Pith. sign in

REVIEW 6 major objections 5 minor 18 references

Learning Distribution-Wise Control in Representation Space for Language Models

T0 review · 6 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 Replacing deterministic intervention nodes with learned Gaussian sampling, especially in early layers, gives consistent gains in language-model steering and robustness across reasoning benchmarks.

desk verdict A simple, plausible stochastic extension of ReFT with a genuinely useful early-layer finding, but the headline comparison is not controlled and the appendix DPI chain is wrong. read the letter →

arxiv 2506.06686 v1 pith:UKQDMDDQ submitted 2025-06-07 cs.CL

Chunyuan Deng , Ruidi Chang , Hanjie Chen This is my paper
classification cs.CL
keywords languagemodelinterventionsrepresentationfine-tuningdistribution-wiseinterventionstochasticreparameterizationconceptsubspacecommonsensereasoningarithmeticsteerabilityandrobustness
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 tries to establish that a language model can be steered more effectively by learning a small distribution over edited representations than by learning a single edited point. It proposes a drop-in change to representation fine-tuning: split the intervention node into two small networks that output a mean and a log-variance, then sample the edited representation from that Gaussian during the forward pass. Across eight commonsense and seven arithmetic reasoning benchmarks on Llama-family models, the stochastic variants consistently beat their pointwise counterparts, with the largest gains (+4% to +6%) coming from early layers and from larger learned standard deviations. If the result holds, adding learned Gaussian noise to early-layer interventions is a cheap, parameter-efficient upgrade to existing steering methods, and the learned variance doubles as a test-time temperature dial for trading reasoning precision against generation diversity.

What carries the argument

The load-bearing object is the stochastic reparameterized intervention node: instead of a deterministic map $\hat{Z} = W^\top Z + b$, the intervention computes $\mu = \mathrm{MLP}_\mu(Z)$, $\log\sigma^2 = \mathrm{MLP}_{\log\sigma^2}(Z)$, and $\hat{Z} = \mu + \sigma \odot \epsilon$ with $\epsilon \sim \mathcal{N}(0,I)$. This is the variational-autoencoder reparameterization trick applied without the KL term, so gradients flow through the sampled edit while the model freely shapes the spread of the distribution. The mean carries the pointwise edit and the learned standard deviation controls how widely the model explores the neighborhood of the concept subspace; clamping the intervention to the range of adjacent layer weights keeps training stable. The same node underlies all four D-variants (D-MLP, D-RED, D-SwiGLU, D-ReFT), and at inference the learned $\sigma$ is scaled by a temperature $\tau$ to modulate stochasticity.

What would settle it

Train D-ReFT on one arithmetic benchmark with the variance network frozen to a fixed constant matched to the final learned $\sigma$, keeping sampling, parameters, and the mean network intact; if this fixed-noise variant reproduces the accuracy gain, then learning the variance is not what drives the result, and the distribution-wise claim collapses to a claim that additive noise helps.

Watch

Extended reading notes

Core claim

The central claim is that distribution-wise intervention—replacing a deterministic transformation of hidden representations with a learnable Gaussian sampler $\hat{Z} = \mu + \sigma \odot \epsilon$—improves both controllability and robustness of language-model steering compared with pointwise intervention. The method reparameterizes the stochastic node as two MLPs, one for $\mu$ and one for $\log\sigma^2$, and drops the usual KL term, so the model freely learns a distribution over the concept subspace rather than a variational posterior. The authors report that the stochastic variants raise accuracy by roughly +4% to +6% when inserted at a single early layer, that accuracy rises with learned $\sigma$, and that the best all-layer configuration uses stochastic nodes in the first 25% of layers and deterministic nodes in the rest, outperforming ReFT, LoRA, DoRA, and RED across 15 benchmarks while using comparable or slightly more parameters (e.g., 0.039% vs 0.026% for Llama-3-8B). They also show that scaling the learned variance by a temperature $\tau$ at inference shifts behavior: low $\tau$ helps reasoning tasks and high $\tau$ helps instruction following, and the stochastic variants tolerate word-deletion perturbations better than pointwise ones.

Load-bearing premise

The load-bearing premise of the paper's explanation is that hidden states cannot regain information about the input once it has been discarded, so later layers cannot recover what early layers lose (the Appendix A data-processing-inequality argument); if that argument is set aside, the account of why early-layer stochasticity beats pointwise intervention reduces to an empirical correlation between accuracy and learned variance.

Editorial extensions

If this is right

  • A mixed configuration with stochastic interventions in the first 25% of layers and pointwise interventions in the remaining layers beats both all-pointwise and all-stochastic configurations on all 15 benchmarks tested.
  • The +4% to +6% early-layer gains are larger than the gains reported for the progression from LoRA through DoRA to ReFT, so distribution-wise control is a disproportionately effective upgrade for early-layer steering.
  • Because D-ReFT is a drop-in replacement for the intervention node, the same distribution-wise treatment can be applied to other intervention architectures (MLP, RED, SwiGLU), and the paper reports gains for each in early layers.
  • A single D-ReFT checkpoint can be switched between reasoning-oriented and instruction-following behavior at inference by scaling the learned variance with temperature $\tau$, without retraining.
  • Distribution-wise interventions degrade more gracefully than pointwise ones when words are randomly deleted from the input, indicating a robustness advantage beyond average accuracy.

Reading between the lines

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

  • If the variance term's role is mainly adaptive noise scheduling, then cheaper mechanisms such as learned per-layer dropout rates or a single noise scale might recover part of the gain without doubling the intervention parameters; the paper does not test this, but its rank ablation already shows that added parameters alone do not explain the improvement.
  • The temperature-scaling results suggest $\tau$ could serve as a general task-style dial, low for precise reasoning and high for open-ended generation, but the paper tests only three task categories, so whether this generalizes is open.
  • Appendix A's Eq. (22) lists mutual-information terms in the opposite order from the standard data-processing inequality for the chain $X \to Z^{(1)} \to \dots \to Z^{(L)} \to \hat{Y}$ (earlier layers should have more information about $X$, not less); a corrected inequality still motivates intervening early, but it does not by itself explain why the stochastic variants beat deterministic ones in ea
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

6 major / 5 minor

Summary. The paper proposes distribution-wise interventions (D-Intervention) as a drop-in replacement for the pointwise interventions used in representation fine-tuning, specifically ReFT. The method replaces a deterministic intervention MLP with two MLPs that predict the mean and log-variance of a Gaussian, and uses the reparameterization trick to sample the modified representation during training. The authors report layer-wise studies on arithmetic reasoning, an all-layer mixed strategy in which the first 25% of intervention layers are stochastic and the rest remain deterministic, experiments on commonsense and arithmetic benchmarks with Llama-family models, a robustness evaluation under random word deletion, and a test-time temperature scaling study on Alpaca-Eval. The central claim is that distribution-wise interventions consistently outperform pointwise interventions in accuracy and robustness, with the largest gains in early layers.

Significance. If the empirical claims hold, the proposed modification is a cheap, parameter-efficient improvement to ReFT and is easy to adopt: it adds a variance network and sampling step, and the reported accuracy gains of +4% to +6% in early layers would be practically useful. The paper also ships a code link, which supports reproducibility. However, the current evidence does not yet establish the headline claim. The baseline comparison is not controlled because ReFT and D-ReFT use different learning rates and training epochs, some baseline numbers are copied from prior work, and no error bars or significance tests are reported. In addition, the information-theoretic justification in Appendix A contains a reversed data-processing-inequality chain, and Figure 4's 'strong correlation' claim rests on only three non-monotonic points. These issues are fixable but require additional experiments and corrections.

major comments (6)
  1. [Appendix B.1, B.2; Tables 1 and 3] The central comparison is not controlled. Appendix B.1 sets ReFT to lr=9e-4 with 12 epochs, while Appendix B.2 sets D-ReFT to lr=1e-3 or 3e-3 with 9 epochs. Appendix B.1 itself states that decreasing ReFT epochs to 9 leads to a performance drop, so the 9-vs-12 epoch difference alone can plausibly account for several points on these benchmarks. In addition, Table 1 marks the ReFT baselines as taken from Wu et al. (2024b) and Liu et al. (2024), so the D-ReFT rows are not produced by the same pipeline. Please rerun ReFT under the identical training budget, learning-rate schedule, and evaluation harness as D-ReFT, or otherwise demonstrate that the reported gains are not an artifact of these discrepancies.
  2. [Section 6.1; Tables 1 and 3] The abstract claims that distribution-wise interventions 'consistently outperform' pointwise interventions, but no variance or significance information is reported. Section 6.1 states that layer-wise results are averaged over three runs with different seeds, yet no standard deviations, per-seed values, or significance tests are given, and Tables 1 and 3 have no error bars. Per-benchmark results already contradict 'consistently': in Table 3, for Llama-3-8B, D-ReFT25% scores 97.4 on MultiArith versus ReFT's 98.5, and 30.3 versus 30.0 on AQuA. Please report full per-seed results and a significance test (or at least error bars), and soften the claim to match the actual pattern.
  3. [Appendix A, Eq. (22)] The data-processing-inequality chain in Eq. (22) is reversed and conflates variables. For the Markov chain X -> Z^(1) -> ... -> Z^(L) -> Y, DPI gives I(X; Z^(1)) >= ... >= I(X; Z^(L)), not the reverse as written. Moreover, I(Y; X) and I(Y; Y-hat) cannot be inserted into a chain of terms I(X; Z^(l)) without additional assumptions about the relationship between X and Y. Consequently, the sentence 'deep layers cannot recover information lost in earlier layers' is not established by Eq. (22). Since Section 6.1 invokes this appendix to explain the early-layer gains, the justification should be corrected or removed.
  4. [Section 6.2, Figure 4] The claimed correlation between accuracy and learned standard deviation is not supported by the evidence. Figure 4 shows three points (accuracy 0.688, 0.723, 0.694) for scaling factors 0.2, 1.0, and 1.8, which are non-monotonic in both accuracy and the reported sigma values. No correlation coefficient, confidence interval, or repeated-run variability is provided. This does not justify the abstract's statement that 'larger standard deviations correlating strongly with improved performance.' Please add more scaling-factor values with error bars or remove the correlation claim.
  5. [Section 8, Table 4] Table 4 is internally inconsistent. By the definitions in Eqs. (17)-(19), tau=1 corresponds to training-time stochasticity, so the row '+tau=1' should match the 'D-ReFT (Baseline)' row if the baseline uses the same inference procedure. The table reports 88.6/76.6/82.4 for the baseline and 86.9/76.8/83.7 for '+tau=1'. Clarify what the baseline row represents, and if the baseline uses a different inference-time sampling or different hyperparameters, the temperature comparison needs to be rerun under a consistent protocol.
  6. [Section 3.3, Eqs. (3)-(7)] The derivation equating cross-entropy loss with conditional entropy H(Y|f_phi(Z^(l))) is not exact in general. Cross-entropy between the data distribution p and the model distribution q is H(p,q) = H(Y|f(Z)) + KL(p||q); it equals H(Y|f(Z)) only when the model distribution matches the true conditional. Therefore the claim in Eq. (7) that minimizing cross-entropy is exactly equivalent to maximizing mutual information I(Y; f_phi(Z^(l))) is not justified. Since the actual training objective is cross-entropy, please present the mutual-information identity as a motivating analogy or derive it under a stated assumption.
minor comments (5)
  1. [Section 5] There are several typographical issues, including 'Abalation' in the Section 6.1 heading, 'pheonmenon' in Section 7.1, and 'evaulate'/'hyperpameter' in Section 5. These should be corrected.
  2. [Appendix B.2] Appendix B.2 is titled 'ReFT Hyperparameter Configuration' but describes D-ReFT hyperparameters; rename it to 'D-ReFT Hyperparameter Configuration' to avoid confusion.
  3. [Section 5.2] The text says 'seven commonsense reasoning benchmarks' but lists eight (BoolQ, PIQA, SIQA, HellaSwag, WinoGrande, ARC-e, ARC-c, OBQA); the abstract also says eight. Please correct the count.
  4. [Table 1] In the LLaMA-13B ReFT row, the ARC-c and OBQA entries appear concatenated as '86.273.784.2'; this should be two separate numbers (likely 73.7 and 84.2).
  5. [Section 5.3] The hyperparameter list mentions 'noise scale (epsilon)' but the text elsewhere distinguishes the learned sigma from the epsilon scaling factor; please define the relationship between these quantities explicitly to avoid ambiguity.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the method is a standard reparameterization applied as a drop-in replacement, and the reported gains are empirical observations rather than fitted predictions.

full rationale

The paper does not fit a parameter and then rename it as a prediction. The core derivation in Section 3.3 (Eqs. 2-7) is the standard identity between cross-entropy, conditional entropy, and mutual information; it introduces no fitted constant and does not presuppose the empirical gains. The stochastic reparameterization in Eqs. (9)-(13) is the standard Kingma and Welling trick applied as a drop-in replacement, and it is not derived from the results it later claims. The early-layer accuracy gains, the noise-scale ablation, and the robustness curves are experimental measurements, not consequences of an equation that already encodes those outcomes. The only self-citation (Deng et al., 2024) appears in Related Work to support the general statement that logical reasoning is linearly encoded; it is not load-bearing for the proposed method or for the central comparison. Two genuine weaknesses are visible in the paper, but neither is circular: Appendix A's DPI chain (Eq. 22) is mis-ordered as written relative to the standard data processing inequality, and the baseline comparisons in Tables 1 and 3 use different hyperparameters (e.g., ReFT uses 12 epochs while D-ReFT uses 9 in Appendix B), with some ReFT numbers marked as copied from prior work. Those are experimental-control and correctness concerns, not circular derivations. No explored step reduces by construction to its own input, so the circularity score is minimal.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method's central claims rest on the reparameterization trick, a continuity assumption about concept space, and an information-theoretic justification. The DPI-based axiom is misstated in the appendix. No invented entities are introduced; the stochastic nodes are a modeling choice, not a new ontological entity.

free parameters (2)
  • epsilon scaling factor lambda = 1.0 (default, ablated 0 to 3.0)
    Hand-picked default noise magnitude; performance degrades beyond roughly 2.4 due to numerical instability, so the method's success is sensitive to this choice.
  • Model-specific clamping bounds (vmin, vmax) = min/max of adjacent layer weight matrices
    Chosen heuristically in Section 4.4 to prevent instability; no derivation or sensitivity analysis is given, and the bounds are fixed before training.
assumptions (4)
  • standard math The reparameterization trick allows gradients to flow through stochastic sampling (Kingma & Welling, 2014).
    Invoked in Section 4.2, Eq. (13), to enable backpropagation through the sampled intervention.
  • domain assumption The concept space of interventions is continuous, so neighborhood exploration around a learned intervention is meaningful.
    Stated in Section 1 and Section 4.1, citing Gandikota et al. (2023); this motivates the distribution-wise extension.
  • domain assumption The data processing inequality implies that deep layers cannot recover information lost in earlier layers, so early-layer intervention is preferable.
    Appendix A, Eq. (22); the equation as written has the direction of the inequalities reversed and conflates X, Y, and Z, so the invoked support is questionable.
  • standard math Minimizing cross-entropy is equivalent to maximizing mutual information I(Y; f_phi(Z)).
    Section 3.3, Eqs. (2)-(7); standard identity, not circular.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Distribution-Wise Control in Representation Space for Language Models." pith.science (2026). https://pith.science/paper/UKQDMDDQ

@misc{pith2026250606686,
  author       = {Pith},
  title        = {Pith review of: Learning Distribution-Wise Control in Representation Space for Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UKQDMDDQ}},
  note         = {Machine review of arXiv:2506.06686}
}
read the original abstract

Interventions in language models (LMs) are applied strategically to steer model behavior during the forward pass. Learnable interventions, also known as representation fine-tuning, aim to apply pointwise control within the concept subspace and have proven effective in altering high-level behaviors. In this work, we extend this approach to the distribution level, enabling the model to learn not only pointwise transformations but also the surrounding regions of the concept subspace. We demonstrate that these methods perform effectively in early layers, with larger standard deviations correlating strongly with improved performance. Across eight commonsense reasoning and seven arithmetic reasoning benchmarks, our distribution-wise interventions consistently outperform pointwise interventions in controllability and robustness. These results illustrate that distribution-wise interventions provide a more comprehensive method for steering model behavior and enabling finer-grained control over language models. The code is at: \href{https://github.com/chili-lab/D-Intervention}{https://github.com/chili-lab/D-Intervention}.

Figures

Figures reproduced from arXiv: 2506.06686 by the authors.

Figure 1
Figure 1. Distribution-wise vs. Pointwise Intervention. This is an intuitive yet effective adaptation, as previous research suggests that the concept space is continuous (Gandikota et al., 2023). tions allow researchers to steer model behavior, often by manipulating representation in the model’s latent space. A core challenge in intervention-based methods lies in advancing from low-level control to high-level control. Low-lev… view at source ↗
Figure 2
Figure 2. Performance of different layer-wise D-interventions with respect to the intervention layer. We report the average score of Llama-3-8B on seven arithmetic benchmarks: AddSub, SingleEQ, MultiArith, AQuA, GSM8K, MAWPS, and SVAMP. Notably, D-intervention at early layers yields the best performance, highlighting a significant discrepancy across layers. methods are tuned on the development set, and we report the average r… view at source ↗
Figure 3
Figure 3. Accuracy with different scaling factors for ϵ in D￾ReFT. When the scaling factor for ϵ = 0, the method reduces to the original ReFT. For scaling factors of ϵ > 2.4, D-ReFT enters a region prone to numerical instability due to large variance. −0.002 0.000 0.002 x 0 500 1000 1500 2000 Probability Density (µ=3.23e-05, σ=1.85e-04) (a) Acc = 0.688 −0.002 0.000 0.002 x 0 500 1000 1500 2000 Probability Density (µ=3.43e-05,… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Correlation between accuracy and std σ. The scaling factors λ of ϵ are 0.2, 1.0, and 1.8 for subfigures (a), (b), and (c). increasing ϵ further introduces higher variance, making training more difficult and causing numerical instability. We explore the correlation betw…
Figure 5
Figure 5. Figure 5: Accuracy with different choices of rank in Llama-3-8B for arithmetic reasoning tasks. Analyzing the results of low-rank settings ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 8 canonical work pages

  1. [1]

    Find the total number of books Sam bought: 13 adventure books + 17 mystery books = 30 books

    work page

  2. [2]

    blackboxnlp-1.2

    URL https://aclanthology.org/2023. blackboxnlp-1.2. Olah, C. et al. Understanding lstm networks. 2015. Park, K., Choe, Y . J., and Veitch, V . The linear rep- resentation hypothesis and the geometry of large lan- guage models. InForty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. UR...

    work page 2023

  3. [10]

    doi: 10.18653/v1/2021.naacl-main.168

    Association for Computational Linguistics. doi: 10.18653/v1/2021.naacl-main.168. URL https:// aclanthology.org/2021.naacl-main.168. Pfeiffer, J., Vuli´c, I., Gurevych, I., and Ruder, S. MAD- X: An Adapter-Based Framework for Multi-Task Cross- Lingual Transfer. In Webber, B., Cohn, T., He, Y ., and Liu, Y . (eds.),Proceedings of the 2020 Conference on Empi...

    doi:10.18653/v1/2021.naacl-main.168 2021

  4. [12]

    URL https://proceedings.mlr.press/ v162/ravfogel22a.html. Roy, S. and Roth, D. Solving general arithmetic word problems. In M `arquez, L., Callison-Burch, C., and Su, J. (eds.),Proceedings of the 2015 Conference on Em- pirical Methods in Natural Language Processing, pp. 1743–1752, Lisbon, Portugal, 2015. Association for Com- putational Linguistics. doi: 1...

    arXiv 2015

  5. [15]

    doi: 10.18653/v1/2020.emnlp-main.101

    Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.101. URL https:// aclanthology.org/2020.emnlp-main.101. Tishby, N. and Zaslavsky, N. Deep learning and the in- formation bottleneck principle, 2015. URL https: //arxiv.org/abs/1503.02406. Tishby, N., Pereira, F. C., and Bialek, W. The informa- tion bottleneck method, 2000. URL htt...

    arXiv 2020

  6. [16]

    org/CorpusID:219978318

    URL https://api.semanticscholar. org/CorpusID:219978318. Wu, M., Liu, W., Wang, X., Li, T., Lv, C., Ling, Z., Zhu, J., Zhang, C., Zheng, X., and Huang, X. Advancing parameter efficiency in fine-tuning via representation editing, 2024a. URL https://arxiv.org/abs/ 2402.15179. Wu, Z., Geiger, A., Icard, T., Potts, C., and Goodman, N. D. Interpretability at s...

    arXiv 2023

  7. [18]

    The answer in Arabic numerals is 15

    Subtract the number of used books from the total: 30 books - 15 used books = 15 new books Therefore, Sam bought 15 new books. The answer in Arabic numerals is 15. SingleEq.The SingleEq (Koncel-Kedziorski et al., 2015) dataset is a specialized resource designed to support research and development in the field of mathematical problem-solving, particularly f...

    work page 2015

  8. [48]

    findings-acl.48

    URL https://aclanthology.org/2022. findings-acl.48. 13 Learning Distribution-Wise Control in Representation Space for Language Models Tian, R., Mao, Y ., and Zhang, R. Learning V AE-LDA models with rounded reparameterization trick. In Webber, B., Cohn, T., He, Y ., and Liu, Y . (eds.),Proceedings of the 2020 Conference on Empirical Methods in Natural Lang...

    work page 2022

Show all 18 references
  1. [106]

    findings-emnlp.106

    URL https://aclanthology.org/2023. findings-emnlp.106. Belrose, N., Schneider-Joseph, D., Ravfogel, S., Cotterell, R., Raff, E., and Biderman, S. LEACE: perfect linear concept erasure in closed form. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. ...

    2023

  2. [353]

    acl-long.353

    URL https://aclanthology.org/2021. acl-long.353. Ling, W., Yogatama, D., Dyer, C., and Blunsom, P. Pro- gram induction by rationale generation: Learning to solve and explain algebraic word problems. In Barzilay, R. and Kan, M.-Y . (eds.),Proceedings of the 55th Annual Meeting ...

    2021 doi

  3. [435]

    emnlp-main.435

    URL https://aclanthology.org/2023. emnlp-main.435. Subramani, N., Suresh, N., and Peters, M. Ex- tracting latent steering vectors from pretrained lan- guage models. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.),Findings of the Association for Computational Linguistic...

    2023 doi

  4. [617]

    emnlp-main.617

    URL https://aclanthology.org/2020. emnlp-main.617. Ravfogel, S., Twiton, M., Goldberg, Y ., and Cotterell, R. Linear adversarial concept erasure. In Chaud- huri, K., Jegelka, S., Song, L., Szepesv ´ari, C., Niu, G., and Sabato, S. (eds.),International Conference on Ma- chine L...

    2020

  5. [744]

    emnlp-main.744

    URL https://aclanthology.org/2023. emnlp-main.744. Hewitt, J. and Manning, C. D. A structural probe for finding syntax in word representations. In Burstein, J., Doran, C., and Solorio, T. (eds.),Proceedings of the 2019 Confer- ence of the North American Chapter of the Associat...

    2023 doi

  6. [2019]

    URL http://proceedings.mlr.press/ v97/houlsby19a.html. Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., Li, Y ., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adapta- tion of large language models. InThe Tenth Interna- tional Conference on Learning Representations, ICLR 2022...

    2022

  7. [2020]

    doi: 10.18653/v1/2020.emnlp-main.347

    Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.347. URL https:// aclanthology.org/2020.emnlp-main.347. Clark, C., Lee, K., Chang, M.-W., Kwiatkowski, T., Collins, M., and Toutanova, K. BoolQ: Exploring the surpris- ing difficulty of natural yes/no ...

    2020 arXiv

  8. [2021]

    Gandikota, R., Materzynska, J., Zhou, T., Torralba, A., and Bau, D

    URL http://proceedings.mlr.press/ v139/fu21a.html. Gandikota, R., Materzynska, J., Zhou, T., Torralba, A., and Bau, D. Concept sliders: Lora adaptors for pre- cise control in diffusion models, 2023. URL https: //arxiv.org/abs/2311.12092. Geiger, A., Lu, H., Icard, T., and Pott...

    2023 arXiv

  9. [2022]

    Hu, Z., Wang, L., Lan, Y ., Xu, W., Lim, E.-P., Bing, L., Xu, X., Poria, S., and Lee, R

    URL https://openreview.net/forum? id=nZeVKeeFYf9. Hu, Z., Wang, L., Lan, Y ., Xu, W., Lim, E.-P., Bing, L., Xu, X., Poria, S., and Lee, R. LLM-adapters: An adapter fam- ily for parameter-efficient fine-tuning of large language models. In Bouamor, H., Pino, J., and Bali, K. (ed...

    2023

  10. [2023]

    doi: 10.18653/v1/2023.emnlp-main.319

    Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.319. URL https:// aclanthology.org/2023.emnlp-main.319. Huang, J., Wu, Z., Potts, C., Geva, M., and Geiger, A. Ravel: Evaluating interpretability methods on disen- tangling language model representatio...

    2023 arXiv

Pith tools

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