Democratic Preference Alignment via Sortition-Weighted RLHF

Whose values should AI systems learn? Preference based alignment methods like RLHF derive their training signal from human raters, yet these rater pools are typically convenience samples that systematically over represent some demographics and under represent others. We introduce Democratic Preference Optimization, or DemPO, a framework that applies algorithmic sortition, the same mechanism used to construct citizen assemblies, to preference based fine tuning. DemPO offers two training schemes. Hard Panel trains exclusively on preferences from a quota satisfying mini public sampled via sortition. Soft Panel retains all data but reweights each rater by their inclusion probability under the sortition lottery. We prove that Soft Panel weighting recovers the expected Hard Panel objective in closed form. Using a public preference dataset that pairs human judgments with rater demographics and a seventy five clause constitution independently elicited from a representative United States panel, we evaluate Llama models from one billion to eight billion parameters fine tuned under each scheme. Across six aggregation methods, the Hard Panel consistently ranks first and the Soft Panel consistently outperforms the unweighted baseline, with effect sizes growing as model capacity increases. These results demonstrate that enforcing demographic representativeness at the preference collection stage, rather than post hoc correction, yields models whose behavior better reflects values elicited from representative publics.

Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models

Generative Reward Models (GenRMs) and LLM-as-a-Judge exhibit deceptive alignment by producing correct judgments for incorrect reasons, as they are trained and evaluated to prioritize Outcome Accuracy, which undermines their ability to generalize during RLHF. We introduce Rationale Consistency, a fine-grained metric that quantifies the alignment between the model's reasoning process and human judgment. Our evaluation of frontier models reveals that rationale consistency effectively discriminates among state-of-the-art models and detects deceptive alignment, while outcome accuracy falls short in both respects. To mitigate this gap, we introduce a hybrid signal that combines rationale consistency with outcome accuracy for GenRM training. Our training method achieves state-of-the-art performance on RM-Bench (87.1%) and JudgeBench (82%), surpassing outcome-only baselines by an average of 5%. Using RM during RLHF, our method effectively improves performance as demonstrated on Arena Hard v2, notably yielding a 7% improvement in creative writing tasks. Further analysis confirms that our method escapes the deceptive alignment trap, effectively reversing the decline in rationale consistency observed in outcome-only training.

SAFE: Stable Alignment Finetuning with Entropy-Aware Predictive Control for RLHF

Optimization (PPO) has been positioned by recent literature as the canonical method for the RL part of RLHF. PPO performs well empirically but has a heuristic motivation and handles the KL-divergence constraint used in LM-RLHF in an ad-hoc manner and suffers form reward oscillations, entropy collapse, value function drift, and sudden policy divergence that require frequent restarts and extensive hyperparameter tuning. In this paper, we develop a new pure on policy actor-critic RL method for the LM-RLHF setting. We present SAFE (Stable Alignment Finetuning with Entropy-aware control),a novel RLHF algorithm that combines a Double Soft-Min Critic for pessimistic value estimation with a new multi-layer stabilization framework combining entropy-gated KL regulation, and PID-controlled adaptive thresholds. Unlike standard PPO's symmetric KL penalties, SAFE distinguishes high-entropy exploration from low-entropy mode collapse and adjusts penalties dynamically based on reward velocity. Experiments on a 3B parameter model show SAFE achieves +5.15\% training-average reward than PPO (0.725 vs 0.689), negligible reward crashes, and superior KL control than ppo . Our method adds minimal computational overhead and provides an interpretable, crash-resistant RLHF framework that maintains aggressive learning speed while ensuring stable long-horizon optimization suitable for production deployment. Code is available at https://github.com/ryyzn9/SAFE

SLIME: Stabilized Likelihood Implicit Margin Enforcement for Preference Optimization

Direct preference optimization methods have emerged as a computationally efficient alternative to Reinforcement Learning from Human Feedback (RLHF) for aligning Large Language Models (LLMs). Latest approaches have streamlined the alignment process by deriving implicit reward functions, yet they often suffer from a critical objective mismatch: optimizing the relative margin between chosen and rejected responses does not guarantee the preservation of the chosen response's absolute likelihood. This can lead to unlearning, where the model degrades the probability of high-quality outputs to satisfy margin constraints, and formatting collapse caused by the over-penalization of rejected sequences. In this work, we introduce SLIME (Stabilized Likelihood Implicit Margin Enforcement), a reference-free alignment objective designed to decouple preference learning from generation quality. SLIME incorporates a three-pronged objective: (1) an anchoring term to maximize the likelihood of preferred responses; (2) a stabilizing penalty that prevents the probabilities of rejected tokens from collapsing to zero; and (3) a dual-margin mechanism that combines hard and soft constraints for precise boundary shaping. Our results demonstrate that SLIME achieves superior performance compared to state-of-the-art baselines while maintaining higher generation stability.

Semantic-aware Wasserstein Policy Regularization for Large Language Model Alignment

Large language models (LLMs) are commonly aligned with human preferences using reinforcement learning from human feedback (RLHF). In this method, LLM policies are generally optimized through reward maximization with Kullback-Leibler (KL) divergence regularization of the reference policy. However, KL and its $f$-divergence variants only compare token probabilities at identical indices, failing to capture semantic similarity. We propose Wasserstein Policy Regularization (WPR), a semantic-aware regularization for the RLHF framework based on the entropy-regularized Wasserstein distance, which incorporates the geometry of the token space. The dual formulation of the distance expresses the regularization as penalty terms applied to the reward via optimal dual variables, which yield a tractable objective compatible with standard RL algorithms. Empirically, our method outperforms KL- and $f$-divergence-based baselines, demonstrating the benefits of semantic-aware policy distances for alignment. Our code is available at https://github.com/aailab-kaist/WPR.

Adversarial Reward Auditing for Active Detection and Mitigation of Reward Hacking

Reinforcement Learning from Human Feedback (RLHF) remains vulnerable to reward hacking, where models exploit spurious correlations in learned reward models to achieve high scores while violating human intent. Existing mitigations rely on static defenses that cannot adapt to novel exploitation strategies. We propose Adversarial Reward Auditing (ARA), a framework that reconceptualizes reward hacking as a dynamic, competitive game. ARA operates in two stages: first, a Hacker policy discovers reward model vulnerabilities while an Auditor learns to detect exploitation from latent representations; second, Auditor-Guided RLHF (AG-RLHF) gates reward signals to penalize detected hacking, transforming reward hacking from an unobservable failure into a measurable, controllable signal. Experiments across three hacking scenarios demonstrate that ARA achieves the best alignment-utility tradeoff among all baselines: reducing sycophancy to near-SFT levels while improving helpfulness, decreasing verbosity while achieving the highest ROUGE-L, and suppressing code gaming while improving Pass@1. Beyond single-domain evaluation, we show that reward hacking, detection, and mitigation all generalize across domains -- a Hacker trained on code gaming exhibits increased sycophancy despite no reward for this behavior, and an Auditor trained on one domain effectively suppresses exploitation in others, enabling efficient multi-domain defense with a single model.

How RLHF Amplifies Sycophancy

Large language models often exhibit increased sycophantic behavior after preference-based post-training, showing a stronger tendency to affirm a user's stated or implied belief even when this conflicts with factual accuracy or sound judgment. We present a formal analysis of how alignment from human feedback can increase this failure mode by identifying an explicit amplification mechanism that causally links optimization against a learned reward to bias in the human preference data used for alignment. We show that the direction of behavioral drift is determined by a covariance under the base policy between endorsing the belief signal in the prompt and the learned reward, and that the first-order effect reduces to a simple mean-gap condition. We then analyze reward learning from pairwise comparisons under random utility models like Bradley-Terry and characterize when bias in human annotators' preferences induces this reward gap. Next, we propose a training-time intervention designed to neutralize the amplification mechanism itself. Among all post-trained policies that prevent sycophantic behavior from increasing, we characterize the unique policy closest in KL divergence to the unconstrained post-trained policy, and derive the corresponding minimal reward correction as a closed-form agreement penalty. Computational experiments find that reward gaps are common and cause behavioral drift in all the configurations considered.

GOPO: Policy Optimization using Ranked Rewards

Standard reinforcement learning from human feedback (RLHF) trains a reward model on pairwise preference data and then uses it for policy optimization. However, while reward models are optimized to capture relative preferences, existing policy optimization techniques rely on absolute reward magnitudes during training. In settings where the rewards are non-verifiable such as summarization, instruction following, and chat completion, this misalignment often leads to suboptimal performance. We introduce Group Ordinal Policy Optimization (GOPO), a policy optimization method that uses only the ranking of the rewards and discards their magnitudes. Our rank-based transformation of rewards provides several gains, compared to Group Relative Policy Optimization (GRPO), in settings with non-verifiable rewards: (1) consistently higher training/validation reward trajectories, (2) improved LLM-as-judge evaluations across most intermediate training steps, and (3) reaching a policy of comparable quality in substantially less training steps than GRPO. We demonstrate consistent improvements across a range of tasks and model sizes.

Omni-RRM: Advancing Omni Reward Modeling via Automatic Rubric-Grounded Preference Synthesis

Multimodal large language models (MLLMs) have shown remarkable capabilities, yet their performance is often capped by the coarse nature of existing alignment techniques. A critical bottleneck remains the lack of effective reward models (RMs): existing RMs are predominantly vision-centric, return opaque scalar scores, and rely on costly human annotations. We introduce \textbf{Omni-RRM}, the first open-source rubric-grounded reward model that produces structured, multi-dimension preference judgments with dimension-wise justifications across \textbf{text, image, video, and audio}. At the core of our approach is \textbf{Omni-Preference}, a large-scale dataset built via a fully automated pipeline: we synthesize candidate response pairs by contrasting models of different capabilities, and use strong teacher models to \emph{reconcile and filter} preferences while providing a modality-aware \emph{rubric-grounded rationale} for each pair. This eliminates the need for human-labeled training preferences. Omni-RRM is trained in two stages: supervised fine-tuning to learn the rubric-grounded outputs, followed by reinforcement learning (GRPO) to sharpen discrimination on difficult, low-contrast pairs. Comprehensive evaluations show that Omni-RRM achieves state-of-the-art accuracy on video (80.2\% on ShareGPT-V) and audio (66.8\% on Audio-HH-RLHF) benchmarks, and substantially outperforms existing open-source RMs on image tasks, with a 17.7\% absolute gain over its base model on overall accuracy. Omni-RRM also improves downstream performance via Best-of-$N$ selection and transfers to text-only preference benchmarks. Our data, code, and models are available at https://anonymous.4open.science/r/Omni-RRM-CC08.

Unifying Adversarial Robustness and Training Across Text Scoring Models

Research on adversarial robustness in language models is currently fragmented across applications and attacks, obscuring shared vulnerabilities. In this work, we propose unifying the study of adversarial robustness in text scoring models spanning dense retrievers, rerankers, and reward models. This motivates adapting both attacks and adversarial training methods across model roles. Unlike open-ended generation, text scoring failures are directly testable: an attack succeeds when an irrelevant or rejected text outscores a relevant or chosen one. Using this principled lens of text scoring, we demonstrate that current adversarial training formulations for language models are often short-sighted, failing to effectively generalize across attacks. To address this, we introduce multiple adversarial training methods for text scoring models and show that combining complementary training methods can yield strong robustness while also improving task effectiveness. We also highlight the practical value of our approach for RLHF, showing that our adversarially trained reward models mitigate reward hacking and support the training of better-aligned LLMs. We provide our code and models for further study.