that LLM performance degrades as relevant information becomes embedded within extended prompts [35].”
Unified Cognitive Consciousness Theory: Formalization, Validation, and Outlook
Chang, E. Y. (2025) · ACM eBooks · doi:10.1145/3749421.3749438
C-P&M: Unified cross-prompt and cross-model adversarial attack on multimodal large language models
Bao, Q., Yan, Y., Ding, X., Qian, Y., Gu, Z., Ji, S., & Wang, B. (2026) · Neurocomputing · doi:10.1016/j.neucom.2026.134821 · Cited by 1
SafeSQL-LLM: A Synthetic Data Approach to Privacy-Preserving Text-to-SQL
Zhang, R., & Zhao, S. (2025) · IEEE International Conference on Parallel and Distributed Systems (ICPADS) · doi:10.1109/icpads67057.2025.11323205 · Cited by 1
“This gap aligns with established observations: relevant evidence buried mid-context is used unreliably in long inputs [29], so direct answering suffers from retrieval and coverage errors; replacing free-form outputs with an executable intermediate (SQL) offloads retrieval and algebra to the…”
Syah, R. A., Santi, O. I., Witjaksono, G., & Kautsar, I. A. (2026) · International Electronics Symposium (IES) · doi:10.1109/ies71033.2026.11702632
“At k =10 , accuracy degrades, consistent with the “lost in the middle” phenomenon reported in long-context LLM evaluations [13].”
Minimal-Context RAG: Structured Context Compression with Hierarchical Quality Control
Chen, Y., Wang, K., Huachao, C., Cao, Z., & Zhao, J. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-3426-4_12
Towards Self-Evolving Agents: A Dual-Process Framework for Continual Context Refinement
Teng, L., Ni, W., Song, L., Shi, J. F., & Li, Y. (2026) · Electronics · doi:10.3390/electronics15061232 · Cited by 1
Teaching Recommenders to Listen: Bidirectional Natural-Language Control for Personalized News
Hasan, M. (2026) · ACM Conference on Recommender Systems (RecSys) · doi:10.1145/3773078.3831940
Retrieval augmented generation for open-set entity alignment with large language models
Yang, L., Wang, X. J., Ding, W., Chen, H., & Chen, L. (2026) · Expert Systems with Applications · doi:10.1016/j.eswa.2026.132489 · Cited by 1
Carroll, M., Kentis, S., Kareff, H., Schechter, C. B., & Jariwala, S. P. (2026) · Applied Clinical Informatics · doi:10.1055/a-2946-7393 · Cited by 1
“40 Indeed, these challenges persist despite advances in context length, as recent evaluations of LLMs for data extraction from unstructured and semistructured electronic health records have shown that performance can degrade significantly according to the position of relevant information with…”
Controlled knowledge updating in memory-enabled AI agents: beyond model editing and recall
Chavira, G., Bravo, E. J. G., García, G. E. R., Espinosa, J. A. A., & Nava, S. W. (2026) · Frontiers in Big Data · doi:10.3389/fdata.2026.1938137
Towards Automated Page Object Generation for Web Testing using Large Language Models
Karagöz, B., Ricca, F., Biagiola, M., & Stocco, A. (2026) · IEEE International Conference on Software Testing, Verification and Validation (ICST) · doi:10.1109/icst69053.2026.00055
“These raw HTML files, however, often contain excessive or redundant data, such as scripts and embedded resources, that could easily overload the context for LLMs [29], [30].”
How Does Regulation Travel with Offshore Financing? Evidence from Initial Coin OfferingsPreprint
Wang, S. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6018394
Shi, P., Shui, Y., Jin, R., & Pan, Q. (2026) · Journal of King Saud University - Computer and Information Sciences · doi:10.1007/s44443-026-00825-0 · Cited by 1
Towards an FA ChatBot with Retrieval-Augmented Language Modeling
Fichtenkamm, M., Kofler, M., Schekotihin, K., & Burmer, C. (2024) · IEEE International Symposium on the Physical and Failure Analysis of Integrated Circuits (IPFA) · doi:10.1109/ipfa61654.2024.10691083 · Cited by 1
“[9] • Semantic re-rankers measure the similarity between the query and retrieved documents leveraging specialized re-ranker models [7].”
BackgroundAgentCite: Task-Isolated Multi-Agent Collaboration for Cost-Effective Long-Context Citations
Cao, Z., Wang, H., Jin, L., Zhu, T., & Ren, L. (2026) · International Symposium on Autonomous Systems (ISAS) · doi:10.1109/icaisisas68969.2026.11567713
“As revealed in Lost in the Middle [9], LLMs exhibit severe performance degradation when extracting facts buried deep within long contexts.”
Repository-Level Code Generation Method Enhanced by Context-Dependent Graph Retrieval
Zhang, H., Li, A., Kuai, Z., Wei, Z., Kuang, L., & Xia, Y. (2025) · Lecture notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering · doi:10.1007/978-3-031-93257-1_3
When it Comes to Scientific Information Extraction and LLMS, Less is More
Shaik, S., Blaiszik, B. J., & Tchoua, R. (2026) · International Journal of Semantic Computing · doi:10.1142/s1793351x26420018
Metadata, Structure, or Strategy? A Decomposition of RAG Context Enrichment
Zerhoudi, S., Granitzer, M., & Mitrović, J. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-37676-3_4 · Cited by 1
Feng, X., Li, H., & Xu, L. (2026) · International Conference on Artificial Intelligence and Industrial Technology Applications (AIITA) · doi:10.1109/aiita69518.2026.11567060
“[3], [4], [5].”
RegMem: Graph-Based Regulatory Memory for Faithful Multi-turn RAG in Safety-Critical Domains
Teixeira, A. L. B. D. A., Presa, J. P. C., & de Oliveira, S. S. T. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-39882-6_33
Basu, S., Patel, S., Sheth, P., Morgan, J., & Batniji, R. (2026) · Research Square · doi:10.21203/rs.3.rs-10153003/v1
SmartCLAI: A Unified Multi-Agent CLI for Automated Network Configuration
HOSSAINI, M. H., Kassler, A., & Hayes, M. (2026) · IEEE/IFIP Network Operations and Management Symposium (NOMS) · doi:10.1109/noms69089.2026.11668335
“Template-driven tools need continual maintenance due to ongoing template evolution, while general-purpose LLMs often hallucinate or produce incomplete commands due to missing domain context [4]–[7].”
A Process for Developing a Traffic Safety Domain Specific Language Model Using Fine-Tuning and RAG
Lee, Y., Kim, Y., Moon, S., Joo, S., Lee, J., & Yun, I. (2025) · The Journal of The Korea Institute of Intelligent Transport Systems · doi:10.12815/kits.2025.24.5.54
Xue, Y., Tsunokake, M., Koreeda, Y., & Sogawa, Y. (2026) · International Journal of Semantic Computing · doi:10.1142/s1793351x26410023
Zhang, X., Du, J., Guo, B., & Xiang, M. (2025) · Research Square · doi:10.21203/rs.3.rs-6288322/v1
de Mattos, E. U., Nascimento, E., Amoedo, D. A., Guimarães, V., Matias, P. V., Oliveira, E., Naja, L., Castro, M., OLIVEIRA, P. R., Uchôa, M. S., et al. (2026) · International Conference on Consumer Electronics · doi:10.1109/icce67443.2026.11449886 · Cited by 1
“Furthermore, these studies often fail to report benefits under real-world scenarios, such as installation and configuration steps and system diagnostics, which require different levels of context and tolerance for formulation variations [10].”
Branchat: A Tree-Structured Interface for Efficient Revisitation in Long-Horizon LLM Conversations
Li, J., Qin, P., & Lee, Y. (2026) · CHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA) · doi:10.1145/3772363.3798792 · Cited by 1
Anomalies in AI Outputs Beyond Input Data Quality: The Significance of Reasoning
Royo, J. [. ]. A., Martín, Ó., Martínez, E., Montalvo, F. J. F., & Blanco‐Fernández, J. (2026) · Applied Sciences · doi:10.3390/app16115491
Fast or Slow? Human-Inspired Self-Evolving Framework for Resilient AI Systems
Qiu, H., Cao, P., Cui, S., Patke, A., & Iyer, R. K. (2026) · IEEE/IFIP International Conference on Dependable Systems and Networks - Supplemental Volume (DSN-S) · doi:10.1109/dsn-s70715.2026.00048
“As interaction histories grow, LLM performance degrades due to context truncation, attention dilution, and lost-in-the-middle effects [17], [28], [33].”
Queryome: orchestrating retrieval, reasoning, and synthesis across biomedical literature
Punuru, P., Ibtehaz, N., Giri, S. J., Srirangam, H., Tugolukova, E. A., & Kihara, D. (2026) · Journal of Biomedical Semantics · doi:10.1186/s13326-026-00366-5 · Cited by 3
Chen, H., 吴贞东, Pang, X., Yang, Y., & Feng, J. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-3435-6_34
Comparing Knowledge Injection Methods for LLMs in a Low-Resource Regime
Abonizio, H., Almeida, T., Lotufo, R., & Nogueira, R. (2025) · doi:10.5753/eniac.2025.14220 · Cited by 3
Liu, L., Yao, S., Liu, Y. P., Wang, Y., Xia, H., & Tang, X. (2026) · Information Processing & Management · doi:10.1016/j.ipm.2026.105027
LLM as a law professor: having a large language model write a commentary on freedom of assembly
Engel, C., & Kruse, J. (2026) · Artificial Intelligence and Law · doi:10.1007/s10506-026-09527-4 · Cited by 2
HAVEN: Hierarchical Attention for Viral protEin-based host iNferencePreprint
Antony, B., Haghani, M., Lauring, A. S., Karpatne, A., & Murali, T. M. (2025) · bioRxiv (Cold Spring Harbor Laboratory) · doi:10.1101/2025.06.09.658367 · Cited by 1
“The performance of transformer-based language models degrades for long contexts, e.g., by paying higher attention to the head and tail portions of “long” input sequences compared to their middle parts [26].”
Benchmarking Serverless AI Architectures: Modular RAG, Serverless RAG, and Long Context Inference
Petrusel, F. A. S., & García-López, P. (2026) · ACM International Conference on Distributed and Event-Based Systems (DEBS) · doi:10.1145/3809481.3816479
Bati, V. S., B, A. C., Sindhya, S. H., & Kalpa, R. (2026) · International Conference on Knowledge Engineering and Communication Systems (ICKECS) · doi:10.1109/ickecs70176.2026.11528014
“The “Lost in the Middle” phenomenon, highlighted in recent RAG literature [19], demonstrates how critical information can be overlooked when documents are fragmented or when key provisions appear far from retrieved passages.”
Vithanage, D., Duong, Q. V., Deng, C., Acharya, R. U., Gu, Z., Song, T., & Yu, P. (2026) · Journal of Healthcare Informatics Research · doi:10.1007/s41666-026-00254-8 · Cited by 1
Attwa, M. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6572624
Николаевич, Г. В. (2026) · Modelirovanie, optimizaciâ i informacionnye tehnologii. · doi:10.26102/2310-6018/2026.60.9.006
Zhou, H., Skalski-Fouts, Z., Nguyen, L., Rivera, C., Spring, B., Miao, H., & Zhang, X. (2026) · Smart Health · doi:10.1016/j.smhl.2026.100694
Risk-Informed Maintenance of Battery Energy Storage Systems with Knowledge GraphsPreprint
Qu, J., Fu, Y., Lu, Y., Zhang, P., Li, M., Li, W., & Wang, Y. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7477235
Zhou, A., Wu, Z., Peng, K., Liu, P., & Li, X. (2026) · Lecture notes in computer science · doi:10.1007/978-981-95-6960-1_22 · Cited by 1
Wang, Y., Li, H., & Chen, L. (2026) · Proceedings of the VLDB Endowment · doi:10.14778/3819518.3819565 · Cited by 1
NumCache: KV Cache Compression and Retrieval for Financial Document QA
Makri, E., Li, P., Jiang, Y., Chen, J., Chen, J., Maatouk, A., Tassiulas, L., Brenner, E., Xiang, B., & Ying, R. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3818131
Talking to Ourselves Through a Smart Mirror: Artificial Confidence in Human–AI Interaction
Hochman, G. (2026) · Systems · doi:10.3390/systems14060627 · Cited by 3
Hussain, R., Ma, Z., Khandelwal, R., Oltmanns, J. R., & Gupta, M. (2026) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2026.1760246
“This limitation is commonly referred to as the lost-in-the-middle phenomenon (An et al., 2024; Beltagy et al., 2020; Liu et al., 2024; Huang et al., 2023).”
Transformer Models and their Cross-Domain Applications: A Systematic Review
Pathania, K., Singh, S., & Sharma, A. (2026) · Archives of Computational Methods in Engineering · doi:10.1007/s11831-026-10742-0
Wie funktioniert ein Sprachmodell?
Lang, F. U. (2025) · Essentials · doi:10.1007/978-3-658-49885-6_2
Task-Aligned Contrastive Learning for Filtering Noise in Multi-label Text Classification
Metzger, A., Dantu, R., Quach, V., & Blackwell, A. (2026) · Lecture notes in networks and systems · doi:10.1007/978-3-032-35078-7_6
Large language models enable consensus-level interpretation in metagenomic diagnosticsPreprint
Steinig, E. J., Krysiak, M., Deo, K., Duncan, A., Prestedge, J., Barr, J. J., Moselen, J. M., Khan, S. F., Fernando, J. A., Savic, I., et al. (2026) · medRxiv · doi:10.64898/2026.07.29.26358751 · Cited by 1
An Evidence-Guided Reasoning Framework for Explainable Chinese Health Misinformation Detection
Liu, Y., & Xu, G. (2026) · Systems · doi:10.3390/systems14101232
“Long contexts can impair the use of relevant information, especially when critical evidence is poorly positioned [31].”
LadRanker: Making LLMs All-Round Rankers for Recommendation
Liu, M., Sun, Y., Lam, C. K., Wang, K., Jiang, H., Wang, Y., Zhang, W., & Lin, J. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-3435-6_1 · Cited by 2
DTopAgent: A Multi-Agent Framework for Dynamic Top-K Chunk Retrieval in Rag Pipeline
Wang, S., Liu, C., Duan, J., Zhao, Z., Yi, Z., & Li, X. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11464642
Xiong, Y.-T., Wang, Y.-C., Sun, Y., Zhan, Z., Liu, W. T., Liu, H.-N., Meng, R.-Y., Tang, W., & Liu, C. (2026) · BMJ Evidence-Based Medicine · doi:10.1136/bmjebm-2025-114055 · Cited by 1
“39 40 Revising prompts through CLP can help models follow complex decision criteria and produce more consistent outputs.”
Semantic Raft: A Fault-Tolerant Multi-Agent Inference Framework for Reliable LLM Services
Su, Y.-D., Wang, Y., Dong, Y., & Xu, Z. (2026) · International Workshop on Quality of Service · doi:10.1109/iwqos70441.2026.11661187
“These errors accumulate and amplify throughout the long chain, leading to Cascading Failures [3], [4], [5].”
Identifying Algorithmic and Domain-Specific Bias in Parliamentary Debate Summarisation
Cunningham, E., Cross, J., & Greene, D. (2026) · Communications in computer and information science · doi:10.1007/978-3-032-19108-3_28
Duan, Z., Zhao, G., Li, C., & Liu, B. (2026) · Biology · doi:10.3390/biology15090708
“Although recent models provide larger context windows, processing massive multi-page historical documents still often requires page-by-page processing or fixed-window chunking strategies, which introduce semantic discontinuities and computational redundancy [22].”
The Omission Gap: Defining the Professional Standard of Care for Human Review of AI Outputs.Preprint
Wright-Zhang, J. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6631319
Garcés, K., Fernandez-Nieto, G. M., Zhao, L., Samaraweera, S., Gašević, D., Martínez‐Maldonado, R., & Echeverría, V. (2026) · ACM Conference on Learning @ Scale (L@S) · doi:10.1145/3774398.3811613
“Our findings also align with broader claims that LLMs struggle with long-context reasoning, particularly when relevant information is not positioned near the beginning or end of the prompt, suggesting that long, batched prompts may dilute salient signals [32].”
Jeon, K., Ko, Y., Jung, W., & Kim, H. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3818040
“…(i) retrieval-augmentation with the severity-aware trajectories, the relevant clinical notes, and similar peer cases, (ii) dynamic context repacking, and (iii) a step-by-step reasoning process, our framework mitigates the “lost-in-the-middle” problem [25] inherent to long-context LLM inputs.”
SWin: A Sliding Window Summarization Approach for Coherent LLM-driven Dialogue Systems
Savant, S., Zhao, Y., Yan, J., & Rengarajan, N. K. K. (2026) · IEEE Conference on Artificial Intelligence (CAI) · doi:10.1109/cai68641.2026.11536307
“[5] The performance of open-source LLMs in summarization methods for long-range dialogue scenarios has not been thoroughly studied despite their increasing efficiency and strong performance on summarization benchmarks.”
Peng, C., Jiang, M., Zhou, Y., & Wu, L. (2026) · Proceedings of the ACM on Software Engineering · doi:10.1145/3808141
“Unlike traditional keyword-based approaches, semantic search understands the conceptual meaning behind queries.”
GHP: Hardening Graph‐Based Reranking Defense for Poisoned Retrieval‐Augmented Generation
She, W., Dong, Z., Yang, Y. W., Tian, Z., Liu, W., & Kong, D. (2026) · Concurrency and Computation Practice and Experience · doi:10.1002/cpe.70824 · Cited by 2
“The issue becomes especially important when the final evidence interface is fixed: Even a small number of highly ranked poisoned passages may still enter the prompt and affect generation, while models are not uniformly robust to the position of critical information within the context [ 9].”
Lin, Z., & Cai, J. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-3426-4_48
Neubig, S., Göhl, L., Radhakrishnan, R., Weirich, M., Hein, A., Keller, R., & Krcmar, H. (2026) · Information Technology & Tourism · doi:10.1007/s40558-026-00363-5
“In contrast to previous observations (Liu et al. 2024), this did not depend on the placement within the prompt.”
Yu, Z., Wang, C., Xu, M., Huang, Y. J., Tian, K., & Qin, Y. (2026) · Computers · doi:10.3390/computers15080487
Chain-of-Detection Enables Robust and Efficient Jailbreak DefensePreprint
Wu, T., & Zhang, H. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5124466
ClauseRoute: Risk-aware clause-guided routing for retrieval-augmented legal reasoning
Du, Z., Wu, C., Lu, J., Gao, Y., & Me, C. (2026) · Expert Systems with Applications · doi:10.1016/j.eswa.2026.133633
Key-augmented neural triggers for knowledge sharing
Wolf, A., Palma, M. E., Rani, P., & Gall, H. C. (2026) · Journal of Systems and Software · doi:10.1016/j.jss.2026.112850 · Cited by 1
Structured Creativity: Enhancing Business Model Innovation with Generative AI
Scheiber, C., Böttcher, T. P., Fielt, E., Kowalkiewicz, M., Weber, M., & Krcmar, H. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-93979-2_13
SCA-Lex: A Lightweight Forensic Triage Framework Based on Few-Shot Semantic Centroid Alignment
da Silva, I., Ronzani, M. D., Gonçalves, A. L., & Panisson, A. R. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-39882-6_4
CRAWLDoc: A System for Contextual Ranking and Bibliographic Metadata Extraction from Web Resources
Karl, F., & Scherp, A. (2026) · Information Retrieval Research · doi:10.54195/irrj.23861
A Retrieval-Augmented Generation-Powered Chatbot for Air Force Policy and Logistics Compliance
Mahawar, V., Bhalla, S., McKee, J., & Boicu, M. (2025) · IEEE MIT Undergraduate Research Technology Conference (URTC) · doi:10.1109/urtc68753.2025.11533029
“Liu et al. [16] details the “lost in the middle problem” where language models can lose accuracy with long input.”
Comfrey: Mitigating Integration Failures in LLM-enabled Software at Run-Time
Shao, Y., Huang, Y., Zou, J., Shi, Y., Yang, L., Ma, L., Su, T., & Wan, C. (2026) · ACM/IEEE International Conference on Software Engineering: Companion Proceedings (ICSE-Companion) · doi:10.1145/3744916.3787847 · Cited by 2
From scoring to workflow: A survey of agentic recommender systems
Xu, Y., Mi, Y., Wu, Y., Shen, J., Wang, C., Zhao, G., Liu, C., & Qian, X. (2026) · Information Processing & Management · doi:10.1016/j.ipm.2026.105164
A Multi-Agent Collaborative and Controllable Reasoning Approach for Ship Building Process
Li, N., Yang, G., Cai, Y., Chen, C., Wang, T., & Cheng, L. (2026) · Youth Academic Annual Conference of Chinese Association of Automation (YAC) · doi:10.1109/yac71005.2026.11616153
“…confident conclusions without citing verifiable source text, making review and accountability difficult [2], [3]; (2) Long documents and noise sensitivity : technical manuals are lengthy and include substantial irrelevant content, which may cause unstable retrieval or off-topic generation [4], [5].”
Lin, Y., & Lee, Y.-T. (2026) · American Journal of Geriatric Psychiatry · doi:10.1016/j.jagp.2026.08.016
Designing Prompts for Shared-Experience-Focused Utterance Generation in LLM-Based Dialogue Systems
Fukushima, R., & Hashimoto, M. (2025) · Journal of Machine Intelligence and Data Science · doi:10.11159/jmids.2025.009
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.