Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

The paper claims that long-context reliability is governed by fact distribution and that anti-hallucination prompts can impose a measurable accuracy tax.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 12:37 UTC pith:4MHSVWTA

load-bearing objection Useful evaluation protocol and a credible safety-tax finding, but the headline 'Distributional Collapse' claim contradicts its own evidence — collapse happens under concentrated, not dispersed, distributions. the 4 major comments →

arxiv 2601.02023 v2 pith:4MHSVWTA submitted 2026-01-05 cs.CL cs.AI

Not All Needles Are Found: How Fact Distribution and Don't Make It Up Prompts Shape Retrieval, Reasoning, and Hallucination in Long-Context LLMs

Amirali Ebrahimzadeh , Seyyed M. Salili This is my paper
classification cs.CL cs.AI
keywords long-context LLMsneedle-in-a-haystack benchmarkhallucinationanti-hallucination promptingfact distributionliteral extractionlogical inferencesafety tax
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper sets out to show that a long-context language model's reliability is shaped less by the advertised context window and more by where and how the relevant facts sit inside it. Using an extended needle-in-a-haystack benchmark that scores literal extraction, logical inference, and faithfulness separately, the authors find two failure modes: Distributional Collapse, where accuracy falls sharply when evidence is arranged in realistic non-uniform patterns, and a Safety Tax, where explicit 'Don't Make It Up' prompts push models into over-conservative refusals of facts that are actually present. The paper argues that nominal context size therefore overstates usable context, and that model choice and prompt phrasing matter as much as token count. A sympathetic reader would care because organizations increasingly paste large unfiltered documents into prompts and treat the answers as authoritative; these results say that arrangement of evidence and safety instructions can change accuracy massively, even to zero in some conditions.

Core claim

The central claim is that long-context reliability is governed by effective context length and fact distribution rather than by the nominal context window, and that anti-hallucination prompting carries a measurable accuracy cost. The paper demonstrates this with an extended needle-in-a-haystack design in which story-congruent facts are injected into a long narrative corpus, contexts are contracted to fractions of each model's maximum, facts are placed according to nine probabilistic distributions, and the same 30-question quiz is run under a standard prompt and a 'Don't Make It Up' prompt. On this test, Gemini-2.5-flash and Deepseek-v3.2-chat stay near their capacity limits across positions

What carries the argument

The central instrument is the extended needle-in-a-haystack benchmark. A realistic narrative corpus (nineteenth-century French novels) is processed with Recursive Context Contraction, which summarizes segments piece-wise to hit target token counts while preserving narrative flow. Fact 'needles' are injected either as one dense paragraph at controlled depths (Protocol A) or as ten scattered sentences following nine probability distributions—Uniform, Normal, Exponential, Exponential Flipped, Bimodal Gaussian Mixture, Arcsine, Lorentzian, Rayleigh, and Rayleigh Flipped (Protocol B). A fixed 30-question quiz with an answer key measures literal extraction, logical inference, and faithfulness, and

Load-bearing premise

The load-bearing premise is that the measuring pipeline is sound—that the context-contraction method keeps the target facts intact, the 30-question answer key is unambiguous, the LLM judge grades correctly, and a single 30-question quiz per condition is representative; the paper's own Limitations section concedes that exhaustive statistical significance testing was not performed, so the 0% and 100% cells in Table A1 could be sample noise.

What would settle it

Re-run each (context length, depth, fact distribution, prompt) cell with several independent 30-question quizzes built from different story texts, and check whether ChatGPT-5-mini's 0% Literal Extraction scores under Normal and Lorentzian distributions with the 'Don't Make It Up' prompt persist. If those scores become moderate across replications, the collapse is an artifact of the single quiz or judge; if they stay near zero, the failure mode is real. Also inspect the raw answer sheets to see whether the 0% comes from refusals, wrong answers, or grading errors.

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

If this is right

  • Nominal context windows overstate usable context: effective context length, the range where literal extraction stays reliable, is often much shorter than the advertised maximum.
  • Moving the same facts from uniform placement to clustered distributions such as Normal or Lorentzian can collapse accuracy to zero for some models, so benchmarks that test only uniform placement overestimate real-world performance.
  • Anti-hallucination prompts reduce fabrications but can also trigger refusals of present facts; the safety tax is largest near a model's context limit and for inference-heavy questions.
  • Reliability is model-specific: under identical conditions, some models remain stable near their capacity while others show cliffs or U-shaped failure, so context-window size alone is not a safe purchasing or deployment criterion.
  • The results imply that expanding context capacity does not remove the need for retrieval or grounding; distributional robustness and effective context management remain necessary.

Where Pith is reading between the lines

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

  • Editorial inference: the 0% collapse under centrally clustered distributions suggests the model's safety filter may be treating dense central evidence as a hallucination risk; checking whether the failures are refusals ('Not mentioned') versus wrong answers would pin down the mechanism, a split the paper does not report.
  • Editorial inference: a natural testable extension is to vary the width and centrality of the fact cluster to map the boundary of Distributional Collapse; if collapse tracks the distance from the context edges, the failure is likely a positional-attention effect interacting with the prompt.
  • Editorial inference: the Safety Tax might be mitigated by softening the anti-hallucination instruction to allow answering when evidence is present and refusing only when it is truly absent, or by calibrating refusal thresholds; the paper does not test these fixes.
  • Editorial inference: because each condition rests on a single 30-question quiz from one narrative corpus, the exact percentages are less trustworthy than the qualitative pattern; reusing the protocol across multiple corpora and quiz instances would tell whether the collapse boundaries are stable.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces an extended needle-in-a-haystack benchmark that separates literal extraction, logical inference, and faithfulness, and applies it to four long-context LLMs (Gemini-2.5-flash, ChatGPT-5-mini, Claude-4.5-haiku, Deepseek-v3.2-chat). It sweeps context length and fact depth under uniform placement and additionally scatters ten facts according to nine probabilistic distributions, under both a standard prompt and an anti-hallucination ('Don't Make It Up') prompt. The main reported findings are two failure modes: a 'Safety Tax,' in which anti-hallucination instructions cause over-conservative refusal and degrade accuracy, and a 'Distributional Collapse,' described in the abstract and conclusion as degraded performance when evidence is dispersed. The paper also argues that nominal context windows overstate the 'effective context length' that models can actually use. The empirical basis is a set of accuracy tables and heatmaps, with only one 30-question quiz per condition in the distribution protocol and no significance testing.

Significance. If the findings are valid, the paper makes a useful practical contribution by showing that long-context reliability depends on fact distribution and prompt regime, not just nominal context size, and by quantifying a safety/accuracy trade-off. The benchmark design has strengths: it evaluates three capabilities separately, uses a public-domain narrative corpus rather than purely synthetic text, includes nine fact distributions and two prompt conditions, and reports deterministic decoding settings and full prompt templates. However, the headline 'Distributional Collapse' claim is defined as a dispersed-evidence phenomenon but is measured under centrally concentrated distributions, and the measurement pipeline (one quiz per cell, LLM-only grading, unvalidated context contraction, no released data) is too thin to support several strong quantitative claims. The 'Safety Tax' finding for ChatGPT-5-mini is more directly supported by the data, but still needs statistical backing and independent grading validation.

major comments (4)
  1. [Abstract, §4.6, §5, §C.3, Table A1] The definition-to-evidence mismatch for 'Distributional Collapse' is load-bearing. The abstract defines it as degradation 'when evidence is dispersed,' and §5/Conclusion repeat that models struggle when evidence is 'dispersed across the corpus rather than concentrated in more favorable locations.' Yet the only systematic collapse in Table A1 is ChatGPT-5-mini under Normal and Lorentzian distributions with AH prompts (Literal 0, Logical 0) and Claude-4.5-haiku under Normal/Uniform logical inference. Section C.3 itself states that Normal and Lorentzian have information 'concentrated heavily in the center of the context window'—the opposite of dispersed. Under the most dispersed distribution, Uniform, ChatGPT-5-mini does not collapse (e.g., 100/90 and 70/90 for Literal/Logical). The measured phenomenon is therefore fragility under central clustering / non-uniform concentration, not disperse
  2. [§3.5.2, Table A1, Limitations] Statistical support is insufficient for the cell-level claims. Protocol B uses one 30-question quiz per (model, distribution, prompt) cell, and the Limitations section explicitly states that no exhaustive statistical significance testing was performed. With n=30, a reported 0% score has a 95% binomial CI of roughly 0–11.6%, and 100% has a CI of roughly 88.4–100%. The text in §C.3 that scores drop 'exactly to 0.0%' overstates precision. Aggregate differences in Table 2 also lack confidence intervals. I recommend reporting CIs or significance tests, and releasing per-item responses so these can be independently computed.
  3. [§3.2] The Recursive Context Contraction method is used to create every non-native context length, but no validation is reported that the contraction preserves the injected facts and the answer key. If summarization paraphrases, omits, or distorts a needle, then the length-scaling and depth curves would partly measure summarization artifacts. The authors should report a fact-preservation check—e.g., human or LLM verification that all injected facts and their corresponding answers remain answerable in the contracted story—and ideally release the contracted corpora.
  4. [§3.4, Appendix B] All scoring is performed by an independent LLM judge with no reported human validation, inter-annotator agreement, or error analysis. Because the grading prompt allows paraphrase matching, judge bias or inconsistency could directly change every numeric result. I recommend validating the LLM judge against human annotations on a representative sample of outputs for each model and prompt condition, and reporting agreement rates.
minor comments (5)
  1. [§3.4, Appendix A] The main text says 'a set of 30 questions,' but the prompt template in Appendix A shows 'Question 1' through 'Question 8.' Please clarify whether the quiz contains 8 or 30 questions, and if the template is truncated, state so explicitly.
  2. [References] References [30] and [59] contain placeholder arXiv identifiers ('arXiv:XXXX.XXXXX') and must be completed before publication. Also check for duplicate entries, e.g., LV-Eval appears as both [7] and [16] with different arXiv IDs.
  3. [Table 2] The 'Capacity' rows report conditional means, but the sample size (number of quizzes/questions) is not stated. Adding the number of questions per cell would help readers interpret the precision of these values.
  4. [Title] The title contains spacing/LaTeX artifacts: 'andDon’t Make It UpPrompts.' Please fix the formatting.
  5. [Figure 5] The caption says saturation is capped at ±30%, which is useful, but the color scale is not shown. Adding a color bar would improve interpretability.

Circularity Check

0 steps flagged

No significant circularity: the paper's findings are direct empirical measurements, not derived from fitted parameters or self-citations.

full rationale

The paper is an empirical benchmark study, not a derivation. Its central quantities—literal extraction, logical inference, faithfulness, Safety Tax, and Distributional Collapse—are measured outcomes of API queries under controlled conditions, not parameters fitted to those same outcomes. The Safety Tax is defined as the observed accuracy delta between Standard and Anti-Hallucination prompts (Section 4.5, Figure 5) and is therefore a direct comparison, not a construction. Distributional Collapse is presented as an observed failure pattern in Section C.3 and Table A1; while the skeptic's objection that the named 'dispersed evidence' collapse actually occurs under central-tendency (Normal/Lorentzian) distributions is a legitimate internal-consistency and validity concern, that is a factual/definitional mismatch, not a circular reduction of the result to its inputs. No load-bearing self-citations appear: the authors do not cite their own prior work to justify the benchmark, and no uniqueness theorem or ansatz is imported from prior work by the same authors. The stated limitation in Section 6—lack of exhaustive statistical significance testing—is a power/sampling concern, not evidence of circularity. The reliance on an LLM judge for grading is a measurement-reliability concern, not a logical equivalence between input and output. Therefore, no circular step can be identified under the required standard of quoting a specific reduction.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The paper contributes an evaluation design rather than a derivation, so the ledger tracks the experimental assumptions its scores rest on. No free parameters are fitted to produce a target; however, the chosen decoding hyperparameters, unreported distribution shapes, and LLM-graded scoring chain are load-bearing choices. No new physical or model entities are introduced; 'Distributional Collapse' and 'Safety Tax' are descriptive labels.

free parameters (3)
  • Decoding temperature = 0.0
    Chosen by hand for determinism; directly influences output variability and refusal behavior, and is not motivated beyond reproducibility.
  • Presence penalty = 0.3
    Chosen by hand; small penalty intended to steer attention, but it could systematically affect extraction and faithfulness scores across all models.
  • Fact-distribution shape parameters = not reported
    Nominal distributions (Normal, Bimodal, Lorentzian, etc.) need concrete means, variances, and mixture weights; these are not given, yet they control where facts land and therefore the measured distributional-collapse effects.
axioms (6)
  • domain assumption The LLM judge's binary grades are a faithful proxy for human correctness.
    All headline accuracies come from the grading prompt in Appendix B; no human validation or inter-annotator agreement is reported (§3.4).
  • domain assumption Recursive Context Contraction summarizes the Balzac corpus without dropping or distorting injected needles.
    The method in §3.2 contracts segments by summarization; if summaries introduce artifacts, apparent long-context failures could actually be summarization artifacts.
  • domain assumption A single 30-question quiz per condition is representative of model performance at that condition.
    Protocol B yields only one quiz per distribution × prompt cell; no repeated runs, confidence intervals, or significance tests are provided (§3.5.2, §6).
  • domain assumption tiktoken token counts approximate the models' native context windows.
    Table 1 admits Claude-4.5-haiku's native limit is about 200k while counts are reported via tiktoken; length percentages may be misaligned.
  • domain assumption The author-written answer key for the 30 evaluation questions is correct and unambiguous.
    All scores are compared against this key; errors in the key would change every headline result (§3.4).
  • domain assumption Balzac's fictional narrative is a valid proxy for real-world enterprise corpora.
    The authors acknowledge in Limitations that specialized domains may change behavior, yet many practical conclusions target enterprise workflows.

pith-pipeline@v1.3.0-alltime-deepseek · 17144 in / 15819 out tokens · 149453 ms · 2026-08-03T12:37:46.618629+00:00 · methodology

0 comments
read the original abstract

As Large Language Models (LLMs) increasingly utilize massive context windows as working memory for autonomous tasks, their reliability fluctuates significantly depending on how information is distributed in real-world corpora. We investigate how fact placement, corpus-level distributions, and anti-hallucination ("Don't Make It Up") prompts influence model behavior by introducing a model-agnostic extended needle-in-a-haystack benchmark designed for scalability, which we apply to evaluate Gemini-2.5-flash, ChatGPT-5-mini, Claude-4.5-haiku, and Deepseek-v3.2-chat. Unlike prior work, we separately evaluate literal extraction, logical inference, and hallucination risk. We identify two critical failure modes: Distributional Collapse, where performance degrades significantly when evidence is dispersed; and a Safety Tax, where anti-hallucination prompts cause over-conservative refusal of present facts and evidence, sharply reducing accuracy. Our results suggest that many failures stem from ineffective context utilization, as models struggle to prioritize relevant information even when it is present. These findings highlight the need for model-specific robustness and effective context management to ensure reliable deployment in long-horizon agentic workflows.

Figures

Figures reproduced from arXiv: 2601.02023 by Amirali Ebrahimzadeh, Seyyed M. Salili.

Figure 1
Figure 1. Figure 1: Overview of the Extended Needle in a Haystack Evaluation Framework. (A) Uniform sweeps [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average performance scaling across context lengths (log scale). For each context length, the reported [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance sensitivity to information depth. The x-axis represents the relative position of the fact [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Literal Extraction accuracy heatmaps comparing (a) Standard Prompts and (b) Anti-Hallucination [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance delta (Δ) heatmaps showing the shift in accuracy when applying Anti-Hallucination prompts. Red indicates performance degradation (over-refusal), while blue indicates improvement. ChatGPT￾5-mini (Column 2) exhibits severe degradation in Literal Extraction and Logical Inference, contrasting with the Faithfulness gains in Panel (c). Saturation is capped at ±30% to highlight subtle performance vari… view at source ↗
Figure 6
Figure 6. Figure 6: Radar charts depicting model robustness across varying fact distributions (Uniform, Normal, [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Where Facts Go Missing: A Layerwise Taxonomy and Per-Layer Attribution of Information Omission in Air-Gapped LLMAgent Pipelines

    cs.MA 2026-07 conditional novelty 6.0

    In a controlled 75,476-trial stress test, about 73% of omitted-fact failures in LLM agent pipelines are traced to deterministic middleware (redaction, pagination, truncation) rather than model behavior.

Reference graph

Works this paper leans on

99 extracted references · 51 linked inside Pith · cited by 1 Pith paper

  1. [1]

    BABILong: Testing the limits of LLMs with long context reasoning-in-a-haystack,

    Y. Kuratov, A. Bulatov, P. Anokhinet al., “BABILong: Testing the limits of LLMs with long context reasoning-in-a-haystack,”arXiv preprint arXiv:2406.10149, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Needlebench: Can llms do retrieval and reasoning in 1 million context window?

    M. Li, S. Zhang, Y. Liuet al., “Needlebench: Can llms do retrieval and reasoning in 1 million context window?”arXiv preprint arXiv:2407.11963, 2024

    arXiv 2024

  3. [3]

    ∞bench: Extending long context evaluation beyond 100k tokens,

    X. Zhang, Y. Chen, S. Huet al., “∞bench: Extending long context evaluation beyond 100k tokens,” arXiv preprint arXiv:2402.13718, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Rethinking context length in large language models,

    S. Wang, Y. Lu, Y. Niu, and J. Lin, “Rethinking context length in large language models,”arXiv preprint arXiv:2402.14488, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Lost in the middle: How language models use long contexts,

    N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang, “Lost in the middle: How language models use long contexts,”Transactions of the Association for Computational Linguistics, 2024

    2024

  6. [6]

    RULER: What’s the real context size of your long-context language models?

    C.-P. Hsieh, S. Sun, S. Krimanet al., “RULER: What’s the real context size of your long-context language models?”arXiv preprint arXiv:2404.06654, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Lv-eval: A balanced long-context benchmark with 5 length levels up to 256k,

    T. Yuan, X. Ning, D. Zhou, Z. Yang, S. Li, M. Zhuang, Z. Tan, Z. Yao, D. Lin, B. Li, G. Dai, S. Yan, and Y. Wang, “Lv-eval: A balanced long-context benchmark with 5 length levels up to 256k,”arXiv preprint arXiv:2402.05136, 2025, arXiv:2402.05136v3 [cs.CL]

    arXiv 2025

  8. [8]

    Lost-in-the-later: Framework for quantifying contextual grounding in large language models,

    Y. Tao, A. Hiatt, R. Seetharamanet al., “Lost-in-the-later: Framework for quantifying contextual grounding in large language models,”arXiv preprint arXiv:2507.05424, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Detectbench: Can large language model detect and piece together implicit evidence?

    Z. Gu, L. Zhang, X. Zhu, J. Chen, W. Huang, Y. Zhang, S. Wang, Z. Ye, Y. Gao, Y. Xiao, and H. Feng, “Detectbench: Can large language model detect and piece together implicit evidence?”arXiv preprint arXiv:2406.12641, 2024, arXiv:2406.12641v2 [cs.CL]

    Pith/arXiv arXiv 2024

  10. [10]

    Evaluating multilingual long- context models for retrieval and reasoning,

    A. Agrawal, A. Dang, S. Bagheri Nezhad, R. Pokharel, and R. Scheinberg, “Evaluating multilingual long- context models for retrieval and reasoning,”arXiv preprint arXiv:2409.18006, 2024, arXiv:2409.18006v3 [cs.CL]

    Pith/arXiv arXiv 2024

  11. [11]

    The two-hop curse: LLMs trained on 𝐴→𝐵,𝐵→𝐶 fail to learn𝐴→𝐶,

    M. Balesni, T. Korbak, and O. Evans, “The two-hop curse: LLMs trained on 𝐴→𝐵,𝐵→𝐶 fail to learn𝐴→𝐶,”arXiv preprint arXiv:2411.16353, 2024

    arXiv 2024

  12. [12]

    Do large language models perform latent multi-hop reasoning without exploiting shortcuts?

    S. Yang, N. Kassner, E. Gribovskayaet al., “Do large language models perform latent multi-hop reasoning without exploiting shortcuts?”arXiv preprint arXiv:2411.16679, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Generating wikipedia by summarizing long sequences,

    P. J. Liu, M. Saleh, E. Pot, B. Goodrich, R. Sepassi, L. Kaiser, and N. Shazeer, “Generating wikipedia by summarizing long sequences,” inInternational Conference on Learning Representations, 2018

    2018

  14. [15]

    FactBench: A dynamic benchmark for in-the-wild language model factuality evaluation,

    F. F. Bayat, L. Zhang, S. Muniret al., “FactBench: A dynamic benchmark for in-the-wild language model factuality evaluation,”arXiv preprint arXiv:2410.22257, 2024

    Pith/arXiv arXiv 2024

  15. [16]

    Lv-eval: A balanced long-context benchmark with 5 length levels up to 256k,

    T. Yuan, X. Ning, D. Zhouet al., “Lv-eval: A balanced long-context benchmark with 5 length levels up to 256k,”arXiv preprint arXiv:2402.05133, 2024. 14

    Pith/arXiv arXiv 2024

  16. [17]

    Faitheval: Can your language model stay faithful to context,

    Y. Ming, S. Purushwalkam, S. Panditet al., “Faitheval: Can your language model stay faithful to context,” arXiv preprint arXiv:2410.03727, 2024

    Pith/arXiv arXiv 2024

  17. [18]

    Longbench: A bilingual, multitask benchmark for long context understanding,

    Y. Bai, X. Lv, J. Zhanget al., “Longbench: A bilingual, multitask benchmark for long context understanding,”arXiv preprint arXiv:2308.14508, 2023

    Pith/arXiv arXiv 2023

  18. [19]

    L-eval: Instituting standardized evaluation for long context language models,

    C. An, S. Gong, M. Zhonget al., “L-eval: Instituting standardized evaluation for long context language models,”arXiv preprint arXiv:2307.11088, 2022

    Pith/arXiv arXiv 2022

  19. [20]

    GPT-4 technical report,

    OpenAI, “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  20. [21]

    Llmeval-3: A large-scale longitudinal study on robust and fair evaluation of large language models,

    M. Zhang, Y. Shen, J. Denget al., “Llmeval-3: A large-scale longitudinal study on robust and fair evaluation of large language models,”arXiv preprint arXiv:2508.05452, 2025

    Pith/arXiv arXiv 2025

  21. [22]

    Long-context language models fail in basic retrieval tasks without sufficient reasoning steps,

    Y. Yu, Y. Huang, Z. Qi, W. Wang, W. Liu, R. Chen, and J. Pei, “Long-context language models fail in basic retrieval tasks without sufficient reasoning steps,”arXiv preprint arXiv:2410.04422, 2025, arXiv:2410.04422v9 [cs.CL]

    Pith/arXiv arXiv 2025

  22. [23]

    FACTORY: A challenging human-verified prompt set for long-form factuality,

    M. Chen, Y. Li, X. Chenet al., “FACTORY: A challenging human-verified prompt set for long-form factuality,”arXiv preprint arXiv:2508.00109, 2025

    Pith/arXiv arXiv 2025

  23. [24]

    Investigating factuality in long-form text generation,

    L. Tu, R. Meng, S. Jotyet al., “Investigating factuality in long-form text generation,”arXiv preprint arXiv:2411.15993, 2024

    arXiv 2024

  24. [25]

    Evaluating language model context windows: A “working memory

    A. Dsouza, C. M. Glaze, C. Shinet al., “Evaluating language model context windows: A “working memory” test and inference-time correction,”arXiv preprint arXiv:2407.03651, 2024

    Pith/arXiv arXiv 2024

  25. [26]

    LLMs get lost in multi-turn conversation,

    P. Laban, H. Hayashi, Y. Zhou, and J. Neville, “LLMs get lost in multi-turn conversation,”arXiv preprint arXiv:2505.06120, 2025

    Pith/arXiv arXiv 2025

  26. [27]

    Longins: A challenging long-context instruction-based exam for llms,

    S. Gavin, T. Zheng, J. Liu, Q. Que, N. Wang, J. Yang, C. Zhang, W. Huang, and G. Zhang, “Longins: A challenging long-context instruction-based exam for llms,”arXiv preprint arXiv:2406.17588, 2025, arXiv:2406.17588v3 [cs.CL]

    Pith/arXiv arXiv 2025

  27. [28]

    Needle in a haystack - pressure testing LLMs,

    G. Kamradt, “Needle in a haystack - pressure testing LLMs,” https://github.com/gkamradt/LLMTest NeedleInAHaystack, 2023

    2023

  28. [29]

    The needle in a haystack test: Evaluating the performance of LLM RAG systems,

    E. Jolley and A. Dhinakaran, “The needle in a haystack test: Evaluating the performance of LLM RAG systems,”Arize AI Blog, 2024

    2024

  29. [30]

    Sequential-NIAH: A needle-in-a-haystack benchmark for extracting sequential needles from long contexts,

    Y. Yu, Q.-W. Zhang, L. Qiao, D. Yin, F. Li, J. Wang, C. Z. Xi, S. Zheng, X. Liang, and X. Sun, “Sequential-NIAH: A needle-in-a-haystack benchmark for extracting sequential needles from long contexts,”arXiv preprint arXiv:XXXX.XXXXX, 2025, in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2025

  30. [31]

    NoLiMa: Long-context evaluation beyond literal matching,

    A. Modarressi, H. Deilamsalehy, F. Dernoncourt, T. Bui, R. A. Rossi, S. Yoon, and H. Sch¨ utze, “NoLiMa: Long-context evaluation beyond literal matching,”arXiv preprint arXiv:2502.05167, 2025

    Pith/arXiv arXiv 2025

  31. [32]

    LongGenBench: Benchmarking long-form generation in long context LLMs,

    Y. Wu, M. S. Hee, Z. Hu, and R. K.-W. Lee, “LongGenBench: Benchmarking long-form generation in long context LLMs,”arXiv preprint arXiv:2409.02076, 2024

    Pith/arXiv arXiv 2024

  32. [33]

    When context leads but parametric memory follows in large language models,

    Y. Tao, A. Hiatt, E. Haakeet al., “When context leads but parametric memory follows in large language models,”arXiv preprint arXiv:2409.08435, 2024. 15

    Pith/arXiv arXiv 2024

  33. [34]

    Multilingual needle in a haystack: Investigating long-context behavior of multilingual large language models,

    A. Hengle, P. Bajpai, S. Dan, and T. Chakraborty, “Multilingual needle in a haystack: Investigating long-context behavior of multilingual large language models,”arXiv preprint arXiv:2408.10151, 2024, arXiv:2408.10151 [cs.CL]

    Pith/arXiv arXiv 2024

  34. [35]

    Premise order matters in reasoning with large language models,

    X. Chen, R. Chi, X. Wang, and D. Zhou, “Premise order matters in reasoning with large language models,” inInternational Conference on Machine Learning (ICML), 2024

    2024

  35. [36]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin, and T. Liu, “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,”ACM Trans. Inf. Syst., vol. 43, no. 2, jan 2025

    2025

  36. [37]

    Long context RAG performance of large language models,

    Q. Leng, J. Portes, S. Havenset al., “Long context RAG performance of large language models,”arXiv preprint arXiv:2411.03538, 2024

    Pith/arXiv arXiv 2024

  37. [38]

    Understanding and addressing ai hallucinations in healthcare and life sciences,

    G. Aditya, “Understanding and addressing ai hallucinations in healthcare and life sciences,”International Journal of Health Sciences, vol. 7, no. 3, pp. 1–11, 2024

    2024

  38. [39]

    A survey on hallucination in large language and foundation models,

    P. Ahadian and Q. Guan, “A survey on hallucination in large language and foundation models,” Preprints.org, 2025, 202504.1236.v1

    arXiv 2025

  39. [40]

    Creating trustworthy llms: Dealing with hallucinations in healthcare ai,

    M. A. Ahmad, I. Yaramis, and T. D. Roy, “Creating trustworthy llms: Dealing with hallucinations in healthcare ai,”arXiv preprint arXiv:2311.01463, 2023

    Pith/arXiv arXiv 2023

  40. [41]

    Unravelling the mysteries of hallucination in large language models: Strategies for precision in artificial intelligence language generation,

    A. R. Ahmadi, “Unravelling the mysteries of hallucination in large language models: Strategies for precision in artificial intelligence language generation,”Asian Journal of Computer Science and Technology, vol. 13, no. 1, pp. 1–10, 2024

    2024

  41. [42]

    Halo: Hallucination analysis and learning optimization to empower llms with retrieval-augmented context for guided clinical decision making,

    S. Anjum, H. Zhang, W. Zhou, E. J. Paek, X. Zhao, and Y. Feng, “Halo: Hallucination analysis and learning optimization to empower llms with retrieval-augmented context for guided clinical decision making,”arXiv preprint arXiv:2409.10011, 2024

    Pith/arXiv arXiv 2024

  42. [43]

    Dual process theory for large language models: An overview of using psychology to address hallucination and reliability issues,

    S. C. Bellini-Leite, “Dual process theory for large language models: An overview of using psychology to address hallucination and reliability issues,”Adaptive Behavior, 2023

    2023

  43. [44]

    Factchd: Benchmarking fact-conflicting hallucination detection,

    X. Chen, D. Song, H. Guiet al., “Factchd: Benchmarking fact-conflicting hallucination detection,” in Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024

    2024

  44. [45]

    Explainable hallucination mitigation in large language models: A survey,

    W. Deng, J. Li, H. Zhanget al., “Explainable hallucination mitigation in large language models: A survey,”Preprints.org, 2025, 202505.0456.v1

    arXiv 2025

  45. [46]

    Zero-resource hallucination detection for text generation via graph- based contextual knowledge triples modeling,

    X. Fang, Z. Huang, Z. Tianet al., “Zero-resource hallucination detection for text generation via graph- based contextual knowledge triples modeling,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, 2025, pp. 23 868–23 877

    2025

  46. [47]

    Chainpoll: A high efficacy method for llm hallucination detection,

    R. Friel and A. Sanyal, “Chainpoll: A high efficacy method for llm hallucination detection,”arXiv preprint arXiv:2310.18344, 2023

    Pith/arXiv arXiv 2023

  47. [48]

    Zero-knowledge llm hallucination detection and mitigation through fine-grained cross-model consistency,

    A. Goel, D. Schwartz, and Y. Qi, “Zero-knowledge llm hallucination detection and mitigation through fine-grained cross-model consistency,”arXiv preprint arXiv:2508.14314, 2025

    arXiv 2025

  48. [49]

    Detecting and preventing hallucinations in large vision language models,

    A. Gunjal, J. Yin, and E. Bas, “Detecting and preventing hallucinations in large vision language models,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 18 135–18 143. 16

    2024

  49. [50]

    Beyond probabilities: Unveiling the delicate dance of large language models (llms) and ai-hallucination,

    O. H. Hamid, “Beyond probabilities: Unveiling the delicate dance of large language models (llms) and ai-hallucination,” in2024 IEEE International Conference on Cognitive and Innovative Military Applications (CogSIMA), 2024

    2024

  50. [51]

    Kea explain: Explanations of hallucinations using graph kernel analysis,

    R. Haskins and B. Adams, “Kea explain: Explanations of hallucinations using graph kernel analysis,” arXiv preprint arXiv:2507.03847, 2025

    Pith/arXiv arXiv 2025

  51. [52]

    Mitigating hallucinations in large language models for educational application,

    H.-T. Ho, D.-T. Ly, and L. V. Nguyen, “Mitigating hallucinations in large language models for educational application,” in2024 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia), 2024, pp. 1–4

    2024

  52. [53]

    The illusion of progress: Re-evaluating hallucination detection in llms,

    D. Janiak, J. Binkowski, A. Sawczyn, B. Gabrys, R. Shwartz-Ziv, and T. Kajdanowicz, “The illusion of progress: Re-evaluating hallucination detection in llms,”arXiv preprint arXiv:2508.08285, 2025

    Pith/arXiv arXiv 2025

  53. [54]

    Hallucinations in large language models (llm’s): challenges in mitigation, trust, and future directions,

    R. Karne, P. K. Pativada, and A. Dudhipala, “Hallucinations in large language models (llm’s): challenges in mitigation, trust, and future directions,”Indian Journal of Computer Science and Engineering, vol. 16, no. 3, pp. 17–26, 2025

    2025

  54. [55]

    Detecting llm hallucinations using monte carlo simulations on token probabilities,

    G. Ledger and R. Mancinni, “Detecting llm hallucinations using monte carlo simulations on token probabilities,”TechRxiv, 2024, 171822396.61518693/v1

    arXiv 2024

  55. [56]

    Halueval: A large-scale hallucination evaluation benchmark for large language models,

    J. Li, X. Cheng, W. X. Zhaoet al., “Halueval: A large-scale hallucination evaluation benchmark for large language models,”arXiv preprint arXiv:2305.11747, 2023

    Pith/arXiv arXiv 2023

  56. [57]

    Loki’s dance of illusions: A comprehensive survey of hallucination in large language models,

    C. Li, P. Wang, C. Wanget al., “Loki’s dance of illusions: A comprehensive survey of hallucination in large language models,”arXiv preprint arXiv:2507.02870, 2025

    Pith/arXiv arXiv 2025

  57. [58]

    Learning to trust your feelings: Leveraging self-awareness in llms for hallucination mitigation,

    Y. Liang, Z. Song, H. Wang, and J. Zhang, “Learning to trust your feelings: Leveraging self-awareness in llms for hallucination mitigation,”arXiv preprint arXiv:2401.15449, 2024

    Pith/arXiv arXiv 2024

  58. [59]

    Attention-guided self-reflection for zero-shot hallucination detection in large language models,

    Q. Liu, X. Chen, Y. Ding, B. Song, W. Wang, S. Wu, and L. Wang, “Attention-guided self-reflection for zero-shot hallucination detection in large language models,”arXiv preprint arXiv:XXXX.XXXXX, 2025, in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2025

  59. [60]

    Roberta with low-rank adaptation and hierarchical attention for hallucination detection in llms,

    J. Lu and S. Li, “Roberta with low-rank adaptation and hierarchical attention for hallucination detection in llms,”Preprints.org, 2025, 202504.0465.v1

    arXiv 2025

  60. [61]

    Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models,

    P. Manakul, A. Liusie, and M. J. F. Gales, “Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models,”arXiv preprint arXiv:2303.08896, 2023

    Pith/arXiv arXiv 2023

  61. [62]

    Hallucination detox: Sensitivity dropout (send) for large language model training,

    S. Mohammadzadeh, J. D. Guerra, M. Bonizzato, R. Rabbany, and G. Farnadi, “Hallucination detox: Sensitivity dropout (send) for large language model training,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 2025, pp. 5538–5554

    2025

  62. [63]

    Fakes of varying shades: How warning affects human perception and engagement regarding llm hallucinations,

    M. Nahar, H. Seo, E.-J. Lee, A. Xiong, and D. Lee, “Fakes of varying shades: How warning affects human perception and engagement regarding llm hallucinations,”arXiv preprint arXiv:2404.03745, 2024

    Pith/arXiv arXiv 2024

  63. [64]

    Leveraging graph structures to detect hallucinations in large language models,

    N. Nonkes, S. Agaronian, E. Kanoulas, and R. Petcu, “Leveraging graph structures to detect hallucinations in large language models,”arXiv preprint arXiv:2407.04485, 2024. 17

    Pith/arXiv arXiv 2024

  64. [65]

    Erbench: An entity-relationship based automatically verifiable hallucination benchmark for large language models,

    J. Oh, S. Kim, J. Seo, J. Wang, R. Xu, X. Xie, and S. E. Whang, “Erbench: An entity-relationship based automatically verifiable hallucination benchmark for large language models,”arXiv preprint arXiv:2403.05266, 2024

    Pith/arXiv arXiv 2024

  65. [66]

    Llms know more than they show: On the intrinsic representation of llm hallucinations,

    H. Orgad, M. Toker, Z. Gekhman, R. Reichart, I. Szpektor, H. Kotek, and Y. Belinkov, “Llms know more than they show: On the intrinsic representation of llm hallucinations,”arXiv preprint arXiv:2410.02707, 2024

    Pith/arXiv arXiv 2024

  66. [67]

    Mitigating hallucinations in large language models via semantic enrichment of prompts: Insights from biobert and ontological integration,

    S. Penkov, “Mitigating hallucinations in large language models via semantic enrichment of prompts: Insights from biobert and ontological integration,” inCLIB 2024, 2024, pp. 272–276

    2024

  67. [68]

    Hallusafe at semeval-2024 task 6: An nli-based approach to make llms safer by better detecting hallucinations and overgeneration mistakes,

    Z. Rahimi, H. Amirzadeh, A. Sohrabiet al., “Hallusafe at semeval-2024 task 6: An nli-based approach to make llms safer by better detecting hallucinations and overgeneration mistakes,” inProceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), 2024

    2024

  68. [69]

    A survey of hallucination in large foundation models,

    V. Rawte, A. P. Sheth, and A. Das, “A survey of hallucination in large foundation models,”arXiv preprint arXiv:2309.05922, 2023

    Pith/arXiv arXiv 2023

  69. [70]

    Delucionqa: Detecting hallucinations in domain-specific question answering,

    M. Sadat, Z. Zhou, L. Langeet al., “Delucionqa: Detecting hallucinations in domain-specific question answering,” inFindings of the Association for Computational Linguistics (EMNLP), 2023, pp. 822–835

    2023

  70. [71]

    Mitigation of hallucinations in language models in education: A new approach of comparative and cross-verification,

    W. d. A. d. Silva, L. C. C. Fonseca, S. Labidi, and J. C. L. Pacheco, “Mitigation of hallucinations in language models in education: A new approach of comparative and cross-verification,” in2024 IEEE International Conference on Advanced Learning Technologies (ICALT), 2024, pp. 207–209

    2024

  71. [72]

    Unsupervised real-time hallucination detection based on the internal states of large language models,

    W. Su, C. Wang, Q. Aiet al., “Unsupervised real-time hallucination detection based on the internal states of large language models,”arXiv preprint arXiv:2403.06448, 2024

    Pith/arXiv arXiv 2024

  72. [73]

    Confabulation: The surprising value of large language model hallucinations,

    P. Sui, E. Duede, S. Wu, and R. J. So, “Confabulation: The surprising value of large language model hallucinations,”arXiv preprint arXiv:2406.04175, 2024

    Pith/arXiv arXiv 2024

  73. [74]

    A comprehensive survey of hallucination mitigation techniques in large language models,

    S. Tonmoy, S. M. M. Zaman, V. Jain, A. Rani, V. Rawte, A. Chadha, and A. Das, “A comprehensive survey of hallucination mitigation techniques in large language models,”arXiv preprint arXiv:2401.01313, 2024

    Pith/arXiv arXiv 2024

  74. [75]

    Investigating hallucination tendencies of large language models in japanese and english,

    H. Tsuruta and R. Sakaguchi, “Investigating hallucination tendencies of large language models in japanese and english,”Research Square, 2024, 4521710/v1

    2024

  75. [76]

    A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation,

    N. Varshney, W. Yao, H. Zhang, J. Chen, and D. Yu, “A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation,”arXiv preprint arXiv:2307.03987v2, 2023

    Pith/arXiv arXiv 2023

  76. [77]

    Mitigating hallucinations in large vision-language models with instruction contrastive decoding,

    X. Wang, J. Pan, L. Ding, and C. Biemann, “Mitigating hallucinations in large vision-language models with instruction contrastive decoding,”arXiv preprint arXiv:2403.18715, 2024

    Pith/arXiv arXiv 2024

  77. [78]

    Alleviating hallucinations in large language models with scepticism modeling,

    Y. Wu, Y. Wang, T. Chenet al., “Alleviating hallucinations in large language models with scepticism modeling,”arXiv preprint arXiv:2409.06601, 2024

    Pith/arXiv arXiv 2024

  78. [79]

    Detecting and reducing the factual hallucinations of large language models with metamorphic testing,

    W. Wu, Y. Cao, N. Yi, R. Ou, and Z. Zheng, “Detecting and reducing the factual hallucinations of large language models with metamorphic testing,”Proceedings of the ACM on Software Engineering, vol. 2, no. FSE, pp. 1432–1453, 2025

    2025

  79. [80]

    Efuf: Efficient fine-grained unlearning framework for mitigating hallucinations in multimodal large language models,

    S. Xing, F. Zhao, Z. Wuet al., “Efuf: Efficient fine-grained unlearning framework for mitigating hallucinations in multimodal large language models,”arXiv preprint arXiv:2402.09801, 2024. 18

    Pith/arXiv arXiv 2024

  80. [81]

    Interrogatellm: Zero-resource hallucination detection in llm-generated answers,

    Y. Yehuda, I. Malkiel, O. Barkan, J. Weill, R. Ronen, and N. Koenigstein, “Interrogatellm: Zero-resource hallucination detection in llm-generated answers,”arXiv preprint arXiv:2403.02889v3, 2024

    Pith/arXiv arXiv 2024

Showing first 80 references.