Direct Preference Optimization

Also known as DPO

Overview

Direct Preference Optimization (DPO) studies how to align a model with preferences by training it to favor chosen responses over rejected ones, usually relative to a reference model. Research extends the basic pairwise objective with adaptive margins and weighting, token- or segment-level credit, methods for noisy or conflicting labels, and online or curriculum-based data selection. DPO is used for language-model alignment in helpfulness, safety, and reasoning, and is increasingly adapted to multimodal models and diffusion-based image, video, and audio generation. A central tension is improving preference alignment and robustness without sacrificing general capabilities, diversity, or reliable performance when preferences and inputs shift.

Latest papers

174 papers23 papers in the last four weeks, up 28% on the four weeks before. 0.2% of all new papers.

Nov 8, 2026

TLRS-DPO: Token-Level Risk Scoring for Better Direct Preference Optimization

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.
Nov 6, 2026

On the Reliability of Majority-Vote Pseudo-Labels in Test-Time Reinforcement Learning

Test-time reinforcement learning (TTRL) adapts pretrained language models to unlabeled target-domain data using self-generated pseudo-labels, whose reliability remains poorly understood. We analyze majority-vote supervision for TTRL with online Direct Preference Optimization (DPO) and decompose its error into population-majority bias and prompt-dependent finite-rollout variation. Because population-majority bias is generally unidentifiable without external supervision, we focus on mitigating the downstream effects of finite-rollout variation. We propose Prompt-Conditional Distributionally Robust DPO (PC-Dr.DPO), which conditions within-prompt robust reweighting on observed vote reliability to make updates from sampling-sensitive prompts more conservative. We further introduce a cache-based mechanism to reduce self-supervised forgetting during online adaptation. Across six mathematical and knowledge-intensive reasoning benchmarks and five model backbones, PC-Dr.DPO matches or outperforms the strongest TTRL baseline in 24 of 30 settings.
Nov 5, 2026

When Scores Conflict with Preferences: Pre-Alignment Diagnosis and Adaptive Correction of Constraint Drift in DPO

Direct Preference Optimization (DPO) is increasingly used in settings where pairwise preferences coexist with pointwise criteria such as response length, answer correctness, or safety scores. When the two disagree, standard DPO treats pairs whose preferred output scores worse on the pointwise criterion (conflict pairs) in the same way as the remaining (matching) pairs. We characterize a systematic failure mode, constraint drift: for sufficiently small updates, conflict pairs tend to increase constraint cost while matching pairs tend to repair it, and the net change depends on their balance. This balance yields a closed-form pre-alignment diagnostic, ρeq∗ρ^{*}_{\mathrm{eq}}, computable from dataset statistics, that indicates whether a dataset is self-correcting or requires intervention. Building on this diagnostic, we propose Conflict-aware DPO (CoDPO), which down-weights conflict pairs with a strength set from ρeq∗ρ^{*}_{\mathrm{eq}} rather than tuned, and adds an adaptive margin shift that addresses residual drift during training. Empirically, drift grows monotonically with conflict exposure on the conflict-dominated domains, and the intervention decision given by ρeq∗ρ^{*}_{\mathrm{eq}} matches the observed drift on every domain, including one that it identifies as self-correcting. On all four domains of the main comparison, CoDPO reduces drift while largely preserving the preference learning of vanilla DPO. Compared with hard filtering and recent constrained and multi-objective baselines, CoDPO also achieves a more favorable drift-preference trade-off.
Nov 4, 2026

The Alignment Tax Is Graded: Task-Selective Capability Loss Under Length-Normalized DPO with SFT Retention

Setting. This is an empirical characterization study; we propose no new method and run no comparisons against alternative objectives. We fine-tune Qwen3.5-4B with sparse LoRA and length-normalized DPO plus an SFT retention term (the RPO objective) on structured financial-filing tasks, and track downstream capability as the implicit reward margin grows to ∼7\sim 7 ββ-scaled units (nf4 training; ∼3\sim 3 under fp16 measurement on the unquantized base). Question. Does margin growth degrade capability uniformly across tasks, or selectively? Findings. In a pre-registered study (8 seeds, 6 unseen at freeze) with artifact-aware scoring and TOST equivalence testing, the alignment tax is graded and task-selective: extraction is preserved (Δ=−0.55Δ= -0.55 pts, equivalent at ±5\pm5 pts), json loses little (Δ=−3.52Δ= -3.52 pts, equivalence fails on a clustered item-level interval), and router degrades severely (Δ=−30Δ= -30 pts). Router's loss concentrates in a margin 1-3 damage window containing a transient format-break spike. The margin is accounted for almost entirely (≥96%\geq 96\%) by suppression of rejected completions, with chosen reward approximately flat, consistent with the retention term anchoring it; we do not attribute the behavioral damage to any single gradient channel. Gold-logp deflation is task-selective (extraction −0.15-0.15, json −1.04-1.04, router −1.47-1.47 nats/token under fp16), with gold selectively protected relative to incorrect controls on extraction (8/8 seeds). QLoRA reward margins differ up to 3.4×3.4\times between nf4 training and unquantized fp16 measurement, with an axis-dependent gap. Scope. Length-normalized rewards with SFT retention, sparse LoRA on 8/32 layers, one 4B model, easily-separable synthetic preferences, three structured tasks. A no-retention control and a measured format-alignment metric are left to future work.
Nov 3, 2026

Objective-Conditioned Active Preference Learning for Multi-Objective Alignment

Aligning large language models (LLMs) with several objectives, such as helpfulness, harmlessness, and honesty, typically relies on preference judgment for each objective, growing the annotation cost linearly with the number of objectives. Active preference learning can reduce this cost by selecting the comparisons whose labels are expected to change the policy most. Existing remedies either lack an objective-aware query selection mechanism for the multi-objective setting or regard conflicting preferences as harmful, therefore excluding them from the training data. We introduce OGC-DPO (Objective-conditioned Gradient Coverage DPO), an active learning framework that treats each (response pair, objective) combination as a separate query and measures its informativeness through the policy's own implicit gradient, without relying on an external reward model. OGC-DPO then selects the queries that most expand each objective's gradient coverage. Empirical experiments across varying numbers of objectives demonstrate the effectiveness of OGC-DPO over existing active preference learning methods under the same annotation cost. Further analysis demonstrates that informative conflicts, rather than being uniformly harmful, provide a useful policy training signal.
Nov 1, 2026

TopoDPO: Dense Preference Alignment for Structure-Sensitive Aerial Referring Segmentation

Preference optimization for multimodal segmentation faces a granularity mismatch: dense masks contain spatially heterogeneous preference signals, whereas image-level objectives assign one shared nonlinear weight to all spatial margins. We introduce TopoDPO, a topology-aware dense preference alignment framework for structure-sensitive aerial referring segmentation. Topo-Prior mines informative preference pairs using connected-component discrepancy and macro-corner recall, while the dense objective applies the preference penalty to each local margin before spatial aggregation against a frozen supervised reference. On Topo-RS-Hard, a 4590-instance structural stress set, TopoDPO improves intersection over union (IoU) from 50.91% to 51.24% over image-level direct preference optimization (DPO) and reduces Betti-0 error from 1.34 to 0.90; paired image-cluster bootstrap intervals against image-level DPO exclude zero for all five reported metrics. Matched supervised controls and a complete factorial analysis separate the effects of topology-aware pair mining, localized correction, and dense reference-relative preference weighting. An unfiltered stratification over all 39,244 validation records further shows the clearest structural gains on samples with moderate or severe pre-alignment component-count errors. The code is publicly available at: https://github.com/phy709/TopoDPO.
Oct 30, 2026

SecPO: Principled Adversarial Training for Prompt Injection Security

Recognized as a top threat to AI agents, prompt injection allows instructions from untrusted data (email, documents, websites) to manipulate the agent's operation. The existing best-defended LLM, Meta-SecAlign, is trained with the DPO loss but still suffers from attack success rates (ASRs) approaching 100% under adaptive prompt injections. One reason is that DPO anchors the model around the undefended reference model, which is insecure. We show an idealized reference model---an injection-invariant, utility-preserving oracle---can nevertheless be approximated, as its output distribution in training is interestingly accessible. This new observation motivates a theoretically principled SecPO loss. Another Meta-SecAlign shortcoming is that it trains on weak simulated attacks. We implement an efficient adversarial training (AT) pipeline that synthesizes strong injected inputs (best@240), and train the model on them. Our AT reduces ASRs by 1.5-14.5×\times across three SoTA adaptive attacks, compared with Meta-SecAlign; the SecPO loss reduces them by a further 2.4-4.6×\times compared with using DPO loss. In agentic workflows, our SecPO + AT Qwen3.6-27B is much more secure than Meta-SecAlign: 8.8% vs. 78.8% AgentDojo ASR@800 under the AttackerMovesSecond attack. Moreover, our defended model shows no noticeable reduction in benign utility in nine benchmarks covering agentic tool calling, instruction following, science, and math. We release our code, training data (with optimized injections from AT), and model.
Oct 30, 2026

Reliability-Weighted Direct Preference Optimization: Penalized Weights for Robust Preference Learning

Direct Preference Optimization (DPO) treats every reported comparison as correct, yet preference data contain reversed, tied or otherwise noisy comparisons. RW-DPO gives each comparison an unknown reliability weight wi∈[0,1]w_i\in[0,1] that enters the likelihood linearly -- DPO is w≡1w\equiv1, a reversed comparison wi=0w_i=0 -- and estimates the weights jointly with the policy under a minimax-concave penalty on −log⁡wi-\log w_i. The penalty keeps every comparison the model agrees with at exactly wi=1w_i=1 and acts only on those it contradicts, through closed-form weight updates at negligible cost over DPO. We prove that recovering all flipped comparisons from a single annotation per comparison requires margins of order log⁡n\log n. Furthermore, in a linear-index model, starting from an initial fit pointing in the correct direction, our estimator jointly consistently recovers the preference direction and weights, as well as identifies the flipped comparisons above this margin scale, even when a fixed fraction of comparisons is flipped. In simulations RW-DPO has the lowest average estimation error among DPO, noise-rate corrections and recent robust methods, with far more precise flags than recent label-correction methods. In DPO fine-tuning of a language model with 10-30% flipped labels, its win rate does not fall as the flip rate grows and exceeds that of DPO trained on clean labels, matching the strongest recent method with corrections concentrated on the flipped labels.
Oct 27, 2026

Aligning Large Language Models for Safety in African Languages: A Case Study of Hausa

Safety alignment research for large language models (LLMs) remains concentrated on high-resource languages, leaving open how existing alignment methods perform in low-resource, culturally distinct settings. We study this question for Hausa, a language spoken by more than 200 million people across West Africa. We introduce HausaSafeLM, one of the first Hausa safety preference dataset, built from 10,000 prompts derived from PKU-SafeRLHF: 3,000 human-corrected (gold) and 7,000 machine-translated (silver) examples spanning six harm macro-categories. Using Qwen2.5-7B-Instruct, we compare supervised fine-tuning (SFT) and SFT followed by Direct Preference Optimization (DPO) under both data regimes, evaluating on 500 harmful and 200 benign Hausa prompts.We find that the untuned base model appears highly safe by refusal-rate metrics alone, but this reflects limited Hausa generation ability rather than genuine alignment: its refusals are almost uniformly low quality. SFT improves both refusal fluency and precision and benefits substantially from the larger, noisier silver-tier data, while DPO sharpens refusal precision but comes at a steep, largely data-invariant cost to linguistic quality. No single configuration jointly maximizes safety, helpfulness, and fluency. Our results suggest that current alignment techniques teach models when to refuse more readily than how to do so fluently in underrepresented languages, and that refusal-rate metrics alone are an unreliable signal of safety in low-resource settings.
Oct 27, 2026

On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK Regime

Direct Preference Optimization (DPO) has been widely adopted for aligning language models with human preferences. Recent empirical work has revealed that DPO induces remarkably sparse parameter updates, yet no theoretical explanation for this phenomenon has been offered. We analyze DPO's optimization dynamics under a neural tangent kernel linearization and show that the logistic structure of the DPO loss gives rise to an implicit maximum-margin bias. Building on this observation, we introduce the concept of dual sparsity: the DPO update direction is a sparse linear combination (data-level sparsity, driven by support vectors) of sparse gradient feature vectors (parameter-level sparsity, arising from gradient cancellation between preferred and dispreferred responses). We formalize this structure in our main theorem, bound the number of support pairs by a margin-dependent spectral quantity, and present a framework relating dual sparsity to weight disentanglement---a condition favorable for model merging. Experiments on GPT-2 Small and Medium and on 7--9B instruction-tuned models (Llama-3-8B, Mistral-7B, Gemma-2-9B) with the Anthropic HH-RLHF and UltraFeedback datasets confirm both sources of sparsity, and experiments on GPT-2 show that DPO reduces cross-task interference by up to 35×35{\times} compared to supervised fine-tuning.
Oct 27, 2026

Exploring the Limits of the Delta Learning Hypothesis: Single-Pair Generator Biases in Direct Preference Optimization

Direct Preference Optimization (DPO) typically relies on preference pairs generated from diverse model pools and ranked by an external LLM judge. The recently proposed delta-learning hypothesis suggests that comparable supervision can instead be obtained from a single pair of models by treating the larger model's response as preferred. We investigate the limits of this approach across model families and scales. We show, both theoretically and empirically, that fixed model pairs introduce systematic, generator-dependent biases that are predictive of the preference label independent of response quality. Consequently, DPO trained on single-pair data consistently underperforms multi-model preference data, despite the larger generator producing higher-quality responses, and we trace this gap to training dynamics consistent with reward hacking on these biases. We further show that simply increasing the number of generators, even without an external judge, is enough to partially mitigate this effect. Our results highlight generator diversity as a critical factor in preference-data construction.
Oct 23, 2026

Preference Optimization Improves Definition Generation: Semantic Change in EU Institutional Sustainable Finance Disclosures

Dictionary-like definition generators are known to be useful for lexical semantic change detection (LSCD). However, they suffer from overly granular sense distribution, producing excessive number of definitions for one and the same lexical sense. In this work, we show that fine-tuning with Direct Preference Optimization (DPO) helps to overcome this problem. We test our approach both quantitatively on a curated academic LSCD benchmark and qualitatively on real-world data for a practical task of auditing diachronic changes in sustainability terminology within project documentation of EU development finance. All models, code and predictions produced in this study are publicly available at https://github.com/ltgoslo/dpo-for-definition-generation.
Oct 22, 2026

Aligning Large Language Models with Human Feedback: Mathematical Foundations and Algorithm Design

This article provides an introduction to the mathematical foundations and algorithmic frameworks used to align Large Language Models (LLMs) with human intentions, preferences, and values. We discuss standard alignment techniques, such as fine-tuning (SFT), reinforcement learning with human feedback (RLHF), and direct preference optimization (DPO). We also explore the theoretical underpinnings of learning from human preferences, drawing connections to inverse reinforcement learning (IRL) and discrete choice models. We present state-of-the-art algorithms in a tutorial style, discuss their advantages and limitations, and offer insights into practical implementation. Our exposition is intended to serve as a comprehensive resource for researchers and practitioners, providing both a foundational understanding of alignment methodologies and a framework for developing more robust and scalable alignment techniques.
Oct 21, 2026

Stochastic Pruning of the DPO Computation Graph: 1.4x Faster Training and Lower Peak Memory

Direct preference optimization (DPO) incurs training costs in output scoring, sequence execution, and backpropagation, so reducing supervision alone does not necessarily save computation. We introduce a unified framework for stochastic pruning of the DPO computation graph that independently samples response positions for scoring, response prefixes for execution, and preference branches for backpropagation. By making these selections before the corresponding operations, the framework omits computation at each stage of a training step. Our analysis characterizes the objective changes induced by position and sequence pruning and uses inverse-probability weighting to obtain a conditionally unbiased single-branch gradient estimator for the sampled objective. Controlled measurements show that position and sequence pruning primarily reduce peak memory, with most output-memory savings attributable to an implementation that retains full response supervision. Branch pruning yields a 1.426×1.426\times step-time speedup over a concatenated DPO baseline and reduces training time by about 30% relative to stock DPO across two model families, with slightly lower mean accuracy and benchmark-dependent changes in open-ended scores. These findings relate the resource benefits of stochastic graph pruning to the computation omitted and the resulting changes in the training objective.
Oct 21, 2026

ECHO: Emergent Common-Mode Hidden Offsets Drive Vocabulary Collapse in Preference Optimization

Preference optimization can improve every ranking metric while building a hidden, shared representation shift that the relative loss barely sees, one that can eventually destroy generation. We trace this failure to the common-mode offset, a shared directional shift in the model's hidden states that appears across token positions, prompts, and both chosen and rejected responses. Because it moves chosen and rejected representations together, the relative ranking loss barely reacts to its accumulation. As this offset grows, it projects onto the output layer to form a vocabulary well, elevating a single token's probability until it dominates decoding and causes generation collapse. The isolated offset alone reconstructs the vocabulary well, injecting the learned offset into an untreated base model recreates the collapse where norm-matched random directions do not, and masking this common component from the loss prevents the well in our flagship setting. Under the per-token ranking loss we study, the same collapse appears in 14 of 17 base-model lineages and under full-parameter fine-tuning. Moreover, the offset is not merely an artifact of collapse, since it is already present in textbook DPO and SimPO runs that form no well. Textbook DPO keeps it weakly aligned with any single vocabulary direction even with three times our base budget. Trained from base models for that budget, however, full-parameter SimPO drives it into collapse on both Llama-2 and Llama-3.1, and the same kind of offset grows, without a well, from a public SFT checkpoint under the official SimPO recipe. Across 17 UltraFeedback runs, the better a run ranks held-out responses, the less often two public reward models prefer its samples to the base model's. These findings expose a blind spot in standard ranking metrics and introduce a representation-level signal that can warn of generation collapse before it occurs.
Oct 20, 2026

Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training

People adapt how they talk to the person in front of them. However, chat models tend to respond in Standard English regardless of the dialect used by the user. A uniform answer can hide whether training treats these users alike, and output audits cannot tell as well. We trace how each post-training stage, from supervised fine-tuning (SFT) to preference optimization (DPO) and reinforcement learning, changes a model's tendency to match the user, and read what each update prefers from two content-matched replies, one matching the user and one not. Across 10 open model lineages, post-training standardizes replies to every informal cue, yet after adjusting for how likely each reply already was, the DPO update still favours matching British and Australian spelling over AAE and Singlish on 10 lineages, and models still produce AAE features when asked. We then introduce calibrated accommodation pairs (CAP), preference data that reward matching users' dialect density while preserving Standard English for Standard-English users. On a fully open model, CAP raises convergence on features it was not trained on without a detectable increase in dialect for Standard-English users or a cost in instruction following. In contrast, one-sided pairs reach a similar gain only by overshooting the user and imposing dialect. Our findings show that uniform replies can hide selective updates and audits should read the update as well as the output. Code: https://github.com/kouzhizhuo/cap-accommodation.
Oct 20, 2026

TPO: Tri-level Distributionally Robust Learning for OOD Direct Preference Optimization

Direct Preference Optimization (DPO) minimizes empirical risk over observed preference comparisons, implicitly assuming that the training preference distribution matches the deployment preference distribution. In practice, this distributional assumption often fails, as preference distribution shifts can arise at three levels: group-level shift, prompt-level shift, and preference-level shift. Such multi-level distributional mismatch poses a critical challenge for safe LLM deployment, yet existing robust DPO methods typically either model uncertainty over the joint preference distribution or focus on individual sources of shift, leaving the hierarchical structure of OOD preference shifts underexplored. We propose TPO, a tri-level distributionally robust framework for OOD preference alignment, which jointly models all three levels of distributional uncertainty through a hierarchically structured rectangular ambiguity set. By instantiating it with Wasserstein distance, we derive a tractable and interpretable objective that combines a preference-level margin penalty, prompt-level adversarial embedding perturbation, and worst-case group reweighting, each governed by an independent radius. We establish finite-sample convergence guarantees and empirically validate the effectiveness of TPO across diverse OOD environments.
Oct 20, 2026

The Shape of the DPO Hyperparameter Landscape: What to Measure, How It Moves with Scale, and When Not to Transfer It

Direct preference optimisation (DPO) is widely used for language model post-training, but its performance depends strongly on the choice of its hyperparameters, such as the learning rate and the reference-policy regularisation parameter ββ. To better understand the DPO hyperparameter landscape, we conduct extensive experiments using a dense hyperparameter grid across different model scales from the same family. This allows us to analyse how the optimal DPO hyperparameters evolve with model size and the number of update steps, with intermediate checkpoints providing an inexpensive signal to accelerate the hyperparameter optimisation process. Furthermore, by evaluating our models on both static and dynamic benchmarks, we identify which benchmarks provide the strongest signal for hyperparameter selection. Finally, we investigate how different state-of-the-art hyperparameter optimisation strategies perform in this setting. Our experiments show that the optimal learning rate varies with model size and that hyperparameters are not transferable across training datasets. We also find that validation loss is a poor proxy for hyperparameter optimisation, as hyperparameters that minimize validation loss only weakly correlate with optimal hyperparameters for downstream benchmarks. Therefore, DPO hyperparameters need to be tuned specifically for the target model, dataset, and evaluation objective. We also show that short candidate training runs can substantially reduce the cost of this selection. We release all our training configurations and evaluation results to support reproducible analysis of selection quality and training cost.
Oct 19, 2026

Statistical Learning Theory for Direct Preference Optimization

Direct Preference Optimization (DPO) learns from pairwise preferences through a logistic objective whose score-difference parameterization follows the Bradley--Terry model. We study how failure of this modeling assumption limits DPO, distinguishing structural misspecification from limited model capacity and finite-sample estimation. Our main contribution is a characterization of structural approximation error through the triangle-curl of the preference log-odds, which vanishes exactly when preferences admit a Bradley--Terry representation. We upper bound the achievable excess logistic risk by a score-class approximation term and a triangle-curl term, and establish a lower bound proportional to triangle-curl under additional boundedness assumptions. For ReLU score models, including clipped models, we derive explicit approximation rates showing that increasing capacity can eliminate the score-approximation term while leaving the structural obstruction unchanged. We further establish finite-sample bounds for clipped empirical risk minimizers that separate network approximation, structural misspecification, and estimation error. We place these results within a surrogate-risk framework by identifying contextual preference accuracy (CPA) as a natural measure of agreement between preference judgments and changes in pairwise probability ratios relative to the reference policy. Standard logistic consistency and calibration results connect DPO's objective to this policy property and transfer our excess-risk upper bounds to CPA, without treating CPA as a complete measure of generated-response quality.
Oct 18, 2026

Over-Defense in Coding Agents Begins at Silent Forks

Language models behind coding agents are post-trained with outcome rewards, and recent methods concentrate credit on the turns that decide the outcome. We find that defensive actions beyond what a task requires escape both. They begin at silent forks, the turns where every next action leads to the same outcome. Such actions leave the success rate unchanged, and frontier models take the disproportionate branch there in up to 91% of decisive actions under implicit budgets. The most direct fix, a checker in the trajectory reward, reaches these forks only through fully proportionate trajectories, leaving 74% of GRPO groups with identical rewards and behavior unchanged. This motivates a rethink of where learning signals belong, namely where behavior diverges, which a behavioral checker can locate and the outcome cannot. Our method, ForkDPO, contrasts a proportionate and a disproportionate action at the same prefix, at silent forks on both sides of the required behavior. On a test set of unseen repositories, it lowers a 4B model's disproportionate behavior from 70.3% to 58.9%, against 67.7% for DPO at randomly chosen turns, establishing behavioral divergence as a learning signal for how agents act beyond whether they finish. The code is publicly available at https://github.com/Astarojth/silent-forks.
Oct 18, 2026

Robust Direct Preference Optimization under Preference Label-Flip Attacks

Direct Preference Optimization (DPO) is widely used for offline Reinforcement Learning from Human Feedback. However, recent works have shown that DPO can be easily manipulated by adversarial preference label flip attacks, which are specifically designed to mislead the learner toward an attacker-chosen target policy. To address this issue, we design robust DPO methods that can defend against such attacks. Building on the property that lable flips induce parameter-independent gradient shifts in log-linear DPO, we propose Trim-KK Robust-DPO, which removes the KK highest-impact pairs, and Soft-Trim-KK Robust-DPO, which smoothly downweights them. We establish deterministic gradient perturbation bounds uniformly over all attack sets containing at most KK flips, explicitly separating defense-induced bias from surviving attack effects. Under bounded-region assumptions, these bounds yield time-uniform parameter and policy deviation guarantees. We further quantify how quadratic regularization improves curvature while introducing additional bias. Experiments on Stanford Human Preferences, HH-RLHF, and UltraFeedback-Binarized validate our analysis.
Oct 14, 2026

Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge

Brazilian Portuguese remains underserved by open-source language models, and the few that are safety-aligned rarely document what did not work or validate comparisons with statistical significance. We present Manacá-1B-Instruct-2, the second-generation post-training alignment of Manacá-1B (1.72B parameters, trained from scratch for PT-BR), with an SFT corpus of 291,223 examples. We document three DPO attempts, all negative, including a case where the training reward margin reaches 1.0 while free generation does not change at all, and show that safety came from a dedicated safety-SFT (LLM-jp AnswerCarefully style), refined over three iterations until closing the safety-helpfulness frontier. A controlled scan of the fraction of general data in the safety-SFT mix (10% to 60%) reveals a non-monotonic over-refusal curve with a knee point at 40-50%. The final model is compared against four external PT-BR models on the same items, with Wilson confidence intervals and paired McNemar tests with Holm-Bonferroni correction: the lower bound of its safety interval (94.1%) lies above the upper bound of all four baselines. We also report a methodological finding for LLM-judge evaluation: on two of six safety items in MT-Bench-PT, the judge categorically refuses to assign a score regardless of the evaluated response, silently reducing the effective N of that category in every prior report of the project. We complement this with an n-gram decontamination check and a Chinchilla-anchored compute-budget table that attributes the gap in general capability to pre-training scale rather than alignment methodology. All code, raw data, logs, and scripts that produce every reported number are publicly released.
Oct 13, 2026

DPOison-VL: Clean Preference Poisoning in Vision-Language Direct Preference Optimization

Direct Preference Optimization (DPO) has become a widely adopted approach for aligning vision-language models (VLMs), yet the security of multimodal preference data remains largely unexplored. Existing preference poisoning attacks either corrupt explicit supervision signals or require visible triggers, making them detectable under human inspection. We introduce DPOison-VL, the first clean preference poisoning attack against VLM-DPO, which manipulates model alignment solely through imperceptible perturbations on a small fraction of training images, while leaving all prompts, responses, and preference annotations semantically intact. We formulate poison construction as a bilevel optimization and derive a gradient alignment surrogate that efficiently steers image perturbations toward the attacker-specified preference. We further introduce a preference displacement score for principled poison selection with provable first-order optimality under a fixed poison budget, and derive a sufficient condition for preference inversion. Extensive experiments across diverse attack tasks, different VLM architectures and scales, and baseline comparisons demonstrate that DPOison-VL achieves up to 98% ASR with only 2% poisoned data and exhibits cross-model transferability, while largely preserving clean accuracy and general utility. Our findings reveal a critical vulnerability in VLM-DPO, where imperceptible image perturbations in preference training data can silently redirect alignment.
Oct 9, 2026

Personalized Federated Direct Preference Optimization via LoRA Adapters

Federated Direct Preference Optimization (FedDPO) enables privacy-preserving alignment of language models across distributed agents without centralizing sensitive preference data. However, existing FedDPO analysis assumes all clients converge to a single shared policy, which is fundamentally at odds with the heterogeneous preference distributions present in practice. We propose Personalized FedDPO (pFedDPO), in which each client maintains a small private LoRA adapter atop a shared global base model. In pFedDPO, the base model is synchronized via FedAvg each round, while LoRA adapters remain local and capture client-specific preference patterns. We prove that this decomposition replaces the gradient diversity κ2κ^2 in the FedDPO bound with a residual heterogeneity κeff2κ^2_{\mathrm{eff}}, which is no larger than κ2κ^2 when adapters absorb client-specific directions, at the cost of an additive adapter-bias term that vanishes as adapters converge. Experiments on preference datasets show that pFedDPO consistently attains a lower gradient norm than vanilla FedDPO while remaining competitive in training loss and per-client reward accuracy.
Oct 6, 2026

CM-DPO: Constraint-Margin Direct Preference Optimization for LLM Planning

Direct Preference Optimization (DPO) treats all constraint violations equally: a 1budgetovershootanda1 budget overshoot and a 1,000 overshoot induce the same training signal. It is also susceptible to length and style bias when preference pairs come from different model families. We introduce Constraint-Margin DPO (CM-DPO), which replaces DPO's binary preference signal with a continuous margin derived from a deterministic symbolic verifier and scaled by violation severity. Hard and soft constraints are separated through a lexicographic objective, ensuring hard constraints are never traded off against preferences. To supply CM-DPO with bias-reduced training pairs, we generate preference data through procedurally generated constraint profiles (DCCG) and minimal-edit distillation from a reasoning teacher (RT-MED), within a framework we call SynPlan-R. On TravelPlanner, NaturalPlan, and out-of-distribution PlanBench, an 8B model fine-tuned with CM-DPO achieves 89.2% pass rate and 93.4% solve rate, matching multi-agent systems at 13x lower latency while outperforming GPT-4o on unseen Blocksworld by 9.2 points.
Oct 6, 2026

LSC-DPO: Learning-Signal-Controlled Direct Preference Optimization

Direct Preference Optimization (DPO) has become a standard reward-model-free approach for aligning language models with preference data. However, as the scaled preference margin grows during training, the logistic DPO loss becomes progressively less sensitive to further changes. We study DPO from a loss-level geometric perspective and identify the sigmoid factor as a learning signal that characterizes the local sensitivity of the objective. Based on this view, we propose Learning-Signal-Controlled Direct Preference Optimization (LSC-DPO), which dynamically regulates the learning signal near a target regime. A log-space analysis establishes conditions for stable tracking of the target learning-signal regime. Experiments on AlpacaEval 2, MT-Bench, and Anthropic-HH show that LSC-DPO consistently improves over DPO and strong preference-optimization baselines. We further find that different coefficient initializations induce distinct transient learning-signal trajectories even when their later signal levels become similar. Based on this observation, we derive a signal-budget compensation rule that adjusts the target learning signal to compensate for these transient differences. The resulting compensation substantially reduces performance variation across coefficient initializations.
Oct 4, 2026

Geometry-Aware Preference Optimization for Text-to-Image Diffusion Models

Preference alignment has become a standard practice for text-to-image diffusion models. Direct Preference Optimization (DPO) simplifies this process by eliminating explicit reward modeling. Its diffusion variant, Diffusion-DPO, has become a widely adopted baseline. Diffusion-DPO essentially encourages the likelihood of preferred samples while suppressing dispreferred ones. In this paper, we revisit DPO-style alignment methods for diffusion models from the perspective of the manifold hypothesis. Under this view, natural images concentrate near a low-dimensional manifold embedded in the high-dimensional ambient space, whereas DPO directly optimizes preference distributions in the full space without accounting for this geometric structure. This creates a mismatch in the optimization dynamics: it suppresses geometry-preserving tangential updates, while insufficiently restricting hazardous normal-direction updates. This mismatch gradually degrades image quality and diversity. To address this issue, we propose Anisotropic Geometry-Aware Preference Optimization (APO), which replaces the uniform Euclidean treatment of prediction errors with a geometry-aware anisotropic metric derived from the reference model. Concretely, APO adaptively strengthens regularization in directions where the reference denoising function is highly sensitive, while relaxing constraints in directions that permit safe semantic adjustment. This recalibrates preference optimization according to the local manifold geometry, and maintains the original manifold structure. Experiments show that APO achieves strong performance and an average win rate exceeding 60% against various existing alignment methods across diverse benchmarks. It requires significantly fewer training steps than prior methods, and preserves generation diversity throughout training.
Oct 2, 2026

Hallucination Reduction for LLM-Based Audio Understanding via Multimodal Direct Preference Optimization

Large Audio Language Models (LALMs) are prone to hallucinating and over-relying on text priors when simultaneously presented with audio and text inputs. To mitigate these hallucinations, we propose utilizing the multimodal Direct Preference Optimization (mDPO) objective, which forces the model to ground its generation in the acoustic input by contrasting intact and distorted audio counterparts. We extend this preference learning framework to the audio domain by applying a variety of acoustic perturbations. Evaluating the Qwen2-Audio backbone across the DCASE 2025 Challenge and AH Existence datasets, we demonstrate that extending mDPO to LALMs significantly enhances temporal reasoning in the complex DCASE dataset, and improves performance on basic existence verification in the AH benchmark. We identify temporal reversal, frequency masking, and random noise as the most effective perturbations. Ultimately, our approach achieves an absolute accuracy improvement of 14.0% on the DCASE 2025 dataset and 27.4% on AH Existence.
Oct 1, 2026

GAW-PO: Preference Optimization with Gradient-Aligned Token Weights

Most preference optimization methods, such as Direct Preference Optimization (DPO), apply preference supervision at the response level, although autoregressive language models are optimized token by token. As a result, all tokens in a rejected response contribute to the negative training signal, including tokens that may encode behavior that is useful for the preferred response. We introduce GAW-PO, a gradient-aligned token reweighting method for DPO that estimates, for each rejected token, whether penalizing it would interfere with the preferred update directions. Tokens whose gradients are strongly aligned with the preferred behavior receive a weaker negative contribution, while conflicting tokens retain a stronger penalty. Our method achieves the highest average performance among the evaluated preference-optimization methods, improving by 0.97 points over standard DPO and 0.65 points over the strongest competing baseline across 11 benchmarks spanning mathematics, reasoning, coding, and question answering. We further show that gradient-aligned weighting is substantially more robust to aggressive preference optimization: as the DPO regularization parameter ββ decreases, standard DPO degrades sharply, whereas GAW-PO continues to improve. These results suggest that accounting for the interaction between rejected-token updates and preferred behavior provides an effective form of token-level credit assignment for preference optimization.
Sep 30, 2026

Smaller Models, Better Rejects: Preference Distillation Scaling

Preference distillation typically treats a teacher response as preferred and the student's own response as rejected. This assumes that self-generated failures are the most informative negatives and that rejects must come from a model at least as large as the student, making generation costly at scale. We find neither assumption holds: across students from 7B to 72B, smaller frozen models generate rejects with less inference compute yet train stronger students than self-generated rejects, before and after sequence-level knowledge distillation, on code generation and mathematical reasoning. To explain this result, we derive a finite-horizon utility bound for Direct Preference Optimization in a linearized feature model. The bound characterizes favorable reject distributions and motivates three interventions. First, mixing rejects from smaller and student-scale models improves performance as the smaller model's share increases. Second, reassigning rejects to other prompts and shuffling their code tokens still outperform length-matched gibberish, showing that task structure contributes to reject utility. Third, selecting candidates with lower likelihood under the reference policy improves net transfer when higher-likelihood candidates provide less useful contrast. Lower-likelihood selections outperform higher-likelihood ones for every source. These results suggest that effective rejects preserve task structure while limiting coupling to the reference policy, and that smaller frozen models can provide them at low cost.