The Differences Between Direct Alignment Algorithms are a Blur

• URL: http://arxiv.org/abs/2502.01237v1
• Date: Mon, 03 Feb 2025 10:54:14 GMT
• Title: The Differences Between Direct Alignment Algorithms are a Blur
• Authors: Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daniil Gavrilov,
• Abstract summary: Direct Alignment Algorithms (DAAs) simplify language model alignment by replacing reinforcement learning (RL) and reward modeling (RM)<n>DAAs can be classified by their ranking losses (pairwise vs. pointwise), by the rewards used in those losses (e.g., likelihood ratios of policy and reference policy, or odds ratios), or by whether a Supervised Fine-Tuning phase is required (two-stage vs. one-stage)<n>These results highlight the importance of careful evaluation to avoid premature claims of performance gains or overall superiority in alignment algorithms.
• Score: 3.0059120458540383
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Direct Alignment Algorithms (DAAs) simplify language model alignment by replacing reinforcement learning (RL) and reward modeling (RM) in Reinforcement Learning from Human Feedback (RLHF) with direct policy optimization. DAAs can be classified by their ranking losses (pairwise vs. pointwise), by the rewards used in those losses (e.g., likelihood ratios of policy and reference policy, or odds ratios), or by whether a Supervised Fine-Tuning (SFT) phase is required (two-stage vs. one-stage). We first show that one-stage methods underperform two-stage methods. To address this, we incorporate an explicit SFT phase and introduce the $\beta$ parameter, controlling the strength of preference optimization, into single-stage ORPO and ASFT. These modifications improve their performance in Alpaca Eval 2 by +$3.46$ (ORPO) and +$8.27$ (ASFT), matching two-stage methods like DPO. Further analysis reveals that the key factor is whether the approach uses pairwise or pointwise objectives, rather than the specific implicit reward or loss function. These results highlight the importance of careful evaluation to avoid premature claims of performance gains or overall superiority in alignment algorithms.

Related papers

• Towards Bridging the Reward-Generation Gap in Direct Alignment Algorithms [6.844727437767206]
  Direct Alignment Algorithms (DAAs) have emerged as efficient alternatives to Reinforcement Learning from Human Feedback (RLHF) algorithms.<n>DAAs suffer from a misalignment between optimization objectives during training and actual generation performance during inference.<n>We introduce a simple yet effective approach called Prefix-Oriented Equal-length Training (POET)<n>POET truncates both preferred and dispreferred responses to match the shorter one's length.
  arXiv  Detail & Related papers  (2025-06-11T07:02:18Z)
• ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Preference Optimization [48.50761200321113]
  We introduce ConfPO, a method for preference learning in Large Language Models (LLMs)<n>It identifies and optimize preference-critical tokens based solely on the training policy's confidence, without requiring any auxiliary models or compute.<n> Experimental results on challenging alignment benchmarks, including AlpacaEval 2 and Arena-Hard, demonstrate that ConfPO consistently outperforms uniform DAAs.
  arXiv  Detail & Related papers  (2025-06-10T11:54:22Z)
• ABoN: Adaptive Best-of-N Alignment [19.22348775001393]
  We propose a prompt-adaptive strategy for Best-of-N alignment that allocates inference-time compute more efficiently.<n>Our method is simple, practical, and compatible with any LM/RM combination.
  arXiv  Detail & Related papers  (2025-05-17T15:24:48Z)
• One-Stage Top-$k$ Learning-to-Defer: Score-Based Surrogates with Theoretical Guarantees [3.6787328174619254]
  We introduce the first one-stage Top-$k$ Learning-to-Defer framework.<n>We learn a shared score-based model that selects the $k$ most cost-effective entities-labels or experts-per input.<n>Experiments on CIFAR-10 and SVHN confirm that our one-stage Top-$k$ method strictly outperforms Top-1 deferral.
  arXiv  Detail & Related papers  (2025-05-15T10:41:16Z)
• ComPO: Preference Alignment via Comparison Oracles [36.81379432115315]
  We propose a new preference alignment method based on comparison oracles and provide the convergence guarantee for its basic scheme.<n>A highlight of our work is that we evidence the importance of designing specialized methods for preference pairs with distinct likelihood margin.
  arXiv  Detail & Related papers  (2025-05-08T17:56:57Z)
• Achieving $\widetilde{\mathcal{O}}(\sqrt{T})$ Regret in Average-Reward POMDPs with Known Observation Models [56.92178753201331]
  We tackle average-reward infinite-horizon POMDPs with an unknown transition model. We present a novel and simple estimator that overcomes this barrier.
  arXiv  Detail & Related papers  (2025-01-30T22:29:41Z)
• $f$-PO: Generalizing Preference Optimization with $f$-divergence Minimization [54.94545757220999]
  $f$-PO is a novel framework that generalizes and extends existing approaches. We conduct experiments on state-of-the-art language models using benchmark datasets.
  arXiv  Detail & Related papers  (2024-10-29T02:11:45Z)
• $α$-DPO: Adaptive Reward Margin is What Direct Preference Optimization Needs [45.46582930202524]
  $alpha$-DPO is an adaptive preference optimization algorithm for large language models. It balances the policy model and the reference model to achieve personalized reward margins. It consistently outperforms DPO and SimPO across various model settings.
  arXiv  Detail & Related papers  (2024-10-14T04:29:57Z)
• SeRA: Self-Reviewing and Alignment of Large Language Models using Implicit Reward Margins [30.767203592231496]
  Self-Reviewing and Alignment (SeRA) is a cost-efficient and effective method that can be readily combined with existing DAAs. SeRA comprises of two components: (1) sample selection using implicit reward margins, which helps alleviate over-fitting to some undesired features, and (2) preference bootstrapping using implicit rewards to augment preference data with updated policy models.
  arXiv  Detail & Related papers  (2024-10-12T04:17:28Z)
• TIS-DPO: Token-level Importance Sampling for Direct Preference Optimization With Estimated Weights [73.9088920210495]
  We propose a token-level importance sampling DPO objective named TIS-DPO that assigns importance weights to each token based on its reward.<n>TIS-DPO significantly outperforms various baseline methods on harmlessness and helpfulness alignment and summarization tasks.
  arXiv  Detail & Related papers  (2024-10-06T04:03:00Z)
• Bridging and Modeling Correlations in Pairwise Data for Direct Preference Optimization [75.1240295759264]
  We propose an effective framework for Bridging and Modeling Correlations in pairwise data, named BMC.<n>We increase the consistency and informativeness of the pairwise preference signals through targeted modifications.<n>We identify that DPO alone is insufficient to model these correlations and capture nuanced variations.
  arXiv  Detail & Related papers  (2024-08-14T11:29:47Z)
• Correcting the Mythos of KL-Regularization: Direct Alignment without Overoptimization via Chi-Squared Preference Optimization [78.82586283794886]
  We present a new offline alignment algorithm, $chi2$-Preference Optimization ($chi$PO) $chi$PO implements the principle of pessimism in the face of uncertainty via regularization. It is provably robust to overoptimization and achieves sample-complexity guarantees based on single-policy concentrability.
  arXiv  Detail & Related papers  (2024-07-18T11:08:40Z)
• A Finite-Sample Analysis of an Actor-Critic Algorithm for Mean-Variance Optimization in a Discounted MDP [1.0923877073891446]
  We analyze a Temporal Difference (TD) learning algorithm with linear function approximation (LFA) for policy evaluation. We derive finite-sample bounds that hold (i) in the mean-squared sense and (ii) with high probability under tail iterate averaging. These results establish finite-sample theoretical guarantees for risk-sensitive actor-critic methods in reinforcement learning.
  arXiv  Detail & Related papers  (2024-06-12T05:49:53Z)
• Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer [52.09480867526656]
  We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences.<n>To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model.<n>Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
  arXiv  Detail & Related papers  (2024-05-26T05:38:50Z)
• $i$REPO: $i$mplicit Reward Pairwise Difference based Empirical Preference Optimization [12.266207199002604]
  Large Language Models (LLM) can sometimes produce outputs that deviate from human expectations. We propose a novel framework named $i$REPO, which utilizes implicit Reward pairwise difference regression for Empirical Preference Optimization. We show that $i$REPO effectively achieves self-alignment using soft-label, self-generated responses and the logit of empirical AI annotators.
  arXiv  Detail & Related papers  (2024-05-24T05:42:11Z)
• Policy Gradient with Active Importance Sampling [55.112959067035916]
  Policy gradient (PG) methods significantly benefit from IS, enabling the effective reuse of previously collected samples. However, IS is employed in RL as a passive tool for re-weighting historical samples. We look for the best behavioral policy from which to collect samples to reduce the policy gradient variance.
  arXiv  Detail & Related papers  (2024-05-09T09:08:09Z)
• Stability-penalty-adaptive follow-the-regularized-leader: Sparsity, game-dependency, and best-of-both-worlds [46.30750729936261]
  Follow-the-regularized-leader (FTRL) has recently emerged as one of the most promising approaches for obtaining various types of adaptivity in bandit problems. We establish several algorithms with three types of adaptivity: sparsity, game-dependency, and best-of-both-worlds (BOBW)
  arXiv  Detail & Related papers  (2023-05-26T23:20:48Z)
• Towards Model-Agnostic Post-Hoc Adjustment for Balancing Ranking Fairness and Algorithm Utility [54.179859639868646]
  Bipartite ranking aims to learn a scoring function that ranks positive individuals higher than negative ones from labeled data. There have been rising concerns on whether the learned scoring function can cause systematic disparity across different protected groups. We propose a model post-processing framework for balancing them in the bipartite ranking scenario.
  arXiv  Detail & Related papers  (2020-06-15T10:08:39Z)
• ADAHESSIAN: An Adaptive Second Order Optimizer for Machine Learning [91.13797346047984]
  We introduce ADAHESSIAN, a second order optimization algorithm which dynamically incorporates the curvature of the loss function via ADAptive estimates. We show that ADAHESSIAN achieves new state-of-the-art results by a large margin as compared to other adaptive optimization methods.
  arXiv  Detail & Related papers  (2020-06-01T05:00:51Z)
• Provably Efficient Exploration in Policy Optimization [117.09887790160406]
  This paper proposes an Optimistic variant of the Proximal Policy Optimization algorithm (OPPO) OPPO achieves $tildeO(sqrtd2 H3 T )$ regret. To the best of our knowledge, OPPO is the first provably efficient policy optimization algorithm that explores.
  arXiv  Detail & Related papers  (2019-12-12T08:40:02Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.