Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Next Token Perception Score: Analytical Assessment of your LLM Perception Skills

T0 review · 3 major / 5 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 A closed-form subspace-overlap score predicts when a frozen language model will benefit from fine-tuning.

desk verdict A genuinely new closed-form subspace-overlap score with a clean linear bound, but the headline correlations are maxima over a per-model configuration search and need re-analysis before the empirical claims hold. read the letter →

arxiv 2505.17169 v1 pith:Z55MCEOD submitted 2025-05-22 cs.CL cs.AI

Yu-Ang Cheng , Leyang Hu , Hai Huang , Randall Balestriero This is my paper
classification cs.CLcs.AI
keywords nexttokenpredictionlinearprobingsubspacealignmentgeneralizedeigenvalueproblemLoRAfine-tuningrepresentationanalysisperceptiontasksmodeltransfer
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

This paper introduces the Next Token Perception Score (NTPS), a single number computed from a frozen model's hidden states and labeled data that measures how much of the information a perception task needs is already present after autoregressive pretraining. The score is defined as the squared projection of the perception subspace onto the autoregressive subspace, and the authors prove that in a linear setting it bounds the excess loss a next-token-prediction encoder incurs on a regression task. Empirically, NTPS correlates with linear-probe accuracy across twelve NLP datasets and eight models, and it increases after LoRA fine-tuning. The practical payoff is a lightweight prescreening tool: low NTPS on a task signals that LoRA adaptation has room to improve accuracy.

What carries the argument

The machinery is a pair of generalized eigenvalue problems: the autoregressive encoder $V$ solves one built from cumulative-context token covariances, while the perception encoder $U$ solves another built from the covariance between summed token representations and labels. NTPS is the normalized squared Frobenius norm of the projection of $U$ onto $V$'s column space, so it equals 1 when the subspaces coincide and approaches 0 when they are orthogonal. Theorem 2 derives the two-sided bound on excess loss by whitening the encoders with the token covariance $N$ and using extremal eigenvalues of $N$.

What would settle it

On a held-out model and task, compute NTPS from the training split and then measure the actual LoRA accuracy gain on the test split; if the Spearman correlation between NTPS and gain is not substantially negative, the paper's central predictive claim fails.

Watch

Extended reading notes

Core claim

The central claim is that NTPS, defined as $\|P U\|_F^2 / \|U\|_F^2$ where $P$ is the orthogonal projector onto the column space of the autoregressive encoder $V$, captures the alignment between what next-token prediction learns and what a perception task needs. In a linear setting, the optimal encoders $U$ and $V$ are obtained from two different generalized eigenvalue problems, and Theorem 2 states that the excess regression loss of $V$ relative to $U$ satisfies $C_{\min}(1-\mathrm{NTPS}) \le \Delta L \le C_{\max}(1-\mathrm{NTPS})$. The paper argues that this linear-regime identity transfers to real LLMs: across eight pretrained models from 270M to 8B parameters and 12 datasets, NTPS correlates with linear probe accuracy, and models with lower NTPS show larger accuracy gains after LoRA fine-tuning.

Load-bearing premise

The load-bearing premise is that real transformer hidden states can be treated as the linear encoder of the theory, so that U and V computed from finite samples are the same objects whose subspace overlap bounds the excess loss.

Editorial extensions

If this is right

  • NTPS can be computed from a frozen model and a small labeled sample without any fine-tuning, giving a cheap pre-screening test for whether LoRA is worth applying.
  • Tasks where a model scores low on NTPS are precisely the tasks where LoRA yields the largest accuracy gains, so the metric identifies where parameter-efficient adaptation has the most headroom.
  • The divergence between $U$ and $V$ spaces explains the variability in linear probing: next-token prediction emphasizes syntactic structure, which can be orthogonal to the label-relevant semantics of downstream perception tasks.
  • Because NTPS increases after LoRA in 71 of 96 runs and most strongly in large models, the metric offers a mechanistic account of why LoRA works: it rotates the predictive subspace toward the perception subspace.

Reading between the lines

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

  • A natural extension, not explored in the paper, is to use NTPS as a layer-selection criterion: the authors sweep all layers and $k$ values, but a principled rule for choosing the configuration could make the score fully automatic.
  • The same subspace-overlap logic could be applied to other pretraining objectives, such as masked language modeling or contrastive learning, to see whether their encoders also leave perception-relevant directions unoccupied.
  • The strong correlation on training splits suggests NTPS might also serve as a diagnostic for distribution shift: comparing NTPS computed on train versus test representations could reveal when a task's perceptual structure changes between splits.
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

3 major / 5 minor

Summary. The paper introduces the Next Token Perception Score (NTPS), a closed-form subspace overlap metric between the optimal linear encoder for a downstream perception task (U) and the optimal linear encoder for autoregressive next-token prediction (V). In a linear-regime model, the authors prove (Theorem 2, eq. 10) that the excess regression loss of V relative to U is bounded above and below by affine functions of 1 - NTPS, with a derivation developed in Section 3.2 and Appendix A. The paper then computes NTPS from pretrained LLM hidden states for 12 NLP datasets and 8 models, reports Spearman correlations between NTPS and linear-probe accuracy/MSE (Figure 2), shows that NTPS increases after LoRA fine-tuning in 71 of 96 runs (Table 2), and reports inverse correlations between NTPS and LoRA accuracy gains (Figure 3). The paper concludes that NTPS is a lightweight prescreening tool for deciding when LoRA fine-tuning will be beneficial.

Significance. If the empirical correlations survive proper validation, NTPS would be a valuable and inexpensive diagnostic: it is computed in closed form from pretrained representations and labeled data, it comes with a coherent linear-theory derivation that is not circular, and it makes a falsifiable prediction about LoRA gains (low NTPS predicts larger gains). The paper also provides code and a detailed appendix with the proofs. The main weakness is the maximum-over-configuration reporting in the core empirical sections, which currently prevents the correlational claims from being accepted at face value.

major comments (3)
  1. [Section 4.1, Figure 2] The reported Spearman correlations are maxima over a per-model search over all layers and all k proportions from 0.05 to 0.95 in increments of 0.05, with the text stating 'we report for each model the alignment score corresponding to the configuration that yields the strongest Spearman's r.' Because the same 12 datasets are used to compute NTPS, to evaluate linear-probe accuracy, and to select the configuration, the reported r is an in-sample maximum over roughly (number of layers) x 19 configurations and is not an unbiased estimate of the metric's predictive validity. The authors' own limitation note in Section 5 acknowledges the lack of a principled configuration strategy, but it does not address the selection-inflation problem. The revision should either prespecify a configuration, use nested or held-out splits, or report the distribution of correlations across configurations with appropriate multiple-comparison control.
  2. [Section 4.3, Figure 3] The same maximization procedure is applied when predicting LoRA accuracy gains, and here the selection is performed on the outcome being predicted: the text says 'Again we (1) compute NTPS over all layers and ks exactly as before' and defines ACC gain 'as the difference between accuracy after LoRA adaptation and the baseline linear-probe accuracy (both on the test split).' Thus the Spearman r values between 0.40 and 0.90 reflect the best of many configurations chosen to maximize the correlation with the test gain on the same 12 datasets. This does not support the claim that NTPS 'reliably predicts' LoRA gains; an out-of-sample protocol is needed.
  3. [Section 3.2, Appendix A.3, Section 4] The theoretical bound in eq. (41) is derived for linear encoders from population generalized eigenproblems, but the empirical NTPS in Algorithm 1 is computed from finite samples of nonlinear transformer hidden states. The paper repeatedly acknowledges this gap (Section 5), but the empirical claims in Section 4 would be considerably strengthened by a control analysis, such as a permutation test on the dataset labels or a comparison with a random-subspace overlap, to show that the observed correlations are specific to the perception subspace U and not an artifact of the subspace estimation or configuration search. Without such a control, the correlation results remain purely phenomenological even if the linear theory is correct.
minor comments (5)
  1. [Table 2, Section 4.2] The Table 2 caption states that 'NTPS is universally increased after LoRA finetuning,' but the text says it 'increases in 71 out of 96 runs' and the table itself contains many negative entries for OpenELM-270M and OpenELM-450M; the caption should be corrected to match the data.
  2. [Appendix A.2, between eq. (14) and (15)] The derivation of W contains an incorrect intermediate step: it writes W = E[(V^T X1 X1^T V)^{-1}] E[V^T X1 X2^T], which does not follow from the stationarity condition. The correct route is W = (V^T E[X1 X1^T] V)^{-1} V^T E[X1 X2^T]; the final formula is right, but the proof as written is invalid.
  3. [Section 3.1] The text refers to 'As shown in table 12' where it should refer to Table 1; this appears to be a typo.
  4. [Figures 2 and 3, Table 2] The spelling of the model name is inconsistent: 'LlaMA-3-8B' in Table 2 and Figure 3 versus 'Meta-Llama-3-8B' in Figure 2; unify the nomenclature.
  5. [Section 4.1] No confidence intervals or p-values are reported for the Spearman correlations; with only 12 datasets per model, a statement of uncertainty (e.g., bootstrap or permutation intervals) would substantially aid interpretation.

Circularity Check

2 steps flagged · score 6.0 of 10

Empirical 'predictions' are maxima over an in-sample layer/k search; the theoretical derivation itself is not circular.

  1. fitted input called prediction [Section 4.1, configuration-search paragraph and Figure 2]
    "For each model across all datasets, we compute alignment scores over all layers (from the word-embedding layer through the penultimate layer) and every k proportion value from 0.05 to 0.95 in increments of 0.05. ... To summarize each model succinctly, we report for each model the alignment score corresponding to the configuration that yields the strongest Spearman's r."

    The reported Spearman correlations (r = 0.42-0.82 in Figure 2) are not evaluations of a fixed NTPS configuration: they are maxima over a per-model grid of layers and 19 k-values, selected using the same training-set accuracy/MSE values they are then correlated with. The configuration is therefore fitted to the target quantity, and the reported 'NTPS correlates with linear probe performance' is an in-sample maximum statistic rather than an unbiased predictive claim. Both NTPS's perception subspace U and the linear probe are also fit to the same labeled training set, compounding the selection-driven inflation.

  2. fitted input called prediction [Section 4.3, 'Predicting LoRA finetuning gain' and Figure 3]
    "Again we (1) compute NTPS over all layers and ks exactly as before, (2) train a linear probe under CE to get baseline accuracy (AdamW optimizer, learning rate of 10−4; 50 epochs), and (3) apply LoRA (rank 32, α=32, 5000 steps, Adafactor, 5% warm-up) and record the adapted accuracy. Finally, we correlate NTPS with the observed LoRA gains using Spearman's r."

    The LoRA-gain 'prediction' inherits the same per-model search over layers and k values used in Section 4.1, where the configuration was explicitly chosen to maximize the correlation with the target. Because the same data determine both the selected configuration and the reported Spearman r (0.40-0.90), the claim that NTPS 'reliably predicts' accuracy gains is partly an artifact of in-sample selection: the metric's configuration is fitted to the very gains it is supposed to forecast, rather than fixed a priori or chosen on a separate validation set.

full rationale

The theoretical chain in Sections 3.2-3.3 and Appendix A is a genuine derivation: Theorem 1 solves the two MSE problems by generalized eigenproblems, and Theorem 2 algebraically bounds the excess loss ΔL by (1 - NTPS) with constants derived from the task covariances. Those results are not circular; they do not require the empirical correlation they are used to explain. The citation to Balestriero and LeCun [19] is background material and is not load-bearing for the NTPS derivation. The circularity is confined to the empirical validation strategy. In Section 4.1, the paper computes NTPS for every layer and 19 values of k, then reports, per model, the configuration yielding the strongest Spearman correlation with in-sample linear-probe accuracy; the reported r is therefore a maximum statistic fitted to the target it is supposed to validate. Section 4.3 repeats the same search when claiming NTPS 'predicts' LoRA accuracy gains, so those correlations are also selected in-sample. This does not invalidate the mathematical definition of NTPS, but it means the paper's headline empirical claims are partially forced by the selection procedure: the same data were used to choose layer and k and to compute the reported correlation. The 'universally increased' caption of Table 2 conflicts with the '71 out of 96' statement and the many negative entries in that table, but that is an internal inconsistency rather than a circular step.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entity; NTPS is a function of existing hidden states and labels. The main burden is the linear modeling of both next-token and perception tasks, and the finite-sample substitution of population covariances. The empirical validation adds post-hoc selection of layer and subspace dimension that acts as a fitted configuration.

free parameters (2)
  • subspace dimension ratio k = searched from 0.05 to 0.95 in 0.05 increments
    The score is a function of k; the paper reports the configuration with strongest Spearman r, so the plotted correlations are maxima over k.
  • layer index l = chosen per model as the layer giving strongest Spearman r
    NTPS is computed over all layers and only the best layer is reported, a post-hoc choice that inflates the correlation.
assumptions (5)
  • domain assumption Autoregressive prediction is well approximated by a linear map from cumulative token sums to the next token (eq. 2).
    The autoregressive encoder V and its loss are defined under this linear model, while real LLMs are nonlinear and use attention.
  • domain assumption Downstream perception tasks are well approximated by linear regression on the sum of token representations (eq. 3).
    The perception encoder U and loss L* are defined for a linear map on the summed token representation, which ignores the nonlinear sentence composition in actual LLMs.
  • domain assumption Vocabulary embeddings satisfy a positive margin Delta > 0 (Lemma 1).
    Lemma 1 requires a positive margin between correct and competing token embeddings to convert vanishing MSE into correct decoding.
  • domain assumption Population expectations can be replaced by finite-sample empirical averages in Algorithm 1.
    The theory uses population covariances, but NTPS is computed from one training set without confidence intervals or bias correction.
  • domain assumption The covariance matrix N is nonsingular enough for nontrivial constants in Theorem 2.
    The bounds in eq. (41) use lambda_min(N) and lambda_max(N); if N is singular, the lower bound degenerates to zero.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Next Token Perception Score: Analytical Assessment of your LLM Perception Skills." pith.science (2026). https://pith.science/paper/Z55MCEOD

@misc{pith2026250517169,
  author       = {Pith},
  title        = {Pith review of: Next Token Perception Score: Analytical Assessment of your LLM Perception Skills},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z55MCEOD}},
  note         = {Machine review of arXiv:2505.17169}
}
read the original abstract

Autoregressive pretraining has become the de facto paradigm for learning general-purpose representations in large language models (LLMs). However, linear probe performance across downstream perception tasks shows substantial variability, suggesting that features optimized for next-token prediction do not consistently transfer well to downstream perception tasks. We demonstrate that representations learned via autoregression capture features that may lie outside the subspaces most informative for perception. To quantify the (mis)alignment between autoregressive pretraining and downstream perception, we introduce the Next Token Perception Score (NTPS)-a score derived under a linear setting that measures the overlap between autoregressive and perception feature subspaces. This metric can be easily computed in closed form from pretrained representations and labeled data, and is proven to both upper- and lower-bound the excess loss. Empirically, we show that NTPS correlates strongly with linear probe accuracy across 12 diverse NLP datasets and eight pretrained models ranging from 270M to 8B parameters, confirming its utility as a measure of alignment. Furthermore, we show that NTPS increases following low-rank adaptation (LoRA) fine-tuning, especially in large models, suggesting that LoRA aligning representations to perception tasks enhances subspace overlap and thus improves downstream performance. More importantly, we find that NTPS reliably predicts the additional accuracy gains attained by LoRA finetuning thereby providing a lightweight prescreening tool for LoRA adaptation. Our results offer both theoretical insights and practical tools for analytically assessing LLM perception skills.

Figures

Figures reproduced from arXiv: 2505.17169 by the authors.

Figure 2
Figure 2. Correlation between NTPS and downstream MSE loss (rows 1 and 2), and between NTPS [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Correlation between NTPS and accuracy gain (LoRA finetuning accuracy-linear probing [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 13 canonical work pages

  1. [1]

    CARER: Contextualized affect representations for emotion recognition

    Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. CARER: Contextualized affect representations for emotion recognition. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3687–3697, Brussels, Belgium, October-November 2018. Association for Computational Linguistics. doi: 10.18653/ v...

    work page 2018

  2. [2]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018

    2018

  3. [3]

    Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners.OpenAI blog, 1(8):9, 2019

    2019

  4. [4]

    Language models are few-shot learners.arXiv preprint arXiv:2005.14165, 1:3, 2020

    Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sas- try, A Askell, S Agarwal, et al. Language models are few-shot learners.arXiv preprint arXiv:2005.14165, 1:3, 2020

    arXiv 2005

  5. [5]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  6. [6]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timo- thée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    arXiv 2023

  7. [7]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    arXiv 2023

  8. [8]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    arXiv 2024

Show all 36 references
  1. [9]

    Meta llama 4: The future of multimodal ai.Available at SSRN 5208228, 2025

    Ajit Singh. Meta llama 4: The future of multimodal ai.Available at SSRN 5208228, 2025

    2025

  2. [10]

    Fine- tuning can distort pretrained features and underperform out-of-distribution.arXiv preprint arXiv:2202.10054, 2022

    Ananya Kumar, Aditi Raghunathan, Robbie Jones, Tengyu Ma, and Percy Liang. Fine- tuning can distort pretrained features and underperform out-of-distribution.arXiv preprint arXiv:2202.10054, 2022

    2022 arXiv

  3. [11]

    Same pre-training loss, better downstream: Implicit bias matters for language models

    Hong Liu, Sang Michael Xie, Zhiyuan Li, and Tengyu Ma. Same pre-training loss, better downstream: Implicit bias matters for language models. InInternational Conference on Machine Learning, pages 22188–22214. PMLR, 2023

    2023

  4. [12]

    From words to numbers: Your large language model is secretly a capable regressor when given in-context examples.arXiv preprint arXiv:2404.07544, 2024

    Robert Vacareanu, Vlad-Andrei Negru, Vasile Suciu, and Mihai Surdeanu. From words to numbers: Your large language model is secretly a capable regressor when given in-context examples.arXiv preprint arXiv:2404.07544, 2024

    2024 arXiv

  5. [13]

    What do you learn from context? probing for sentence structure in contextualized word representations.arXiv preprint arXiv:1905.06316, 2019

    Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R Thomas McCoy, Najoung Kim, Benjamin Van Durme, Samuel R Bowman, Dipanjan Das, et al. What do you learn from context? probing for sentence structure in contextualized word representations.arXiv preprint arXiv:1905....

    1905 arXiv

  6. [14]

    What does bert learn about the structure of language? InACL 2019-57th Annual Meeting of the Association for Computational Linguistics, 2019

    Ganesh Jawahar, Benoît Sagot, and Djamé Seddah. What does bert learn about the structure of language? InACL 2019-57th Annual Meeting of the Association for Computational Linguistics, 2019

    2019

  7. [15]

    Understanding llm embeddings for regression

    Ethan Tang, Boyu Yang, and Xinyi Song. Understanding llm embeddings for regression. In ICLR Workshop on Foundation Models, 2024

    2024

  8. [16]

    Language models scale reliably with over-training and on downstream tasks.arXiv preprint arXiv:2403.08540, 2024

    Samir Yitzhak Gadre, Georgios Smyrnis, Vaishaal Shankar, Suchin Gururangan, Mitchell Wortsman, Rulin Shao, Jean Mercat, Alex Fang, Jeffrey Li, Sedrick Keh, et al. Language models scale reliably with over-training and on downstream tasks.arXiv preprint arXiv:2403.08540, 2024. 11

    2024 arXiv

  9. [17]

    Scaling laws for downstream task performance in machine translation

    Berivan Isik, Natalia Ponomareva, Hussein Hazimeh, Dimitris Paparas, Sergei Vassilvitskii, and Sanmi Koyejo. Scaling laws for downstream task performance in machine translation. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  10. [18]

    Scaling laws for predicting downstream performance in llms.arXiv preprint arXiv:2410.08527, 2024

    Yangyi Chen, Binxuan Huang, Yifan Gao, Zhengyang Wang, Jingfeng Yang, and Heng Ji. Scaling laws for predicting downstream performance in llms.arXiv preprint arXiv:2410.08527, 2024

    2024 arXiv

  11. [19]

    Learning by reconstruction produces uninformative features for perception.arXiv preprint arXiv:2402.11337, 2024

    Randall Balestriero and Yann LeCun. Learning by reconstruction produces uninformative features for perception.arXiv preprint arXiv:2402.11337, 2024

    2024 arXiv

  12. [20]

    Connecting pre-trained language model and down- stream task via properties of representation.Advances in Neural Information Processing Systems, 36:47216–47238, 2023

    Chenwei Wu, Holden Lee, and Rong Ge. Connecting pre-trained language model and down- stream task via properties of representation.Advances in Neural Information Processing Systems, 36:47216–47238, 2023

    2023

  13. [21]

    Qwen2 technical report, 2024

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze...

    2024 arXiv

  14. [22]

    OpenELM: An Efficient Language Model Family with Open Training and Inference Framework.arXiv.org, April 2024

    Sachin Mehta, Mohammad Hossein Sekhavat, Qingqing Cao, Maxwell Horton, Yanzi Jin, Chenfan Sun, Iman Mirzadeh, Mahyar Najibi, Dmitry Belenko, Peter Zatloukal, and Mohammad Rastegari. OpenELM: An Efficient Language Model Family with Open Training and Inference Framework.arXiv.or...

    2024 arXiv

  15. [23]

    Intent classification

    Bhuvaneshwari. Intent classification. https://huggingface.co/datasets/ Bhuvaneshwari/intent_classification. Accessed: 2025-04-29

    2025

  16. [24]

    Stop clickbait: Detecting and preventing clickbaits in online news media

    Abhijnan Chakraborty, Bhargavi Paranjape, Sourya Kakarla, and Niloy Ganguly. Stop clickbait: Detecting and preventing clickbaits in online news media. InAdvances in Social Networks Analysis and Mining (ASONAM), 2016 IEEE/ACM International Conference on, pages 9–16. IEEE, 2016

    2016

  17. [25]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. InProceedings of the 2013 Conference on Empirical Methods in Natural Language Process...

    2013

  18. [26]

    Bias identification

    Priya Patel. Bias identification. https://huggingface.co/datasets/PriyaPatel/Bias_ identification. Accessed: 2025-04-29

    2025

  19. [27]

    Efficient intent detection with dual sentence encoders

    Iñigo Casanueva, Tadas Temcinas, Daniela Gerz, Matthew Henderson, and Ivan Vulic. Efficient intent detection with dual sentence encoders. InProceedings of the 2nd Workshop on NLP for ConvAI - ACL 2020, mar 2020. URL https://arxiv.org/abs/2003.04807. Data available at https://g...

    2020 arXiv

  20. [28]

    Almeida, Jose Maria Gomez Hidalgo, and Akebo Yamakami

    Tiago A. Almeida, Jose Maria Gomez Hidalgo, and Akebo Yamakami. Contributions to the study of sms spam filtering: New collection and results. InProceedings of the 2011 ACM Symposium on Document Engineering (DOCENG’11), 2011

    2011

  21. [29]

    McCreery, Namit Katariya, Anitha Kannan, Manish Chablani, and Xavier Amatriain

    Clara H. McCreery, Namit Katariya, Anitha Kannan, Manish Chablani, and Xavier Amatriain. Effective transfer learning for identifying similar questions: Matching user questions to covid-19 faqs, 2020

    2020

  22. [30]

    Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales

    Bo Pang and Lillian Lee. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. InProceedings of the ACL, 2005. 12

    2005

  23. [31]

    CommonsenseQA: A question answering challenge targeting commonsense knowledge

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. CommonsenseQA: A question answering challenge targeting commonsense knowledge. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Langu...

    2019 doi

  24. [32]

    How cheap talk in climate disclosures relates to climate initiatives, corporate emissions, and reputation risk

    Julia Bingler, Mathias Kraus, Markus Leippold, and Nicolas Webersinke. How cheap talk in climate disclosures relates to climate initiatives, corporate emissions, and reputation risk. Working paper, Available at SSRN 3998435, 2023

    2023

  25. [33]

    Maas, Raymond E

    Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y . Ng, and Christopher Potts. Learning word vectors for sentiment analysis. InProceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142– 150...

    2011

  26. [34]

    Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks.arXiv preprint arXiv:2006.07322, 2020

    Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks.arXiv preprint arXiv:2006.07322, 2020

    2006 arXiv

  27. [35]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. InInternational conference on machine learning, pages 5156–5165. PMLR, 2020

    2020

  28. [36]

    Eigenvalue and generalized eigenvalue problems: Tutorial.arXiv preprint arXiv:1903.11240, 2019

    Benyamin Ghojogh, Fakhri Karray, and Mark Crowley. Eigenvalue and generalized eigenvalue problems: Tutorial.arXiv preprint arXiv:1903.11240, 2019. 13 A Appendix A.1 Proof of Lemma 1 Lemma 1(Equivalence between MSE and CE).Let X∈R d×ℓ be the token representations Denote h∗ =X L...

    1903 arXiv

Pith tools

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