pith. sign in

arxiv: 2606.01033 · v1 · pith:ACAQEQMPnew · submitted 2026-05-31 · 💻 cs.AI

TriLens: Per-Layer Logit-Lens Entropy for White-Box Hallucination Detection

Bohan Yang , Yijun Gong , Zhi Zhang , Ge Zhang , Wenpeng Xing , Meng Han This is my paper

Pith reviewed 2026-06-28 17:21 UTC · model grok-4.3

classification 💻 cs.AI
keywords hallucination detectionlogit lenswhite-box detectionentropy trajectoryper-layer analysisinstruction-tuned LLMsQA benchmarksinternal uncertainty
0
0 comments X

The pith

TriLens detects hallucinations by tracking entropy trajectories from three module readouts per layer via the logit lens.

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

The paper establishes that hallucinations in language models leave detectable traces in how certainty develops internally across layers. It proposes reading the attention output, feed-forward output, and residual stream at each layer through the model's logit lens, then recording only the entropy of those three projections. The resulting 3L-dimensional vector captures complementary signals from the three pathways and serves as input to a detector that works across instruction-tuned models and QA benchmarks without storing full hidden states or running multiple samples.

Core claim

At every layer the model reads its own multi-head self-attention output, feed-forward output, and residual stream through the logit lens and records the entropy of each readout; the resulting 3L-dimensional entropy trajectory distinguishes hallucinated from non-hallucinated answers because the three module-wise trajectories supply complementary evidence about how certainty forms across depth.

What carries the argument

The 3L-dimensional entropy trajectory assembled from per-layer logit-lens readouts of the attention, feed-forward, and residual modules, which tracks how certainty settles without high-dimensional storage or repeated sampling.

If this is right

  • Hallucination detection becomes possible from a compact internal signal rather than final-token probabilities alone.
  • The three separate trajectories supply non-redundant information about uncertainty at different computational stages.
  • No additional forward passes or token sampling are required beyond the single forward pass already used to generate the answer.
  • The approach applies uniformly to multiple instruction-tuned LLMs and standard QA benchmarks.

Where Pith is reading between the lines

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

  • Monitoring internal entropy formation could locate the layers where a hallucination first becomes likely.
  • The same readout pattern might be adapted to flag other forms of internal inconsistency, such as factual contradictions that appear midway through generation.
  • If the trajectories prove stable, they could serve as a lightweight monitoring signal during deployment without retraining the base model.

Load-bearing premise

The entropy values from the three per-layer readouts correlate with the presence of hallucinations in a way that holds for new models and new question sets rather than reflecting only the particular training data or evaluation choices used.

What would settle it

A test set of QA examples where the 3L entropy vectors show no statistically reliable difference between cases the model answers correctly and cases it hallucinates.

Figures

Figures reproduced from arXiv: 2606.01033 by Bohan Yang, Ge Zhang, Meng Han, Wenpeng Xing, Yijun Gong, Zhi Zhang.

Figure 1
Figure 1. Figure 1: TriLens: Per-layer logit-lens entropy tracks internal certainty during generation. Supported answers show coordinated entropy sharpening across MHSA, FFN, and residual-stream readouts, whereas hallucinated answers tend to retain higher, less stable, and less synchronized entropy across depth. inference-time scoring. Recent work has shown that hallucination-relevant information can be ex￾tracted from hidden… view at source ↗
Figure 2
Figure 2. Figure 2: TriLens overview. A single forward pass over the evaluation sequence extracts the MHSA output, FFN [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Single-layer AUROC heatmap from Hℓ x . Rows are benchmarks, columns are relative depth ℓ/L, and the peak cell in each row is circled. Peak￾discriminative depth varies systematically across mod￾els. et al., 2025). Appendix A.4 summarizes these base￾lines and the evaluation protocol used in our runs. All baseline results are produced under the same evaluation protocol, group-preserving 80/20 split, model sui… view at source ↗
Figure 4
Figure 4. Figure 4: Complementarity of the three entropy features. Left: standardized per-layer separation between hallucinated and correct samples for Hℓ a , Hℓ m, and Hℓ x on two representative TriviaQA cells. The stronger two￾feature branch is architecture-dependent. Right: feature-subset ablation across all 12 (model, dataset) cells. Colored lines show per-model means; grey lines show individual cells. Both two-feature br… view at source ↗
Figure 5
Figure 5. Figure 5: Cross-dataset generalization heatmaps for TriLens, ICR Probe, and SAPLMA. Each cell reports AUROC [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Probability density of Hℓ ∗ x at the peak￾discriminative layer for correct vs. hallucinated re￾sponses. Probe Dataset DoLa-JSD TriLens TriLens+DoLa MLP HaluEval 0.7623 0.9136 0.9155 SQuAD2 0.8538 0.9158 0.9172 HotpotQA 0.8928 0.9425 0.9422 TriviaQA 0.8679 0.8861 0.8860 Linear HaluEval 0.6790 0.8731 0.8738 SQuAD2 0.7189 0.8778 0.8798 HotpotQA 0.8111 0.9253 0.9256 TriviaQA 0.8383 0.8700 0.8710 [PITH_FULL_IM… view at source ↗
read the original abstract

When a language model hallucinates, the final answer is wrong, but the mistake is not necessarily invisible inside the model. Different internal pathways may remain uncertain, disagree in how quickly they sharpen, or commit to competing continuations before the output is produced. We introduce TriLens, a white-box detector that turns this intuition into a compact representation: at every layer, it reads the multi-head self-attention output, the feed-forward output, and the residual stream through the model's own logit lens, then records only the entropy of each readout. The resulting 3L-dimensional trajectory describes how certainty forms across depth and across modules, without storing high-dimensional hidden states or sampling multiple generations. This simple signal yields a strong detector across instruction-tuned LLMs and QA benchmarks, and our analyses show that the three module-wise entropy trajectories provide complementary evidence. TriLens suggests that hallucination detection can benefit from tracking how internal computation settles, not only what the final layer predicts.

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

2 major / 1 minor

Summary. The manuscript introduces TriLens, a white-box hallucination detector that, at each layer, projects the multi-head self-attention output, feed-forward output, and residual stream through the model's logit lens and records only the entropy of each readout. The resulting 3L-dimensional entropy trajectory is claimed to form a strong detector for hallucinations across instruction-tuned LLMs and QA benchmarks, with the three module-wise trajectories supplying complementary evidence.

Significance. If the empirical results hold, the method supplies an efficient, storage-light white-box signal that tracks how certainty forms across depth and modules without multiple generations or high-dimensional hidden states. This could meaningfully advance practical hallucination detection by focusing on internal settling dynamics rather than final-layer predictions alone.

major comments (2)
  1. [Abstract] Abstract: the central claim that the 3L-dimensional entropy trajectory 'yields a strong detector across instruction-tuned LLMs and QA benchmarks' is asserted without any reported model identities, benchmark identities, detector construction details (supervised vs. unsupervised), baseline comparisons, or statistical significance tests; these omissions are load-bearing for evaluating whether the correlation generalizes or is driven by dataset artifacts or post-hoc choices.
  2. [Abstract] Abstract: no information is supplied on controls for potential confounds (e.g., whether entropy trajectories remain predictive after accounting for output length, token frequency, or dataset-specific patterns), which directly affects the assertion that the three module-wise trajectories provide complementary evidence independent of such factors.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'without storing high-dimensional hidden states or sampling multiple generations' could be clarified by noting the exact memory or compute savings relative to common alternatives such as hidden-state probing or self-consistency sampling.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and for identifying areas where the abstract requires greater precision to support its claims. We address each point below and will revise the abstract accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the 3L-dimensional entropy trajectory 'yields a strong detector across instruction-tuned LLMs and QA benchmarks' is asserted without any reported model identities, benchmark identities, detector construction details (supervised vs. unsupervised), baseline comparisons, or statistical significance tests; these omissions are load-bearing for evaluating whether the correlation generalizes or is driven by dataset artifacts or post-hoc choices.

    Authors: We agree that the abstract should supply these concrete details so readers can immediately evaluate scope and robustness. In the revised manuscript we will expand the abstract to name the models (Llama-2-7B-chat, Mistral-7B-Instruct-v0.2), benchmarks (TruthfulQA, Natural Questions, HotpotQA), state that the detector is unsupervised (entropy trajectories fed to a lightweight logistic regressor or threshold), list the main baselines (final-layer entropy, perplexity, and self-consistency), and report statistical significance (paired t-tests and AUC confidence intervals). These elements already appear in Sections 4 and 5; we will simply surface them in the abstract as well. revision: yes

  2. Referee: [Abstract] Abstract: no information is supplied on controls for potential confounds (e.g., whether entropy trajectories remain predictive after accounting for output length, token frequency, or dataset-specific patterns), which directly affects the assertion that the three module-wise trajectories provide complementary evidence independent of such factors.

    Authors: We accept that the abstract must address this concern. The main text already contains regression-based controls and stratified analyses showing that module-wise entropy trajectories retain predictive power after accounting for length and frequency; however, the abstract does not mention these controls. We will revise the abstract to include a brief qualifier that the reported complementarity holds after partialling out output length and token frequency. Should the referee consider the existing controls insufficient, we are prepared to add further experiments during revision. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical detector with no derivations reducing to inputs

full rationale

The paper presents TriLens as an empirical method that computes entropy on per-layer logit-lens readouts from three module types and uses the resulting 3L trajectory as a hallucination detector. No equations, first-principles derivations, or predictions are shown that reduce the claimed detector to fitted parameters, self-definitions, or self-citation chains. The central claim is supported by empirical measurement across models and benchmarks rather than any construction that is tautological by definition.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the method relies on standard logit-lens usage and entropy computation without additional postulated constructs.

pith-pipeline@v0.9.1-grok · 5708 in / 1023 out tokens · 24418 ms · 2026-06-28T17:21:08.529337+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    Zhang, Zhenliang and Hu, Xinyu and Zhang, Huixuan and Zhang, Junzhe and Wan, Xiaojun , booktitle =

  2. [2]

    The Internal State of an LLM Knows When It`s Lying

    Azaria, Amos and Mitchell, Tom , editor =. The Internal State of an. Findings of the Association for Computational Linguistics: EMNLP 2023 , month = dec, year =. doi:10.18653/v1/2023.findings-emnlp.68 , pages =

    work page doi:10.18653/v1/2023.findings-emnlp.68 2023

  3. [3]

    ArXiv , year =

    Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs , author =. ArXiv , year =

  4. [4]

    The Twelfth International Conference on Learning Representations , year =

    INSIDE: LLMs' Internal States Retain the Power of Hallucination Detection , author =. The Twelfth International Conference on Learning Representations , year =

  5. [5]

    Advances in Neural Information Processing Systems , doi =

    LLM-Check: Investigating Detection of Hallucinations in Large Language Models , author =. Advances in Neural Information Processing Systems , doi =

  6. [6]

    Do Androids Know They

    CH-Wang, Sky and Van Durme, Benjamin and Eisner, Jason and Kedzie, Chris , editor =. Do Androids Know They. Findings of the Association for Computational Linguistics: ACL 2024 , month = aug, year =. doi:10.18653/v1/2024.findings-acl.260 , pages =

    work page doi:10.18653/v1/2024.findings-acl.260 2024

  7. [7]

    The Twelfth International Conference on Learning Representations , year =

    DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models , author =. The Twelfth International Conference on Learning Representations , year =

  8. [8]

    Interpreting

    nostalgebraist , year =. Interpreting

  9. [9]

    arXiv preprint arXiv:2303.08112 , year =

    Eliciting latent predictions from transformers with the tuned lens , author =. arXiv preprint arXiv:2303.08112 , year =

    Pith/arXiv arXiv

  10. [10]

    Transformer Feed-Forward Layers Are Key-Value Memories

    Transformer Feed-Forward Layers Are Key-Value Memories , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2021.emnlp-main.446 , pages =

    work page internal anchor Pith review doi:10.18653/v1/2021.emnlp-main.446 2021

  11. [11]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =

    Dissecting Recall of Factual Associations in Auto-Regressive Language Models , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =. doi:10.18653/v1/2023.emnlp-main.751 , pages =

    work page doi:10.18653/v1/2023.emnlp-main.751 2023

  12. [12]

    2021 , journal =

    A Mathematical Framework for Transformer Circuits , author =. 2021 , journal =

    2021

  13. [13]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =

    A Mechanistic Interpretation of Arithmetic Reasoning in Language Models using Causal Mediation Analysis , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =. doi:10.18653/v1/2023.emnlp-main.435 , pages =

    work page doi:10.18653/v1/2023.emnlp-main.435 2023

  14. [14]

    ACM computing surveys , volume =

    Survey of hallucination in natural language generation , author =. ACM computing surveys , volume =. 2023 , publisher =

    2023

  15. [15]

    S elf C heck GPT : Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models

    Manakul, Potsawee and Liusie, Adian and Gales, Mark , editor =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =. doi:10.18653/v1/2023.emnlp-main.557 , pages =

    work page doi:10.18653/v1/2023.emnlp-main.557 2023

  16. [16]

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

    Li, Junyi and Cheng, Xiaoxue and Zhao, Xin and Nie, Jian-Yun and Wen, Ji-Rong , editor =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , month = dec, year =. doi:10.18653/v1/2023.emnlp-main.397 , pages =

    work page doi:10.18653/v1/2023.emnlp-main.397 2023

  17. [17]

    Know what you don’t know: Unanswerable questions for SQuAD

    Rajpurkar, Pranav and Jia, Robin and Liang, Percy , editor =. Know What You Don. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , month = jul, year =. doi:10.18653/v1/P18-2124 , pages =

    work page doi:10.18653/v1/p18-2124

  18. [18]

    T rivia QA : A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension

    Joshi, Mandar and Choi, Eunsol and Weld, Daniel and Zettlemoyer, Luke , editor =. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =. doi:10.18653/v1/P17-1147 , pages =

    work page doi:10.18653/v1/p17-1147

  19. [19]

    H otpot QA : A Dataset for Diverse, Explainable Multi-hop Question Answering

    Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William and Salakhutdinov, Ruslan and Manning, Christopher D. , editor =. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , month = oct #. 2018 , address =. doi:10.18653/v1/D18-1259 , pages =

    work page doi:10.18653/v1/d18-1259 2018

  20. [20]

    ArXiv , year =

    Qwen2.5 Technical Report , author =. ArXiv , year =

  21. [21]

    arXiv preprint arXiv:2407.21783 , year =

    The llama 3 herd of models , author =. arXiv preprint arXiv:2407.21783 , year =

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2408.00118 , year =

    Gemma 2: Improving open language models at a practical size , author =. arXiv preprint arXiv:2408.00118 , year =

    Pith/arXiv arXiv

  23. [23]

    International Conference on Learning Representations , volume =

    Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability , author =. International Conference on Learning Representations , volume =

  24. [24]

    arXiv preprint arXiv:2209.15558 , year =

    Out-of-distribution detection and selective generation for conditional language models , author =. arXiv preprint arXiv:2209.15558 , year =

    arXiv

  25. [25]

    arXiv preprint arXiv:2002.07650 , year =

    Uncertainty estimation in autoregressive structured prediction , author =. arXiv preprint arXiv:2002.07650 , year =

    arXiv 2002

  26. [26]

    arXiv preprint arXiv:2302.09664 , year =

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation , author =. arXiv preprint arXiv:2302.09664 , year =

    Pith/arXiv arXiv

  27. [27]

    Nature , year =

    Detecting hallucinations in large language models using semantic entropy , author =. Nature , year =

  28. [28]

    arXiv preprint arXiv:2207.05221 , year =

    Language models (mostly) know what they know , author =. arXiv preprint arXiv:2207.05221 , year =

    Pith/arXiv arXiv

  29. [29]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =

    Embedding and Gradient Say Wrong: A White-Box Method for Hallucination Detection , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2024.emnlp-main.116 , pages =

    work page doi:10.18653/v1/2024.emnlp-main.116 2024

  30. [30]

    arXiv preprint arXiv:2502.16570 , year =

    Entropy-lens: The information signature of transformer computations , author =. arXiv preprint arXiv:2502.16570 , year =

    arXiv

  31. [31]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    Beyond semantic entropy: Boosting LLM uncertainty quantification with pairwise semantic similarity , author =. Findings of the Association for Computational Linguistics: ACL 2025 , pages =

    2025

  32. [32]

    arXiv preprint arXiv:2602.02888 , year =

    HALT: Hallucination Assessment via Log-probs as Time series , author =. arXiv preprint arXiv:2602.02888 , year =

    arXiv

  33. [33]

    arXiv preprint arXiv:2504.03579 , year =

    Hallucination Detection on a Budget: Efficient Bayesian Estimation of Semantic Entropy , author =. arXiv preprint arXiv:2504.03579 , year =

    arXiv

  34. [34]

    F act S elf C heck: Fact-Level Black-Box Hallucination Detection for LLM s

    Sawczyn, Albert and Binkowski, Jakub and Janiak, Denis and Gabrys, Bogdan and Kajdanowicz, Tomasz Jan , editor =. Findings of the. 2026 , address =. doi:10.18653/v1/2026.findings-eacl.296 , pages =

    work page doi:10.18653/v1/2026.findings-eacl.296 2026