pith. machine review for the scientific record. sign in

arxiv: 2604.04743 · v1 · submitted 2026-04-06 · 💻 cs.CL · cs.AI· cs.SY· eess.SY

Recognition: 3 theorem links

· Lean Theorem

Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations

Kalyan Cherukuri , Lav R. Varshney
Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:52 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.SYeess.SY
keywords LLM hallucinationslatent space basinsdynamical systemsgeometry-aware steeringtransformer hidden statesautoregressive generationtask-dependent behavior
0
0 comments X

The pith

Hallucinations in large language models arise from task-dependent basin structures in their latent space.

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

The paper proposes that when LLMs generate incorrect facts, their internal state trajectories fall into attracting regions or basins whose shape depends on the task at hand. Tracking hidden states during autoregressive generation across open-source models reveals clearer separation between truthful and hallucinated paths in simple factoid settings than in summarization or misconception-heavy ones. The authors formalize this with theorems on task complexity and multi-basin dynamics in L-layer transformers, then show that geometry-aware adjustments to the state space can steer outputs toward lower hallucination rates without any retraining.

Core claim

Hallucinations emerge from task-dependent basin structure in latent space. Autoregressive hidden-state trajectories exhibit separability that varies strongly with task type, formalized through task-complexity and multi-basin theorems that characterize basin emergence across transformer layers. Geometry-aware steering then reduces hallucination probability by manipulating these structures.

What carries the argument

Task-dependent basin structure in latent space, identified via separability of autoregressive hidden-state trajectories and manipulated through geometry-aware steering.

If this is right

  • Geometry-aware steering lowers hallucination rates on factoid tasks while leaving the underlying model weights unchanged.
  • Basin separability weakens on summarization and complex-reasoning tasks, limiting the immediate reach of steering.
  • Task-complexity and multi-basin theorems predict how basins form across the layers of an L-layer transformer.
  • The same geometric view can be used to compare hallucination behavior across different open-source model families.

Where Pith is reading between the lines

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

  • If the basins prove causal, the same steering approach could be tested on other generation failures such as inconsistency or bias.
  • Accessing hidden states in closed models would require new interfaces before the method can be applied at scale.
  • Training objectives that penalize basin formation might yield more robust models from the start.
  • The dynamical-systems framing invites direct comparisons with basin analyses in other sequential models such as those used in reinforcement learning.

Load-bearing premise

The separated regions visible in hidden-state trajectories reflect causal basin structures that can be steered reliably rather than mere correlations or model-specific artifacts.

What would settle it

Apply the proposed geometry-aware steering to a held-out model and task and measure whether hallucination rates stay the same or increase instead of decreasing.

Figures

Figures reproduced from arXiv: 2604.04743 by Kalyan Cherukuri, Lav R. Varshney.

Figure 2
Figure 2. Figure 2: Causal Intervention: Factual → Basin. (Left) Dose￾response curve fold increase in hallucination probability as factual hidden states are in-model steered toward the hallucination centroid (interpolation strength α on the horizontal axis). Right: bar plot comparing the maximum fold increase produced by steering along the basin direction versus two controls (random direction and an orthogonal direction). See… view at source ↗
Figure 1
Figure 1. Figure 1: Task-Dependent Basin Geometry. Llama-3.2-3b’s performance on various tasks and 3D PCA projected outputs. (a) shows performance on MuSiQue, (b) shows performance on HaluEvalQA, (c) shows performance on HaluEvalSummarization, (d) shows performance on TruthfulQA. fulQA and summarization the AUROC value lingers be￾tween 0.5 across all models, indicating a near random per￾formance [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 3
Figure 3. Figure 3: Multi-basin Voronoi structure across models on TruthfulQA. Each panel shows distinct hallucination basins corresponding to different misconception modes. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Efficacy of Algorithm 2 in hallucination reduction as a function of the steering strength λ. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Irreversibility summary under autoregressive decoding (HaluEval QA, Llama-3.2-1B, best layer). We report basin-entry, conditional irreversibility, escape-after-entry, and factual entry rates. This verifies Theorem 5.9. D.2. Layer-Wise Attention Entropy 0 2 4 6 8 10 12 14 Layer 0.2 0.3 0.4 0.5 0.6 Attention Entropy llama-3.2-1b Factual Hallucinated 0 5 10 15 20 25 Layer llama-3.2-3b Layer-wise Attention Ent… view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise attention entropy for factual versus hallucinated generations under uninformative contexts (autoregressive extraction). Entropy trends provide a complementary signal to basin-separation metrics. Supports the uniform attention assumption. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Causal Intervention Paths: Llama-3.2-1B (HaluEval QA) 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Causal Intervention Paths: Llama-3.2-3B (HaluEval QA) 20 [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Causal Intervention Paths: Qwen-2.5-1.5B (HaluEval QA) 21 [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: 2D PCA Evolution: Llama-3.2 1B (QA) [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: 2D PCA Evolution: Llama-3.2 1B (Summarization) 22 [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: 2D PCA Evolution: Qwen-2.5 1.5B (QA) [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: 2D PCA Evolution: Gemma-2 2B (Summarization) 23 [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: 3D PCA Evolution: Llama-3.2 1B (QA) [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: 3D PCA Evolution: Llama-3.2 1B (Summarization) [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: 3D PCA Evolution: Qwen-2.5 1.5B (QA) [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: 3D PCA Evolution: Gemma-2 2B (Summarization) 24 [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
read the original abstract

Large language models (LLMs) hallucinate: they produce fluent outputs that are factually incorrect. We present a geometric dynamical systems framework in which hallucinations arise from task-dependent basin structure in latent space. Using autoregressive hidden-state trajectories across multiple open-source models and benchmarks, we find that separability is strongly task-dependent rather than universal: factoid settings can show clearer basin separation, whereas summarization and misconception-heavy settings are typically less stable and often overlap. We formalize this behavior with task-complexity and multi-basin theorems, characterize basin emergence in L-layer transformers, and show that geometry-aware steering can reduce hallucination probability without retraining.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a geometric dynamical systems framework in which LLM hallucinations arise from task-dependent basin structures in latent space, identified via autoregressive hidden-state trajectories. It reports that basin separability varies by task (clearer in factoid settings, less stable in summarization or misconception-heavy ones), formalizes this via task-complexity and multi-basin theorems, characterizes basin emergence in L-layer transformers, and shows that geometry-aware steering of hidden states during generation reduces hallucination probability without retraining, across multiple open-source models and benchmarks.

Significance. If the central claims hold, the framework would supply a new dynamical-systems lens on hallucinations and a practical, training-free control method via latent-space steering. The task-dependent separability finding and the theorems on basin emergence could influence both mechanistic interpretability and deployment strategies for reliable generation.

major comments (3)
  1. [Steering Experiments] The steering experiments (described after the theorems) do not include controls that preserve magnitude and directional statistics of the hidden-state updates while removing the specific basin-derived geometry; without such isolation it is impossible to attribute the reported drop in hallucination probability to basin navigation rather than generic distributional shifts.
  2. [Formal Theorems] The task-complexity and multi-basin theorems are stated without proof sketches, derivation steps, or explicit assumptions on the hidden-state dynamics; this leaves the formalization of separability unverified and makes it difficult to assess whether the reported task dependence follows from the geometry or is partly definitional.
  3. [Abstract and Experimental Results] The abstract and experimental sections assert results across multiple models and benchmarks, yet supply no methods subsection detailing trajectory extraction, basin identification procedure, evaluation metrics, or error analysis; the central empirical claims therefore cannot be reproduced or evaluated from the provided information.
minor comments (2)
  1. [Notation] Notation for basin boundaries and the separability metric is introduced without a dedicated definitions subsection, making cross-references to the theorems harder to follow.
  2. [Results] The manuscript would benefit from a table summarizing separability statistics (e.g., overlap measures) per task and model rather than only qualitative statements.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. We address each major comment below with clarifications and commit to targeted revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: The steering experiments (described after the theorems) do not include controls that preserve magnitude and directional statistics of the hidden-state updates while removing the specific basin-derived geometry; without such isolation it is impossible to attribute the reported drop in hallucination probability to basin navigation rather than generic distributional shifts.

    Authors: We agree that isolating the contribution of basin-derived geometry from generic distributional shifts would strengthen causal attribution. In the revised manuscript we will add control conditions that generate hidden-state updates matching the empirical magnitude and directional statistics of the steering vectors but drawn from non-basin directions; we will report the resulting hallucination rates alongside the original geometry-aware results. revision: yes

  2. Referee: The task-complexity and multi-basin theorems are stated without proof sketches, derivation steps, or explicit assumptions on the hidden-state dynamics; this leaves the formalization of separability unverified and makes it difficult to assess whether the reported task dependence follows from the geometry or is partly definitional.

    Authors: The theorems formalize observed separability patterns under the autoregressive trajectory model; however, we acknowledge that explicit assumptions and derivation steps are needed for verification. We will append a supplementary section containing the full statements with proof sketches, the precise dynamical assumptions on hidden-state evolution, and a discussion of how task dependence emerges from the geometry rather than by definition. revision: yes

  3. Referee: The abstract and experimental sections assert results across multiple models and benchmarks, yet supply no methods subsection detailing trajectory extraction, basin identification procedure, evaluation metrics, or error analysis; the central empirical claims therefore cannot be reproduced or evaluated from the provided information.

    Authors: We recognize that a self-contained methods subsection is required for reproducibility. Although the main text describes the overall pipeline, we will expand the experimental section with a dedicated methods subsection that specifies the exact trajectory extraction procedure, basin identification algorithm, evaluation metrics, statistical tests, and error analysis protocol, including any preprocessing steps applied to the hidden states. revision: yes

Circularity Check

1 steps flagged

Basin separability observed in trajectories is used both to define and to evidence the claimed causal basin structure

specific steps
  1. self definitional [Abstract]
    "Using autoregressive hidden-state trajectories across multiple open-source models and benchmarks, we find that separability is strongly task-dependent rather than universal: factoid settings can show clearer basin separation, whereas summarization and misconception-heavy settings are typically less stable and often overlap. We formalize this behavior with task-complexity and multi-basin theorems, characterize basin emergence in L-layer transformers, and show that geometry-aware steering can reduce hallucination probability without retraining."

    Separability is measured in the trajectories and immediately labeled 'basin separation'; the same separability is then cited as evidence that hallucinations arise from the basin structure. Because the basin structure is defined by the observed separability patterns in the identical data, the causal attribution reduces to a restatement of the input observation rather than an independent derivation.

full rationale

The paper's core claim is that hallucinations arise from task-dependent basin structure in latent space, with separability in autoregressive hidden-state trajectories presented as evidence. However, the basins appear to be characterized directly from the separability patterns in those same trajectories (factoid vs. summarization settings), after which the framework formalizes the behavior via theorems and attributes causality. This makes the reported separability partly definitional rather than an independent test of an a priori basin model. Steering results are not shown to isolate the specific geometry from generic distributional shifts. No load-bearing self-citations or external uniqueness theorems are invoked in the provided text, so the circularity is limited to the observation-to-framework step rather than a full self-referential loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities with supporting detail. The central concept of 'basin structure' functions as a postulated explanatory entity whose independent evidence is not described.

axioms (1)
  • domain assumption Autoregressive hidden-state trajectories capture the relevant dynamical structure that determines factual correctness.
    Invoked when the paper states that separability is observed in these trajectories.
invented entities (1)
  • task-dependent hallucination basins no independent evidence
    purpose: To explain the geometric origin of hallucinations and enable steering
    New postulated structure in latent space introduced to account for observed behavior.

pith-pipeline@v0.9.0 · 5415 in / 1278 out tokens · 46537 ms · 2026-05-10T18:52:48.452831+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

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

  1. Hallucination as Trajectory Commitment: Causal Evidence for Asymmetric Attractor Dynamics in Transformer Generation

    cs.LG 2026-04 unverdicted novelty 6.0

    Hallucination is an early trajectory commitment in transformers governed by asymmetric attractor dynamics, with prompt encoding selecting the basin and correction needing multi-step intervention.

Reference graph

Works this paper leans on

35 extracted references · 19 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Large language models hallucination: A comprehensive survey.arXiv preprint arXiv:2510.06265, 2025

    Alansari, A. and Luqman, H. Large language models hallucination: A comprehensive survey. arXiv:2510.06265, 2025

    work page arXiv 2025

  3. [3]

    and Zhang, H

    Chen, B. and Zhang, H. High-order rotor H opfield neural networks for associative memory. Neurocomputing, 616: 0 128893, 2025. doi:10.1016/j.neucom.2024.128893

    work page doi:10.1016/j.neucom.2024.128893 2025

  4. [4]

    INSIDE: LLMs’ internal states retain the power of hallucination detection.arXiv preprint arXiv:2402.03744,

    Chen, C., Liu, K., Chen, Z., Gu, Y., Wu, Y., Tao, M., Fu, Z., and Ye, J. INSIDE : LLMs' internal states retain the power of hallucination detection. arXiv:2402.03744, 2024 a

    work page arXiv 2024

  5. [5]

    In-context sharpness as alerts: An inner representation perspective for hallucination mitigation

    Chen, S., Xiong, M., Liu, J., Wu, Z., Xiao, T., Gao, S., and He, J. In-context sharpness as alerts: An inner representation perspective for hallucination mitigation. In Proceedings of the 41st International Conference on Machine Learning, pp.\ 7553--7567, 2024 b

    2024

  6. [6]

    E., Janson, N

    Essex, A. E., Janson, N. B., Norris, R. A., and Balanov, A. G. Memorisation and forgetting in a learning H opfield neural network: bifurcation mechanisms, attractors and basins. arXiv:2508.10765, 2025

    work page arXiv 2025

  7. [7]

    Nature , year =

    Farquhar, S., Kossen, J., Kuhn, L., and Gal, Y. Detecting hallucinations in large language models using semantic entropy. Nature, 630: 0 625--630, 2024. doi:10.1038/s41586-024-07421-0

    work page doi:10.1038/s41586-024-07421-0 2024

  8. [8]

    Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79 0 (8): 0 2554--2558, 1982. doi:10.1073/pnas.79.8.2554

    work page doi:10.1073/pnas.79.8.2554 1982

  9. [9]

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions , volume=

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., and Liu, T. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43 0 (2): 0 42, 2025. doi:10.1145/3703155

    work page doi:10.1145/3703155 2025

  10. [10]

    Associative memory model with neural networks: Memorizing multiple images with one neuron

    Inazawa, H. Associative memory model with neural networks: Memorizing multiple images with one neuron. arXiv:2510.06542, 2025

    work page arXiv 2025

  11. [11]

    S., Krotov, D., Bicknell, B

    Kafraj, M. S., Krotov, D., Bicknell, B. A., and Latham, P. E. A biologically plausible associative memory network. In ICLR 2025 Workshop on New Frontiers in Associative Memories, 2025. URL https://openreview.net/forum?id=u4YzOzEMfR

    2025

  12. [12]

    (Im) possibility of automated hallucination detection in large language models

    Karbasi, A., Montasser, O., Sous, J., and Velegkas, G. (Im) possibility of automated hallucination detection in large language models. In ICML 2025 Workshop on Reliable and Responsible Foundation Models, 2025. URL https://openreview.net/forum?id=B4SFmNvBNz

    2025

  13. [13]

    Halueval: A large-scale hallucination evaluation benchmark for large language models.arXiv preprint arXiv:2305.11747, 2023

    Li, J., Cheng, X., Zhao, W. X., Nie, J.-Y., and Wen, J.-R. HaluEval : A large-scale hallucination evaluation benchmark for large language models. arXiv:2305.11747, 2023

    work page arXiv 2023

  14. [14]

    Dynamic analysis and implementation of a multi-stable H opfield neural network

    Li, X., Luo, M., Zhang, B., and Liu, S. Dynamic analysis and implementation of a multi-stable H opfield neural network. Chaos, Solitons & Fractals, 199: 0 116657, 2025. doi:10.1016/j.chaos.2025.116657

    work page doi:10.1016/j.chaos.2025.116657 2025

  15. [15]

    TruthfulQA : Measuring how models mimic human falsehoods

    Lin, S., Hilton, J., and Evans, O. TruthfulQA : Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pp.\ 3214--3252, 2022

    2022

  16. [16]

    Least Squares Quantization in PCM,

    Lloyd, S. Least squares quantization in PCM . IEEE Transactions on Information Theory, 28 0 (2): 0 129--137, 1982. doi:10.1109/TIT.1982.1056489

    work page doi:10.1109/tit.1982.1056489 1982

  17. [17]

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

    Manakul, P., Liusie, A., and Gales, M. SelfCheckGPT : Zero-resource black-box hallucination detection for generative large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp.\ 9004--9017, 2023

    2023

  18. [18]

    The Meta Llama 3.2 collection of multilingual language models

    Meta AI . The Meta Llama 3.2 collection of multilingual language models. https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/MODEL_CARD.md, 2024. Accessed: 2026-01-15

    2024

  19. [19]

    Steer LLM latents for hallucination detection

    Park, S., Du, X., Yeh, M.-H., Wang, H., and Li, Y. Steer LLM latents for hallucination detection. In Proceedings of the Forty-second International Conference on Machine Learning, pp.\ 47971--47990, 2025

    2025

  20. [20]

    Bistability of somatic pattern memories: stochastic outcomes in bioelectric circuits underlying regeneration

    Pezzulo, G., LaPalme, J., Durant, F., and Levin, M. Bistability of somatic pattern memories: stochastic outcomes in bioelectric circuits underlying regeneration. Philosophical Transactions of the Royal Society B: Biological Sciences, 376 0 (1821): 0 20190765, 2021. doi:10.1098/rstb.2019.0765

    work page doi:10.1098/rstb.2019.0765 2021

  21. [21]

    K., Klambauer, G., Brandstetter, J., and Hochreiter, S

    Ramsauer, H., Sch \" a fl, B., Lehner, J., Seidl, P., Widrich, M., Gruber, L., Holzleitner, M., Adler, T., Kreil, D., Kopp, M. K., Klambauer, G., Brandstetter, J., and Hochreiter, S. Hopfield networks is all you need. In Proceedings of the 9th International Conference on Learning Representations, 2021

    2021

  22. [22]

    Gemma 2: Improving Open Language Models at a Practical Size

    Riviere, M., Pathak, S., Sessa, P. G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al. Gemma 2: Improving open language models at a practical size. arXiv:2408.00118, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    R., Saxena, A., Maharaj, K., Ahmad, A

    Sahoo, N. R., Saxena, A., Maharaj, K., Ahmad, A. A., Mishra, A., and Bhattacharyya, P. Addressing bias and hallucination in large language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, pp.\ 73--79, 2024

    2024

  24. [24]

    S., Saha, S., Kattakinda, P., and Feizi, S

    Sriramanan, G., Bharti, S., Sadasivan, V. S., Saha, S., Kattakinda, P., and Feizi, S. LLM-Check : Investigating detection of hallucinations in large language models. In Advances in Neural Information Processing Systems, volume 37, pp.\ 34188--34216. 2024

    2024

  25. [25]

    Why and how LLM s hallucinate: Connecting the dots with subsequence associations

    Sun, Y., Gai, Y., Chen, L., Ravichander, A., Choi, Y., Dziri, N., and Song, D. Why and how LLM s hallucinate: Connecting the dots with subsequence associations. In Advances in Neural Information Processing Systems, volume 37, pp.\ 34188--34216. 2025 a

    2025

  26. [26]

    Associative transformer

    Sun, Y., Ochiai, H., Wu, Z., Lin, S., and Kanai, R. Associative transformer. In Proceedings of the IEEE/CVF Computer Vision and Pattern Recognition Conference, pp.\ 4518--4527, 2025 b

    2025

  27. [27]

    Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability,

    Sun, Z., Zang, X., Zheng, K., Song, Y., Xu, J., Zhang, X., Yu, W., and Li, H. ReDeEP : Detecting hallucination in retrieval-augmented generation via mechanistic interpretability. arXiv:2410.11414, 2024

    work page arXiv 2024

  28. [28]

    InterpDetect : Interpretable signals for detecting hallucinations in retrieval-augmented generation

    Tan, L., Huang, K.-W., Shi, J., and Wu, K. InterpDetect : Interpretable signals for detecting hallucinations in retrieval-augmented generation. arXiv:2510.21538, 2025

    work page arXiv 2025

  29. [29]

    FEVER: a large-scale dataset for Fact Extraction and VERification

    Thorne, J., Vlachos, A., Christodoulopoulos, C., and Mittal, A. FEVER : a large-scale dataset for F act E xtraction and VER ification. In Walker, M., Ji, H., and Stent, A. (eds.), Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics , pp.\ 809--819, June 2018. doi:10.18653/v1/N18-1074

    work page internal anchor Pith review doi:10.18653/v1/n18-1074 2018

  30. [30]

    M u S i Q ue: Multihop questions via single-hop question composition

    Trivedi, H., Balasubramanian, N., Khot, T., and Sabharwal, A. MuSiQue : Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10: 0 539--554, 2022. doi:10.1162/tacl_a_00475

    work page doi:10.1162/tacl_a_00475 2022

  31. [31]

    Varshney, K. R. Generalization error of linear discriminant analysis in spatially-correlated sensor networks. IEEE Transactions on Signal Processing, 60 0 (6): 0 3295--3301, 2012. doi:10.1109/TSP.2012.2190063

    work page doi:10.1109/tsp.2012.2190063 2012

  32. [32]

    Adaptive activation steering: A tuning-free LLM truthfulness improvement method for diverse hallucinations categories

    Wang, T., Jiao, X., Zhu, Y., Chen, Z., He, Y., Chu, X., Gao, J., Wang, Y., and Ma, L. Adaptive activation steering: A tuning-free LLM truthfulness improvement method for diverse hallucinations categories. In Proceedings of the ACM on Web Conference 2025, pp.\ 2562--2578, 2025

    2025

  33. [33]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report. arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Alleviating hallucinations in large language models through multi-model contrastive decoding and dynamic hallucination detection

    Zhu, C., Liu, Y., Zhang, H., Wang, A., Chen, G., Wang, L., Luo, W., Zhang, K., et al. Alleviating hallucinations in large language models through multi-model contrastive decoding and dynamic hallucination detection. In Advances in Neural Information Processing Systems, volume 39. 2025

    2025

  35. [35]

    Look twice before you answer: Memory-space visual retracing for hallucination mitigation in multimodal large language models

    Zou, X., Wang, Y., Yan, Y., Lyu, Y., Zheng, K., Huang, S., Chen, J., Jiang, P., Liu, J., Tang, C., and Hu, X. Look twice before you answer: Memory-space visual retracing for hallucination mitigation in multimodal large language models. In Proceedings of the 42nd International Conference on Machine Learning, pp.\ 80873--80899, 2025

    2025