Direct Preference Optimization (DPO)
direct-preference-optimization-dpo--4e5a161d·17 events·first seen 1mo agoAliases: Direct Preference Optimization (DPO), Direct Preference Optimisation, Direct Preference Optimization
Co-occurring entities
More like this (12)
Guides (1)
Recent events (17)
Preference Optimization for Vision Language Models
This Hugging Face blog post covers the application of Direct Preference Optimization (DPO) to vision-language models (VLMs). It likely discusses how preference learning techniques originally developed for text-only LLMs can be adapted to multimodal settings. The post addresses training methodology for aligning VLMs with human preferences across both visual and textual modalities.
Preference Tuning LLMs with Direct Preference Optimization Methods
A Hugging Face blog post surveys Direct Preference Optimization (DPO) and related preference tuning methods for aligning large language models. The post covers the landscape of DPO variants and their practical application via the TRL library. It serves as a technical reference for practitioners implementing RLHF alternatives.
Fine-tune Llama 2 with DPO
This Hugging Face blog post provides a practical guide to fine-tuning Llama 2 using Direct Preference Optimization (DPO) via the TRL library. It covers the alignment technique that bypasses the need for a separate reward model compared to RLHF, walking through dataset preparation, training configuration, and implementation details. The post targets practitioners looking to apply preference-based alignment to open-weights models.
Direct Preference Optimization Beyond Chatbots
A Hugging Face blog post explores applications of Direct Preference Optimization (DPO) outside of conversational AI contexts. The post appears to survey or analyze how DPO, a technique for aligning language models with human preferences, can be applied to non-chatbot domains. The body content is unavailable, limiting assessment of specific claims or findings.
AMRS: Rollout-Based World Model for Offline Affective Music Recommendation with DPO
LUCID's Affective Music Recommendation System (AMRS) uses a causal transformer world model trained on logged listening data to jointly predict engagement, ratings, and self-reported valence/arousal, enabling offline policy optimization without ethically problematic online experimentation. A recommender policy is initialized via behavior cloning and fine-tuned with Direct Preference Optimization (DPO) against a multi-objective utility function. The system is deployed on LUCID's health-and-wellness platforms serving clinical users (older adults with neurocognitive conditions) and consumer-wellness users across four modes. Under cold-start conditions, DPO improves predicted affective signals over the cloned baseline while maintaining diversity and avoiding distributional collapse.
LLUMI: Fine-Tuning Open-Source LLMs for Mental Health Writing Assistance Using Reddit Community Feedback
LLUMI is a two-component system (a generation model and an improvement model) designed to provide mental health writing assistance using smaller open-source LLMs hosted in privacy-preserving, on-premise environments. The system leverages Reddit community endorsement signals (upvotes/downvotes) to construct preference pairs for SFT and DPO training, then further aligns outputs via human evaluation across readability, empathy, connection, actionability, and safety dimensions. Results show LLUMI achieves performance comparable to proprietary GPT-based models on linguistic and human evaluations, suggesting community-derived preference signals can substitute for expensive expert labeling in sensitive domains.
Self-improving VLMs can silently regress when verifier quality is task-mismatched
A new arXiv paper demonstrates that verifier-driven self-DPO, a common recipe for self-improving visual-language models, can silently degrade student model performance when the verifier's task-rubric accuracy is insufficient for the target task. Experiments on Qwen-3-VL-2B and Qwen-2.5-VL-3B across MathVista, MMMU, and BLINK show regressions of 3.4–10.9 percentage points below frozen baselines, with the counterintuitive finding that more accurate-but-still-wrong verifiers cause larger regressions than near-random ones. The authors provide a mechanistic explanation via a variance theorem for progress-gated replay and offer operational guidance: measure target-task rubric accuracy before running any verifier-driven loop and rank verifiers by task-specific quality rather than parameter count.
Vision Language Model Alignment in TRL
Hugging Face's TRL library has added support for aligning Vision Language Models (VLMs), extending existing RLHF and preference optimization tooling to multimodal settings. The blog post covers the new capabilities for training VLMs with alignment techniques such as DPO and related methods. This expands the open-source ecosystem for multimodal model fine-tuning and alignment.
The Matching Principle: A Geometric Theory Unifying Robustness, Domain Adaptation, and Alignment via Nuisance Covariance
This paper proposes the 'matching principle': a unified geometric framework arguing that robustness methods (CORAL, IRM, adversarial training, augmentation, metric learning, Jacobian penalties, alignment constraints) are all estimators of the same object—the covariance of label-preserving deployment nuisance—and that regularizing the encoder Jacobian along this covariance's range is the core statistical problem. The authors prove closed-form optimality results in a linear-Gaussian model, introduce the Trajectory Deviation Index (TDI) as a label-free embedding sensitivity probe, and validate predictions across 13 pre-registered experimental blocks including Qwen2.5-7B. At 7B scale, matched style-PMH improves selective honesty while standard DPO degrades Style TDI, connecting the theory to alignment safety.
IFLLM dataset uses mouse and eye-tracking signals to improve LLM alignment via implicit feedback
Researchers introduce IFLLM, a dataset of 1,336 multi-turn interactions from 59 Mechanical Turk workers capturing mouse trajectories and webcam-derived eye gaze to study implicit user feedback for LLM alignment. A reward model trained on this implicit feedback improves text-based reward model accuracy from 55% to 64% and nearly triples relative response quality improvements when combined with DPO across eight LLMs. The work addresses the scarcity and cost of explicit preference annotations by mining behavioral signals already present in user interactions.
Language models linearly encode a 'value axis' tracking expected goal success, study finds
Researchers construct a 'value axis' in Qwen3-8B's activation space using synthetic in-context RL data, finding that this axis distinguishes high vs. low confidence, backtracking vs. non-backtracking rollouts, and correct vs. corrupted code. Steering along this axis causally modulates self-correction behavior and verbosity, while DPO training shifts the internal value of rewarded behaviors. Applied to real-world settings, the axis reveals that Qwen assigns low internal value to politically sensitive queries post-training and that SFT increases domain-specific confidence. The findings suggest LLMs linearly encode an estimate of expected goal success that shapes their generative behavior.
Defending against Prompt Injection with Structured Queries (StruQ) and Preference Optimization (SecAlign)
Researchers from BAIR propose two fine-tuning-based defenses against prompt injection attacks: StruQ (Structured Instruction Tuning) and SecAlign (Special Preference Optimization). Both methods use a Secure Front-End with special delimiter tokens to separate trusted prompts from untrusted data, then fine-tune LLMs to ignore injected instructions. SecAlign, which uses DPO-style preference optimization, reduces attack success rates to under 15% against strong optimization-based attacks—more than 4x better than prior SOTA—while preserving model utility on AlpacaEval2.
Drifting Preference Optimization (DrPO) for One-Step Text-to-Image Generators
DrPO is a new online preference fine-tuning method designed specifically for deterministic one-step text-to-image generators like SD-Turbo and SDXL-Turbo, which are difficult to align with standard RLHF methods that require policy likelihoods or differentiable reward gradients. The method samples candidates per prompt, ranks them with a target reward, and synthesizes a feature-space update direction via a non-parametric dipole preference field plus a reference drift from the frozen base model. Because the reward is used only for ranking, DrPO supports black-box and non-differentiable reward functions while keeping inference as a single forward pass. Evaluations on HPSv3 and GenEval show improved alignment over reward-gradient-free baselines and a 3.51× reduction in training compute by eliminating reward-model backpropagation.
Mixtral 8x7B: Mistral AI Releases Sparse Mixture-of-Experts Open-Weight Model
Mistral AI has released Mixtral 8x7B, a sparse mixture-of-experts (SMoE) model with 46.7B total parameters but only 12.9B active parameters per token, enabling inference speed and cost equivalent to a 12.9B model. Licensed under Apache 2.0, Mixtral outperforms Llama 2 70B on most benchmarks and matches or exceeds GPT-3.5, with support for 32k context, five European languages, and strong code generation. An instruction-tuned variant (Mixtral 8x7B Instruct) achieves 8.3 on MT-Bench, claimed best among open-source models at release. The model is deployed behind Mistral's mistral-small API endpoint and supported via vLLM with Megablocks CUDA kernels.
Gravity-Weighted DPO enforces multi-level instruction hierarchies in LLMs
Researchers introduce Gravity-Weighted DPO (GW-DPO), a preference-optimization objective that scales per-sample loss offsets by the structural distance between conflicting instruction levels, addressing the problem of uniform architectural privilege across trust levels in production LLMs. The work formalizes a 5-level instruction hierarchy with ten pairwise priority relations and combines GW-DPO with hierarchy-specific delimiter tokens and Instructional Segment Embeddings (ISE). Evaluated on Llama-3.1-8B-Instruct, the bilateral GW-DPO schedule Pareto-improves over standard DPO on macro pairwise priority adherence while cutting over-refusal rates in half. The approach directly targets prompt injection vulnerabilities arising from models' inability to resolve competing instructions by privilege level.
Mistral AI Launches La Plateforme: First API Endpoints in Early Access
Mistral AI opened beta access to its first developer platform, La Plateforme, offering three generative text endpoints (mistral-tiny, mistral-small, mistral-medium) and an embedding endpoint. Mistral-tiny serves Mistral 7B Instruct v0.2, mistral-small serves Mixtral 8x7B, and mistral-medium serves an unreleased prototype model scoring 8.6 on MT-Bench. The platform also introduces Mistral-embed with a 1024-dimension embedding model achieving 55.26 on MTEB. The API follows OpenAI-compatible chat interface specifications and is ramping toward general availability.
BayLing-Duplex: Native full-duplex speech dialogue using a single autoregressive LLM
Researchers introduce BayLing-Duplex, a speech language model that achieves native full-duplex interaction — simultaneous listening and speaking — using a single autoregressive LLM with no auxiliary VAD or turn-taking module. Built by fine-tuning GLM-4-Voice on 400K samples plus a lightweight DPO stage, it reaches 92% turn-taking success and 100% interruption success on InstructS2S-Eval, and improves speech-response quality substantially over Moshi. The approach adds only special tokens to the standard vocabulary, making it portable across LLM architectures without architectural changes.