Shadows in the Attention: Contextual Perturbation and Representation Drift in the Dynamics of Hallucination in LLMs

He Li; Shuo Wang; Xiaohui Rong; Xuemin Liu; Zeyu Wei

REVIEW 4 major objections 4 minor 2 cited by

Shadows in the Attention: Contextual Perturbation and Representation Drift in the Dynamics of Hallucination in LLMs

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LLM hallucinations harden once attention distributions stop changing, and two drift statistics mark that threshold.

desk verdict The 'attention-locking threshold' is a padding artifact: JS-Drift saturates at log 2 because Eq. (14) zero-pads and renormalizes, and the paper's own Table 1 contradicts the claimed monotonic rise in hallucination rates. read the letter →

arxiv 2505.16894 v1 pith:4HKG74CZ submitted 2025-05-22 cs.CL

Zeyu Wei , Shuo Wang , Xiaohui Rong , Xuemin Liu , He Li This is my paper · ORCID

classification cs.CL

keywords hallucinationlargelanguagemodelsattentiondriftinternalrepresentationTruthfulQAcontextperturbationdetectionlocking

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that hallucinations in large language models are not random failures but the visible endpoint of a measurable internal process: as context is added round by round, hidden representations and attention distributions drift, hallucination rates climb, and then the system settles into a state the authors call “attention locking.” At that point, two statistics—JS-Drift near 0.69 and Spearman-Drift near zero—stop changing, and the errors they accompany become resistant to correction. This matters because it offers a candidate internal signature for predicting when a model is about to hallucinate and when correction will no longer work, which could support detection and mitigation before errors solidify.

What carries the argument

The load-bearing object is the pair of attention-drift statistics computed from zero-padded attention distributions. JS-Drift is the Jensen–Shannon divergence between baseline and context-injected attention, bounded in $[0,1]$; Spearman-Drift is the Spearman rank correlation between the same distributions, bounded in $[-1,1]$. Zero-padding with renormalization lets the authors compare attention over different token sets as context grows. Convergence of JS-Drift near $0.69$ and Spearman-Drift near zero is read as attention locking, the point where attention re-ranking stabilizes and the model’s error pattern hardens; together with cosine and entropy drift it forms the internal-state monitoring pipeline that links overt hallucination flags to hidden-state dynamics.

What would settle it

Run the same 15-round titration with content-free or randomly shuffled context snippets while keeping token counts identical to the real tracks; if JS-Drift still converges to about 0.69 and Spearman-Drift to about zero across all six models, the attention-locking threshold is an artifact of support-size change and padding rather than a cognitive-state signature. A cheaper check is to compute the drift metrics against a fixed-length reference distribution without zero-padding and see whether the universal saturation values persist.

Watch

Extended reading notes

Core claim

Across six open-source transformer models answering TruthfulQA questions, the authors build two 16-round “titration” tracks per question: one appends relevant but partially flawed snippets and the other injects misleading content. They report that overt hallucination rates rise monotonically with context accumulation and plateau after roughly five to seven rounds, while four drift metrics—cosine and entropy drift of hidden states, plus Jensen–Shannon and Spearman drift of attention distributions—also increase and stabilize. The central discovery is the synchronous convergence of JS-Drift (about $0.69$) and Spearman-Drift (about zero) across all models, which is interpreted as an “attention-locking” threshold: once attention topology stops changing, hallucinations solidify and conventional dialogue-based correction becomes ineffective. The paper also distinguishes two error modes—high-confidence, self-consistent hallucinations under relevant context and topic-drift hallucinations under irrelevant context—and reports a negative correlation between assimilation capacity and attention diffusion across model sizes.

Load-bearing premise

The central claim assumes that the convergence of the two attention-drift measures reflects the model’s internal state, not just the arithmetic of padding and renormalizing attention vectors of different lengths.

Editorial extensions

If this is right

Hallucination frequency rises monotonically with accumulated context and saturates after 5–7 rounds, so the first few injected snippets carry disproportionate risk.
Relevant but partially flawed context drives deep semantic assimilation, producing high-confidence, self-consistent hallucinations, whereas irrelevant context produces topic-drift errors anchored by attention re-routing.
The attention-locking threshold (JS-Drift near 0.69, Spearman-Drift near 0) marks the point where hallucinated content becomes resistant to conventional dialogue correction.
The negative correlation between assimilation capacity and attention diffusion implies larger models absorb biased context more thoroughly while smaller models compensate by scattering attention and producing looser patchwork errors.
Internal drift metrics are candidate intrinsic precursors for hallucination prediction and context-aware mitigation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the threshold is a genuine state property, it suggests a practical early-warning rule: monitor JS-Drift and Spearman-Drift during context construction and stop injecting before convergence, rather than attempting corrections after lock-in.
The near-identical saturation values across models and contexts raise the possibility that they are set by the zero-padding and renormalization procedure rather than by the model’s attention state; a direct test is to run the same protocol with randomly shuffled or content-free context snippets.
The same measurement could transfer to retrieval-augmented generation, where appended retrieved passages form a natural titration track; the paper does not test this setting, but its protocol would apply directly.
Because the hallucination detector uses empirically chosen thresholds such as a BERTScore cutoff of 0.7, the reported rates might shift under other judges; cross-detector replication would tighten the mapping from drift to error.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

The headline result—an 'attention-locking threshold' at JS-Drift ≈ 0.69—is almost certainly an artifact of Eq. (14). Zero-padding the baseline attention vector with ϵ = 10⁻¹² and renormalizing forces the JS divergence toward log 2 = 0.6931 as the context grows and the supports become nearly disjoint. Table 2's asymptotic values (0.6913/0.6916) sit right on that limit. The identical Spearman-Drift columns across different models (e.g., Llama3-8B and Llama3.2-1B) confirm the metric is not capturing model-specific attention; the growing block of padded ϵ entries washes out rank information.

What is genuinely new here is the 16-round relevant/irrelevant context titration protocol, and the idea of watching both overt hallucination rates and internal drift over rounds. That is a reasonable design, and the paper is transparent about citing prior internal-state work (Azaria & Mitchell, Duan et al., Wang et al.). The writing is clear.

But the load-bearing claims do not survive contact with the paper's own tables. Table 1 shows non-monotonic hallucination rates: Qwen2.5-7B relevant goes 0.94 (round 5), 0.92 (round 7), 0.90 (round 11); Llama3-8B relevant goes 0.90 (round 1), 0.86 (round 3). So 'monotonic growth' is false. Corollary 4's 'resistant to correction' is asserted without any correction-resistance experiment. The detector threshold θ_sem = 0.7 is uncalibrated, and the full data are withheld 'upon request.' The seesaw correlation (ρ ≈ −0.71) is computed on the same rounds that define the drift metrics, so it is partly circular.

These are load-bearing flaws, not nits. The protocol could be salvageable if the authors recompute drift on overlapping token sets or with a proper projection, and if they actually test correction resistance. But as it stands, the quantitative core is not evidence for the proposed mechanism.

My take: reject in current form. It is a useful reading-group example of how padding artifacts can manufacture signals, but I would not cite it. If the authors fix the metric and release the data, the protocol deserves another look. I would not send this to peer review as-is; the artifact is diagnosable in a sentence, and the data are not available for verification.

Referee Report

4 major / 4 minor

Summary. The paper studies how incremental context injection affects hallucination rates and internal representations in six open-source LLMs. It constructs two 16-round tracks per TruthfulQA question (relevant but partially flawed snippets versus irrelevant misleading snippets), measures QA-level and sentence-level hallucination rates with a tri-perspective detector, and tracks cosine, entropy, JS, and Spearman drifts of hidden states and attention maps. The central claim is that hallucination rates grow monotonically and saturate once an 'attention-locking threshold' (JS-Drift ≈ 0.69, Spearman-Drift ≈ 0) is reached, after which hallucinations become resistant to correction. The paper also reports a seesaw between semantic assimilation and attention diffusion that varies with model size.

Significance. If the central claims were supported, the paper would supply an interpretable internal-state signature for predicting when hallucinations resist correction, with practical implications for mitigation. The controlled dual-track design across six models and the attempt to couple overt error rates with covert representation drift are appropriate and potentially useful. However, the principal quantitative claims are not supported by the reported data: the headline monotonicity is violated in Table 1, and the attention-locking threshold is numerically consistent with an artifact of the zero-padding procedure in Eq. (14). The absence of any correction-resistance experiment means the abstract's central behavioral conclusion is asserted rather than demonstrated. The paper does not ship code, data, or machine-checked artifacts, and key tables are available only on request, further limiting verification.

major comments (4)

[§5, Table 1] The claim that hallucination rates 'rise monotonically and saturate' is contradicted by the paper's own table. For Qwen2.5-7B relevant QA-HallucRate, the reported values are 0.86 (round 1), 0.90 (round 3), 0.94 (round 5), 0.92 (round 7), 0.92 (round 9), 0.90 (round 11), and 0.92 (round 15); for Llama3-8B relevant, they are 0.90, 0.86, 0.88, 0.88, 0.90, 0.90, 0.90. These trajectories are non-monotonic and include decreases of 0.02–0.04 between consecutive reported rounds. The abstract and Section 5 state monotonic growth, so this load-bearing claim needs to be revised or the data reanalyzed before the paper's behavioral conclusions can be accepted.
[§3.3, Eq. (14); Table 2] The attention-locking threshold is not established because the JS-Drift metric, as defined, is dominated by the padding/support-size artifact. Zero-padding distributions of different lengths and renormalizing compares a baseline distribution over the initial token set with a current distribution over a growing token set; as the overlap of effective supports shrinks, the JS divergence approaches log 2 ≈ 0.6931, which is numerically consistent with the reported saturation values (0.6913–0.6916) across all six models. The near-identical JS-Drift values across models (e.g., Qwen2.5-7B and Qwen2.5-1.5B have identical rounded values 0.6861, 0.6886, 0.6897, 0.6903, 0.6907, 0.6913 from round 3 onward) and the Spearman-Drift values tending to zero are exactly what one expects from a padded-vector comparison with a growing block of near-zero entries, rather than from model-specific attention locking. The paper needs an alignment or renormalization procedure that separates genuine attention change from support growth, or an independent validation of the threshold.
[§4.3, Corollary 4] No experiment tests correction resistance. Corollary 4 asserts that once JS-Drift saturates and Spearman-Drift approaches zero, hallucinations become solidified and conventional dialogue correction methods become ineffective, but the manuscript reports no measurement of correction effectiveness before versus after the threshold. The sentence in Section 4.3 that 'the effectiveness of injected corrective information significantly decreases' is an interpretation, not a result. Since the abstract's third main claim is exactly this correction-resistance property, it requires a direct experiment, for example injecting corrective statements at pre-threshold and post-threshold rounds and measuring whether the model's answers change.
[§4.3, Corollaries 3 and 6; §3.2, Eq. (4)] Several correlational and threshold claims are circular or under-validated. The drift metrics in Table 2 are computed on the same six models and the same rounds that are then used to define the 'seesaw' correlation (ρ ≈ −0.71, p < 0.01), with no independence, confidence intervals, or correction for multiple comparisons across the many reported correlations. Corollary 3 equates high metric variance with 'high-confidence self-consistent hallucinations,' but no confidence or self-consistency measurement is reported. In addition, the hallucination detector uses θsem = 0.7 described as 'empirically chosen' (Eq. 4), with no sensitivity analysis, no human validation of the logical-OR label, and no reported sample size N for the rates in Table 1. These issues affect the paper's behavioral metrics as well as its internal-state claims.

minor comments (4)

[§3.2, Eq. (9); §4.2; Table 1] The metric is defined as Intra-HallucRate in Eq. (9) but is called inter-HallucRate in Section 4.2 and Table 1; please standardize the terminology.
[§3.3, Eq. (14)] Equation (14) is ambiguous: it sets Ppad(i) = P(i) for i < |P| and epsilon otherwise, then says 'followed by renormalization'; it should specify whether epsilon is added to every position before renormalization or only to the padded positions, and how ties are handled for the Spearman rank computation.
[§4.3, Figures 2–3] Figures 2 and 3 show detailed drift curves only for Llama3-8B and Llama3.2-1B; the claims about trends across all six models would be easier to verify if the same plots were provided for the other four models or if the per-round values were released.
[§4.2, Table 1] The complete tables are 'available upon request from the authors'; for reproducibility, the data and generation code should be released alongside the paper.

Circularity Check

3 steps flagged · score 6.0 of 10

The 'attention-locking' threshold is the zero-padding ceiling of Eq. (14), so the central universal-threshold claim reduces to the metric definition.

renaming known result [Section 3.3, Eq. (14); Section 4.3, Table 2 and Corollary 4]
"To enable comparison between distributions of different lengths, we apply zero-padding: Ppad(i) = ( P (i) if i < |P |; ϵ otherwise ) where ϵ is a small positive constant (typically 10−12), followed by renormalization. ... Across all models, JS-Drift typically reaches stable saturation values (approximately 0.690±0.001) within 6-8 rounds ... Corollary 4. Once a model’s internal attention structure enters a "locking threshold" (JS-Drift saturation and Spearman-Drift approaching zero), generated hallucinations enter a solidified state, making conventional dialogue correction methods ineffective."

The saturation value reported as the attention-locking threshold is the standard Jensen-Shannon ceiling log 2 ≈ 0.6931 for two distributions with near-disjoint supports. Equation (14) creates that near-disjointness by zero-padding the baseline attention vector with epsilon while the current-round vector grows with context, so JS-Drift is forced toward ~0.69 regardless of model or cognitive state. Table 2 shows all six models converging to the same 0.6913-0.6916 values. Calling this mathematical ceiling an 'attention-locking threshold' renames a property of the padding operation as an empirical discovery, and the threshold is then used as the independent marker for hallucination solidification.
self definitional [Section 3.3, Eq. (15); Section 4.3]
"Zero-padding is again applied to maintain consistent dimensionality. ... Across all models, JS-Drift typically reaches stable saturation values ... while Spearman-Drift simultaneously converges near zero ... This "synchronous convergence" phenomenon shows significant negative correlation with hallucination correctability (ρ ≈ −0.66)."

The Spearman-Drift convergence to ~0 is built into the same zero-padding: once the baseline vector is padded with a long block of constant epsilon entries, those tied ranks dominate the rank correlation and drive rho toward zero as context length grows. The paper presents this forced convergence as a second, independent confirmation of attention-locking ('synchronous convergence'), but it is generated by the same Eq. (14) padding operation, not by independently observed attention reordering.

1 more flagged steps

self definitional [Abstract; Section 4.3, Corollary 4]
"convergence of JS-Drift (~0.69) and Spearman-Drift (~0) marks an "attention-locking" threshold beyond which hallucinations solidify and become resistant to correction."

The explanatory construct 'attention-locking threshold' is defined by the very saturation values it is then used to explain. Corollary 4 asserts that after this threshold corrective information becomes ineffective, but the threshold is just the metric's padding-induced ceiling; no correction experiment is reported. The behavioral conclusion is therefore a relabeling of the input metric values rather than a prediction derived from an independently established internal state.

full rationale

The paper's other measured quantities — monotonic hallucination-rate growth, relevant/irrelevant differences in Cos-Drift and Ent-Drift, and the seesaw correlation — are self-contained empirical observations and are not circular. The circularity is concentrated in the central 'attention-locking threshold' claim. Equation (14) zero-pads and renormalizes attention distributions; the resulting JS-Drift ceiling (log 2 ≈ 0.6931) and the tied-rank-driven Spearman ~0 are mathematical properties of the padding scheme, not model-dependent discoveries. Table 2's near-identical saturation values across six architectures confirm that the threshold is a property of the measurement definition. Corollary 4 then attaches a behavioral consequence (correction-resistant solidified hallucinations) to this arithmetic ceiling without a correction experiment. This makes the headline universal-threshold claim reduce by construction, while leaving the remaining drift/hallucination correlations as independent but weaker evidence. No load-bearing self-citations are present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central analysis rests on an uncalibrated hallucination detector, an untested assumption about which internal states carry the signal, and a zero-padding scheme that may produce the apparent attention-locking saturation. The 'attention-locking threshold' is defined from the same dataset used to claim its consequences. These choices, rather than new physical entities, are the main things the reader is asked to accept.

free parameters (2)

theta_sem = 0.7
Empirically chosen BERTScore threshold in Eq. (4) for semantic deviation hallucination detection. No calibration against human annotations or sensitivity analysis is reported.
epsilon_padding = 10^-12
Zero-padding constant in Eq. (14) used to compare attention distributions of different lengths. The convergence of JS-Drift depends on this padding and the changing token support.

assumptions (4)

domain assumption Hallucination can be validly labeled by the OR of BERTScore threshold, entity extension, and NLI entailment (Eq. 7).
Invoked in Section 3.2. The detector is not validated against human labels or a gold standard, so all hallucination-rate claims inherit this assumption.
domain assumption The final-token final-layer hidden state and attention distributions capture hallucination-relevant internal drift.
Invoked in Section 3.3. Layers and heads are not specified, and earlier work cited by the paper shows mid-layer signals can be more predictive.
ad hoc to paper Zero-padding and renormalization yield a meaningful JS divergence and Spearman correlation between attention distributions over different token sets.
Eq. (14). The saturation near JS-Drift 0.69 across all models may be an artifact of changing support size rather than a cognitive threshold.
domain assumption GPT-4o generated snippets, with unspecified human evaluation, constitute valid 'relevant' and 'irrelevant' context manipulations.
Section 4.1. The generation and selection procedure is not released, so the two tracks are not independently reproducible.

invented entities (1)

Attention-locking threshold
purpose: Labels the observed saturation of JS-Drift around 0.69 and Spearman-Drift near zero as a state after which hallucinations become resistant to correction.
The threshold is defined from the same data used to assert its behavioral consequence. No intervention or out-of-sample test is provided to show that corrections actually fail after this point.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Shadows in the Attention: Contextual Perturbation and Representation Drift in the Dynamics of Hallucination in LLMs." pith.science (2026). https://pith.science/paper/4HKG74CZ

@misc{pith2026250516894,
  author       = {Pith},
  title        = {Pith review of: Shadows in the Attention: Contextual Perturbation and Representation Drift in the Dynamics of Hallucination in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4HKG74CZ}},
  note         = {Machine review of arXiv:2505.16894}
}

abstract

Hallucinations -- plausible yet erroneous outputs -- remain a critical barrier to reliable deployment of large language models (LLMs). We present the first systematic study linking hallucination incidence to internal-state drift induced by incremental context injection. Using TruthfulQA, we construct two 16-round "titration" tracks per question: one appends relevant but partially flawed snippets, the other injects deliberately misleading content. Across six open-source LLMs, we track overt hallucination rates with a tri-perspective detector and covert dynamics via cosine, entropy, JS and Spearman drifts of hidden states and attention maps. Results reveal (1) monotonic growth of hallucination frequency and representation drift that plateaus after 5--7 rounds; (2) relevant context drives deeper semantic assimilation, producing high-confidence "self-consistent" hallucinations, whereas irrelevant context induces topic-drift errors anchored by attention re-routing; and (3) convergence of JS-Drift ($\sim0.69$) and Spearman-Drift ($\sim0$) marks an "attention-locking" threshold beyond which hallucinations solidify and become resistant to correction. Correlation analyses expose a seesaw between assimilation capacity and attention diffusion, clarifying size-dependent error modes. These findings supply empirical foundations for intrinsic hallucination prediction and context-aware mitigation mechanisms.

Figures

Figures reproduced from arXiv: 2505.16894 by the authors.

**Figure 2.** Internal State Drift Metrics of the Llama3-8B Model 2 4 6 8 10 12 14 Round 0.28 0.30 0.32 0.34 0.36 Cosine Drift Cosine Drift Relevant Irrelevant 2 4 6 8 10 12 14 Round 0.8 1.0 1.2 1.4 1.6 Entropy Drift Entropy Drift Relevant Irrelevant 2 4 6 8 10 12 14 Round 0.6750 0.6775 0.6800 0.6825 0.6850 0.6875 0.6900 JS Drift JS Drift Relevant Irrelevant 2 4 6 8 10 12 14 Round 0.10 0.08 0.06 0.04 0.02 Spearman Drift Spearman … view at source ↗

**Figure 3.** Internal State Drift Metrics of the Llama3.2-1B Model Corollary 1. Internal state drift is a universal phenomenon in large language models under accumulated long-context conditions, and hallucination risk gradually increases with cumulative internal drift. This indicates that long-sequence contexts induce systematic shifts in model hidden states and attention distributions. As dialogue rounds increase, the model’s … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Comparison of JS-Drift and Spearman-Drift under Relevant vs. Irrelevant Context Injection [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

Many-body Tipping Dynamics of ChatGPT-like AIs
cs.AI 2026-07 conditional novelty 6.0 of 10

Tipping of ChatGPT-like AI to undesirable outputs is modeled as first-passage transport of a residual-state spin across an output-basin wall, with attention disorder controlling the crossing.
Temperature-driven inversion and nonlinear dynamics in ChatGPT-like AIs
physics.soc-ph 2026-08 reject novelty 5.0 of 10

A projection of LLM internal states, trained on some runs, predicts repetition on held-out runs and can be steered to change repetition; the headline entropy maximum is a reparameterization of an occupancy split.

Reference graph

Works this paper leans on

30 extracted references · 17 canonical work pages · cited by 2 Pith papers

[1]

In: Findings of the Association for Computational Linguistics: EMNLP 2023

Azaria, A., Mitchell, T.: The internal state of an LLM knows when it‘s lying. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 967–976. Singapore (Dec 2023)

work page 2023
[2]

arXiv preprint arXiv:2504.10063 (2025)

Bazarova, A., Yugay, A., et al.: Hallucination detection in llms via topological divergence on attention graphs. arXiv preprint arXiv:2504.10063 (2025)

arXiv 2025
[3]

In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

Brown, T., Mann, B., Ryder, N., et al.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877–1901. Curran Associates, Inc. (2020)

work page 2020
[4]

Bubeck, S., Chandrasekaran, V., Eldan, R., et al.: Sparks of artificial general in- telligence: Early experiments with gpt-4 (2023), https://arxiv.org/abs/2303.12712

arXiv 2023
[5]

Chen, C., Liu, K., Chen, Z., et al.: Inside: Llms’ internal states retain the power of hallucination detection (2024), https://arxiv.org/abs/2402.03744

arXiv 2024
[6]

In: Proceedings of the 2024 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)

Chen, J., Kim, G., Sriram, A., et al.: Complex claim verification with evidence retrieved in the wild. In: Proceedings of the 2024 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). pp. 3569–3587. Mexico City, Mexico (Jun 2024)

work page 2024
[7]

Journal of Machine Learning Research24(240), 1–113 (2023)

Chowdhery, A., Narang, S., Devlin, J., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research24(240), 1–113 (2023)

work page 2023
[8]

In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

Chuang, Y.S., Qiu, L., Hsieh, C.Y., et al.: Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 1419–1436. Miami, Florida, USA (Nov 2024)

work page 2024

Show all 30 references

[9]

Duan, H., Yang, Y., Tam, K.Y.: Do llms know about hallucination? an empirical investigation of llm’s hidden states (2024), https://arxiv.org/abs/2402.09733

2024 arXiv
[10]

In: Findings of the Association for Computational Linguistics: ACL 2024

Fadeeva, E., Rubashevskii, A., Shelmanov, A., et al.: Fact-checking the output of large language models via token-level uncertainty quantification. In: Findings of the Association for Computational Linguistics: ACL 2024. pp. 9367–9385 (Aug 2024)

2024
[11]

Nature630(8017), 625–630 (2024) Title Suppressed Due to Excessive Length 15

Farquhar, S., Kossen, J., Kuhn, L., Gal, Y.: Detecting hallucinations in large lan- guage models using semantic entropy. Nature630(8017), 625–630 (2024) Title Suppressed Due to Excessive Length 15

2024
[12]

In: Findings of the Association for Computational Linguistics: EMNLP 2020

Filippova, K.: Controlled hallucinations: Learning to generate faithfully from noisy data. In: Findings of the Association for Computational Linguistics: EMNLP 2020. pp. 864–870. Online (Nov 2020)

2020
[13]

arXiv preprint arXiv:2410.03026 (2024)

Flemings, J., Zhang, W., Jiang, B., et al.: Characterizing context influence and hallucination in summarization. arXiv preprint arXiv:2410.03026 (2024)

2024
[14]

Gao, Y., Xiong, Y., et al.: Retrieval-augmented generation for large language mod- els: A survey (2024), https://arxiv.org/abs/2312.10997

2024 arXiv
[15]

arXiv preprint arXiv:2407.21783 (2024)

Grattafiori, A., Dubey, A., Jauhri, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

2024 arXiv
[16]

ACM Transactions on Infor- mation Systems 43(2), 1–55 (2025)

Huang, L., Yu, W., et al.: A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Infor- mation Systems 43(2), 1–55 (2025)

2025
[17]

IEEE Transactions on Software Engineering 51(2), 413–429 (2025)

Huang, Y., Song, J., Wang, Z., et al.: Look before you leap: An exploratory study of uncertainty analysis for large language models. IEEE Transactions on Software Engineering 51(2), 413–429 (2025)

2025
[18]

ACM computing surveys55(12), 1–38 (2023)

Ji, Z., Lee, N., Frieske, R., et al.: Survey of hallucination in natural language generation. ACM computing surveys55(12), 1–38 (2023)

2023
[19]

Kadavath, S., Conerly, T., Askell, A., et al.: Language models (mostly) know what they know (2022), https://arxiv.org/abs/2207.05221

2022 arXiv
[20]

In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

Lin, S., Hilton, J., Evans, O.: TruthfulQA: Measuring how models mimic human falsehoods. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 3214–3252 (May 2022)

2022
[21]

In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

Manakul, P., Liusie, A., Gales, M.: SelfCheckGPT: Zero-resource black-box hallu- cination detection for generative large language models. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 9004–9017. Singapore (Dec 2023)

2023
[22]

In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics

Maynez, J., Narayan, S., Bohnet, B., McDonald, R.: On faithfulness and factuality in abstractive summarization. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 1906–1919. Online (Jul 2020)

1906
[23]

Transactions of the Association for Computational Linguistics12, 484–506 (2024)

Pan, L., Saxon, M., Xu, W., et al.: Automatically correcting large language models: Surveying the landscape of diverse automated correction strategies. Transactions of the Association for Computational Linguistics12, 484–506 (2024)

2024
[24]

arXiv preprint arXiv:2311.16867 (2023)

Penedo, G., et al.: The falcon series of open language models. arXiv preprint arXiv:2311.16867 (2023)

2023 arXiv
[25]

arXiv preprint arXiv:2302.12813 (2023)

Peng, B., Galley, M., He, P., et al.: Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813 (2023)

2023 arXiv
[26]

arXiv preprint arXiv:2309.05922 (2023)

Rawte,V.,Sheth,A.,Das,A.:Asurveyofhallucinationinlargefoundationmodels. arXiv preprint arXiv:2309.05922 (2023)

2023 arXiv
[27]

In: Proceedings of the 2024 Annual International ACM SIGIR Con- ference on Research and Development in Information Retrieval in the Asia Pacific Region

Su, W., Tang, Y., Ai, Q., et al.: Mitigating entity-level hallucination in large lan- guage models. In: Proceedings of the 2024 Annual International ACM SIGIR Con- ference on Research and Development in Information Retrieval in the Asia Pacific Region. pp. 23–31 (2024)

2024
[28]

arXiv preprint arXiv:2302.13971 (2023)

Touvron, H., Lavril, T., Izacard, G., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

2023 arXiv
[29]

psychology

Wang, P., Liu, Y., Lu, Y., et al.: What are models thinking about? understand- ing large language model hallucinations "psychology" through model inner state analysis (2025), https://arxiv.org/abs/2502.13490

2025 arXiv
[30]

arXiv preprint arXiv:2412.15115 (2024)

Yang, A., Yang, B., Zhang, B., et al.: Qwen2.5 technical report. arXiv preprint arXiv:2412.15115 (2024)

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

In: Findings of the Association for Computational Linguistics: EMNLP 2023

Azaria, A., Mitchell, T.: The internal state of an LLM knows when it‘s lying. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 967–976. Singapore (Dec 2023)

work page 2023

[2] [2]

arXiv preprint arXiv:2504.10063 (2025)

Bazarova, A., Yugay, A., et al.: Hallucination detection in llms via topological divergence on attention graphs. arXiv preprint arXiv:2504.10063 (2025)

arXiv 2025

[3] [3]

In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

Brown, T., Mann, B., Ryder, N., et al.: Language models are few-shot learners. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 1877–1901. Curran Associates, Inc. (2020)

work page 2020

[4] [4]

Bubeck, S., Chandrasekaran, V., Eldan, R., et al.: Sparks of artificial general in- telligence: Early experiments with gpt-4 (2023), https://arxiv.org/abs/2303.12712

arXiv 2023

[5] [5]

Chen, C., Liu, K., Chen, Z., et al.: Inside: Llms’ internal states retain the power of hallucination detection (2024), https://arxiv.org/abs/2402.03744

arXiv 2024

[6] [6]

In: Proceedings of the 2024 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)

Chen, J., Kim, G., Sriram, A., et al.: Complex claim verification with evidence retrieved in the wild. In: Proceedings of the 2024 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). pp. 3569–3587. Mexico City, Mexico (Jun 2024)

work page 2024

[7] [7]

Journal of Machine Learning Research24(240), 1–113 (2023)

Chowdhery, A., Narang, S., Devlin, J., et al.: Palm: Scaling language modeling with pathways. Journal of Machine Learning Research24(240), 1–113 (2023)

work page 2023

[8] [8]

In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

Chuang, Y.S., Qiu, L., Hsieh, C.Y., et al.: Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 1419–1436. Miami, Florida, USA (Nov 2024)

work page 2024