Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

PILAF: Optimal Human Preference Sampling for Reward Modeling

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read PILAF claims that choosing preference pairs from an interpolated policy makes DPO training follow the true oracle reward, cutting annotation and compute by over 40 percent in iterative settings.

desk verdict A genuinely novel theoretical sampler whose theorems don't yet cover the simplified algorithm actually evaluated; the empirical results are encouraging but need error bars and honest cost accounting. read the letter →

arxiv 2502.04270 v1 pith:AMSWVV7R submitted 2025-02-06 cs.LG stat.ML

classification cs.LGstat.ML
keywords preferencesamplingrewardmodelingdirectoptimizationRLHFgradientalignmentstatisticalefficiencypolicyinterpolationiterativeDPO
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 the way response pairs are chosen for preference labeling is not a neutral detail: by sampling one response from a reward-tilted 'optimistic' policy and one from a 'pessimistic' policy, the gradient of the DPO/MLE loss aligns with the gradient of the true KL-regularized reward objective. The proposed scheme, PILAF, interpolates the current policy and a reference policy at the logit level, so it needs no extra hyperparameters and costs little more than vanilla sampling. The theoretical variant T-PILAF is shown to make the loss gradient point along the oracle-regularized policy gradient up to a second-order error, and to make the estimation error concentrate along the directions that matter most for the value function. If these claims hold, preference data collection becomes a principled part of alignment rather than an afterthought, and the same data budget buys a better reward-KL trade-off.

What carries the argument

The load-bearing object is the tilted sampling distribution $\mu$ built from the current policy and the DPO implicit reward: $\pi_\theta^{\pm} \propto \pi_\theta \exp(\pm r_\theta)$, mixed with $\pi_\theta$ at probability $p_0(x)=Z_\theta^+ Z_\theta^-/(1+Z_\theta^+ Z_\theta^-)$, plus the weight $w(x)=(1+Z_\theta^+ Z_\theta^-)/\mathcal{Z}_\theta$. The key identity, proven in the appendix, is that the averaged sampling density satisfies $\bar{\mu}(\vec y_a,\vec y_b|x)/(\pi_\theta(\vec y_a|x)\pi_\theta(\vec y_b|x)) = (1+Z_\theta^+ Z_\theta^-)/(2\sigma'(r_\theta(\vec y_a)-r_\theta(\vec y_b)))$. This identity does the work: it cancels the sigmoid derivative in the MLE gradient and replaces the sigmoid gap with a reward gap, which is exactly the integrand of $\nabla_\theta J(\pi_\theta)$. The practical algorithm replaces exact normalization with logit-level interpolation, giving per-token prediction rules $\pi_\theta^+ = \mathrm{softmax}((1+\beta)h_\theta - \beta h_{\mathrm{ref}})$ and $\pi_\theta^- = \mathrm{softmax}((1-\beta)h_\theta + \beta h_{\mathrm{ref}})$.

What would settle it

Train two matched DPO runs on the same prompts and proxy reward, one with PILAF and one with T-PILAF where the partition functions are estimated by sampling and the Bernoulli probability is computed from the tilted policies; if the simplified PILAF does not reproduce T-PILAF's per-step gradient alignment, measured as the cosine similarity between $\nabla_\theta L$ and $\nabla_\theta J(\pi_\theta)$ using the known proxy reward, then the empirical gains are not attributable to the proved mechanism.

Watch

Extended reading notes

Core claim

At its own level of claim, the paper's central discovery is a sampling distribution with a gradient-alignment identity. For T-PILAF, response pairs are drawn from a Bernoulli mixture of the current policy $\pi_\theta$ and the pair of tilted policies $\pi_\theta \exp(r_\theta)$ and $\pi_\theta \exp(-r_\theta)$ (normalized); with the accompanying weight function, the density ratio simplifies to $1/(2\sigma'(\Delta r_\theta))$, and this converts the sigmoid difference $\sigma(\Delta r_\star)-\sigma(\Delta r_\theta)$ in the MLE gradient into the linear difference $\Delta r_\star-\Delta r_\theta$ that appears in $\nabla_\theta J(\pi_\theta)$. The resulting identity is $\nabla_\theta L(\theta) = -\frac{\beta}{Z_\theta}\nabla_\theta J(\pi_\theta) + T_2$, with $T_2$ a second-order error. From the statistical side, the asymptotic covariance of the estimated parameter is bounded by the inverse of the gradient covariance under the optimal policy, and the value gap follows a problem-independent chi-square tail. The practical PILAF replaces the partition functions with 1, fixes the Bernoulli probability at $1/2$, uses the unweighted DPO loss, and implements the tilted policies token-wise as $\mathrm{softmax}((1+\beta)h_\theta - \beta h_{\mathrm{ref}})$ and $\mathrm{softmax}((1-\beta)h_\theta + \beta h_{\mathrm{ref}})$.

Load-bearing premise

The load-bearing premise is that the simplified PILAF used in the experiments—fixed $1/2$ mixing probability, unit partition functions, and the unweighted DPO loss—inherits the proven gradient-alignment and chi-square efficiency of T-PILAF; no theorem in the paper covers that simplified scheme.

Editorial extensions

If this is right

  • DPO updates on PILAF-collected data ascend the oracle KL-regularized reward, so each labeled pair is informative in the direction that matters for downstream value.
  • In iterative DPO, PILAF reaches the endpoint reward of the baselines around halfway through training, which translates to more than 40 percent savings in annotation and computation.
  • The statistical result implies the estimator's error concentrates along directions where the Hessian of $J$ is large, so training remains stable even when initialized from an overfitted policy.
  • Because the sampling mechanism is decoupled from the loss, the same interpolation can be applied to other preference optimization losses, not only DPO; the paper sketches how it extends to PPO-based RLHF.

Reading between the lines

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

  • Editorial extension: the gradient-alignment identity suggests a direct diagnostic the paper does not report—during training, compute the cosine similarity between the DPO loss gradient and the oracle objective gradient on a held-out proxy reward; PILAF should show higher alignment than vanilla sampling.
  • Editorial extension: the claimed cost savings are measured against a reward-model proxy for human labels; with real annotators, inter-annotator noise or label-cost asymmetry could change the observed trade-off, and the theory assumes the Bradley-Terry model holds exactly.
  • Editorial extension: the simplified PILAF fixes the mixing probability at $1/2$ and ignores partition functions, and the experiments use $\beta=0.1$; for substantially larger $\beta$, one could test whether estimating $p_0(x)$ or the partition functions preserves the alignment more faithfully.
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

5 major / 5 minor

Summary. The paper proposes a response-sampling scheme for preference data in RLHF. It first introduces T-PILAF, a theoretically analyzed scheme in which response pairs are drawn either from the current policy or from reward-tilted policies, using a prompt-dependent Bernoulli probability, normalized tilted policies, and a weighted DPO loss. Theorems 4.1-4.3 claim first-order gradient alignment with the oracle-regularized objective and asymptotic covariance control for T-PILAF. The paper then presents PILAF, a simplified variant that sets the Bernoulli probability to 1/2, drops partition functions, uses the unweighted DPO loss, and implements the tilted policies by per-token logit interpolation. PILAF is evaluated in iterative and online DPO experiments with a reward-model oracle, reporting better reward-KL trade-offs at reduced annotation and computation cost.

Significance. If the theoretical claims held for the method actually evaluated, this would be a valuable contribution: it would give a principled, theoretically motivated way to choose preference pairs during iterative RLHF and reduce annotation cost. The construction of T-PILAF is elegant, and the explicit density-ratio identity in Eq. (15) is a genuine design achievement. The empirical results are encouraging and include a detailed cost comparison. However, the theoretical guarantees are proved only for T-PILAF, while the evaluated PILAF omits or changes several components of that scheme, and no analysis quantifies the effect of those changes. The statistical theorems also contain an equality-to-upper-bound gap. With the claims appropriately scoped, the paper would be a useful contribution to the RLHF data-curation literature.

major comments (5)
  1. [Section 5 and Algorithm 1 vs. Section 4] Theorems 4.1-4.3 are stated for T-PILAF, whose sampling distribution uses the prompt-dependent Bernoulli probability p0(x), the normalized tilted policies in Eq. (6), and the weighted loss with w(x) in Eq. (7). The practical PILAF in Algorithm 1 and Section 5 replaces p0 by 1/2, sets Z+_theta and Z-_theta to 1, uses the standard unweighted DPO loss, and replaces the tilted policies by per-token softmax interpolation. No theorem, lemma, or numerical bound quantifies the deviation between these two schemes. The footnote in Section 5 saying that partition functions are approximately 1 for small beta is a heuristic, and the per-token approximation is imported from Proposition 1 of Liu et al. (2024b) without verifying that it produces the tilted autoregressive distribution in this setting. A first-order calculation with p0=1/2, Z+=Z-=1, and w=1 changes the density-ratio identity (15) to mu/pi_theta pi_theta = 1/(4 sigma'), so the alignment direction may survive those three simplifications, but the per-token interpolation and the statistical covariance analysis are not covered. The abstract's claim that PILAF is 'theoretically grounded, demonstrating optimality' therefore overstates what is proved for the implemented algorithm.
  2. [Theorem 4.2 and Appendix B.2.2] The main text states Omega <= C1 * Sigma_star^{-1} with C1 a universal constant and says that under T-PILAF the matrix Sigma_star is given by Eq. (8). The appendix proves only a lower bound, Sigma_star >= (1/(C Z_theta)) E_{rho, pi_star}[Cov(nabla_theta r_star | x)], under Condition (27), where C is problem-dependent. Consequently, the inverse inequality gives an upper bound on Omega in terms of E[Cov], not an equality. The constant C1 is not universal because it inherits the constant from Condition (27), the weight bound ||w||_infty, and Z_theta. The theorem statement should be corrected to an inequality with the correct problem-dependent constants.
  3. [Theorem 4.3 and Appendix B.2.3] The theorem text says the value gap J(pi_star) - J(pi_hat) 'asymptotically follows a chi-square distribution.' The proof in Appendix B.2.3 derives only an upper bound: the limiting quadratic form is dominated by a rescaled chi-square random variable, as in Eq. (10). Without an exact covariance equality, the asymptotic distribution is a weighted quadratic form with problem-dependent weights, not a chi-square distribution. Eq. (10) is a valid tail bound under the stated inequalities, but the distributional claim in the main text should be weakened to a stochastic dominance or tail-bound statement.
  4. [Appendix B.2.2, Condition (27)] The comparability condition on conditional variances under r_theta and r_star is introduced only in the appendix and is load-bearing for Theorems 4.2 and 4.3. It is stated without discussion of when it holds; it involves the learned reward r_theta for arbitrary theta and is not implied by the smoothness or realizability assumptions stated elsewhere. The main statements should either include this condition explicitly or provide a justification or a relaxation.
  5. [Appendix C.2.2, Lemma B.5] The Hessian calculation differentiates the loss expression (14a) while treating the weight w(x) and the sampling distribution mu as independent of theta, although w and mu are defined in terms of pi_theta in Eq. (7) and Section 3. If the expectation is taken over a sampling distribution that varies with theta, the derivatives of mu and w contribute additional terms. If instead mu and w are meant to be frozen at the data-collection policy, this should be stated explicitly, since the Z-estimator argument in Theorem 4.2 relies on the population loss being a fixed objective in theta.
minor comments (5)
  1. [Throughout] There are several typos: 'Specfically' in Section 1, 'poposed' in Theorem B.1, 'Graident' in the Appendix C.2.3 heading, 'T able' in several table captions, and 'respsect' in Appendix C.1.1.
  2. [Section 2.3 and Theorem 4.3] The objective J(pi) is defined inline in Section 2.3 but Theorem 4.3 refers to it as 'equation (6)'; the equation numbering is inconsistent.
  3. [Section 5, Table 1] The cost table reports an average sampling cost of 3 forward passes for PILAF, but the text explains this as a mixture of 2-pass and 4-pass cases. Stating the calculation explicitly in the table caption would improve clarity.
  4. [Section 6, Figures 2-5] The reward-KL curves appear to come from single runs, and no error bars or multiple seeds are reported. Given the variance of online RLHF training, reporting at least a few seeds for the main comparisons would strengthen the empirical claims.
  5. [Section 5] The statement that PILAF involves 'no hyperparameter tuning' is somewhat overstated: the KL regularization coefficient beta is a hyperparameter, even though it is taken from the objective and fixed at 0.1 in all experiments.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the theoretical result is a deliberate construction, not a hidden restatement of its assumptions.

full rationale

The paper's central gradient-alignment theorem (Theorem 4.1) is derived by choosing the T-PILAF sampling distribution and weight function so that the density ratio (15) exactly cancels the sigmoid derivative in the MLE gradient. This is an explicit design derivation: Equation (22) defines the mixture sampling law, Equation (15) shows cancellation of σ′, and Equation (7) defines the weight w(x). The final identity ∇θL(θ) = −β/Zθ ∇θJ(πθ) + T2 is obtained by algebra plus a first-order Taylor expansion, not by assuming the conclusion. The statistical results use standard Z-estimator theory (Kosorok 2008) and clearly stated regularity conditions; no parameter is fitted to data and then reported as a prediction. The only notable gap is that the theorems are proven for T-PILAF (with prompt-dependent p0(x), partition functions Z±θ, and weighted loss), while the implemented PILAF (Algorithm 1, Section 5) simplifies these to p0 = 1/2, Z± = 1, and unweighted DPO. This is an unverified transfer of theory to practice, but it is not a circular reduction: the practical scheme is not obtained by defining it as the thing that makes the theorem true, and the simplification is explicitly acknowledged as a heuristic (Section 5 footnote). The per-token softmax approximation is imported from an external reference (Liu et al., 2024b), not from the authors' own prior results, and is not load-bearing self-citation. No fitted quantity is renamed as a prediction, no uniqueness theorem is invoked, and no known empirical pattern is merely renamed. Accordingly, the paper's derivation chain is self-contained with respect to circularity; the theory–implementation mismatch should be treated as a correctness or scope risk, not as circular reasoning.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central theory assumes a well-specified Bradley-Terry model, smooth bounded rewards, and an unverified variance-comparability condition. The practical algorithm introduces ad hoc simplifications (p=1/2, Z=1, unweighted loss) that are outside the theoretical framework. No new physical or empirical entities are postulated.

free parameters (2)
  • mixing probability p = 1/2 = 1/2
    Replaces the theoretically derived prompt-dependent Bernoulli probability p0(x) = Z+Z-/(1+Z+Z-) in Algorithm 1. Chosen by hand with no derivation; changes the exploration-exploitation balance and is not covered by the theorems.
  • regularization coefficient beta = 0.1
    Fixed in all experiments as the KL regularization coefficient in the objective. Not fitted to data, but it is a manually chosen hyperparameter that controls the strength of logit interpolation and the KL penalty.
assumptions (6)
  • domain assumption Human preferences follow the Bradley-Terry model with an oracle reward r*
    Used throughout, starting at Equation (1), to define the probability that one response is preferred over another.
  • domain assumption The reward model is well-specified: r* = r_theta* for some parameter theta*
    Assumed in Theorem 4.2 and Appendix B.2 to derive the asymptotic distribution of the estimated parameter.
  • standard math The reward function, its gradient, and Hessian are uniformly bounded and Lipschitz continuous in theta
    Invoked in Appendix B.2 to apply the master theorem for Z-estimators and to control Taylor expansion remainders.
  • ad hoc to paper Condition (27): the conditional variance of preference indicators under r_theta is comparable to that under r*
    Introduced in Appendix B.2.2 to prove Theorem 4.2. It is a regularity condition with no verification or discussion of when it might fail, and it is essential for the covariance lower bound.
  • ad hoc to paper The DPO implicit reward decomposes token-wise, so the tilted policies can be sampled autoregressively via logit interpolation
    Used in Section 5 to derive softmax((1+beta)h_theta - beta h_ref) and softmax((1-beta)h_theta + beta h_ref). The paper cites Proposition 1 in Liu et al. (2024b) for this approximation, but it is not generally exact for truncated responses.
  • ad hoc to paper Partition functions Z±_theta(x) are approximately 1
    Section 5 justifies replacing Z± with 1 by arguing beta is small, but no quantitative bound is given and the approximation is not used in the theoretical analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PILAF: Optimal Human Preference Sampling for Reward Modeling." pith.science (2026). https://pith.science/paper/AMSWVV7R

@misc{pith2026250204270,
  author       = {Pith},
  title        = {Pith review of: PILAF: Optimal Human Preference Sampling for Reward Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AMSWVV7R}},
  note         = {Machine review of arXiv:2502.04270}
}
read the original abstract

As large language models increasingly drive real-world applications, aligning them with human values becomes paramount. Reinforcement Learning from Human Feedback (RLHF) has emerged as a key technique, translating preference data into reward models when oracle human values remain inaccessible. In practice, RLHF mostly relies on approximate reward models, which may not consistently guide the policy toward maximizing the underlying human values. We propose Policy-Interpolated Learning for Aligned Feedback (PILAF), a novel response sampling strategy for preference labeling that explicitly aligns preference learning with maximizing the underlying oracle reward. PILAF is theoretically grounded, demonstrating optimality from both an optimization and a statistical perspective. The method is straightforward to implement and demonstrates strong performance in iterative and online RLHF settings where feedback curation is critical.

Figures

Figures reproduced from arXiv: 2502.04270 by the authors.

Figure 1
Figure 1. Overview of our approach. (a) We consider a full RLHF training setup, where a language model (LM) policy is iteratively refined through active data collection. Our goal is to develop an optimal response sampling method for preference labeling. (b) We introduce PILAF, which generates responses by interpolating between the current policy and a reference policy, balancing exploration and exploitation. (c) Our theoretic… view at source ↗
Figure 2
Figure 2. Reward-KL curve for Iterative DPO. All training runs start from the same model obtained at the end of the first iteration via Vanilla Sampling. Each dot represents an evaluation performed every 50 training steps [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Reward-KL curve for Online DPO. Each dot represents an evaluation performed every 50 training steps. Results [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Online DPO with an overfitted initial policy [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Online DPO with an over￾fitted initial policy. Full results of the [PITH_FULL_IMAGE:figures/full_fig_p046_5.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. Inverse Reinforcement Learning Meets Large Language Model Post-Training: Basics, Advances, and Opportunities

    cs.LG 2025-07 unverdicted novelty 1.0 of 10

    A tutorial reviewing LLM alignment through the lens of inverse reinforcement learning, arguing that neural reward models learned from human data are central to post-training.

Reference graph

Works this paper leans on

46 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    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 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    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

  3. [3]

    G., Guo, Z

    Azar, M. G., Guo, Z. D., Piot, B., Munos, R., Rowland, M., Valko, M., and Calandriello, D. A general theoretical paradigm to understand learning from human preferences. In International Conference on Artificial Intelligence and Statistics, pp.\ 4447--4455. PMLR, 2024

  4. [4]

    Training a helpful and harmless assistant with reinforcement learning from human feedback

    Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022

  5. [5]

    Value-incentivized preference optimization: A unified approach to online and offline rlhf

    Cen, S., Mei, J., Goshvadi, K., Dai, H., Yang, T., Yang, S., Schuurmans, D., Chi, Y., and Dai, B. Value-incentivized preference optimization: A unified approach to online and offline rlhf. arXiv preprint arXiv:2405.19320, 2024

  6. [6]

    Das, N., Chakraborty, S., Pacchiano, A., and Chowdhury, S. R. Active preference optimization for sample efficient rlhf. In ICML 2024 Workshop on Theoretical Foundations of Foundation Models, 2024

  7. [7]

    Degris, T., White, M., and Sutton, R. S. Off-policy actor-critic. arXiv preprint arXiv:1205.4839, 2012

  8. [8]

    Raft: Reward ranked finetuning for generative foundation model alignment

    Dong, H., Xiong, W., Goyal, D., Zhang, Y., Chow, W., Pan, R., Diao, S., Zhang, J., KaShun, S., and Zhang, T. Raft: Reward ranked finetuning for generative foundation model alignment. Transactions on Machine Learning Research, 2023

Show all 46 references
  1. [9]

    Rlhf workflow: From reward modeling to online rlhf, 2024

    Dong, H., Xiong, W., Pang, B., Wang, H., Zhao, H., Zhou, Y., Jiang, N., Sahoo, D., Xiong, C., and Zhang, T. Rlhf workflow: From reward modeling to online rlhf, 2024

  2. [10]

    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

  3. [11]

    Kto: Model alignment as prospect theoretic optimization

    Ethayarajh, K., Xu, W., Muennighoff, N., Jurafsky, D., and Kiela, D. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024

  4. [12]

    Scaling laws for reward model overoptimization

    Gao, L., Schulman, J., and Hilton, J. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pp.\ 10835--10866, 2023

  5. [13]

    L., and Baxter, J

    Greensmith, E., Bartlett, P. L., and Baxter, J. Variance reduction techniques for gradient estimates in reinforcement learning. Journal of Machine Learning Research, 5 0 (9), 2004

  6. [14]

    S., Lillicrap, T., Turner, R

    Gu, S. S., Lillicrap, T., Turner, R. E., Ghahramani, Z., Sch \"o lkopf, B., and Levine, S. Interpolated policy gradient: Merging on-policy and off-policy gradient estimation for deep reinforcement learning. Advances in neural information processing systems, 30, 2017

  7. [15]

    Direct language model alignment from online ai feedback

    Guo, S., Zhang, B., Liu, T., Liu, T., Khalman, M., Llinares, F., Rame, A., Mesnard, T., Zhao, Y., Piot, B., et al. Direct language model alignment from online ai feedback. arXiv preprint arXiv:2402.04792, 2024

  8. [16]

    Openrlhf: An easy-to-use, scalable and high-performance rlhf framework

    Hu, J., Wu, X., Zhu, Z., Xianyu, Wang, W., Zhang, D., and Cao, Y. Openrlhf: An easy-to-use, scalable and high-performance rlhf framework. arXiv preprint arXiv:2405.11143, 2024

  9. [17]

    Reinforcement learning from human feedback with active queries

    Ji, K., He, J., and Gu, Q. Reinforcement learning from human feedback with active queries. arXiv preprint arXiv:2402.09401, 2024

  10. [18]

    and Uehara, M

    Kallus, N. and Uehara, M. Statistically efficient off-policy policy gradients. In International Conference on Machine Learning, pp.\ 5089--5100. PMLR, 2020

  11. [19]

    Kosorok, M. R. Introduction to empirical processes and semiparametric inference, volume 61. Springer, 2008

  12. [20]

    Y., Chandu, K., Dziri, N., Kumar, S., Zick, T., Choi, Y., Smith, N

    Lambert, N., Pyatkin, V., Morrison, J., Miranda, L., Lin, B. Y., Chandu, K., Dziri, N., Kumar, S., Zick, T., Choi, Y., Smith, N. A., and Hajishirzi, H. Rewardbench: Evaluating reward models for language modeling. https://huggingface.co/spaces/allenai/reward-bench, 2024

  13. [21]

    Let's verify step by step, 2023

    Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobbe, K. Let's verify step by step, 2023. URL https://arxiv.org/abs/2305.20050

  14. [22]

    Limo, F. A. F., Tiza, D. R. H., Roque, M. M., Herrera, E. E., Murillo, J. P. M., Huallpa, J. J., Flores, V. A. A., Castillo, A. G. R., Pe \ n a, P. F. P., Carranza, C. P. M., et al. Personalized tutoring: Chatgpt as a virtual tutor for personalized learning experiences. Przest...

  15. [23]

    Y., Zeng, L., Liu, J., Yan, R., He, J., Wang, C., Yan, S., Liu, Y., and Zhou, Y

    Liu, C. Y., Zeng, L., Liu, J., Yan, R., He, J., Wang, C., Yan, S., Liu, Y., and Zhou, Y. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451, 2024 a

  16. [24]

    Decoding-time realignment of language models

    Liu, T., Guo, S., Bianco, L., Calandriello, D., Berthet, Q., Llinares, F., Hoffmann, J., Dixon, L., Valko, M., and Blondel, M. Decoding-time realignment of language models. arXiv preprint arXiv:2402.02992, 2024 b

  17. [25]

    J., and Liu, J

    Liu, T., Zhao, Y., Joshi, R., Khalman, M., Saleh, M., Liu, P. J., and Liu, J. Statistical rejection sampling improves preference optimization. In The Twelfth International Conference on Learning Representations, 2024 c

  18. [26]

    S., and Lin, M

    Liu, Z., Chen, C., Du, C., Lee, W. S., and Lin, M. Sample-efficient alignment for llms. arXiv preprint arXiv:2411.01493, 2024 d

  19. [27]

    Sample efficient reinforcement learning from human feedback via active exploration

    Mehta, V., Das, V., Neopane, O., Dai, Y., Bogunovic, I., Schneider, J., and Neiswanger, W. Sample efficient reinforcement learning from human feedback via active exploration. arXiv preprint arXiv:2312.00267, 2023

  20. [28]

    Active preference learning for large language models

    Muldrew, W., Hayes, P., Zhang, M., and Barber, D. Active preference learning for large language models. In Forty-first International Conference on Machine Learning, 2024

  21. [29]

    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

  22. [30]

    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, 2023

  23. [31]

    I., M \'e nard, P., Moulines, E., and Valko, M

    Scheid, A., Boursier, E., Durmus, A., Jordan, M. I., M \'e nard, P., Moulines, E., and Valko, M. Optimal design for reward modeling in rlhf. arXiv preprint arXiv:2410.17055, 2024

  24. [32]

    Proximal policy optimization algorithms

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

  25. [33]

    Shi, R., Zhou, R., and Du, S. S. The crucial role of samplers in online direct preference optimization. arXiv preprint arXiv:2409.19605, 2024

  26. [34]

    Deterministic policy gradient algorithms

    Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., and Riedmiller, M. Deterministic policy gradient algorithms. In International conference on machine learning, pp.\ 387--395. Pmlr, 2014

  27. [35]

    S., McAllester, D., Singh, S., and Mansour, Y

    Sutton, R. S., McAllester, D., Singh, S., and Mansour, Y. Policy gradient methods for reinforcement learning with function approximation. Advances in neural information processing systems, 12, 1999

  28. [36]

    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

  29. [37]

    Williams, R. J. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8: 0 229--256, 1992

  30. [38]

    J., Krishnamurthy, A., Rosset, C., Awadallah, A., and Rakhlin, A

    Xie, T., Foster, D. J., Krishnamurthy, A., Rosset, C., Awadallah, A., and Rakhlin, A. Exploratory preference optimization: Harnessing implicit q*-approximation for sample-efficient rlhf. arXiv preprint arXiv:2405.21046, 2024

  31. [39]

    Iterative preference learning from human feedback: Bridging theory and practice for rlhf under kl-constraint

    Xiong, W., Dong, H., Ye, C., Wang, Z., Zhong, H., Ji, H., Jiang, N., and Zhang, T. Iterative preference learning from human feedback: Bridging theory and practice for rlhf under kl-constraint. In Forty-first International Conference on Machine Learning, 2024

  32. [40]

    Xu, H., Sharaf, A., Chen, Y., Tan, W., Shen, L., Van Durme, B., Murray, K., and Kim, Y. J. Contrastive preference optimization: Pushing the boundaries of llm performance in machine translation. In Forty-first International Conference on Machine Learning, 2024

  33. [41]

    Some things are more cringe than others: Preference optimization with the pairwise cringe loss

    Xu, J., Lee, A., Sukhbaatar, S., and Weston, J. Some things are more cringe than others: Preference optimization with the pairwise cringe loss. arXiv preprint arXiv:2312.16682, 2023

  34. [42]

    Inference scaling for long-context retrieval augmented generation

    Yue, Z., Zhuang, H., Bai, A., Hui, K., Jagerman, R., Zeng, H., Qin, Z., Wang, D., Wang, X., and Bendersky, M. Inference scaling for long-context retrieval augmented generation. arXiv preprint arXiv:2410.04343, 2024

  35. [43]

    Self-exploring language models: Active preference elicitation for online alignment

    Zhang, S., Yu, D., Sharma, H., Zhong, H., Liu, Z., Yang, Z., Wang, S., Hassan, H., and Wang, Z. Self-exploring language models: Active preference elicitation for online alignment. arXiv preprint arXiv:2405.19332, 2024

  36. [44]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  37. [45]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  38. [46]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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