Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

TruthFlow: Truthful LLM Generation via Representation Flow Correction

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

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

Pith's one-line read TruthFlow claims that truthfulness interventions should be query-specific, and uses a flow-matching model to learn a correction vector for each question that moves a language model's representations from hallucinated to truthful states.

desk verdict Geniune new idea and strong TruthfulQA numbers, but the paper never tests the core query-specific claim against a universal vector, and the transferability results are near-zero. read the letter →

arxiv 2502.04556 v1 pith:P27TMMIM submitted 2025-02-06 cs.CL cs.AIcs.LG

Hanyu Wang , Bochuan Cao , Yuanpu Cao , Jinghui Chen This is my paper · ORCID
classification cs.CLcs.AIcs.LG
keywords truthfulnesshallucinationmitigationrepresentationinterventionflowmatchingquery-specificcorrectionvectorTruthfulQAsingularvaluedecompositionopen-endedgeneration
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 the usual one-vector-fits-all approach to representation intervention is the wrong level of abstraction: a question that asks about a fair coin and a question that asks about the happiest place on Earth need different corrections. TruthFlow trains a flow-matching model that takes a query's hidden representation at one middle layer and outputs that query's own correction vector, defined as the contrast between hidden states of a correct answer and an incorrect answer. On TruthfulQA open-ended generation this raises truthfulness scores by about 7 points on average over the base models across six instruction-tuned LLMs, and the authors report the trained flow model carries over to HaluEval, Natural Questions, and TriviaQA without harming informativeness. The practical point is that query-specific steering can be achieved without fine-tuning the LLM, with a small flow model trained in seconds.

What carries the argument

The load-bearing object is the learned flow vector field $v_\phi: [0,1]\times\mathbb{R}^d \to \mathbb{R}^d$, trained by flow matching to predict the displacement $d_q^\ell - h_q^\ell$ along the interpolated path $z_t = t d_q^\ell + (1-t) h_q^\ell$; flow matching is the generative-model technique that learns a vector field pushing one distribution into another along such paths. The query specificity comes from starting the ODE at the query's own last-token hidden state $h_q^\ell$ at a chosen layer (layer 12 for Llama3, for example), so different questions land on different correction vectors. A final SVD projection onto the top-$k$ right singular vectors of all training correction directions removes components the paper treats as noise before the scaled vector $\alpha \hat d_q^\ell$ is added to every token hidden state at that layer.

What would settle it

Take a held-out split of TruthfulQA and compute, for each question, the oracle correction vector $d_q = \bar h_c - \bar h_i$ from the reference answers. If the flow-generated $\hat d_q$ is no closer in cosine similarity to the oracle than the global mean correction vector is, the query-specific mechanism is not doing the work the paper attributes to it. A second check: at inference, replace the flow model with a fixed vector; if TruthfulQA truthfulness scores do not drop, the flow conditioning is superfluous.

Watch

Extended reading notes

Core claim

TruthFlow's central claim is that the distribution of truthfulness-correction directions is query-dependent, so no single "truth vector" exists that moves every prompt from hallucinated to truthful territory. For each training question the paper builds a correction vector $d_q^\ell = \bar h_c^\ell - \bar h_i^\ell$ by subtracting the averaged hidden states of an incorrect answer from those of a correct answer at layer $\ell$, and pairs it with the query's last-token hidden state $h_q^\ell$. A flow-matching model $v_\phi(t,z_t)$ is trained to follow the linear interpolation between $h_q^\ell$ and $d_q^\ell$; at inference, solving the ODE $dz_t = v_\phi(t,z_t)\,dt$ from $z_0 = h_q^\ell$ produces a query-specific correction $\hat d_q^\ell$, which is projected onto the top-$k$ singular vectors of the training correction matrix and added with strength $\alpha$ to every token's hidden state at layer $\ell$. The authors report that this raises TruthfulQA open-ended truthfulness by about seven points on average across six models, improves MC1/MC2 on most of them, and that a model trained on half of TruthfulQA keeps Llama3 at or above base performance on HaluEval, Natural Questions, and TriviaQA.

Load-bearing premise

The whole method rests on the assumption that the final word of a question, read at one middle layer of the model, carries enough information about how to correct that particular question's answer that a flow model trained on 408 TruthfulQA examples generalizes to unseen questions.

Editorial extensions

If this is right

  • A single-layer intervention can raise open-ended truthfulness by about 7 points on average without fine-tuning or external knowledge, so truthfulness editing need not be a fixed-vector operation.
  • Replacing the universal-vector assumption with a learned query-conditioned distribution lets the same intervention pipeline work across Llama2, Llama3, Mistral, and Gemma models.
  • Projecting the generated correction onto a low-rank truthfulness subspace (k around 10-20) is what keeps both truth and informativeness high; without it the gains on Llama3 are much smaller.
  • A flow model trained on 408 TruthfulQA pairs can be deployed on other open-ended QA benchmarks; the paper reports Llama3's True and True*Info scores on HaluEval, NQ, and TriviaQA stay at or above base level, where ITI drops sharply.
  • Because the flow model trains in seconds on a single GPU, the method is inexpensive enough to be retrained for each target model or domain.

Reading between the lines

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

  • A direct test the paper does not run is whether the query conditioning itself matters: replacing the flow output with a single global mean correction vector on held-out queries would separate the contribution of per-query conditioning from the contribution of the SVD projection.
  • The correction vectors are built from reference correct/incorrect answer pairs, so on questions whose truthful answer is absent from the training distribution the method can only fall back on what the base model already knows; a domain-specific variant would need domain answer pairs rather than TruthfulQA pairs.
  • Given how cheap the flow model is to train, a natural extension is to condition the flow not just on the last token's hidden state but on a pooled representation of the whole query, which may matter for multi-sentence or multi-hop questions.
  • The small transfer deltas on HaluEval and NQ leave room for a follow-up study that isolates when query-specific conditioning adds value beyond a benign fixed offset.
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.

Referee Report

4 major / 4 minor

Summary. The paper proposes TruthFlow, a representation intervention method that uses a flow matching model to generate query-specific correction vectors. Correction vectors are derived from paired truthful/hallucinated hidden states on a training half of TruthfulQA, and at inference the flow model maps a query's last-token hidden state to a correction vector that is projected onto a low-rank truthfulness subspace and added to the hidden states of a chosen layer. Experiments on six instruction-tuned LLMs report improved TruthfulQA open-generation and multiple-choice scores over several baselines, and transfer experiments to HaluEval, Natural Questions, and TriviaQA are presented as evidence of generalization.

Significance. If the central claim held, TruthFlow would provide a lightweight, query-conditional alternative to fixed-vector representation intervention, addressing a real limitation of prior methods. The paper is clearly written, the flow-matching framing is novel in this context, and the ablation in Section 5.4 shows that the SVD projection contributes substantially. However, the evaluation does not isolate the query-specific component, the transfer results are numerically weak, and the experimental protocol leaves unresolved questions about hyperparameter selection and variance. These issues currently prevent the paper from substantiating its main claims.

major comments (4)
  1. [Section 5.4, Table 5; Section 3.3] The central claim of the paper is that flow matching produces query-specific correction vectors that outperform a universal correction vector. Yet no experiment applies a single fixed correction vector (for example, the mean of the training directions or the projection of that mean) with the same layer, same alpha, and same SVD projection as TruthFlow. The baselines in Table 1 (ITI, TruthX, DoLa, etc.) use different training/selection protocols and different intervention locations, so they cannot control for the effect of query-specificity. Without such a universal-vector baseline, the reported gains could be entirely due to the SVD projection or to a well-tuned fixed vector, and the paper's central contribution remains undemonstrated.
  2. [Table 3, Section 5.1] The transferability results are described as demonstrating 'remarkable generalizability' and 'significantly enhances both True and True*Info scores,' but the numbers in Table 3 show almost no improvement on HaluEval (True 36.74 to 36.82; True*Info 33.86 to 33.87) and only modest gains on NQ and TriviaQA. These differences are small and could easily be within evaluation noise, especially given that no variance estimates are reported. The claim of strong transferability is not supported by the presented evidence and should be substantially weakened or backed by a more sensitive analysis.
  3. [Section 4.2, Table 7, Appendix B] The hyperparameter selection procedure is not described. The paper states that TruthfulQA is split into 408 training and 409 test examples, but Table 7 lists per-model choices of layer, alpha, and k without explaining how they were chosen. If these values were selected based on the test half of TruthfulQA, the reported results would be optimistically biased. In addition, the transferability experiments in Appendix B state that hyperparameters were 'slightly tuned' for each target benchmark, which further weakens the generality claim. A clear validation split or a prespecified selection rule is needed.
  4. [Table 1, Section 4.2] No error bars, confidence intervals, or multiple-seed results are reported for the main experiments. The flow model is trained on only 408 examples with random initialization, and the evaluation uses GPT-4 judgments, which may be stochastic. The differences between TruthFlow and the second-best baseline are often a few percentage points; without variance estimates, it is not possible to determine whether these differences are statistically meaningful.
minor comments (4)
  1. [Table 1] The column header 'BLERUT' appears to be a typo for 'BLEURT'.
  2. [Section 5.1] 'Natrual Questions' is a typo for 'Natural Questions'.
  3. [Section 3.2, Figure 2] The motivation for query-specific correction is based on a qualitative PCA/KDE visualization for a single model and layer. Adding a quantitative comparison (for example, the average cosine similarity between per-query directions and the universal direction, or a small experiment with a fixed vector) would make the argument more convincing.
  4. [Appendix B] The transferability experiments use different hyperparameter triplets (epochs, layer, alpha) for each target dataset; the paper should state explicitly whether these were chosen on a held-out subset of each target benchmark or on the full target benchmark, since this affects the interpretation of the transfer results.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported TruthfulQA and transferability gains are partly constructed by test-set hyperparameter selection, fitting the fitted-input-called-prediction pattern.

  1. fitted input called prediction [Appendix B (hyperparameters), applied to Section 4.2 / Table 1]
    "For each LLM, we apply the following hyperparameters (see Table 7) to achieve the results reported in Table 1."

    The 409-question TruthfulQA test set is the only held-out split described in Section 4.2 ('we divide the whole TruthfulQA dataset into half: 408 data as the training set and 409 remaining data as the test set'). No validation split is provided. The per-model intervention layer, strength alpha, and SVD rank k in Table 7 are then explicitly chosen 'to achieve the results reported in Table 1', meaning these free parameters are fit to the same test set whose scores are presented as TruthFlow's prediction. The reported TruthfulQA improvement is therefore not an independent predictive result; it is partly the output of a test-set hyperparameter search, i.e., a fitted input is later reported as a prediction on the same data.

  2. fitted input called prediction [Appendix B (transferability settings), applied to Section 5.1 / Table 3]
    "On transferability tasks, we slightly tune the hyperparameters considering the training set changes from half of TruthfulQA to the whole dataset. We fix all the training hyperparameters except the number of epochs and also fix k to 20. The triplets of (epochs, layer, alpha) are (30, 13, 0.5), (30, 13, 1.5), and (35, 13, 0.5) for HaluEval, Natural Questions, and TriviaQA, respectively."

    The paper's transferability claim is that TruthFlow generalizes to 'unseen hallucination benchmarks', but the numbers in Table 3 are produced after separately tuning epochs, layer, and alpha for each target dataset. Thus the transfer scores are fitted to the very benchmarks used to demonstrate transferability, rather than being generated by one fixed model. The conclusion in Section 5.1 that TruthFlow 'maintains the LLM's performance' on unseen domains is therefore not a prediction of a fixed method; it is a per-dataset fitted result, which is a fitted input called a generalization result.

full rationale

The core flow-matching derivation is not circular in itself: the model is trained on 408 TruthfulQA pairs to map a query hidden state h_l_q to a correction vector d_l_q = mean(correct answer hidden states) - mean(incorrect answer hidden states), and then applied to new queries. No equation in the paper equates the test-time correction to the test labels by construction, and no load-bearing self-citation or imported uniqueness theorem is used. The circularity I can exhibit is in the evaluation protocol: for the main TruthfulQA claims, the per-model hyperparameters (layer, alpha, k) are chosen on the same 409-question test set and then used to produce the reported Table 1 scores; for the transferability claims, epochs, layer, and alpha are explicitly tuned for each unseen benchmark before reporting Table 3. Both are instances of fitting parameters to the evaluation data and then presenting the result as a prediction or as generalization, which is the fitted-input-called-prediction pattern. The missing universal-vector baseline is a real attribution confound but is not, by itself, circularity under the given rubric; I therefore do not count it as a separate circular step. Because the central empirical claims partially reduce to test-set and per-dataset hyperparameter selection, a score of 6 is appropriate rather than 0-2.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method relies on several unverified geometric assumptions about LLM representations: that mean correct-minus-incorrect hidden states are truthful directions, that the top singular vectors of 408 training directions span a truthfulness subspace, and that a query's single-layer hidden state predicts its correction. The many per-model tuned hyperparameters (alpha, layer, k) and per-dataset retuning in transfer experiments raise the risk that reported gains are partly test-set artifacts.

free parameters (4)
  • intervention strength alpha = 1.5 to 4.3 per model (Table 7)
    Chosen per LLM to maximize truthfulness; also retuned per transfer dataset (Appendix D).
  • intervention layer l = 12, 13, or 20 depending on model (Table 7)
    Selected per model; no validation procedure described.
  • number of singular vectors k = 10, 12, or 20 per model (Table 7)
    Selected per model to balance truthfulness and informativeness.
  • training epochs = 25-45 depending on model (Table 7/A.2)
    Varied per model; not shown to be critical but is a tuned hyperparameter.
assumptions (5)
  • domain assumption The mean difference between hidden states of correct and incorrect answers (d = h_c - h_i) is a valid truthful correction direction.
    Assumed in Section 3.3 when constructing training targets; no independent evidence beyond the observed improvement.
  • domain assumption The top-k singular vectors of the training correction directions span a truthfulness-related subspace, and projection onto it removes noise.
    Stated in Section 3.4; the choice of k is empirical and the subspace is computed from the same 408 training directions.
  • domain assumption The last-token hidden state of a query at a single layer is sufficient to predict its correction vector.
    Underpins the flow model input; never validated directly, only through downstream truthfulness.
  • domain assumption Adding a single correction vector to all token positions at layer l is a valid intervention.
    Applied in Section 3.4; no per-token analysis is provided.
  • standard math Flow matching ODE integration with the Midpoint method accurately approximates the learned vector field.
    Standard numerical analysis; the flow model is trained with the same interpolation objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TruthFlow: Truthful LLM Generation via Representation Flow Correction." pith.science (2026). https://pith.science/paper/P27TMMIM

@misc{pith2026250204556,
  author       = {Pith},
  title        = {Pith review of: TruthFlow: Truthful LLM Generation via Representation Flow Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P27TMMIM}},
  note         = {Machine review of arXiv:2502.04556}
}
read the original abstract

Large language models (LLMs) are known to struggle with consistently generating truthful responses. While various representation intervention techniques have been proposed, these methods typically apply a universal representation correction vector to all input queries, limiting their effectiveness against diverse queries in practice. In this study, we introduce TruthFlow, a novel method that leverages the Flow Matching technique for query-specific truthful representation correction. Specifically, TruthFlow first uses a flow model to learn query-specific correction vectors that transition representations from hallucinated to truthful states. Then, during inference, the trained flow model generates these correction vectors to enhance the truthfulness of LLM outputs. Experimental results demonstrate that TruthFlow significantly improves performance on open-ended generation tasks across various advanced LLMs evaluated on TruthfulQA. Moreover, the trained TruthFlow model exhibits strong transferability, performing effectively on other unseen hallucination benchmarks.

Figures

Figures reproduced from arXiv: 2502.04556 by the authors.

Figure 1
Figure 1. Comparison of the generated answers from Llama [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of hallucinated hidden states and truthful hidden states of Llama-2-7b-chat at the 13-th trans￾former layer using PCA and KDE. The bold blue arrow in the middle shows the general direction from hallucination to truthfulness. However, each sample has its own direction towards truthfulness as is shown by a light blue arrow. 3.2. Motivation: Universal Correction Vector? Current representation intervention… view at source ↗
Figure 3
Figure 3. Performance comparison on different choices of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Prompt template for GPT-4 to evaluate True Score on TruthfulQA. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Prompt template for GPT-4 to evaluate Info Score on TruthfulQA. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Prompt template for GPT-4 to evaluate True Score on HaluEval given 3 answers from base model, TruthFlow, and [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Prompt template for GPT-4 to evaluate True Score on Natural Questions. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Prompt template for GPT-4 to evaluate True Score on TriviaQA. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Prompt template for GPT-4 to evaluate Info Score on HaluEval given 3 answers from base model, TruthFlow, and [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Focus on What Matters: Enhancing Medical Vision-Language Models with Automatic Attention Alignment Tuning

    cs.CV 2025-05 conditional novelty 6.0 of 10

    A3Tune aligns the visual attention of medical LVLMs to prompt-relevant regions via SAM and BioMedCLIP weak labels plus a Mixture-of-Experts over LoRA, improving VQA and report generation accuracy.

Reference graph

Works this paper leans on

59 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Intrinsic dimensionality explains the effectiveness of language model fine-tuning

    Aghajanyan, A., Zettlemoyer, L., and Gupta, S. Intrinsic dimensionality explains the effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2020

    arXiv 2012

  3. [3]

    and Mitchell, T

    Azaria, A. and Mitchell, T. The internal state of an llm knows when it's lying. arXiv preprint arXiv:2304.13734, 2023

    arXiv 2023

  4. [4]

    Qwen technical report

    Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023

    arXiv 2023

  5. [5]

    F., Liu, X., Jagadish, H., and Wang, L

    Bayat, F. F., Liu, X., Jagadish, H., and Wang, L. Enhanced language model truthfulness with learnable intervention and uncertainty expression. In Findings of the Association for Computational Linguistics ACL 2024, pp.\ 12388--12400, 2024

    work page 2024

  6. [6]

    Burden, R. L. and Faires, J. D. Numerical analysis, 2010

    work page 2010

  7. [7]

    Self-control of llm behaviors by compressing suffix gradient into prefix controller

    Cai, M., Zhang, Y., Zhang, S., Yin, F., Zhang, D., Zou, D., Yue, Y., and Hu, Z. Self-control of llm behaviors by compressing suffix gradient into prefix controller. arXiv preprint arXiv:2406.02721, 2024

    arXiv 2024

  8. [8]

    Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization

    Cao, Y., Zhang, T., Cao, B., Yin, Z., Lin, L., Ma, F., and Chen, J. Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization. arXiv preprint arXiv:2406.00045, 2024

    arXiv 2024

Show all 59 references
  1. [9]

    K., Scheurer, J., Rando, J., Freedman, R., Korbak, T., Lindner, D., Freire, P., et al

    Casper, S., Davies, X., Shi, C., Gilbert, T. K., Scheurer, J., Rando, J., Freedman, R., Korbak, T., Lindner, D., Freire, P., et al. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217, 2023

    2023 arXiv

  2. [10]

    Inside: Llms' internal states retain the power of hallucination detection

    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 preprint arXiv:2402.03744, 2024 a

    2024 arXiv

  3. [11]

    Lower layer matters: Alleviating hallucination via multi-layer fusion contrastive decoding with truthfulness refocused

    Chen, D., Fang, F., Ni, S., Liang, F., Xu, R., Yang, M., and Li, C. Lower layer matters: Alleviating hallucination via multi-layer fusion contrastive decoding with truthfulness refocused. arXiv preprint arXiv:2408.08769, 2024 b

    2024 arXiv

  4. [12]

    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. arXiv preprint arXiv:2403.01548, 2024 c

    2024 arXiv

  5. [13]

    Grath: Gradual self-truthifying for large language models

    Chen, W., Song, D., and Li, B. Grath: Gradual self-truthifying for large language models. arXiv preprint arXiv:2401.12292, 2024 d

    2024 arXiv

  6. [14]

    Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning

    Chen, Z., Sun, X., Jiao, X., Lian, F., Kang, Z., Wang, D., and Xu, C. Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 20967--20974, 2024 e

    2024

  7. [15]

    Dola: Decoding by contrasting layers improves factuality in large language models

    Chuang, Y.-S., Xie, Y., Luo, H., Kim, Y., Glass, J., and He, P. Dola: Decoding by contrasting layers improves factuality in large language models. arXiv preprint arXiv:2309.03883, 2023

    2023 arXiv

  8. [16]

    Haloscope: Harnessing unlabeled llm generations for hallucination detection

    Du, X., Xiao, C., and Li, Y. Haloscope: Harnessing unlabeled llm generations for hallucination detection. arXiv preprint arXiv:2409.17504, 2024

    2024 arXiv

  9. [17]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  10. [18]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024

    2024

  11. [19]

    Institutionum calculi integralis, volume 4

    Euler, L. Institutionum calculi integralis, volume 4. impensis Academiae imperialis scientiarum, 1845

  12. [20]

    Nl-iti: Optimizing probing and intervention for improvement of iti method

    Hoscilowicz, J., Wiacek, A., Chojnacki, J., Cieslak, A., Michon, L., Urbanevych, V., and Janicki, A. Nl-iti: Optimizing probing and intervention for improvement of iti method. arXiv preprint arXiv:2403.18680, 2024

    2024 arXiv

  13. [21]

    Mitigating large language model hallucination with faithful finetuning

    Hu, M., He, B., Wang, Y., Li, L., Ma, C., and King, I. Mitigating large language model hallucination with faithful finetuning. arXiv preprint arXiv:2406.11267, 2024

    2024 arXiv

  14. [22]

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

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 2023

    2023

  15. [23]

    J., Madotto, A., and Fung, P

    Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., and Fung, P. Survey of hallucination in natural language generation. ACM Computing Surveys, 55 0 (12): 0 1--38, 2023

    2023

  16. [24]

    Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D

    Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. d. l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023

    2023 arXiv

  17. [25]

    Exploring concept depth: How large language models acquire knowledge at different layers? arXiv preprint arXiv:2404.07066, 2024

    Jin, M., Yu, Q., Huang, J., Zeng, Q., Wang, Z., Hua, W., Zhao, H., Mei, K., Meng, Y., Ding, K., et al. Exploring concept depth: How large language models acquire knowledge at different layers? arXiv preprint arXiv:2404.07066, 2024

    2024 arXiv

  18. [26]

    S., and Zettlemoyer, L

    Joshi, M., Choi, E., Weld, D. S., and Zettlemoyer, L. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551, 2017

    2017 arXiv

  19. [27]

    Sh2: Self-highlighted hesitation helps you decode more truthfully

    Kai, J., Zhang, T., Hu, H., and Lin, Z. Sh2: Self-highlighted hesitation helps you decode more truthfully. arXiv preprint arXiv:2401.05930, 2024

    2024 arXiv

  20. [28]

    Beitrag zur n \"a herungsweisen Integration totaler Differentialgleichungen

    Kutta, W. Beitrag zur n \"a herungsweisen Integration totaler Differentialgleichungen . Teubner, 1901

    1901

  21. [29]

    Natural questions: a benchmark for question answering research

    Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M., Parikh, A., Alberti, C., Epstein, D., Polosukhin, I., Devlin, J., Lee, K., et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7: 0 453--...

    2019

  22. [30]

    u ttler, H., Lewis, M., Yih, W.-t., Rockt \

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K \"u ttler, H., Lewis, M., Yih, W.-t., Rockt \"a schel, T., et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33: 0 9459--9474, 2020

    2020

  23. [31]

    X., Nie, J.-Y., and Wen, J.-R

    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 preprint arXiv:2305.11747, 2023

    2023 arXiv

  24. [32]

    Inference-time intervention: Eliciting truthful answers from a language model

    Li, K., Patel, O., Vi \'e gas, F., Pfister, H., and Wattenberg, M. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024

    2024

  25. [33]

    L., Holtzman, A., Fried, D., Liang, P., Eisner, J., Hashimoto, T., Zettlemoyer, L., and Lewis, M

    Li, X. L., Holtzman, A., Fried, D., Liang, P., Eisner, J., Hashimoto, T., Zettlemoyer, L., and Lewis, M. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097, 2022

    2022 arXiv

  26. [34]

    Truthfulqa: Measuring how models mimic human falsehoods

    Lin, S., Hilton, J., and Evans, O. Truthfulqa: Measuring how models mimic human falsehoods. arXiv preprint arXiv:2109.07958, 2021

    2021 arXiv

  27. [35]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  28. [36]

    Deepseek-v3 technical report

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024 a

    2024 arXiv

  29. [37]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  30. [38]

    Fantastic semantics and where to find them: Investigating which layers of generative llms reflect lexical semantics

    Liu, Z., Kong, C., Liu, Y., and Sun, M. Fantastic semantics and where to find them: Investigating which layers of generative llms reflect lexical semantics. arXiv preprint arXiv:2403.01509, 2024 b

    2024 arXiv

  31. [39]

    Probing llms for logical reasoning

    Manigrasso, F., Schouten, S., Morra, L., and Bloem, P. Probing llms for logical reasoning. In International Conference on Neural-Symbolic Learning and Reasoning, pp.\ 257--278. Springer, 2024

    2024

  32. [40]

    and Lewis, M

    O'Brien, S. and Lewis, M. Contrastive decoding improves reasoning in large language models. arXiv preprint arXiv:2309.09117, 2023

    2023 arXiv

  33. [41]

    Training language models to follow instructions with human feedback

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

    2022

  34. [42]

    K., and Sankarasubbu, M

    Pal, A., Umapathi, L. K., and Sankarasubbu, M. Med-halt: Medical domain hallucination test for large language models. arXiv preprint arXiv:2307.15343, 2023

    2023 arXiv

  35. [43]

    Panickssery, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., and Turner, A. M. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023

    2023 arXiv

  36. [44]

    D., Ermon, S., and Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024

    2024

  37. [45]

    A survey of hallucination in large foundation models

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

    2023 arXiv

  38. [46]

    Ren, J., Luo, J., Zhao, Y., Krishna, K., Saleh, M., Lakshminarayanan, B., and Liu, P. J. Out-of-distribution detection and selective generation for conditional language models. In The Eleventh International Conference on Learning Representations, 2022

    2022

  39. [47]

    U-net: Convolutional networks for biomedical image segmentation

    Ronneberger, O., Fischer, P., and Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, ...

    2015

  40. [48]

    U ber die numerische aufl \

    Runge, C. \"U ber die numerische aufl \"o sung von differentialgleichungen. Mathematische Annalen, 46 0 (2): 0 167--178, 1895

  41. [49]

    Sellam, T., Das, D., and Parikh, A. P. Bleurt: Learning robust metrics for text generation. arXiv preprint arXiv:2004.04696, 2020

    2004 arXiv

  42. [50]

    S., Love, J., et al

    Team, G., Mesnard, T., Hardin, C., Dadashi, R., Bhupatiraju, S., Pathak, S., Sifre, L., Rivi \`e re, M., Kale, M. S., Love, J., et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024

    2024 arXiv

  43. [51]

    D., and Finn, C

    Tian, K., Mitchell, E., Yao, H., Manning, C. D., and Finn, C. Fine-tuning language models for factuality. arXiv preprint arXiv:2311.08401, 2023

    2023 arXiv

  44. [52]

    Llama 2: Open foundation and fine-tuned chat models

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  45. [53]

    Attention is all you need

    Vaswani, A. Attention is all you need. Advances in Neural Information Processing Systems, 2017

    2017

  46. [54]

    W., Salakhutdinov, R., and Manning, C

    Yang, Z., Qi, P., Zhang, S., Bengio, Y., Cohen, W. W., Salakhutdinov, R., and Manning, C. D. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600, 2018

    2018 arXiv

  47. [55]

    Truthx: Alleviating hallucinations by editing large language models in truthful space

    Zhang, S., Yu, T., and Feng, Y. Truthx: Alleviating hallucinations by editing large language models in truthful space. arXiv preprint arXiv:2402.17811, 2024

    2024 arXiv

  48. [56]

    Alleviating hallucinations of large language models through induced hallucinations

    Zhang, Y., Cui, L., Bi, W., and Shi, S. Alleviating hallucinations of large language models through induced hallucinations. arXiv preprint arXiv:2312.15710, 2023

    2023 arXiv

  49. [57]

    Representation engineering: A top-down approach to ai transparency

    Zou, A., Phan, L., Chen, S., Campbell, J., Guo, P., Ren, R., Pan, A., Yin, X., Mazeika, M., Dombrowski, A.-K., et al. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405, 2023

    2023 arXiv

  50. [58]

    Z., Fredrikson, M., and Hendrycks, D

    Zou, A., Phan, L., Wang, J., Duenas, D., Lin, M., Andriushchenko, M., Kolter, J. Z., Fredrikson, M., and Hendrycks, D. Improving alignment and robustness with circuit breakers. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  51. [59]

    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 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.