Preference Learning for LLMs

Preference optimization has become a central approach for aligning large language models with human values, but its effectiveness depends heavily on the quality of preference annotations. In practice, preference data is often noisy due to annotation errors and ambiguity. Existing robust preference optimization methods primarily operate at the sequence level, implicitly treating entire responses as uniformly correct or incorrect. However, rejected responses may still contain informative reasoning steps, while preferred responses can include subtle errors. In this work, we propose Preference Optimization via Robust Token-level reweighting (PORT), a fine-grained framework for robust alignment under noisy preferences. PORT performs token-level reweighting using the empirical cumulative distribution function (CDF) of token log-probabilities, yielding an efficient forward-pass-only proxy to gradient-norm penalization that selectively suppresses corrupted tokens without additional backward-pass computation. We provide theoretical analysis showing that PORT reduces the gradient bias from noisy preference labels by minimizing an upper bound on the bias term. Extensive experiments across diverse noise settings demonstrate that PORT consistently improves robustness and outperforms existing sequence-level baselines. The code is available at https://github.com/osu-srml/PORT

Direct Preference Optimization (DPO) is a simple and widely used alternative to RLHF for aligning large language models (LLMs), but its sequence-level objective cannot precisely locate or suppress the local patterns that make an otherwise acceptable response unsafe or undesirable. We propose TLRS-DPO (Token-Level Risk Scoring for Better Direct Preference Optimization), which augments DPO with fine-grained token-level control while remaining fully offline and reference-based. TLRS-DPO introduces a token-level robust compensation term that converts the policy-reference log-ratio on rejected responses into localized risk scores, a curriculum-style schedule that gradually increases the strength of this compensation during training, and a risk-adaptive screening mechanism that selectively penalizes only high-risk tokens instead of entire sequences. This design bridges sequence-level preference fitting and token-level behavior control, and can be implemented as a drop-in modification to standard DPO trainers. Empirically, TLRS-DPO improves alignment quality and safety-relevant behavior with minimal impact on helpfulness compared to strong DPO baselines.

We introduce a proximal preference optimization framework that unifies the implicit reward formulation of Direct Preference Optimization (DPO) with the trust-region stability of Proximal Policy Optimization (PPO). Our framework yields two complementary algorithms. P$^2$O (Proximal Preference Optimization) introduces a clipped importance ratio directly into the DPO preference margin, producing a reward-model-free algorithm that supports safe dataset reuse, bounded policy divergence, and monotonic preference improvement guarantees. PKTO (Proximal Kahneman-Tversky Optimization) extends P$^2$O by replacing the DPO margin with KTO's prospect-theoretic utility margins, adding a self-regulating adaptive baseline that tightens the effective trust region as the policy drifts, and an asymmetric weighting that handles imbalanced or asymmetric preference datasets. We provide a unified theoretical analysis, including Kullback Leibler (KL) trust-region bounds, a policy-improvement lower bound, and convergence guarantees under multi-epoch proximal updates. Empirically, P$^2$O and PKTO match or exceed DPO and KTO in alignment quality while maintaining better stability under repeated dataset reuse across multiple LLMs on three preference benchmarks. The code base is available here.

Aligning Large Language Models (LLMs) with human preferences via Reinforcement Learning from Human Feedback (RLHF) has driven substantial improvements across a wide range of tasks, but the cost of acquiring preference labels remains the dominant bottleneck in the alignment pipeline. While many recent alignment methods improve sample efficiency through active exploration over candidate responses, they typically query the oracle uniformly across contexts, even for contexts where the learned reward estimates are already reliable enough to provide synthetic labels. We propose Adaptive Ensemble-disagreement Routing for Oracle feedback (AERO), a fully online RLHF algorithm built on a Dyna-style model-based RL architecture, in which an ensemble-based Epistemic Reward Model (ERM) serves both as a reward posterior for exploration and as a learned preference-feedback model for synthetic labeling. AERO uses ensemble vote entropy, a Query-by-Committee disagreement measure, combined with an adaptive thresholding mechanism to decide which contexts require oracle feedback and which can instead receive ERM-based synthetic preference labels, concentrating expensive labels where they are most informative. Empirical results across different model families and scales show that AERO achieves highly competitive alignment performance while reducing the required oracle-query budget compared to online DPO and strong active-exploration baselines.

Post-training language models for reasoning typically rely on gold answers, task-specific verifiers, or learned reward models, limiting scalability when such signals are scarce. Multi-agent debate offers a self-supervised alternative, but converting debate into reliable training supervision remains challenging: majority voting weights agreeing responses equally, lacks a principled tie-breaking rule, and is ill-defined for open-ended tasks without a canonical answer; likelihood-based scoring can conflate fluency with correctness; and homogeneous committees may reinforce correlated errors. We introduce Alignment by Confident Debate (AbCD), which converts debate traces into preference data without gold answer labels. A sampling-diverse committee debates each problem, after which final answers are clustered by task-level semantic equivalence. We use verifiable tasks (mathematical reasoning, multiple-choice quantitative QA, and executable code-generation) as a controlled test bed: their answer keys or executable checkers let us evaluate the quality of debate-derived supervision against ground truth, even though the oracle itself never uses gold labels. Post-training on the resulting preferences outperforms inference-only ensembling and debate, as well as SFT on the same traces, across five model families spanning 1B to 8B parameters and four datasets, improving consensus accuracy, with a maximum gain of $21.5_{pp}$. Ablations show that the approach transfers to a larger model and to code generation, while also revealing limitations of heterogeneous committees and agreement-based weighting. Overall, the results suggest that debate traces can yield useful preference supervision when answer equivalence is combined with confidence-weighted disagreement.

AI sycophancy, the tendency of large language models (LLMs) to flatter users and affirm their beliefs, can undermine accuracy and contribute to misinformation, dependence, and overconfidence. Although prior work has benchmarked sycophancy in released models and proposed mitigations, its origins during training remain poorly understood. Preference alignment methods such as RLHF and DPO are often blamed, but their effects are rarely compared with earlier training stages. We investigate sycophancy across the training pipelines of three model families and three benchmarks. We find that sycophancy is largely established by the end of supervised fine-tuning (SFT), rather than introduced by preference alignment. Across all nine model benchmark combinations, SFT checkpoints attain at least 92% of preference aligned models' sycophancy scores and at least 85% of final models' scores; on average, they are about as sycophantic as final models. Preference alignment increases sycophancy in only four of nine cases, by an average of 4.2 percentage points when it does. To probe the underlying causes, we isolate model representations associated with sycophancy using mechanistic interpretability and measure their activation on training datasets from each stage. SFT datasets show roughly the same average activation as other training data but contribute substantially more examples, and the highest-scoring individual datasets all come from SFT. Our findings suggest that addressing sycophancy requires interventions throughout the training pipeline, particularly in data selection and augmentation for pre-training and SFT.

Fact-checking explanations help Web users understand why a claim receives its verdict, yet the same explanation may not satisfy readers with different backgrounds and preferences. We formulate Faithful Personalized Fact-Checking Explanation Rewriting, which adapts an existing explanation to individual readers while preserving factual fidelity. Rather than relying on reader-written reference explanations, we learn personalized rewriting from the relative preferences of LLM-simulated readers. However, this presents two challenges: distinguishing persona-specific preferences from general explanation quality and learning beyond a limited set of predefined explanation styles. To address these challenges, we propose Personalized Explanation Rewriting via Preference Learning (PERPL), a framework consisting of two components. Simulated Preference Construction derives persona-specific supervision from simulated reader preferences through cross-persona reciprocal pairing. Progressive Preference Learning first trains a lightweight rewriter on the constructed supervision and then refines it through exploration-guided preference optimization, enabling it to discover preferred explanations beyond the predefined styles. Extensive experiments demonstrate that PERPL outperforms much larger general-purpose LLMs in personalized preference alignment while maintaining high factual fidelity.

Interactive dialogue games test a capability that static benchmarks largely leave implicit: a model must carry state across turns, interpret feedback, and choose valid actions under changing constraints. We study this setting in the LM Playschool Challenge with a 2B open-weight model, and find that many failures are not only broad knowledge failures but also local decision failures: repeated guesses, malformed actions, and violations of feedback that the model has just seen. These diagnostics motivate a training recipe organized around three steps: acquire broad game participation through supervised fine-tuning, repair mechanically verifiable failures within one targeted dialogue-game family using turn-local preference pairs, and preserve general capabilities beyond these dialogue games. In the official final evaluation, our submission improves public clemscore from 10.67 to 38.92 and closed in-domain score from 13.41 to 41.17, while approximately preserving aggregate static performance (44.14 vs. 44.24 for the baseline). Out-of-domain clemscore remains low at 7.88, with the largest gains concentrated in unseen variants of the targeted family. Our results suggest that broad SFT brings most of the model's capability improvement; turn-local supervision can be effective when failure detection is precise, with observed transfer concentrated primarily within-family.

Smart home assistants interpret a wide range of user commands, from explicit device control to underspecified and preference dependent requests. While recent systems based on Large Language Models (LLMs) improve this capability, they often rely on heavyweight reasoning pipelines and cloud-based deployment, limiting their efficiency and suitability for resource-constrained environments, and raising privacy concerns. In addition, existing approaches provide limited support for stable long-term personalization. To address these issues, we present AdaHome, an adaptive smart home assistant designed for locally deployed small language models in smart home environments. Rather than applying complex reasoning uniformly, AdaHome introduces an intent-aware planning framework that dynamically routes commands either to straightforward prompt-based or lightweight reasoning-based components. For commands requiring interpretation, we adopt a Chain-of-Draft strategy to enable efficient and stable decision-making. To support personalization, we further propose a preference adaptation mechanism that learns from user feedback over time without requiring prompt augmentation or model retraining. We evaluate AdaHome against representative LLM-based baselines under a unified small model setting. AdaHome achieves substantially higher accuracy on direct commands (86.7%) while reducing latency by up to 3$\times$. Furthermore, it maintains competitive performance on ambiguous inputs with lower computational cost. In multi-turn scenarios, AdaHome achieves 88% preference consistency, compared to 52.5% for a prompt augmentation baseline.

Reinforcement learning with evaluation metrics as rewards is widely used to enhance specific capabilities of language models. However, for tasks such as factually consistent summarisation, existing metrics remain underdeveloped, limiting their effectiveness as signals for shaping model behaviour.While individual factuality metrics are unreliable, their combination can more effectively capture diverse factual errors. We leverage this insight to introduce an automated training pipeline that improves factual consistency in summaries by aggregating scores from different weak metrics. Our approach avoids the need for complex reward shaping by mapping scores to preferences and filtering out cases with high disagreement between metrics. For each source document, we generate lexically similar summary pairs by varying decoding strategies, enabling the model to learn from factual differences caused by subtle lexical differences. This approach constructs a high-quality preference dataset using only source documents.Experiments demonstrate consistent factuality gains across models, ranging from early encoder-decoder architectures to modern large language models, with smaller models reaching comparable factuality to larger ones.

Learning from preference-based feedback has become an effective approach for aligning LLMs across diverse tasks. However, high-quality human-annotated preference data remains expensive and scarce. Existing methods address this challenge through either self-rewarding, which scales by using purely AI-generated labels but risks unreliability, or active learning, which ensures quality through oracle annotation but cannot fully leverage unlabeled data. In this paper, we present CoAct, a novel framework that synergistically combines self-rewarding and active learning through strategic human-AI collaboration. CoAct leverages self-consistency to identify both reliable self-labeled data and samples that require oracle verification. Additionally, oracle feedback guides the model to generate new instructions within its solvable capability. Evaluated on three reasoning benchmarks across two model families, CoAct achieves average improvements of +13.25% on GSM8K, +8.19% on MATH, and +13.16% on WebInstruct, consistently outperforming all baselines.

Reinforcement Learning from Human Feedback (RLHF), using algorithms like Proximal Policy Optimization (PPO), aligns Large Language Models (LLMs) with human values but is costly and unstable. Alternatives have been proposed to replace PPO or integrate Supervised Fine-Tuning (SFT) and contrastive learning for direct fine-tuning and value alignment. However, these methods still require voluminous data to learn preferences and may weaken the generalization ability of LLMs. To further enhance alignment efficiency and performance while mitigating the loss of generalization ability, this paper introduces Distribution-guided Efficient Fine-Tuning (DEFT), an efficient alignment framework incorporating data filtering and distributional guidance by calculating the differential distribution reward based on the output distribution of language model and the discrepancy distribution of preference data. A small yet high-quality subset is filtered from the raw data using a differential distribution reward, which is then incorporated into existing alignment methods to guide the model's output distribution. Experimental results demonstrate that the methods enhanced by DEFT outperform the original methods in both alignment capability and generalization ability, with significantly reduced training time.

Current Large Language Model (LLM) preference learning methods such as Proximal Policy Optimization and Direct Preference Optimization learn from direct rankings or numerical ratings of model outputs, these rankings are subjective, and a single numerical rating chosen directly by a judge is a poor proxy for the quality of natural language, we introduce the What Is Missing (WIM) rating system to produce rankings from natural-language feedback, WIM integrates into existing training pipelines, can be combined with other rating techniques, and can be used as input to any preference learning method without changing the learning algorithm, to compute a WIM rating, a human or LLM judge writes feedback describing what the model output is missing, we embed the output and the feedback with a sentence embedding model and compute the cosine similarity between the resulting vectors, we empirically observe that, compared to discrete numerical ratings, WIM yields fewer ties and larger rating deltas, which improves the availability of a learning signal in pairwise preference data, we use interpretable in the following limited sense: for each scalar rating, we can inspect the judge's missing-information text that produced it, enabling qualitative debugging of the preference labels.

Aligning large language models (LLMs) with human preferences is critical to recent advances in generative artificial intelligence. Reinforcement learning from human feedback (RLHF) is widely applied to achieve this objective. A key step in RLHF is to learn the reward function from human feedback. However, human feedback is costly and time-consuming, making it essential to collect high-quality conversation data for human teachers to label. Additionally, different human teachers have different levels of expertise. It is thus critical to query the most appropriate teacher for their opinions. In this paper, we use reinforcement learning (RL) to formulate the alignment problem. Motivated by the idea of $D$-optimal design, we first propose a dual active reward learning algorithm for the simultaneous selection of conversations and teachers. Next, we apply pessimistic RL to solve the alignment problem, based on the learned reward estimator. Theoretically, we show that the reward estimator obtained through our proposed adaptive selection strategy achieves minimal generalized variance asymptotically, and prove that the sub-optimality of our pessimistic policy scales as $O(1/\sqrt{T})$ with a given sample budget $T$. Through simulations and experiments on LLMs, including annotator-level preference data and downstream Best-of-$N$ response-selection evaluations, we demonstrate the effectiveness of our algorithm and its superiority over existing baselines.