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 601–700 of 1,597
CHASE: LLM Agents for Dissecting Malicious PyPI Packages
Toda, T., & Mori, T. (2025) · ACM International Conference on AI-Powered Software (AIware) · doi:10.1109/aiware69974.2025.00008 · Cited by 2
“1) State-Based Communication for Context Integrity: Conventional multi-turn conversations that accumulate message history are susceptible to context bloat and the “Lost in the Middle” problem [32].”
Kim, Y., Kim, M., Kim, S., Cho, S., & Kim, J. (2026) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3772318.3790971
Evaluating the utility of large language models for detecting and simulating language dysfunction
Cong, Y., Lee, J., Rajput, N., Tumacder, E., & LaCroix, A. N. (2026) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2026.1857106 · Cited by 1
“From a computational perspective (Liu et al., 2024; Touvron et al., 2023; Grattafiori et al., 2024), model performance tends to be strongest when relevant information appears at the beginning or end of the input, but declines substantially when key details are embedded in the middle of longer…”
Large Language Models for API Classification: An Explorative Study
Morais, G., Lemelin, E., Adda, M., & Bork, D. (2025) · International Conference on Evaluation and Assessment in Software Engineering (EASE) · doi:10.1145/3756681.3756997 · Cited by 2
Prompt Compression based on Key-Information Density
Lin, Y., Guo, W., Zhang, Y., Yang, C., & Li, Z. (2025) · Expert Systems with Applications · doi:10.1016/j.eswa.2025.127738 · Cited by 4
Venn Diagram Prompting: Accelerating Comprehension with Scaffolding Effect
Mahendru, S., & Pandit, T. (2024) · World Symposium on Artificial Intelligence (WSAI) · doi:10.1109/wsai62426.2024.10828919 · Cited by 1
Inspectable AI for Science: A Research Object Approach to Generative AI Governance
Binkytė, R., Abuaddba, S., Mahawaga, C., Ding, M., Fernandes, N., & Fritz, M. (2026) · IEEE Security and Privacy Workshops (SPW) · doi:10.1109/spw72489.2026.00039 · Cited by 6
“Regardless of configuration, it is advisable to place hard constraints at the beginning or end of the template, as LLMs tend to exhibit reduced attention to instructions embedded in the middle of longer prompts [19].”
Liu, Z., Huang, P., Xu, Z., Li, X., Liu, S., Peng, C., Xin, H., Yan, Y., Wang, S., Han, X., et al. (2026) · AI Open · doi:10.1016/j.aiopen.2026.02.002 · Cited by 11
EviMMQA: Multimodal question answering for medical evidence extraction in systematic reviews
Ji, C., Li, Y., Wang, Y., He, W., Cheng, Y., Zhang, Y., Feng, R., & Zhang, X. (2026) · Pattern Recognition · doi:10.1016/j.patcog.2026.113441 · Cited by 2
Zhang, D., & Chen, T. (2026) · Expert Systems with Applications · doi:10.1016/j.eswa.2026.133291 · Cited by 2
Attend and Replay: Efficient Action Understanding in Long Videos via Mechanistic Interpretability
Hou, P., Zhang, J., & Huang, D. (2025) · IEEE International Conference on Computer Vision Workshops · doi:10.1109/iccvw69036.2025.00688 · Cited by 1
“Early work on mechanistic interpretability examined attention patterns in LMs [1, 16, 28, 34, 35].”
TLR: Codebase-Level C Memory Management Error Repair with Large Language Models
Cheng, X. Y., Guo, Z., Huo, H., & Sui, Y. (2026) · Proceedings of the ACM on software engineering. · doi:10.1145/3797085 · Cited by 1
“This presents a fundamental challenge for LLMs, which are constrained by token limitations and exhibit performance degradation with extensive prompts—a phenomenon known as “lost in the middle” [33].”
FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory
Wei, L., Dong, X., Peng, X., Xie, N., & Wang, B. (2026) · ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp55912.2026.11462095 · Cited by 1
ReCompGPT: An NL2NL Framework for Automated Requirements Completeness
Sheng, T., Wang, L., & Liu, H. (2025) · IEEE Transactions on Software Engineering · doi:10.1109/tse.2025.3613507 · Cited by 2
“This configura-tion risks causing the LLM to overlook intermediate content (instructions) while disproportionately focusing on the concluding part (content including the action sequence) [49].”
Too Noisy To Learn: Enhancing Data Quality for Code Review Comment Generation
Liu, C., Lin, H. Y., & Thongtanunam, P. (2025) · International Conference on Mining Software Repositories (MSR) · doi:10.1109/msr66628.2025.00043 · Cited by 29
“This may be because the additional details of P A UXILIARY and code context increase the prompt length and distract the models, potentially leading them to overlook crucial criteria in longer prompts [47].”
Busch, M., Tacke, M., Lamaka, S. V., Zheludkevich, M. L., Cyron, C. J., Feiler, C., & Aydin, R. C. (2026) · Journal of Chemical Information and Modeling · doi:10.1021/acs.jcim.6c00961 · Cited by 6
A Pilot Empirical Study on When and How to Use Knowledge Graphs as Retrieval Augmented Generation
Yuan, X., Liu, Y., Di, S., Wu, S., Zheng, L., Meng, R., Chen, L., Zhou, X., & Yin, J. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-0372-7_29 · Cited by 6
CoRe-DoS: Inference-time denial-of-service attack against retrieval-augmented generation
Sun, H., Yuan, D., Li, M., & Deng, Y. (2026) · Computer Networks · doi:10.1016/j.comnet.2026.112567 · Cited by 1
Pinzuti, E., Tüscher, O., & Castro, A. F. (2025) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2025.1706090 · Cited by 4
“Prior research has shown that small models exhibit rapid degradation in long-context retention, losing attention to early examples even within the formal context limit Liu et al. (2023).”
Confidence-Conflict Routing Between Parametric Memory, RAG, and Long-Context LLMs
Fang, R., Zhou, J., Zhu, Z., Hu, R., & Zheng, Y. (2026) · International Conference on Electronic Technology and Information Science (ICETIS) · doi:10.1109/icetis70504.2026.11633419 · Cited by 1
“When retrieved passages contain conflicting or misleading information, even a well-calibrated RAG system can produce incorrect answers [7].”
Cammarota, S., Ferrante, M., Carosi, A., D’Angelillo, R. M., & Toschi, N. (2026) · Medical Physics · doi:10.1002/mp.70258 · Cited by 3
“This could be attributable to the limited context window of GPT-4 (which may restrict the model’s ability to fully exploit additional spatial information in the Multi-View case 11 ) and the fact that the axial view alone often sufficed to inform good beam choices for these cases.”
Can AI match professional analysts? Evidence from a multi-agent system
Francis, A., & Zhang, C. X. (2026) · Finance research letters · doi:10.1016/j.frl.2026.110127 · Cited by 1
Generating formal smart-contract specifications: comparing few-shot learning and fine-tuned LLMs
Leite, G., Arruda, F., Antonino, P., Sampaio, A. C. A., & Roscoe, A. W. (2026) · Science of Computer Programming · doi:10.1016/j.scico.2026.103464 · Cited by 2
Improving few-shot named entity recognition with distilled knowledge from large language model
Li, Q., Xie, T., Su, J., Zhang, J., & Wang, H. (2026) · Neurocomputing · doi:10.1016/j.neucom.2026.133657 · Cited by 1
Salvatore, N., & Zhang, Q. (2025) · Communications Psychology · doi:10.1038/s44271-025-00322-6 · Cited by 2
“There is evidence that pre-trained LLMs exhibit striking similarities to human episodic recall, including serial position effects 63,64 and temporal contiguity 65,66 , without being explicitly trained on memory tasks.”
Are They All Good? Evaluating the Quality of CoTs in LLM-Based Code Generation
Zhang, B., Zhang, L., Luo, Z., Du, Y., Liu, F., Wang, S., & Shi, L. (2026) · IEEE Transactions on Software Engineering · doi:10.1109/tse.2026.3676295 · Cited by 10
“This limitation is consistent with the lost in the middle phenomenon identified by Liu et al. [28], which reveals that LLMs struggle to access relevant information when it is not positioned at the beginning or end of the context.”
Beyond More Context: How Granularity and Order Drive Code Completion Quality
Yusuf, U., Caumartin, G., & Costa, D. E. (2025) · IEEE/ACM International Conference on Automated Software Engineering Workshops (ASEW) · doi:10.1109/asew67777.2025.00075 · Cited by 1
“Moreover, overloading a model’s context window can lead to noise or affect its performance [6].”
GuARD: Effective Anomaly Detection through a Text-Rich and Graph-Informed Language Model
Pang, Y., Chen, B., Zhang, F., Rao, Y., Kharlamov, E., & Tang, J. (2025) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3711896.3736993 · Cited by 9
GeAR: Generation Augmented RetrievalPreprint
Liu, H., Huang, S., Liu, J., Zhan, Y., Sun, H., Deng, W., Sun, F., Wei, F., & Zhang, Q. (2025) · Qeios · doi:10.32388/gwq6lr · Cited by 1
From jargon to clarity: bridging understanding through graded simplification of legal data
Adibhatla, H. S., Mukherjee, A., & Shrivastava, M. (2026) · Artificial Intelligence and Law · doi:10.1007/s10506-026-09503-y · Cited by 1
GenAI and the Mirage of Personalised Learning for All
Alimardani, A., & Jane, E. A. (2025) · Law Technology and Humans · doi:10.5204/lthj.3764 · Cited by 5
DePriest, K. N., Feher, J., III, Gore, K., et al. (2025) · Journal of Public Health Management and Practice · doi:10.1097/phh.0000000000002148 · Cited by 1
Dynamic and Parametric Retrieval Augmented Generation
Su, W., Dong, Q., Ai, Q., & Liu, Y. (2025) · Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (SIGIR-AP) · doi:10.1145/3767695.3769672 · Cited by 12
Learning from Earnings Calls: Graph-Based Conversational Modeling for Financial Prediction
Yang, Y., Tang, Y., Fan, Y., & Zhang, K. (2026) · Information Systems Research · doi:10.1287/isre.2023.0519 · Cited by 6
Retrieval-Based Knowledge Consistency Validation for Fake News Detection
An, R., Wang, J., Zhang, R., Zhang, J., & Luan, R. (2026) · Lecture notes in computer science · doi:10.1007/978-981-95-5719-6_28 · Cited by 1
Enhancing Legal Question Answering in Brazilian Portuguese through Domain-Specific Embedding Models
Novais, A. M. A., Ferreira, D. O. C., Silva, J. P. C., Brakes, M. F. C., Presa, J. P. C., & de Oliveira, S. S. T. (2025) · doi:10.5753/eniac.2025.13841 · Cited by 1
“However, the effectiveness of RAG is susceptible to challenges such as the ”lost in the middle” problem, where models struggle to utilize information located in the middle of long contexts [Liu et al. 2024].”
POI Extraction From Digital City: An Engineering Exploration With Large‐Language Models
Sun, M., Li, Q., Pan, J., Wei, H., Liu, C., Xu, X., & Li, Y. (2025) · Expert Systems · doi:10.1111/exsy.70001 · Cited by 1
“This method, inspired by (Liu et al. 2024), suggests that placing key information at both the beginning and end of prompts improves LLMs accuracy.”
: A Voting-Based Paradigm for Enhancing Retrieval Augmented Generation
Guan, W., Li, X., Lu, J., & Zhou, J. (2024) · Lecture notes in computer science · doi:10.1007/978-3-031-78119-3_9 · Cited by 1
Addie, Y. O., & McNealy, J. E. (2025) · Oxford University Press eBooks · doi:10.1093/9780198945215.003.0195 · Cited by 1
Peykani, P., Charles, V., Emrouznejad, A., Ghanidel, S., & Sabour, D. (2026) · Archives of Computational Methods in Engineering · doi:10.1007/s11831-026-10720-6 · Cited by 1
PatchPorter: LLM-Driven Security Patch Porting via Version Tracing and Context Selection for NPM
Yu, Z., Wen, M., Wei, Z., Wu, Y., Zou, D., & Jin, H. (2026) · Proceedings of the ACM on software engineering. · doi:10.1145/3832287
“Fixed granularity context can be excessive for self-contained patches and insufficient for dependency introducing patches, which might increase attention noise and the risk of hallucination [15, 23].”
Koraag, T., Wagner, N., Dobslaw, F., & Gren, L. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5489581 · Cited by 2
Quantifying Uncertainty: A New Era of Measurement Through Large Language ModelsPreprint
Audrino, F., Gentner, J., & Stalder, S. (2024) · SSRN Electronic Journal · doi:10.2139/ssrn.4998319 · Cited by 4
Predicate Renaming via Large Language Models
Gentili, E., Ribeiro, T., Riguzzi, F., & Inoue, K. (2026) · Machine Learning · doi:10.1007/s10994-026-07075-1 · Cited by 1
“This behavior is referred to as primacy effect and recency effect, respectively (Alammar and Grootendorst (2024),Liu et al. (2024a)).”
Uma, S., Ghosh, S., & Mustaqim, D. D. A. (2026) · ACM International Conference on Intelligent User Interfaces (IUI) · doi:10.1145/3742413.3789130
A State-Aware Dual-Tool Architecture for Improved Tool Use in Multi-Turn LLM Dialogue
Pałgan, B., Wawryszczuk, M., Chwał, J., Tkacz, E., & Dzik, R. (2026) · Applied Sciences · doi:10.3390/app16094329 · Cited by 1
GPT-5 vs Other LLMs in Long Short-Context Performance
Esmi, N., Nezhad-Moghaddam, M., Borhani, F., Shahbahrami, A., Daemdoost, A., & Gaydadjiev, G. (2025) · 2025 3rd International Conference on Foundation and Large Language Models (FLLM) · doi:10.1109/fllm67465.2025.11391194 · Cited by 1
“…highlights how models exhibit primacy and recency biases, performing best when relevant information is at the beginning or end of the input context, while struggling to access and utilize data in the middle, as evidenced in tasks like multi-document question answering and key-value retrieval [1].”
PReMM: LLM-Based Program Repair for Multi-method Bugs via Divide and Conquer
Xie, L., Li, Z., Pei, Y., Wen, Z., Liu, K., Zhang, T., & Li, X. (2025) · Proceedings of the ACM on Programming Languages · doi:10.1145/3763097 · Cited by 13
“This observation aligns with our earlier concerns that excessive context may mislead LLMs, mirroring findings by Liu et al [33].”
Scoring Structured Academic Documents with Large Language Models: Impact Case Studies
Thelwall, M. A., Kousha, K., & He, G. (2026) · Journal of Data and Information Science · doi:10.1515/jdis-2025-0465 · Cited by 1
“Whilst it is known that LLMs can sometimes struggle to fully access information in long documents (Du et al. 2025; Liu et al. 2024) and may therefore not be e ff ective at processing full ICSs, this has not given weaker correlations.”
DA-RAG: Dynamic Attributed Community Search for Retrieval-Augmented Generation
Zeng, X. L., Wu, Z., Wang, Y., Zhang, C., Yao, Q., Zheng, L., & Yin, J. P. (2026) · International World Wide Web Conference (WWW) · doi:10.1145/3774904.3792430 · Cited by 2
AixelAsk: A Stepwise-Guided Retrieval and Reasoning Framework for Large Table QA
Zhang, C., Zhang, M., Yang, Y., Chen, T., & Luo, Z. (2025) · Proceedings of the ACM on Management of Data · doi:10.1145/3769831 · Cited by 8
Research on a Dual‐Mode Human–Robot Interaction Method Based on a Large Language Model
Guo, J., & Han, X. (2026) · IET Cyber-Systems and Robotics · doi:10.1049/csy2.70037 · Cited by 2
Park, J., Min, B., Son, K., Song, J. Y., Ma, X., & Kim, J. (2026) · ACM International Conference on Intelligent User Interfaces (IUI) · doi:10.1145/3742413.3789107 · Cited by 52
“For both quantitative and self-reported survey measures, we used the Friedman test with the Wilcoxon signed-rank test for post-hoc analysis to statistically compare the three conditions.”
Tzikas, I., Didovic, S., Gerschner, F., Rössle, M., Theissler, A., & Klaiber, M. (2025) · Procedia Computer Science · doi:10.1016/j.procs.2025.09.309 · Cited by 4
Barbosa, A., Silveira, I. C., & Mauá, D. D. (2025) · Journal of the Brazilian Computer Society · doi:10.5753/jbcs.2025.5817 · Cited by 5
“In contrast, the smaller Phi3 collapses on C2 (QWK –0.03), a pattern consistent with the lost-in-the-middle effect — where long-context models struggle to retrieve information located far from the input boundaries [ Liu et al ., 2024].”
InfluentialFaridi, A. R., Masood, F., Keshvi, & Yunus, T. (2026) · Wiley Interdisciplinary Reviews Data Mining and Knowledge Discovery · doi:10.1002/widm.70122 · Cited by 2
“Controlled studies reveal a strong sensitivity to the location of relevant evidence in long inputs (a U-shaped, “primacy/recency” pattern), with impaired performance in the case of important information in the middle of the context (Liu et al. 2024).”
InfluentialA Quantitative Evaluation of Natural Language Retrieval Methods for OpenAPI Specifications
Palamootil, S. M., & Grobler, J. (2025) · Lecture notes in computer science · doi:10.1007/978-3-032-05179-0_23 · Cited by 1
Clouseau: A Hierarchical Multi-Agent Approach for Autonomous Attack Investigation
Aldaihan, A., Alotaibi, F., & Maffeis, S. (2025) · Annual Computer Security Applications Conference · doi:10.1109/acsac67867.2025.00051 · Cited by 11
“Managing the context size of LLMs is crucial: unbounded growth leads to increased hallucination rates and degrades overall performance [58], [59].”
CodeMind: Evaluating Large Language Models for Implicit and Explicit Code Execution Reasoning
Liu, C., Chen, Y., & Jabbarvand, R. (2026) · IEEE Transactions on Software Engineering · doi:10.1109/tse.2026.3691250 · Cited by 87
“3-b), challenging the attention span of LLMs [40].”
Automatic Minds: Cognitive Parallels Between Hypnotic States and Large Language Model Processing
Riva, G., Wiederhold, B. K., & Mantovani, F. (2025) · Cyberpsychology Behavior and Social Networking · doi:10.1177/21522715251400733 · Cited by 2
“52 This is analogous to state-dependent memory, where information learned in one state of consciousness is harder to access in another.”
Liu, Y., Jia, Z., Gao, X., Xu, K., & Xiong, Y. (2026) · International World Wide Web Conference (WWW) · doi:10.1145/3774904.3792700 · Cited by 8
Global Business NetworksPreprint
Breitung, C., & Müller, S. (2023) · SSRN Electronic Journal · doi:10.2139/ssrn.4395079 · Cited by 1
CyGPT: Knowledge Graph-Based Enhancement Techniques for Large Language Models in Cybersecurity
Ou, L., Ni, X., Wu, W., & Tian, Z. (2024) · IEEE International Conference on Data Science in Cyberspace · doi:10.1109/dsc63484.2024.00036 · Cited by 4
“While document retrieval provides relevant information, its effectiveness is often hindered by inaccuracies in retrieval and document redundancy [17].”
KGMP: Augmenting retrieval knowledge graph with multi-hop perceptron
Yang, Z., Weng, L.-Y., Zhang, L., Tong, R.-J., Xie, J., Zeng, Z., & Chen, D.-B. (2025) · PLoS ONE · doi:10.1371/journal.pone.0333037 · Cited by 3
“It is worth noting that Liu et al. [43] empirically found that there is a significant “Lost in the middle” phenomenon in the knowledge retrieval process of large-scale language models, which is manifested in the following two aspects: (1) Document position sensitivity: When the document containing…”
InfluentialIntegrating Large Language Models and knowledge graphs to capture political viewpoints in news media
Fadda, M., Motta, E., Osborne, F., Recupero, D. R., & Salatino, A. A. (2026) · EPJ Data Science · doi:10.1140/epjds/s13688-026-00678-7 · Cited by 2
“In particular, our main objective is to assess the value derived from i) enhancing claims through a semantic representation of the actors and ii) fine-tuning the models on high-quality data.”
Interpretation Gaps in Large Language Model-Assisted Comprehension of Privacy Documents
Dewri, R. (2025) · Computer · doi:10.1109/mc.2025.3575345 · Cited by 2
AGON: Automated Design Framework for Customizing Processors From ISA Documents
Li, C., Huang, D. S., Jin, P., Ma, T., Han, H., Cheng, S., Hao, Y., Zhao, Y., Xu, G., Du, Z., et al. (2025) · IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems · doi:10.1109/tcad.2025.3617851 · Cited by 2
“LLMs struggle to process information within long contexts effectively, especially when their length exceeds their training data [40].”
QUITO: Accelerating Long-Context Reasoning Through Query-Guided Context Compression
Wang, W., Wang, Y., Fan, Y., Liao, H., & Guo, J. (2025) · Lecture notes in computer science · doi:10.1007/978-981-96-1710-4_11 · Cited by 7
Exploring the use of open-source LLMs for Game Learning Analytics: an empirical study
Filho, D., Honda, F., Pires, F., & Pessoa, M. (2025) · Anais do ... Simpósio Brasileiro de Informática na Educação/Anais do Simpósio Brasileiro de Informática na Educação · doi:10.5753/sbie.2025.12885 · Cited by 2
“In addition, models tend to ignore information in the middle of the input when the prompt is long [Liu et al. 2024].”
Leveraging LLMs Towards Assistant-based Support for Industrial Threat Models
Fregnan, E., Göttel, C., Maag, B., Dawoud, A., & Nakas, G. (2025) · IEEE International Conference on Emerging Technologies and Factory Automation (ETFA) · doi:10.1109/etfa65518.2025.11205762 · Cited by 3
“However, a large k will also increase the number of tokens of the query and, hence, may increase the cost of running the assistant when using a paid LLM service and may decrease the LLMs ability to accurately answer the query due to effects like lost-in-the-middle [38].”
Alwakeel, M., Buck, E., Martin, J. G., Aslam, I., Rajagopal, S., Pei, J., Podgoreanu, M. V., Lindsell, C. J., & Wong, A.-K. I. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-09569-5_21 · Cited by 1
Reasoning Capabilities of Large Language Models on Dynamic Tasks
Wong, A., Bäck, T., Plaat, A., van Stein, B., & Kononova, A. V. (2025) · International Conference on Foundation and Large Language Models (FLLM) · doi:10.1109/fllm67465.2025.11391251 · Cited by 7
“This is consistent with the idea that extraneous context reduces signal-to-noise ratio and adds distraction [17], [18].”
Myers, S., Dligach, D., Miller, T. A., Barr, S., Landefeld, J., Gao, Y., Churpek, M. M., Mayampurath, A. M., & Afshar, M. (2026) · Journal of the American Medical Informatics Association · doi:10.1093/jamia/ocag139 · Cited by 10
“The medications in both the generated predictions and the gold data are normalized to their ingredients using the RxNorm [23] API provided by the National Library of Medicine and a handful of manual rules for edge cases, such as typos.”
InfluentialFrom Redundancy to Efficiency: Exploiting Shared UI Interactions towards Efficient LLM-Based Testing
Wang, X., Wang, Y., Hu, Y., Zhang, Y., Jin, H., Guo, S., Yuan, J., & Zhou, Y. (2025) · IEEE/ACM International Conference on Automated Software Engineering (ASE) · doi:10.1109/ase63991.2025.00284 · Cited by 1
“This submodule serves two purposes: First, MLLMs perform poorly when processing multiple screenshots simultaneously [36].”
LLM-based Contextual Query Generation on Relational Databases
Tapashetti, V., Motagi, A. U., & Azam, F. (2025) · International Conference on Intelligent Technologies (CONIT) · doi:10.1109/conit65521.2025.11167865 · Cited by 1
“However, the “middle forgetting” phenomenon highlighted by Liu et al. [7], in which LLM loses track of information poses a significant challenge.”
Akbar, N. A., Dembani, R., Airlangga, G., Wibowo, R. M., Lenzitti, B., & Tegolo, D. (2026) · ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) · doi:10.1145/3770855.3816477 · Cited by 1
Sheng, C. Y., Xin, C. Y., & Jie, T. G. (2026) · Cureus Journal of Computer Science. · doi:10.7759/s44389-026-00047-5 · Cited by 2
Biomedical Relation Extraction Using LLMs and Knowledge Graphs
Chellagurki, P., Kumaru, S. P. K., Peela, R. R., Yeluri, N., Rojas, C., & Jetcheva, J. G. (2024) · IEEE International Conference on Big Data Computing Service and Applications (BigDataService) · doi:10.1109/bigdataservice62917.2024.00015 · Cited by 3
“Moreover, the research detailed in ”Lost in the Middle: How Language Models Use Long Contexts” explores how models from the Llama2-7b series (7b, 13b, and 70b) [21] manage tasks that require processing long texts.”
BackgroundGuyeux, C., Laiymani, D., & Sola, C. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-82484-5_8 · Cited by 1
“Although recent LLMs have significantly large context sizes—far exceeding that of a single article, such as GPT-4 at 128k tokens or Gemini/Claude at 1M tokens—performance degradation with increased context size has been documented [13].”
BackgroundBenchmarking International Construction Contract Knowledge of Large Language ModelsPreprint
Xia, H., Chen, Y., Qi, X., & Wang, Y. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5247570 · Cited by 1
Understanding Hallucinations in Large Visual and Language Models
Ho, Z. Y., Liang, S., & Tao, D. (2026) · ACM Computing Surveys · doi:10.1145/3811409 · Cited by 11
“A practical diagnostic is that these failures often attenuate substantially when the input is clarified, decomposed, or better grounded.”
ChatGPT is not bullshit, nor is it not not bullshit
Fitts, J. (2026) · Ethics and Information Technology · doi:10.1007/s10676-026-09903-3 · Cited by 1
“Given the background rate of hallucinations combined with 1 Roig ( 2026 ) finds that “fabrication rises steeply with context length,” and Liu et al. ( 2024 ) finds that LLM performance degrades when considering information in the middle of large contexts (as opposed to the beginning or end).”
A Survey of Personalization: From RAG to Agent
Li, X., Jia, P., Xu, D., Wen, Y., Zhang, Y., Zhang, W., Wang, W., Wang, Y., Du, Z., Li, X., et al. (2026) · ACM Transactions on Information Systems · doi:10.1145/3802586 · Cited by 63
Patent Figure Classification Using Large Vision-Language Models
Awale, S., Müller-Budack, E., & Ewerth, R. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-88711-6_2 · Cited by 1
Lv, C., Zhou, X., Hu, G., et al. (2026) · IEEE Transactions on Industrial Informatics · doi:10.1109/tii.2025.3650225 · Cited by 1
“This not only increases computational overhead but also makes the models prone to distraction from irrelevant information [15].”
Efficient Input Data Strategies for LLMs (Large Language Models)-Based Bearing Fault Diagnosis
Jeong, P.-G., Hu, Y., Kong, J., Song, K., Yoo, B., & Kang, K. (2026) · International Journal of Precision Engineering and Manufacturing-Green Technology · doi:10.1007/s40684-026-00880-y · Cited by 2
“Textual transformation of large-scale time-series data often leads to excessive input length, which degrades model performance even within the context window [20 ], and makes it difficult for the model to attend to information located mid-sequence due to limitations in long-context attention [21].”
Dewantoro, M. F., Abdullah, F., Xia, Y., Khan, I., Thawonmas, R., & Ouyang, W. (2025) · IEEE Conference on Games (CoG) · doi:10.1109/cog64752.2025.11114183 · Cited by 1
“This finding aligns with research highlighting the degradation of LLM performance as context length increases, as discussed in Gao et al. [25] and Nelson et al. [26].”
PKFAR: psychiatry knowledge-fused augmented reasoning with large language models
Wang, R., Yu, C., Dong, Q., Qiu, J., Wen, T., Zhang, W., & Qin, K. (2026) · Health Information Science and Systems · doi:10.1007/s13755-026-00447-w · Cited by 4
Zhao, X., Jin, Y., Wang, Y. S., & Zhang, M. (2025) · Information Processing & Management · doi:10.1016/j.ipm.2025.104574 · Cited by 1
Exploring Robust Multi-Agent Workflows for Environmental Data Management
Guan, B., Liu, J., Wu, Y., & Bahreini, K. (2026) · Practice and Experience in Advanced Research Computing · doi:10.1145/3785462.3815810 · Cited by 2
“…experience surfaced the stressors illustrated in Figure 1 (right): heterogeneous inputs weaken constraints as contexts grow (“lost in the middle” [14]); long-running workflows span days or personnel transitions, yet agents cannot reliably retain prior rationale; stochastic generation drifts…”
Context is key: context engineering as the next frontier of medical AI
Wu, D. J., Chen, J. H., & Goh, E. (2026) · BMJ Digital Health & AI · doi:10.1136/bmjdh-2026-000032 · Cited by 3
“Conversely, LLM performance improves when more relevant and salient details are added, enriching the information-yield density of the context 5 6 ; while small perturbations in context, such as appending an irrelevant sentence, ‘ Interesting fact: cats sleep most of their lives’ to a math problem,…”
Tian, W., & Zhou, Y. (2026) · Expert Systems · doi:10.1111/exsy.70341 · Cited by 3
Yajnik, A., & Pushkala, S. A. (2026) · IEEE Access · doi:10.1109/access.2026.3719291 · Cited by 1
“Lost in the Middle shows that models can fail to use relevant information effectively even when it is present within a long context window, while LongBench v2 argues that realistic long-context evaluation should emphasize deeper reasoning and synthesis rather than simple evidence lookup [13], [14].”
Jang, D.-S., Yi, J.-S., Jeon, H., & Hong, Y. (2025) · Sustainability · doi:10.3390/su18010396 · Cited by 2
Zhao, Y., Zhou, Y., Xie, J., Huang, Z., & Gong, L. (2026) · IET Software · doi:10.1049/sfw2/6602283 · Cited by 2
Pándy, Á., Lakatos, R., & Hajdú, A. (2025) · IEEE International Conference on Intelligent Engineering Systems (INES) · doi:10.1109/ines67149.2025.11078188 · Cited by 1
Optimal RoPE extension via Bayesian Optimization for training-free length generalization
Zhang, X., Hu, S., Zhao, W., Wang, H., Han, X., He, C., Zeng, G., Liu, Z., & Sun, M. (2025) · AI Open · doi:10.1016/j.aiopen.2025.01.002 · Cited by 2
Uhm, J., Kim, M., Polychronakis, M., & Koo, H. (2026) · Proceedings of the ACM on software engineering. · doi:10.1145/3797116 · Cited by 4
Nyirongo, B., Jiang, Y., Zhang, Y., & Liu, H. (2026) · IEEE Transactions on Software Engineering · doi:10.1109/tse.2026.3700827
“While this design choice was intended to maintain reasoning precision, minimize inference latency, and mitigate the context rot associated with excessively large inputs [66], it may limit the model’s ability to detect certain cross-functional or long-range data dependencies.”
GeoVisA11y: An AI-based Geovisualization Question-Answering System for Screen-Reader Users
Li, C., Pang, R. Y., Chheda-Kothary, A., Sharif, A., Assalif, H., Heer, J., & Froehlich, J. E. (2026) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3772318.3790334 · Cited by 10
“Additionally, we designed separate prompts for each component as LLM performance degrades with longer input length [37, 39].”
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.